Modelagem da dinâmica de herbicidas no solo e palha utilizando redes neurais artificiais
Carregando...
Data
2020-10-06
Autores
Nascentes, Renan Fonseca [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A compreensão do comportamento ambiental de um herbicida é de extrema importância para garantir a sua aplicação correta, permitindo acertar o momento e a modalidade adequada. Portanto, o desenvolvimento de modelos matemáticos capazes de estimar a transposição do herbicida na palha e sua disponibilidade no solo tem se tornado cada vez mais crucial. Nesse contexto, as redes neurais artificiais (RNAs) têm se mostrado uma abordagem bem-sucedida para modelar relações, especialmente aquelas envolvendo séries temporais e comportamentos não-lineares. Uma das principais vantagens das redes neurais é a sua capacidade de aprendizado por meio de um conjunto de exemplos (padrões), permitindo que posteriormente possam fornecer respostas precisas. Assim, o objetivo deste trabalho é desenvolver um modelo utilizando redes neurais artificiais para estimar a transposição de diferentes herbicidas na palha, bem como a disponibilidade desses herbicidas no solo. Para o modelo RNA palha, foi utilizado as seguintes variáveis como entradas de dados do modelo: massa molecular, solubilidade em água (mg L- 1), pka, kow, pressão de vapor (mmHg), koc, tonelada de palha por hectare, simulação de chuva (mm) e dose do herbicida em m kg ha-1. O banco de dados foi preparado com 1872 dados que foi aleatoriamente divididos em dois conjuntos: 70% (1310) foram utilizados para o treinamento do modelo, enquanto os 30% restantes (562) foram utilizados para a validação do modelo. Para o modelo RNA solo, a amostra total utilizada consistiu de 124 tipos de solo e 15 herbicidas, totalizando 1860 observações. Também dividimos os dados em conjunto de treinamento, com 1302 amostras, e conjunto de validação, com 558 amostras. As entradas utilizadas para o modelo foram as características físico-química dos herbicidas massa molar, pressão de vapor, solubilidade em água, constante de dissociação (pka), coeficiente de partição octanol-água (kow), meia vida, coeficiente de sorção normalizado para MO (koc), e do solo área, argila, silte, pH, matéria orgânica, soma de bases (SB), capacidade de troca de cátions (CTC) e percentual da CTC com SB (V%). Para o desenvolvimento da RNA foi utilizado o software estatístico R com o pacote H2O. O treinamento foi feito variando o número de camadas e o número de neurônios em cada camada, e também o número de épocas de aplicação. Como os pesos são elementos iniciados com valores aleatórios, para cada treinamento obteve-se valores diferentes. Desta forma, treinou-se cada rede 10 vezes com a mesmaarquitetura para selecionar o melhor resultado. Na avaliação do desempenho da RNA, utilizou-se 30% dos dados para a validação. Os parâmetros adotados para avaliar o modelo de RNA foram o Erro Quadrático Médio Amostral (EQM), a Raiz do Erro Quadrático Médio (REQM) e o Percentual da Raiz do Erro Quadrático Médio. Quanto menor for esse erro no conjunto de validação, melhor será a arquitetura da RNA associada a ele. Além disso, foram utilizados dois índices de desempenho adicionais: o coeficiente de determinação (r2) e o índice de confiança (c) para avaliar os resultados obtidos com as RNAs desenvolvidas. O estudo concluiu que todas as arquiteturas de redes neurais tiveram desempenho ótimo, com correlação positiva (r2) acima de 0,95 em relação aos dados de validação. Destacou-se a arquitetura com 1 camada oculta e 9 neurônios, que proporcionou previsões precisas com baixo REQM de 8,9% para a transposição do herbicida na palha. O modelo de RNA para estimar a disponibilidade de herbicidas no solo também teve resultados promissores, com a arquitetura de três camadas ocultas e mais neurônios permitindo estimativas mais precisas. O percentual de importância na análise indicou que a "dose do herbicida" e a "pressão de vapor" foram cruciais para as previsões na dinâmica do herbicida na palha, enquanto no solo, o pH e a matéria orgânica foram os fatores mais relevantes. A forte correlação entre as estimativas da RNA e os dados de validação demonstra a capacidade de generalização do modelo para novos dados. Essas descobertas têm implicações importantes para a aplicação prática da tecnologia de redes neurais na estimativa da transposição de herbicidas em palha e disponibilidade de herbicida no solo.
Understanding the environmental behavior of a herbicide is extremely important to ensure its correct application, allowing for the right moment and the appropriate modality. Therefore, the development of mathematical models capable of estimating the transposition of the herbicide in the straw and its availability in the soil has become increasingly crucial. In this context, artificial neural networks (ANNs) have proven to be a successful approach for modeling relationships, especially involving time series and non-linear behavior. One of the main advantages of neural networks is their ability to learn through a set of examples (patterns), allowing them to later provide accurate answers. Thus, the objective of this work is to develop a model using artificial neural networks to estimate the transposition of different herbicides in the straw, as well as the availability of these herbicides in the soil. For the RNA straw model, the following variables were used as model data inputs: molecular mass, water solubility (mg L-1), pka, kow, