RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Sorocaba 2017 ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Sorocaba IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO RE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR RECONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Tese apresentada como requisito para a obtenção do título de Doutor em Ciências Ambientais da Universidade Estadual Paulista “Júlio de Mesquita Filho” na Área de Concentração Recuperação Ambiental Orientador: Prof. Dr. José Arnaldo Frutuoso Roveda Coorientador: Prof. Dr. Antonio Cesar Germano Martins Sorocaba 2017 ADRIANO BRESSANE IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Tese apresentada como requisito para a obtenção do título de Doutor em Ciências Ambientais da Universidade Estadual Paulista “Júlio de Mesquita Filho” na Área de Concentração Recuperação Ambiental Orientador: Prof. Dr. José Arnaldo Frutuoso Roveda Coorientador: Prof. Dr. Antonio Cesar Germano Martins Sorocaba IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Tese apresentada como requisito para a obtenção do título de Doutor em Ciências Ambientais da Universidade Estadual Paulista “Júlio de Mesquita Filho” na Área de Concentração Diagnóstico, Tratamento e Recuperação Ambiental Prof. Dr. José Arnaldo Frutuoso Roveda Prof. Dr. Antonio Cesar Germano Martins IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO USANDO INTELIGÊNCIA COMPUTACIONAL Tese apresentada como requisito para a obtenção do título de Doutor em Ciências Ambientais da Universidade Estadual Paulista “Júlio de Mesquita Filho” na Área Diagnóstico, Tratamento e Prof. Dr. José Arnaldo Frutuoso Roveda Prof. Dr. Antonio Cesar Germano Martins IDENTIFICAÇÃO DE ESPÉCIES ARBÓREAS APOIADA POR CONHECIMENTO DE PADRÕES DE TEXTURA NO TRONCO Tese apresentada como requisito para a obtenção do título de Doutor em Ciências Ambientais da Universidade Estadual Paulista “Júlio de Mesquita Filho” na Área Diagnóstico, Tratamento e Prof. Dr. José Arnaldo Frutuoso Roveda Prof. Dr. Antonio Cesar Germano Martins Ficha catalográfica elaborada pela Biblioteca da Unesp Instituto de Ciência e Tecnologia – Câmpus de Sorocaba Bressane, Adriano. Identificação de espécies arbóreas apoiada por reconhecimento de padrões de textura no tronco usando inteligência computacional / Adriano Bressane, 2017. 112 f.: il. Orientador: José Arnaldo Frutuoso Roveda Coorientador: Antonio Cesar Germano Martins Tese (Doutorado) – Universidade Estadual Paulista "Júlio de Mesquita Filho". Instituto de Ciência e Tecnologia (Câmpus de Sorocaba), 2017. 1. Bioinformática. 2. Processamento de imagens. 3. Aprendizado do computador. I. Universidade Estadual Paulista "Júlio de Mesquita Filho". Instituto de Ciência e Tecnologia (Câmpus de Sorocaba). II. Título. http://www.sorocaba.unesp.br/#!/pos Instituto de Ciência e Tecnologia http://www.sorocaba.unesp.br/#!/pos Instituto de Ciência e Tecnologia – Av. Três de Março 511, 18087180 http://www.sorocaba.unesp.br/#!/pos-graduacao/pos – Campus de Sorocaba Av. Três de Março 511, 18087180 graduacao/pos-ca/pagina-inicial/CNPJ: 48031918003573 Campus de Sorocaba – inicial/CNPJ: 48031918003573 AGRADECIMENTOS Há muitas pessoas a quem devo agradecimento e dedico essa conquista que, na realidade, é de todos nós, familiares, amigos e professores. À Universidade Estadual Paulista e ao Programa de Pós-Graduação em Ciências Ambientais do Instituto de Ciência e Tecnologia, campus de Sorocaba. À Coordenação de Aperfeiçoamento de Pessoal de Nível Superior, pelo amparo financeiro à pesquisa durante o desenvolvimento da tese. Aos amigos e familiares, em especial aos meus pais Reginaldo Bressane e Lucelena da Cruz Bressane, pessoas de valor, exemplos de caráter, superação e bondade. Aos professores, em particular ao José Arnaldo F. Roveda e a Sandra R. M. M. Roveda, por todo ensinamento, mas principalmente pelo apoio e amizade. Agradeço a todos sem exceção, mas dedico especialmente à pessoa que dá sentido a minha vida, Patricia Satie Mochizuki, por tudo que palavras não seriam capazes de expressar. Bressane A. Identificação de espécies arbóreas apoiada por reconhecimento de padrões de textura no tronco usando inteligência computacional. 2017. 112f. Tese (Doutorado em Ciências Ambientais) - Campus de Sorocaba, UNESP - Univ Estadual Paulista, Sorocaba, 2017. RESUMO Embora fundamental para diversas finalidades, a identificação de espécies arbóreas pode ser complexa e até mesmo inviável em determinadas condições, motivando o desenvolvimento de métodos assistidos por inteligência computacional. Nesse sentido, estudos têm se concentrado na avaliação de características extraídas a partir de imagens da folha e, apesar dos avanços, não são aplicáveis a espécies caducifólias em determinadas épocas do ano. Logo, o uso de características baseadas na textura em imagens do tronco poderia ser uma alternativa, mas ainda há poucos resultados reportados na literatura. Portanto, a partir da revisão de trabalhos anteriores, foram realizados experimentos para avaliar o uso de métodos de inteligência computacional no reconhecimento de padrões de textura em imagens do tronco arbóreo. Para tanto, foram consideradas espécies arbóreas caducifólias nativas da flora brasileira. As primeiras análises experimentais focaram na avaliação de padrões. Como resultado, verificou- se que a melhor capacidade de generalização é alcançada combinando o uso de estatísticas de primeira e segunda ordem. Contudo, o aumento de variáveis preditoras demandou uma abordagem capaz de lidar com informação redundante. Entre as técnicas avaliadas para essa finalidade, a análise fatorial exploratória proporcionou redução na taxa de erros durante o aprendizado de máquina e aumento da acurácia durante a validação com dados de teste. Por fim, constatando que a variabilidade natural da textura no tronco arbóreo causa uma ambiguidade no reconhecimento de padrões, o uso da modelagem fuzzy foi avaliado. Em comparação com outros algoritmos de aprendizagem de máquina, a abordagem fuzzy proporcionou resultados competitivos e, assim, pode ser considerada uma alternativa promissora para novos avanços no apoio a identificação de espécies arbóreas usando inteligência computacional. Palavras-chave: bioinformática, processamento de imagens, aprendizagem de máquina. Bressane A. Arboreal species identification supported by texture pattern recognition in trunk using computational intelligence. 2017. 112f. Thesis (Doctoral’s degree in Environmental Sciences) - Campus de Sorocaba, UNESP - Univ Estadual Paulista, Sorocaba, 2017. ABSTRACT Although the arboreal identification is mandatory for several purposes, it can be complex and infeasible under certain conditions, motivating the development of computer-aided methods. In this sense, studies have focused on the assessment of features extracted from leaf images and, despite advancements, they are not applicable for deciduous species in some periods of year. Therefore, the usage of features based on texture in trunk images could be an alternative, but there are still few outcomes reported in the literature. Thus, from the review on previous studies, experiments have been performed for evaluating the use of computational intelligence methods for texture patterns recognition in trunk images. For that, native species from the deciduous Brazilian forest were considered. Firstly, the experimental analyzes focused on the evaluation of patterns. As a result, it was noted that the best generalization ability is reached using the first-order statistics in combination with second-order descriptors. Nevertheless, the increase of predictor variables required an approach capable of dealing with redundant information. Among the techniques assessed for this purpose, the exploratory factor analysis provided an error rate reduction during the machine learning, and an accuracy improvement in the validation over testing dataset. Finally, taking into account that the natural variability of texture in arboreal trunk causes an ambiguity in the pattern recognition, the usage of fuzzy modeling has been evaluated. In comparison with other machine learning algorithms, the fuzzy approach afforded competitive results, and hence it can be a promising alternative for further progress in the arboreal identification supported by computational intelligence. Key words: bioinformatics, image processing, machine learning. LISTA DE FIGURAS INTRODUÇÃO Figura 1. Estrutura organizacional na composição da tese ..................................................... 18 CAPÍTULO 1 Figure 1. Overview of the approach based on computational intelligence for supporting the arboreal species identification ................................................................................................. 22 Figure 2. Bark features with influence on texture in arboreal trunk images: (a) smooth, (b) striated, (c) fissured, (d) cancerous, (e) with protrusions, (f) with lenticels, (g) spines or aculeus, (h) powdery, detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) thick plaques ............................................................................................................................ 28 CAPÍTULO 2 Figure 1. Location of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. ...................................................................................................................................... 38 Figure 2. Trunk images with 512 x 512 pixels from: (a) Chorisia speciosa, (b) Schizolobiun parahyba, (c) Gochnatia polymorpha, (d) Cedrela fissilis, (e) Anadenanthera falcata ......... 38 Figure 3. Histogram for uniformity values from the trunk images ......................................... 42 Figure 4. Histogram for smoothness values from the trunk images ........................................ 43 Figure 5. Histogram for third moment (asymmetry) values from the trunk images................ 43 Figure 6. Histogram for entropy values from the trunk images .............................................. 44 Figure 7. Representation of the identification system for Anadenanthera falcata (Af); Cedrela fissilis (Cf); Gochnatia polymorpha (Gp); Schizolobiun parahyba (Sp); Chorisia speciosa (Cs), with threshold selection for: U - Uniformity; e - Entropy; R - Smoothness; and 3 – Asymmetry ............................................................................................................................... 45 CAPÍTULO 3 Figure 1. Outer bark images of the tree trunk from: (a) Chorisia speciosa, (b) Schizolobiun parahyba, (c) Gochnatia polymorpha, (d) Cedrela fissilis, (e) Anadenanthera falcata, (f) Hymenaea courbaril, and (g) Inga vera. ................................................................................. 54 Figure 2. Decision tree built for species Anadenanthera falcata (Af), Cedrela fissilis (Cf), Chorisia speciosa (Cs), Gochnatia polymorpha (Gp), Hymenaea courbaril (Hc), Inga vera (Iv), and Schizolobiun parahyba (Sp), based on both first-order statistics and co-occurrence descriptors (DTS+C). ................................................................................................................. 60 Figure 3. Area under the ROC curve for the decision trees (DT) based on statistical parameters (S), co-occurrence descriptors (G) and both (S+G), in supporting the identification of species: Anadenanthera falcata (Af), Cedrela fissilis (Cf), Chorisia speciosa (Cs), Gochnatia polymorpha (Gp), Hymenaea courbaril (Hc), Inga vera (Iv), and Schizolobiun parahyba (Sp) .......................................................................................................................... 63 CAPÍTULO 4 Figure 1. Outer bark images (512 x 512 pixels) of the tree trunk from: (a) Anadenanthera falcata, (b) Cedrela fissilis, (c) Ceiba speciosa, (d) Centrolobium tomentosum, (e) Erythrina speciosa, (f) Gochnatia polymorpha, (g) Hymenaea courbaril, (h) Inga vera, (i) Schizolobiun parahyba, (j) Tibouchina granulosa, and (k) Zanthoxylum kleinii (Zk) .................................. 68 Figure 2. Synthetic variables from linear combination of the original variavles ( and ), correspondent to principal components - P (a) and discriminant functions - D (b), even as their directions with the largest total scatter ( ) projected by PCA (a’), and maximum ( ) given by FDA (b’) ................................................................................................................... 71 Figure 3. Causal relationships between synthetic variables ( ) and original ones ( ) in Principal Component Analysis (PCA), Fischer Discriminant Analysis (FDA), and Exploratory Factor Analysis (EFA) ............................................................................................................. 71 Figure 4. Cumulative variability explained by synthetic variables produced by Principal Component Analysis (a), Fischer Discriminant Analysis (b), and Exploratory Factor Analysis (c), even as the respective projections from the three first principal components (a’), discriminant functions (b’), and principal factors (c’) ............................................................. 74 CAPÍTULO 5 Figure 1. Tree trunk images (512x512 pixels) from: Anadenanthera falcata (Af), Anadenanthera macrocarpa (Am), Bauhinia forficate (Bf), Caesalpinia peltophoroides (Ca), Caesalpinia echinata (Ce), Cedrela fissilis (Cf), Caesalpinia peltophoroides (Cp), Ceiba speciosa (Cs), Centrolobium tomentosum (Ct), Enterolobium contortisiliquum (Ec), Erythrina speciosa (Es), Gochnatia polymorpha (Gp), Guazuma ulmifolia (Gu), Hymenaea courbaril (Hc), Inga vera (Iv), Piptadenia gonoacantha (Pg), Schizolobiun parahyba (Sp), Tibouchina granulosa (Tg), Tabebuia roseoalba (Tr), and Zanthoxylum kleinii (Zk) ................................................................................................................. 87 Figure 2. Grid-type fuzzy partition: (a) partitioning of the predictor variable - into regions correspondent to the antecedents terms - using trapezoidal-shaped membership functions; (b) intervals of certainty and uncertainty that comprises the fuzzy region of the antecedent term. ......................................................................................................................................... 91 Figure 3. Split of database for the learning process and to assessing the generalization ability based on testing dataset. .......................................................................................................... 92 Figure 4. Eigenvalues and cumulative variability explained by the first 20 latent variables (principal factors) produced from the Exploratory Factor Analysis........................................ 94 Figure 5. Performance of different setings of the fuzzy rule-based classification model, from the variations of antecedent terms number in combination with minimum and product t-norm ................................................................................................................................................. 97 Figure 6. Aggregation process of predictor variables ( ) in the rules 1 ( ) and 2 ( ), using minimum and product t-norm .................................................................................................. 98 Figure 7. Increase in decision areas formed by the fuzzy if-then rules as consequence of the increment of the antecedent terms numbers. ........................................................................... 99 LISTA DE TABELAS CAPÍTULO 2 Table 1. Correct hit estimates for the classification system based on: - System output; - Dominant output; - Sample coefficient for species; - Probability for a sample belongs to species; Hrate - Hit rate for each species; rate - Average hit rate ............................................ 47 Table 2. Identified class as the dominant output ..................................................................... 47 Table 3. Confusion matrix for the testing image classification outcomes, with measures: - Total number of samples actually belonging to species; - Total number of samples identified as belonging to species; - Ratio of correctly classified samples; - Ratio of samples miss classified ............................................................................................................ 47 Table 4. Performance assessment for the identification system, using: - Precision rate; - Error rate; - Sensitivity or hit rate; - Accuracy or rate of overall accuracy; K - Kappa or agreement index ....................................................................................................... 48 CAPÍTULO 3 Table 1. Performance from decision trees (DT) based on: first-order statistics (S), co- occurrence descriptors (C), and both (S+C) ............................................................................ 59 Table 2. Texture pattern importance as indicator in the DTS+C ............................................... 61 Table 3. Confusion matrix for the classification results based on testing dataset achieved by DTS+C ....................................................................................................................................... 62 Table 4. Performance metrics based on testing dataset achieved by DTS+C ............................ 62 CAPÍTULO 4 Table 1. Original variables based on first and second order statistics, considering: grey levels number (L), pixel intensity (φi), image histogram (p(φi)), matrix dimension ( ), relative position (∅), probability of satisfying ∅ (pij), mean of rows (mr) and columns (mc) ............... 69 Table 2. Performance based on original variables ( ) and synthesized ones by principal components analysis (PC), PCA-based oblique rotation (OC), Fischer discriminant analysis (DF), and Exploratory Factor Analysis (DF) ............................................................................ 75 Table 3. Performance metrics afforded by the predicting models with the best overall accuracies based on 3-NN classifier, according to: precision ( ), sensitivity ( ), specificity ( ), and area under the curve (AUC) ............................................................. 77 CAPÍTULO 5 Table 1. Machine learning algorithms considered for performance comparison and control parameters settings adjusted during the learning process, which provide the best results in the cross-validation over the checking dataset .............................................................................. 93 Table 2. Performance of the machine learning algorithms in the benchmarking experiments, based on the settings that reach the best accuracy over checking dataset during the learning process, using the first 20 principal factors as predictor variables .......................................... 94 LISTA DE SIGLAS E ABREVIATURAS ACH - angle code histogram AUC - area under the curve ANN - artificial neural network ACM - auto-correlation method BPN - back propagation neural network BDT - binary decision tree C5- boosted rule-based model CDA - canonical discriminant analysis CNN - cascade-correlation neural network CNN - cellular neural network CCD - centroid-contour distance CR - centroid-radii CI - computational intelligence CM - contour moment COMM - cooccurrence matrices method CSS - curvature scale space DT - decision trees EFA - exploratory factor analysis FDA - Fisher discriminant analysis FMT - Fourier moment technique FRBCS - fuzzy rule-based classification system GF - geometrical features GLCM - grey level co-occurrence matrix HM - Hu moment invariants HSV - hue-saturation-value IVM - import vector machine SIFT - invariant feature transform KMO - Kaiser-Meyer-Olkin k-NN - k-nearest neighbor LSH - level-saturation-hue LDA - linear discriminant analysis MDP - modified dynamic programming MFD - modified Fourier descriptor MMC - move median centers MCH - moving center hypersphere MLP - multi-layer perceptron network MWM - multi-resolution wavelet method PDE - partial differential equations PFT - polar Fourier transform PCA - principal component analysis PNN - probabilistic neural network RBF - radial basis function neural network ROC - receiver operating characteristic RGB - red-green-blue ROI - regions of interest SDT - single decision tree SVM - support vector machine VFD - volumetric fractal dimension SUMÁRIO INTRODUÇÃO ................................................................................................................... 16 1 Objetivos ............................................................................................................................ 17 1.1 Geral ................................................................................................................................ 17 1.2 Específicos ...................................................................................................................... 17 2 Estrutura da tese ............................................................................................................... 18 CAPÍTULO 1 ARBOREAL SPECIES IDENTIFICATION USING COMPUTATIONAL INTELLIGENCE ................................................................................................................ 20 Abstract ............................................................................................................................... 20 1 Introduction ...................................................................................................................... 21 2 Morphological based identification and computer-aided approach ..................................... 21 3 Advances and limits of using leaf-based approach ............................................................... 23 4 Texture patterns recognition from the arboreal trunk images ............................................. 27 5 Conclusion .......................................................................................................................... 29 References ............................................................................................................................ 30 CAPÍTULO 2 STATISTICAL ANALYSIS OF TEXTURE IN TRUNK IMAGES FOR BIOMETRIC IDENTIFICATION OF TREE SPECIES ................................................ 36 Abstract ............................................................................................................................... 36 1 Introduction ...................................................................................................................... 37 2 Materials and methods ..................................................................................................... 37 2.1 Construction of the system .............................................................................................. 39 2.2 Classification performance of the constructed system .................................................... 40 3 Results and discussion ...................................................................................................... 42 3.1 Statistical properties of texture from trunk images ......................................................... 42 3.2 Construction of the classification system ........................................................................ 44 3.3 Validation of the classification system ............................................................................ 46 4 Conclusions ....................................................................................................................... 49 References ............................................................................................................................ 49 CAPÍTULO 3 CO-OCCURRENCE PATTERNS ANALYSIS ON THE TRUNK TEXTURE AS INDICATOR FEATURES FOR COMPUTER–AIDED TREE IDENTIFICATION .. 52 Abstract ................................................................................................................................ 52 1 Introduction ...................................................................................................................... 53 2 Methods ............................................................................................................................. 54 2.1 Data sampling and collection .......................................................................................... 54 2.2 Bark texture patterns in tree trunk images ....................................................................... 55 2.3 Predictive modeling procedure ........................................................................................ 57 2.4 Recognition performance assessment .............................................................................. 57 3 Results and discussion ...................................................................................................... 59 4 Conclusions ....................................................................................................................... 64 References ............................................................................................................................ 64 CAPÍTULO 4 MULTIVARIATE ANALYSES OF TRUNK TEXTURE PATTERNS FOR SUPPORTING TREE SPECIES IDENTIFICATION USING COMPUTATIONAL INTELLIGENCE ................................................................................................................ 66 Abstract ................................................................................................................................ 66 1 Introduction ...................................................................................................................... 67 2 Methods ............................................................................................................................. 68 2.1 Data collection for the experimental analysis.................................................................. 68 2.2 Original variables extraction based on trunk texture patterns ......................................... 69 2.3 Synthetic variables generation from multivariate analyses ............................................. 70 2.4 Predicitive modeling and performance assessment ......................................................... 72 3 Results and discussion ...................................................................................................... 74 4 Conclusions ....................................................................................................................... 78 References ............................................................................................................................ 78 CAPÍTULO 5 ARBOREAL IDENTIFICATION SUPPORTED BY FUZZY MODELING FOR TRUNK TEXTURE RECOGNITION ................................................................................ 84 Abstract .................................................................................................................................. 84 1 Introduction ........................................................................................................................ 85 2 Methods ............................................................................................................................... 86 2.1 Data collection and feature extraction ............................................................................... 86 2.2 Fuzzy modeling for the pattern recognition ...................................................................... 90 2.3 Benchmarking experiment ................................................................................................. 92 3 Results and discussion ........................................................................................................ 94 4 Conclusions ......................................................................................................................... 99 References ............................................................................................................................ 100 CONSIDERAÇÕES FINAIS .............................................................................................. 104 REFERÊNCIAS .................................................................................................................. 107 16 INTRODUÇÃO As espécies arbóreas possuem características inatas que lhes atribuem vocações funcionais próprias no ambiente. Logo, particularidades que tornam uma espécie apta para certo fim podem ser prejudiciais à outras finalidades. Algumas espécies possuem madeira pesada e resistente, outras são moles e com baixa durabilidade (LORENZI, 1992). Há espécies com aspectos ornamentais e porte adequado para arborização urbana, já outras podem ser tóxicas, ter raízes que danificam edificações, ou altura e copa que interferem com o sistema viário e de iluminação pública (SOUZA et al., 2011, SILVA, 2009; SANTOS; TEIXEIRA, 2001). Na dinâmica de sucessão primária, assim como nos processos de regeneração natural ou assistida pelo homem, as espécies arbóreas podem compor grupos ecológicos com funções e comportamentos distintos (MACIEL et al., 2003; GANDOLFI et al., 1995; BUDOWSKI, 1970). Espécies pioneiras e secundárias iniciais são mais tolerantes a certas condições, possuem crescimento mais rápido e ciclo de vida curto. Assim, as pioneiras atuam como colonizadoras, enquanto as secundárias iniciais criam condições para proliferação de espécies tardias e climácicas. Existem espécies que atraem e sustentam a fauna, fornecendo abrigo e alimento, mas também espécies exóticas invasoras que desequilibram o ecossistema e ameaçam a biodiversidade, além de causar impactos econômicos severos e, portanto, precisam ser reconhecidas e controladas (SAKAI et al., 2001; WIT; CROOKES; WILGEN, 2001). Pelo exposto nesses exemplos, fica evidente que a identificação de espécies arbóreas é fundamental para diversas finalidades, tanto para o aproveitamento econômico, que inclui a produção madeireira, de alimentos e extração de produtos medicinais, quanto para fins ecológicos, como a conservação da biodiversidade, a recuperação de áreas degradadas, e o manejo da arborização urbana. Contudo, em certos casos essa identificação pode ser complexa, morosa, imprecisa e até mesmo impraticável (BACKES; CASANOVA; BRUNO, 2011; GOUVEIA et al., 1997). Nesse contexto, o estudo de métodos computacionais para apoiar a identificação de espécies arbóreas vem se desenvolvendo nos últimos anos. Contudo, dada a complexidade biológica ainda há questões a serem superadas (YANIKOGLU; APTOULA; TIRKAZ, 2014; MACHADO et al., 2013; PRIYA; THANAMANI, 2012; KAUR; MONGA, 2012; BACKES; CASANOVA; BRUNO, 2011). As técnicas atuais têm focado no reconhecimento de características extraídas de 17 imagens da folha (GOUVEIA et al., 1997; IM et al., 1998; IM et al., 1999; FU e CHI, 2003; QI e YANG, 2003; WANG et al., 2003; YE et al., 2004; LI et al., 2005; PLOTZE et al., 2005; DU et al., 2005; GU et al., 2005; LEE; CHEN, 2006; DU et al., 2006; DU et al., 2007; WU et al., 2007; NAM et al., 2008; WANG et al., 2008; BRUNO et al., 2008; CASANOVA et al., 2009; BACKES et al., 2009; SINGH et al., 2010; BACKES; BRUNO, 2010; BACKES et al., 2011; KADIR et al., 2011a, 2011b; ROSSATTO et al., 2011; CHAKI; PAREKH, 2011; MACHADO et al., 2013; YANIKOGLU et al., 2014). Apesar dos avanços, a abordagem baseada em imagens da folha não atende a determinadas demandas, como no caso das espécies caducifólias em certos períodos do ano. Além das espécies caducifólias, que perdem suas folhas sazonalmente, há ainda os casos de árvores que foram cortadas e estruturas como as folhas não se preservaram. Logo, o reconhecimento de padrões baseados na textura em imagens do tronco arbóreo pode ser uma alternativa para apoiar a identificação de espécies usando inteligência computacional, mas ainda há poucos resultados reportados na literatura (CHI et al., 2003; WAN et al., 2004, SONG et al., 2004; HUANG et al., 2006, HUANG, 2006, POREBSKI et al., 2007, FIEL; SABLATNIG, 2011, KIM et al., 2011, BOMAN (2013). 1. Objetivos 1.1 Geral Avaliar o uso de métodos de inteligência computacional no reconhecimento de padrões de textura em imagens do tronco arbóreo de espécies caducifólias nativas da flora brasileira. 1.2 Específicos 1.2.1 Revisar estudos sobre o uso da inteligência computacional no apoio à identificação de espécies arbóreas; 1.2.2 Avaliar o uso de propriedades estatísticas para reconhecimento de padrões de textura em imagens do tronco arbóreo; 1.2.3 Analisar o uso de padrões de coocorrência, individualmente e em conjunto com estatísticas de primeira ordem; 18 1.2.4 Estudar o uso de análises multivariadas para reforçar o desempenho de padrões de textura no tronco como características indicadoras de espécies arbóreas; 1.2.5 Analisar o uso da modelagem fuzzy para o reconhecimento da textura em imagens do tronco, em comparação com outros algoritmos de aprendizagem de máquina. 2. Estrutura da tese A estrutura da tese foi organizada em sete seções. A primeira seção corresponde à introdução, na qual é apresentada uma contextualização, os objetivos geral e específicos, além da estrutura em si. As cinco seções intermediárias são compostas por capítulos, entre os quais artigos publicados ou submetidos para publicação em periódicos científicos. Por fim, a sétima seção apresenta as considerações finais (Figura 1). Figura 1. Estrutura organizacional na composição da tese. Considerações finais Capítulo V Arboreal identification supported by fuzzy modeling Capítulo IV Multivariate analyses of trunk texture patterns Capítulo III Co-occurrence patterns analysis on the trunk texture Capítulo II Statistical analysis of texture in trunk images Capítulo I Arboreal species identification using computational intelligence Introdução 19 O capítulo 1 apresenta uma visão geral sobre identificação arbórea baseada em caracteres morfológicos e assistida por computador. Em seguida, são discutidos os avanços e limites do uso da abordagem computacional baseada em características foliares. Então, o reconhecimento de padrões de textura em imagens de tronco arbóreo é abordado como uma alternativa para superar limitações no uso de técnicas atuais (Objetivo 1.2.1). A partir dessa revisão foram identificados estudos anteriores, norteando a etapa experimental da pesquisa que buscou contribuir com análises originais, reportadas nos capítulos 2 a 5. No capítulo 2 é desenvolvida a primeira análise experimental da pesquisa, dedicada a avaliar o uso de propriedades estatísticas no reconhecimento de padrões de textura em imagens do tronco (Objetivo 1.2.2). Para tanto, foram usadas 540 amostras de cinco espécies caducifólias nativas da flora brasileira. O capítulo 3 traz uma análise comparativa do desempenho proporcionado por descritores de coocorrência e estatísticas de primeira ordem (Objetivo 1.2.3), por meio de uma análise experimental com 756 amostras de sete espécies arbóreas. No capítulo 4, considerou-se que o uso de um número maior de características requer uma abordagem capaz de tratar informação redundante e, para isso, foi avaliado o uso de técnicas de análise multivariada (Objetivo 1.2.4). Para os procedimentos experimentais foram usadas 1188 amostras de onze espécies arbóreas. A partir dos resultados dos experimentos anteriores, o capítulo 5 apresenta o estudo da modelagem fuzzy como uma alternativa para lidar com a incerteza no reconhecimento da textura em imagens do tronco, em comparação com outros algoritmos de aprendizagem (Objetivo 1.2.5). Para tanto, foram utilizadas 2160 amostras pertencentes a vinte espécies arbóreas. As referências bibliográficas de cada capítulo são apresentadas ao final de cada seção, exceto aquelas relacionadas a essa seção introdutória, que se encontram após as considerações finais. 20 CAPÍTULO 1 ARBOREAL SPECIES IDENTIFICATION USING COMPUTATIONAL INTELLIGENCE Adriano Bressane1, José Arnaldo Frutuoso Roveda2, Antonio Cesar Germano Martins3, Maurício Tavares da Mota4, Minoru Iwakami Beltrão5 1 Environmental engineer, São Paulo State University (UNESP), Brazil 2 Mathematician, University of Brasília (UnB), Brazil 3 Physicist, University of Campinas (Unicamp), Brazil 4 Biologist, Pontifical Catholic University of São Paulo (PUC), Brazil 5 Agronomist, University of São Paulo (USP), Brazil Abstract The computational intelligence has been used for dealing with complex issues in several fields of application. However, the use of computer-aided methods in some areas, as the biometric identification of arboreal species, still requires studies for achieving greater performance and acceptance. In this context, the present study aims to review the use of computational intelligence for that proposal, i.e., for supporting the arboreal species identification. After an overview of morphological and computer-aided identification, we discuss the advances and limits of using leaf-based approach. Then, the texture pattern recognition in arboreal trunk images is addressed as an alternative to overcome limitations in the use of current techniques. Finally, we conclude pointing out possibilities for approaching in future studies, as the analysis of more features, of multivariate analysis techniques, and of soft boundaries for further improve the arboreal species identification using computational intelligence. Keywords: pattern recognition; image processing; taxonomy; computing techniques; bioinformatics. 21 1 Introduction Over the years, the advancement of computational intelligence has provided new approaches to face old challenges. From performing a simple task up to supporting complex decision-making, the computer-aided methods have an ever-increasing number of applications in several fields, including precision crop management, industrial automation, medical procedures, and environmental assessment. The computational intelligence (CI) comprises soft computing methods able to deal with complexity issues, common in the most practical applications. Thus, CI is considered a promising approach that may outperform methods of classical artificial intelligence, based on rigid inference mechanisms or hard computing techniques (Bittermann, 2011a, 2011b). The digital image processing to extract features and the machine learning for pattern recognition are among the areas related to the usage of CI in the bioinformatics, i.e, in analysis of biological data using computationally intelligent systems (Saeys et al., 2007). Notwithstanding, although the computer intelligence is more consolidated in some fields of application, with specialized methods for solving well-defined tasks, other areas still require studies to achieve greater performance and acceptance, as in the case of the biometric identification of arboreal species (Machado et al., 2013; Aptoula and Yanikog, 2013; Priya and Thanamani, 2012). In this context, this paper aims to review the use of computational intelligence for supporting the arboreal species identification. For that, an initial set of studies has been obtained through manual search on Google Scholar. Then, additional studies were identified considering the referenced publications and also studies that were citing the papers already identified. Thus, in section 2 we start from a brief overview of morphological characters and computer-aided identification. Section 3 discusses the advances and limits of using leaf image processing. Then, in section 4 we approach the texture recognition in arboreal trunk images as an alternative to overcome limitations in the use of current techniques. Finally, the last section presents our conclusions and perspective for future studies. 2 Morphological based identification and computer-aided approach The arboreal identification process can be carried out by means of similarities comparison with specimens classified and stored in a herbarium (Bridson and Forman, 1998). Nevertheless, experts often uses an identification key, which takes into account similarities 22 and differentiations in terms of the form and structure of the arboreal features (Urbanetz et al., 2010). It is not unusual to combine both approaches, starting with an identification key to reach a broader classification and then to conclude at a herbarium. The usage of reproductive morphological structures is more common because they suffer less alteration with habitat changes (Marchiori, 1995). However, in the absence of these reproductive characters the vegetative ones can be determinant, with the advantage of being used whatever time of the year (Batalha et al., 1998). Therefore, identifying arboreal species may be a complex task and time-consuming, by requiring analysis of fertile branches, seeds, structure of the flower and fruit, leaf type and shape, bark features, shape and size of treetop, among other morphological characters, and also consider environmental conditions in the area of occurrence (Backes et al., 2011; Rossatto et al., 2011; Gouveia et al., 1997). In this context, it is worth to highlight that the computational intelligence can be useful for supporting arboreal species identification. Nevertheless, it should not be intended to a self- sufficient system or completely autonomous for replacing the specialist’s experience and interpretation. Instead, the usage of computational intelligence methods aims to extract and classify features which can be analyzed together with the morphological characters, and hence to support the identification by experts. The computer-aided identification may be described by means of five main steps: data collection, digital image processing, feature extraction, pattern recognition, and machine learning. In turn, the last one is composed by the machine learning process (training and checking), and validation test. Thus, this approach starts with a data collection, obtaining pictures of the arboreal structure (Figure 1). Figure 1. General steps of the computer aided arboreal species identification. Data collection Image processing Statistical analysis Pattern recognition capturing images measuring variables extracting features finding regularities Machine learning building classifiers 23 Then, a digital image processing is used to perform mathematical operations that produce data related to variables measured over the pictures. To improve preciseness, such variables can be statistically treated to find new dimensions generated by the features extraction with higher discriminant power. From that, the patterns recognition finds regularities in the dataset used for training and checking, allowing a classification into different categories. Finally, this classification is usually performed through a predictive model built with use of machine learning algorithms (Bishop, 2006). 3 Advances and limits of using leaf-based approach The study of the leaf-based approach can be found in findings reported in the literature from the late 1990s, among which the study of Gouveia et al. (1997), which analyzed measures of area, dimensions of the enclosing rectangle, number of teeth and secondary veins, in order to use them for distinguishing different varieties of species. Im et al. (1998) also studied leaves structural properties, from a polygonal approximation of their contour. Then, taking into account that the leaves are subject to undesirable deformations, in a subsequent survey the authors found that techniques of normalization to reduce variations allowed improving the recognition of the species in some cases (Im et al., 1999). Fu and Chi (2003, 2006) used a segmentation approach based on histogram of pixels intensity over the leaf image, to extracted geometric parameters related to the leaves vascular system, then used them as predictor variables in an artificial neural network (ANN). Qi and Yang (2003) focused on the feature extraction present in the edge of the leaf. For that, the authors applied a machine learning algorithm based on support vector machine (SVM) to classify sawtooth and nonsawtooth samples, obtained by a rectangular windows sliding along the leaf edge. Wang et al. (2003) evaluated shape characterization functions for leaf image retrieval. Thus, using shape features referred to as centroid-contour distance (CCD), object eccentricity, and angle code histogram (ACH), the authors achieved better results than ones provided by modified Fourier descriptor (MFD) and Curvature Scale Space (CSS) methods. Ye et al. (2004) presented a computerized plant species recognition system composed by two retrieval methods, the text-based information and features extracted from leaf image. In the last on the CCD method has been used for obtaining the leaf contour. Then, the leaf 24 apex, base and width-height ratio were calculated and applied to equate a similarity metric used for leaf retrieval. Li et al. (2005) studied a method of segmentation known as snakes technique in combination with cellular neural networks (CNN) for improving preciseness and robustness in the extraction of vascular system and outlines, to subsequent leaf modeling and recognition. In Plotze et al. (2005), measures of leaf vein and outline was also evaluated. Notwithstanding, in this study the authors applied a multiscale function of fractal dimension based on the Minkowski method for extracting theses morphometric characteristics from leaves image. Du et al. (2005) compared the performance from different types of artificial neural networks in the classification of features based on the leaf shape, descripts by using a modified Fourier method. Thus, the authors found the probabilistic neural network (PNN) provided better results than radial basis function neural network (RBF), back propagation neural network (BPN), and multi-layer perceptron network (MLP). Gu et al. (2005) proposed a combination of wavelet transform with gaussian interpolation for leaves retrieval based on run-length features extracted from the leaf skeleton, using as classifiers the k-nearest neighbor (k-NN), and radial basis probabilistic neural network (RBPNN). Lee and Chen (2006) analyzed the use of region-based features extracted from leaf image, which included aspect ratio, horizontal and vertical projections, compactness, and centroid. By extracting features from the regions of interest (ROI), the authors achieved better results than ones provided by contour-based methods. Du et al. (2006) adopted an accelerated Douglas-Peucker algorithm for a leaf shape polygonal approximation. Then, the authors assessed the use of modified dynamic programming (MDP) method for leaf recognition based on shape matching, in comparison with other methods, as modified Fourier descriptors (MFD), Hu moment invariants (HM), contour moment (CM), curvature scale space (CSS), and geometrical features (GF). In Du et al. (2007) a method described as move median centers (MMC) hypersphere classifier is introduced for recognizing leaf using contour-based approach. By considering Hu moment invariants and geometrical features, as rectangularity, convexity, circularity, eccentricity, and form factor, the authors reached some improvement in more complex cases. 25 Wu et al. (2007) also used a probabilistic neural network (PNN) for leaf retrieval based on shape information, already assessed in comparison with other algorithms by previous studies. Nevertheless, before the machine learning process, the authors performed a principal component analysis (PCA), in order to reduce the data dimensionality. Nam et al. (2008) studied the use of morphological information (external shape) in combination with (nervure) vascular system features for leaf retrieval, using a scheme based on the similarity degree between images. For that, the authors implemented an adaptive grid- based matching algorithm that outperformed other existing methods. Wang et al. (2008) analyzed the use of marker-controlled watershed method in combination with pre-segmentation and morphological operation to segment leaf images with complicated background. After image segmentation, shape features based on Hu geometric moments and sixteen Zernike moments has been extracted and then used for leaf retrieval with a moving center hypersphere (MCH) classifier. Bruno et al. (2008) compared the box-counting and multiscale Minkowski–Sausage methods for estimating fractal dimensions in analyzes of the leaf complexity. Thus, by taking into account internal and external morphological features, the authors discuss the best approach for supporting the species identification from the leaf shape. Casanova et al. (2009) assessed the usage of Gabor wavelet filters for extracting and discriminating texture patterns in the foliar surface, in order to improve the species identification accuracy by adding these texture features to other leaf morphological attributes. Backes et al. (2009) presented the usage of the volumetric fractal dimension (VFD) method for analyzing, describing, and characterizing the complexity related to the leaf texture patterns that presents a huge variation. By using this approach the authors produced a texture signature able to improve traditional techniques as Gabor filters and Fourier descriptors. Singh et al. (2010) evaluated the performance of Support Vector Machine utilizing Binary Decision Tree (SVM-BDT) in comparison with Probabilistic Neural Network (PNN) and Fourier moment technique (FMT). By using leaf morphological features, the authors found that the SVM-BDT classifier provided the best accuracy. Backes and Bruno (2010) proposed an approach based on color texture analysis for leaf classification using fractal dimension. For that, the authors modeled each color channel from the foliar image as a surface. Then, the complexity of the surfaces has been analyzed using Bouligand-Minkowski and multiscale fractal dimension, overcoming other methods as chromaticity moments and Gabor EEE descriptors. 26 In subsequent study, Backes et al. (2011) also assessed the approach based on multiscale fractal dimension, using the box counting in combination with the Otsu method, in comparison with other techniques. As a result, the fractal analysis and the co-occurrence matrices method achieved higher performances than Gabor filters and Fourier descriptors. Kadir et al. (2011a) performed a comparative experiment of methods for recognizing species using morphological features extracted from leaf images. Thus, the authors verified that polar Fourier transform (PFT) method outperformed approaches based on geometric features, Zernike orthogonal moments, and Hu moment invariants. By using a PNN as classifier, these authors also studied the use of shape features captured by PFT method in combination with color moments, vein and texture features (Kadir et al., 2011b). Rossatto et al. (2011) also studied the volumetric fractal dimension (VFD) method for recognizing species based on leaf-texture properties. Nevertheless, for that the authors used a naive Bayes classifier that assumes a conditional independence hypothesis. Then, a canonical discriminant analysis (CDA) has been performed for removing the correlations among features and maximizing the separation among classes. By using artificial neural networks as classifiers for recognizing leaf images, Chaki and Parekh (2011) discuss the use of the centroid-radii (CR) model for estimating shape-based features, in comparison with Hu moments invariant (MI) method, already assessed in studies aforementioned. In addition, the combined use of features from both methods (CR and MI) has been explored in order to find the best performance. Before using the Bouligand-Minkowski method to estimate fractal dimensions over the leaf image, Machado et al. (2013) analyzed the application of non-linear partial differential equations (PDE) of Perona-Malik for enhancing the texture components. Thus, based on classification experiments with usage of linear discriminant analysis (LDA), the authors found that the proposed approach allows improving the performance in the leaf identification. Yanikoglu et al. (2014) implemented an approach based on a large set of features. As color-based features were used color moments, even as RGB histogram, LSH histogram, and the saturation-weighted hue histogram. The texture futures consisted of orientation histogram and Gabor wavelets. In turn, shape-based features included Fourier descriptors, perimeter and area convexity, compactness, elongation, basic shape statistics, area width factor, regional moments of inertia, angle code and contour point distribution histogram. These features have been assessed individually and in group using a support vector machine (SVM) as classifier. 27 Thus, the authors achieved promising result in recognizing isolated leaves images, but for unconstrained photographs, with complex background, the performance was considered unsatisfactory. Although the species identification supported by computer methods is considered a recent area with outcomes still insufficient to completely solve the involved issues, we verified a significant number of studies focused on leaf properties analyzes, such as color and texture, but mainly shape-based features, both external and internal. By analyzing these studies, we can found the analysis of several techniques for extracting features and recognizing patterns, in order for improving the leaf information retrieval and, consequently, supporting the species identification. In this sense, the results reported in the literature presents important advancements to face the biological complexity that requires, among other challenges, to deal with a huge variation in the morphological features. Moreover, the leaf features are also quite sensible to foliar maturity level, sun exposure, soil properties, and other environmental influences, as weather, pollution and diseases. Despite this, it was noted that the achievement of higher performances demands a leaves collect and preparation, for an image acquisition by scanning or photographing of the isolated sample. Thus, the samples are subject to manipulation and decomposition process that starts after collection and may impair the conservation of characteristics, and hence make hardier their recognizing. 4 Texture patterns recognition from the arboreal trunk images In general, texture patterns in arboreal trunk images are related to the presence, arrangement and dimension of bark features. According to Martins-da-Silva (2002) the arboreal bark can be smooth (without salience or depression), striated (with small grooves), fissured (with deep grooves), cancerous (with small craters kind of rounded), powdery (covered with dust), with protrusions (saliences kind of rounded but without openings), with lenticels, spines or aculeus, and detaching themselves as fine pieces, coriaceous pieces, and thick plaques (Figure 2). Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protr aculeus, (h) thick plaques espécimes botânicos (p. and Wessels, 2011; Vaucher, 2010), alternative for further improving the structures are not available or are insufficient, as the case of deciduous species in certain seasons of the year. texture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, achieving promising results. auto- Using the authors conclude that COMM features were superior methods. Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protr aculeus, (h) powdery, thick plaques. Source: adapted from Martins espécimes botânicos (p. Taking into account that the bark features are relatively uniform by species (Wojtech and Wessels, 2011; Vaucher, 2010), alternative for further improving the structures are not available or are insufficient, as the case of deciduous species in certain seasons of the year. Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, achieving promising results. Wan et al. (2004) compared the texture features perfo -correlation (ACM), histogram (HM), and co Using the -nearest neighbor ( authors conclude that COMM features were superior methods. Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protr powdery, detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Source: adapted from Martins espécimes botânicos (p. 28). Taking into account that the bark features are relatively uniform by species (Wojtech and Wessels, 2011; Vaucher, 2010), alternative for further improving the structures are not available or are insufficient, as the case of deciduous species in certain seasons of the year. Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, achieving promising results. Wan et al. (2004) compared the texture features perfo correlation (ACM), histogram (HM), and co nearest neighbor ( authors conclude that COMM features were superior Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protr detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Source: adapted from Martins 28). Belém-PA: Embrapa (Série Documentos 143). Taking into account that the bark features are relatively uniform by species (Wojtech and Wessels, 2011; Vaucher, 2010), alternative for further improving the structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, achieving promising results. Wan et al. (2004) compared the texture features perfo correlation (ACM), histogram (HM), and co nearest neighbor ( -NN) and moving median centers hypersphere classifiers, the authors conclude that COMM features were superior Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protr detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Source: adapted from Martins-da-Silva, R.C.V. 2002. Coleta e identificação de PA: Embrapa (Série Documentos 143). Taking into account that the bark features are relatively uniform by species (Wojtech and Wessels, 2011; Vaucher, 2010), the arboreal trunk texture recognition can be an alternative for further improving the computer-aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, Wan et al. (2004) compared the texture features perfo correlation (ACM), histogram (HM), and co NN) and moving median centers hypersphere classifiers, the authors conclude that COMM features were superior Figure 2. Bark features with influence on texture in arboreal trunk images: (a) striated, (c) fissured, (d) cancerous, (e) with protrusions, (f) with detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Silva, R.C.V. 2002. Coleta e identificação de PA: Embrapa (Série Documentos 143). Taking into account that the bark features are relatively uniform by species (Wojtech the arboreal trunk texture recognition can be an aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, Wan et al. (2004) compared the texture features perfo correlation (ACM), histogram (HM), and co-occurrence matrices method (COMM). NN) and moving median centers hypersphere classifiers, the authors conclude that COMM features were superior to the ones afforded by the other Figure 2. Bark features with influence on texture in arboreal trunk images: (a) usions, (f) with lenticels, (g) spines or detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Silva, R.C.V. 2002. Coleta e identificação de PA: Embrapa (Série Documentos 143). Taking into account that the bark features are relatively uniform by species (Wojtech the arboreal trunk texture recognition can be an aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, Wan et al. (2004) compared the texture features performance extracted by run occurrence matrices method (COMM). NN) and moving median centers hypersphere classifiers, the to the ones afforded by the other Figure 2. Bark features with influence on texture in arboreal trunk images: (a) smooth, (b) lenticels, (g) spines or detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Silva, R.C.V. 2002. Coleta e identificação de PA: Embrapa (Série Documentos 143). Taking into account that the bark features are relatively uniform by species (Wojtech the arboreal trunk texture recognition can be an aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, rmance extracted by run occurrence matrices method (COMM). NN) and moving median centers hypersphere classifiers, the to the ones afforded by the other 28 smooth, (b) lenticels, (g) spines or detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Silva, R.C.V. 2002. Coleta e identificação de Taking into account that the bark features are relatively uniform by species (Wojtech the arboreal trunk texture recognition can be an aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, rmance extracted by run-length, occurrence matrices method (COMM). NN) and moving median centers hypersphere classifiers, the to the ones afforded by the other 28 smooth, (b) lenticels, (g) spines or detaching themselves (i) as fine pieces, (j) as coriaceous pieces, and (k) Silva, R.C.V. 2002. Coleta e identificação de Taking into account that the bark features are relatively uniform by species (Wojtech the arboreal trunk texture recognition can be an aided identification, mainly when the foliar structures are not available or are insufficient, as the case of deciduous species in certain Chi et al. (2003) evaluated the use of Gabor filter banks to characterize different ture patterns based on central frequencies and normalized rations of amplitudes, extracted by discrete Fourier Transform method, and modeled as multiple narrowband signals, length, occurrence matrices method (COMM). NN) and moving median centers hypersphere classifiers, the to the ones afforded by the other 29 By combining grayscale features extracted by COMM and binary texture patterns, Song et al. (2004) achieved better results than that when each features set was used individually. In turn, Huang et al. (2006) also used features obtained by COMM in combination with fractal dimension descriptors, in order to compare the performance of different artificial neural networks topologies. In Huang (2006), texture features were studied in combination with color information, both extracted using multi-resolution wavelet method and classified by radial basis probabilistic neural network (RBPNN) and support vector machine (SVM). As a result, the author highlights that by combining color and texture features the RBPNN was better than past performances using only features extracted by COMM, ACM and HM. In this context, the use of color information seemed another interesting research direction, but the total number of candidate features also became very high. Then, Porebski et al. (2007) proposed an iterative procedure for selecting the most discriminating texture features extracted by COMM in different color spaces, and classified using -NN method. Fiel and Sablatnig (2011), even as Kim et al. (2011), experienced the combination of features (shape, color and texture) of different tree parts, including leaf, needles, flower and bark texture in the tree trunk. In Kim et al. (2011) the trunk texture was characterized by grayscale and binary features recognized using multi-resolution wavelet method (MWM). Besides the MWM, Fiel and Sablatnig (2011) also used COMM for extracting features and SVM as classifying model. In both studies the features combination afforded better results than when they were used individually. Boman (2013) performed a comparison between SVM and the import vector machine (IVM) classifiers, in the pattern recognition based on features extracted by grey level co- occurrence matrix (GLCM), scale invariant feature transform (SIFT), wavelet with GLCM (WGLCM), and wavelet co-occurrence histogram (WCH) methods. As a result, the best performance was obtained by SVM with GLCM. 5 Conclusions The usage of computational intelligence for identifying arboreal species is considered relatively recent, but it has been developed over the last 20 years. By reviewing previous studies, it was found that the outcomes reported in the literature have focused on the leaf image processing. 30 Indeed, in the absence of reproductive morphological structures, the leaf features are among the most important vegetative characters used by experts for characterizing species. Besides that, the shape approximately two-dimensional of leaves is other factor that encourages the digital image processing for subsequent features extraction and retrieval. On the other hand, when there is no physical collection of leaves for photographing or digital scanning, i.e, in the cases where the images are captured in field, its segmentation for removing overleaping or background elements in the image, make its use more complex and less efficient. Moreover, deciduous species lose their leaves seasonally, making the use of leaf-based approach impractical certain periods of year. However, analyzes of texture in arboreal trunk images are still understudied, with fewer outcomes reported in the literature. From the foregoing, in future studies we intended to perform experiments for analyzing more texture features in trunk images, as the use of first-order statistics, individually, and in combination with co-occurrence descriptors. Furthermore, multivariate analysis techniques will be also evaluated for optimizing the computational effort and improving preciseness during the machine learning. After that, the usage of the soft boundaries, by means of fuzzy modeling, will be experienced for dealing with ambiguity in the pattern matching based on texture in trunk images. References Aptoula, E., Yanikoglu, B. 2013. Morphological features for leaf based plant recognition. IEEE International Conference on Image Processing, p. 1496–1499. Backes, A.R., Casanova, D., Bruno, O.M. 2011. Identificação de plantas por análise de textura foliar. Anais do VI Workshop de Visão Computacional, Presidente Prudente. Backes, A.R., Bruno, O.M. 2010. Plant leaf identification using color and multi-scale fractal dimension. Lecture notes on computer science. 6134: 463-470. Backes, A.R., Casanova, D., Bruno, O.M. 2009. Plant leaf identification based on volumetric fractal dimension. International. Journal of pattern recognition and artificial intelligence. 23(6): 1145-1160. 31 Batalha, M.A., Aragaki, S., Mantovani, W. 1998. Chave de identificação das espécies vasculares do cerrado em Emas (Pirassununga, SP), baseada em caracteres vegetativos. Boletim de Botânica da Universidade de São Paulo. 17, 85-108. Bishop, C.M. 2006. Pattern recognition and machine learning. Cambridge: Springer. Bittermann, M.S. 2011a. A computational design system with cognitive features based on multi-objective evolutionary search with fuzzy information processing. In John S. Gero (ed), Design Computing and Cognition '10, p. 505-524. Netherlands: Springer. Bittermann, M.S. 2011b. Artificial intelligence versus computational intelligence for treatment of complexity in design. In: Proceedings of Workshop Assessing the Impact of Complexity Science in Design at Design Computing and Cognition '10 – DCC’10, Stuttgart 1- 8. Boman, J. 2013. Tree species classification using terrestrial photogrammetry. Umeå: Umeå University. Bridson, D., Forman, L. 1998. The herbarium handbook. Royal Botanic Gardens. Bruno, O. M., Plotze, R., O., Falvo, M., Castro, M. 2008. Fractal Dimension applied to plant identification. Information Sciences, 178: 2722-2733. Casanova, D., Bruno, O.M. 2009. Plant leaf identification using Gabor wavelets. International journal of imaging systems and technology, 19(3): 236-243. Chaki, J., Parekh, R. 2011. Plant leaf recognition using shape based features and neural network classifiers. International journal of advanced computer science and applications, 2(10): 41-47. Chi, Z., Houqiang, L., Chao, W. 2003. Plant species recognition based on bark patterns using novel Gabor filter banks. In: Proceedings of the 2003 International conference on neural networks and signal processing. 32 Du, J., Huang, D., Wang, X., Gu, X. 2005. Shape recognition based on radial basis probabilistic neural network and application to plant species identification. Lecture Notes in Computer Science, 3497: 281-285. Du, J. X., Wang, X. F., Gu, X. 2006. Computer-aided plant species identification (CAPSI) based on leaf shape matching technique, Transactions of the Institute of Measurement and Control, 28 (3): 275-284. Du, J.X., Wang, X.F., Zhang, G.J. 2007. Leaf shape based plant species recognition. Applied mathematics and computation, 185: 883-893. Fiel, S., Sablatnig, R. 2011. Automated identification of tree species from images of the bark, leaves and needles. In: Proceedings of the 2011 Computer vision winter workshop. Fu, H., Chi, Z. 2003. A two-stage approach for leaf vein extraction. In: IEEE International Conference on Neural Networks and Signal Processing, Proceedings…, Nanjing, 208 - 211, 2003. Gouveia, F., Filipe, V., Reis, M., Couto, C., Bulas-Cruz, J. 1997. Biometry: the characterisation of chestnut-tree leaves using computer vision. In: IEEE International Symposium on Industrial Electronics, Proceedings…, Guimarães, p. 757-760. Gu, X., Du, J.X., Wang, X.F. 2005. Leaf recognition based on the combination of wavelet transform and gaussian interpolation. Lecture notes in computer science. 3644: 253-262. Huang, Z.K., Huang, D.S., Du, J., Quan, Z.H., Guo, S.B. 2006. Bark classification based on textural features using Artificial Neural Networks. Lecture notes in computer science, 3972: 355-360. Huang, Z.K. 2006. Bark classification using RBPNN based on both color and texture feature. International journal of computer science and network security. 6(10):100-103. Im, C., Nishida, H., Kunii, T.L. 1998. Recognizing plant species by leaf shapes-a case study of the Acer family. Pattern recognition, 2:1171-1173. 33 _______. 1999. Recognizing plant species by normalized leaf shapes. Vision Interface, 19(21): 397-404. Kadir, A., Nugroho, L.E., Susanto, A., Santosa, P.I. 2011a. A comparative experiment of several shape methods in recognizing plants. International Journal of Computer Science & Information Technology, 3(3): 256-263. _______.2011b. Leaf classification using shape, color, and texture. International Journal of Computer Trends and Technology, 2(1): 225-230. Kim, S.J., Kim, B.W., Kim, D.P. 2011. Tree recognition for landscape using by combination of features of its leaf, flower and bark. In: Proceedings of the 2011 Society of Instrument and Control Engineers Annual Conference. Lee, C.L., Chen, S.Y. 2006. Classification of leaf images. International journal of imaging systems and technology. 16(1): 15-23. Li, Y., Zhu, Q., Cao, Y., Wang, C. 2005. A leaf vein extraction method based on snakes technique. In: IEEE International Conference on Neural Networks and Brain, Proceedings…. Beijing, p. 885-888. Machado, B.B., Casanova, D., Gonçalves, W.N., Bruno, O.M. 2013. Partial differential equations and fractal analysis to plant leaf identification. Journal of Physics 410, 1-4. doi:10.1088/1742-6596/410/1/012066 Marchiori, J.N.C. 1995. Elementos de Dendrologia. Santa Maria: Editora UFSM. Nam, Y., Hwang, E.J., Kim, D.Y. 2008. A similarity-based leaf image retrieval scheme: joining shape and venation features. Computer vision image understanding. 110:245-259. Plotze, R.O., Falvo, M., Pádua, J.G., Bernacci, L.C., Vieira, M.L.C., Oliveira, G.C.X., Bruno, O.M. 2005. Leaf shape analysis by the multiscale minkowski fractal dimension, a new morphometric method: a study in Passiflora L. (Passifloraceae). Canadian journal of botany. 34 83(3): 287-301. Porebski, A., Vandenbroucke, N., Macaire, L. 2007. Iterative feature selection for color texture classification. In: Proceedings of the 2007 IEEE International conference on image processing. Priya, A.C., Thanamani, A.S. 2012. A survey on species recognition system for plant classification. International Journal Computer Technology & Applications, 3(3): 1132-1136. Qi, H., Yang, J.G. 2003. Sawtooth feature extraction of leaf edge based on support vector machine. Machine learning and cybernetics. 5:3039-3044. Rossatto, D.R., Casanova, D., Kolb, R.M., Bruno, O.M. 2011. Fractal analysis of leaf-texture properties as a tool for taxonomic and identification purposes: a case study with species from Neotropical Melastomataceae. Plant Systematics and Evolution 291, 103-116. doi:10.1007/s00606-010-0366-2. Saeys, Y., Inza, I., Larrañaga, P. (2007). A review of feature selection techniques in bioinformatics. Bioinformatics, 23 (19): 2507-2517. Martins-da-Silva, R.C.V. 2002. Coleta e identificação de espécimes botânicos. Belém-PA: Embrapa (Série Documentos 143). Singh, K., Gupta, I., Gupta, S. 2010. SVM-BDT PNN and Fourier Moment Technique for classification of leaf shape. International journal of signal processing, image processing and pattern recognition. 3(4):67-78. Song, J., Chi, Z., Liu, J., Fu, H. 2004. Bark classification by combining grayscale and binary texture features. In: Proceedings of the 2004 Intelligent multimedia, video and speech processing. Urbanetz, C.; Tamashiro, J.Y.; Kinoshita, L. S. 2010. Chave de identificação de espécies lenhosas de um trecho de Floresta Ombrófila Densa Atlântica, no Sudeste do Brasil, baseada em caracteres vegetativos. Biota Neotropica. 10: 349-398. 35 Vaucher, H. 2010. Tree bark: a color guide. Portland: Timer press. Wan, Y.Y., Xiang, D.J., Huang, D.S., Chi, Z., Cheung, Y., Wang, X.F., Zhang, G.J. 2004. Bark texture feature extraction based on statistical texture analysis. In: Proceedings of the 2004 Intelligent multimedia, video and speech processing. Wang, X., Huang, D.S., Du, J.X., Xu, H., Heutte, L. 2008. Classification of plant leaf images with complicated background. Applied mathematics and computation. 205:916-926. Wang, Z., Chi, Z., Feng, D., Wang, Q. 2003. Leaf image retrieval with shape features. Lecture notes in computer science. 1929 (2000):477-487. Wojtech, M., Wessels, T. 2011. Bark: a field guide to trees of the northeast. New England: UPNE. Wu, S.G., Bao, F.S., Xu, E.Y., Wang, YX., Chang, Y.F., Xiang, Q.L. 2007. A leaf recognition algorithm for plant classification using probabilistic neural network. The computing research repository. 1:11-16. Yanikoglu, B., Aptoula, E., Tirkaz, C. 2014. Automatic plant identification from photographs. Machine vision and applications. 25(6): 1369-1383 Ye, Y., Chen, C., Li, C.-T., Fu, H., Chi, Z. 2004. A computerized plant species recognition system. In: International Symposium on Intelligent Multimedia, Machine vision and applications, Proceedings…, p. 723 - 726, Hong Kong. 36 CAPÍTULO 2 STATISTICAL ANALYSIS OF TEXTURE IN TRUNK IMAGES FOR BIOMETRIC IDENTIFICATION OF TREE SPECIES a Adriano Bressane1, José Arnaldo Frutuoso Roveda2, Antonio Cesar Germano Martins3 1 Environmental engineer, São Paulo State University (UNESP), Brazil 2 Mathematician, University of Brasília (UnB), Brazil 3 Physicist, University of Campinas (Unicamp), Brazil Abstract The identification of tree species is a key step for sustainable management plans of forest resources, as well as for several other applications that are based on such surveys. However, the present available techniques are dependent on the presence of tree structures, such as flowers, fruits and leaves, limiting the identification process to certain periods of the year. Therefore, this article introduces a study on the application of statistical parameters for texture classification of tree trunk images. For that, 540 samples from 5 Brazilian native deciduous species were acquired and measures of entropy, uniformity, smoothness, asymmetry (third moment), mean and standard deviation were obtained from the presented textures. Using a decision tree, a biometric species identification system was constructed and resulted a 0.84 average precision rate for species classification with 0.83 accuracy and 0.79 agreement. Thus, it can be considered that the use of texture presented in trunk images can represent an important advance in tree identification, since the limitations of the current techniques can be overcome. Key words: Image processing; Statistical parameters; Image texture, Tree identification. a Published in Environmental Monitoring and Assessment. 2015, 187(4): 212. DOI 10.1007/s10661-015-4400-2. 37 1 Introduction In constant development, image processing has been widely used in various areas, with several applications, including multidisciplinary studies, such as vegetation (Zehm, Nobis and Schwabe, 2003) and agricultural analysis (Vibhute and Bodhe, 2012) or for conservation and environmental management purposes (Yemshanov, McKenney and Pedlar, 2012; Pu, 2011;Ge et al., 2006; Weber and Glenn, 2001). However, the study of computer vision methods to identify species of plants is still a new area, with significant growth potential (Machado et al., 2013). For instance, with texture fractal analysis, Casanova, Florindo and Bruno (2011) achieved promising results with almost 50% accuracy in identifying plants from European countries, highlighted by the authors as a rate much higher than that of related works. Also based on the analysis of images from leaves, there are several other works in the literature Silva et al. (2014), Sá Júnior et al. (2013), Rossato et al. (2011), Sá Júnior et al. (2011), Oliveira and Bruno (2009), Casanova and Bruno (2009), Backes, Casanova and Bruno (2009). In general, species identification based on analysis of images from leaves is considered an advance over conventional morphological techniques which are limited to certain times of year, as they are commonly done from flowers and fruits. However, in the case of deciduous species, that lose their leaves during cold and dry seasons and may not present flowers or fruits, analyzes based on images of leaves are not suitable. The requirement on the presence of leaves also restricts the application of such techniques in the case of tree identification when they were cut and those morphological structures had been removed. Therefore, this paper introduces a study on the application of statistical properties for classification of texture patterns in images from the trunk that can lead to a biometric tree species identification system. 2 Materials and methods For the present study, images of 5 tree species from the Brazilian native deciduous forest were acquired at Biquinha Municipal Natural Park, a nature conservation unit composed of forest remnants and isolated arboreal individuals, located in the city of Sorocaba, São Paulo, Brazil (see Figure 1). Figure 1. digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following species: and Schizolobiun parahyba the trunk, all around the trees, discarding shaded areas or with other int presence of insects. mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Figure parahyba, space Figure 1. Location of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following species: Anadenanthera falcata Schizolobiun parahyba the trunk, all around the trees, discarding shaded areas or with other int presence of insects. From each image a central area of mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. ure 2. Trunk images with 512 x 512pixels from: (a) parahyba, (c) Gochnatia polymorpha For the present investigation, the images were transformed from the RGB to the HSV space (Gonzales and cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Anadenanthera falcata Schizolobiun parahyba the trunk, all around the trees, discarding shaded areas or with other int presence of insects. From each image a central area of mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Trunk images with 512 x 512pixels from: (a) Gochnatia polymorpha For the present investigation, the images were transformed from the RGB to the HSV onzales and Woods São Paulo Brazil Sorocaba cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Anadenanthera falcata, Gochnatia polymorpha Schizolobiun parahyba. The images from each species were taken at different heights of the trunk, all around the trees, discarding shaded areas or with other int From each image a central area of mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Trunk images with 512 x 512pixels from: (a) Gochnatia polymorpha, (d) For the present investigation, the images were transformed from the RGB to the HSV oods, 2008), and the V channel was used in São Paulo Washington Arthur Fonseca Street cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Gochnatia polymorpha . The images from each species were taken at different heights of the trunk, all around the trees, discarding shaded areas or with other int From each image a central area of 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Trunk images with 512 x 512pixels from: (a) , (d) Cedrela fissilis For the present investigation, the images were transformed from the RGB to the HSV , 2008), and the V channel was used in Biquinha Municipal Washington Luiz Avenue Arthur Fonseca Street 7397120 mS cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Gochnatia polymorpha, Cedrela fissilis, Chorisia speciosa . The images from each species were taken at different heights of the trunk, all around the trees, discarding shaded areas or with other int 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Trunk images with 512 x 512pixels from: (a) Chorisia spec Cedrela fissilis, (e) For the present investigation, the images were transformed from the RGB to the HSV , 2008), and the V channel was used in Biquinha Municipal Natural Park Com endador P. Inácio Luiz Avenue Arthur Fonseca Street Avenue 24 92 50 m E cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Cedrela fissilis, Chorisia speciosa . The images from each species were taken at different heights of the trunk, all around the trees, discarding shaded areas or with other int 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Chorisia speciosa, , (e) Anadenanthera falcata For the present investigation, the images were transformed from the RGB to the HSV , 2008), and the V channel was used in the study. Biquinha Municipal Com endador P. Inácio Avenue Juv ena l d e N SAD 69 cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Cedrela fissilis, Chorisia speciosa . The images from each species were taken at different heights of the trunk, all around the trees, discarding shaded areas or with other interference, such as the 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 fo species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. iosa, (b) Schizolobiun Anadenanthera falcata For the present investigation, the images were transformed from the RGB to the HSV the study. 200 m Juv ena l d e C am po s A ve nu e N SAD 69 38 cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Cedrela fissilis, Chorisia speciosa . The images from each species were taken at different heights of erference, such as the 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, nine images were further obtained, generating a total of 540 images, with 108 for each species, from which 80 were used for the construction of the classification system and 28 for performance tests. Figure 2 shows examples of images for each of the 5 tree species. Schizolobiun Anadenanthera falcata. For the present investigation, the images were transformed from the RGB to the HSV 38 cation of Biquinha Municipal Natural Park, in the city of Sorocaba, São Paulo, Brazil. The sample set was obtained from isolated trees (open grown) using a conventional digital camera, recording 12 images with 2560 x 1920 pixels, for each of the following Cedrela fissilis, Chorisia speciosa . The images from each species were taken at different heights of erference, such as the 1024 x 1024 pixels was cut, and, using a moving mask of 512 x 512 pixels displaced by 128 pixels in the horizontal and vertical directions, r each species, from which 80 were used for the construction of the classification system and 28 for For the present investigation, the images were transformed from the RGB to the HSV 39 2.1 Construction of the system Texture from the 400 images (80 for each species) was analyzed in the spatial domain based on four statistical parameters: uniformity, entropy, asymmetry, and smoothness. Uniformity (U) that measures how close gray level intensity are in the image (Gonzales, Woods and Eddins, 2009), is obtained by: = ( ) (1) where is the number of gray levels, is the intensity value of pixel i, and ( )is the image histogram. Based on the randomness of the gray levels in the image that brings the information of the structure organization presented (Jain, Rangachar and Schunck, 1995), entropy (e) can be calculated by: = − ( ) log ( ) (2) Third moment, also known as Asymmetry ( ), and smoothness (R), that takes in to account the transition between the shades of gray in the image, are respectively obtained by: = ( − ) ( ) (3) = 1 − 1 1 + (4) where, is the average intensity (or first moment), and is the standard deviation (or second moment), obtained by: = ( ) (5) = ∑ ( − ) − 1 (6) where n is the number of pixels in the image. The output from each of the four statistical parameters were normalized to the range [0;1]. Histograms with those values were constructed, from which, thresholds limits between distributions associated with different texture patterns were obtained in an attempt to separate species. 40 With those limits, the classification system was built in the form of a decision tree in which a given statistical property permitted the implementation of a logical operation that separates the samples according to a binary rule based on relevant ranges and thresholds. Those operations are sequentially integrated forming the branches of the decision tree. When the addition of a new operation did not provide significant gains in the classifying ability of the system, the growth of the tree in that direction was terminated and an output Sj was obtained. 2.2 Classification performance of the constructed system The classifying ability of the system was evaluated through a hit rate for each species (Hrate) and an average hit rate ( rate), given by: Hrate(spi) =∑ ( . . ) (7) rate= ∑ ( . ) (8) where k is the number of outputs for the classifying system, is the probability that a sample from species (spi) is in the output , is the sample coefficient for in , is the dominant output for , and is the ratio of samples from species i by the total number of samples used in the construction of the system. The probability that a sample from spi is part of the output was estimated by the ratio of the number of samples from spi in by the number of output samples in : ( ) = / (9) To obtain the coefficient of species samples in a given output, it was used the ratio of the number of samples of that species (spi) contained in the output set ( ) by the total initial samples for the same specie : ( ) = / (10) To study the dominant output for each species spi, a binary value defined as a function of the probability that the sample belongs to spi was obtained in such a way that: ( ) = 1, if ( ) is maximum for all the in the output 0, otherwise (11) It is emphasized that the classification system is probabilistic, once it estimates the probability that a sample from a give species is in each output (Sj). To evaluate the system performance (validation), th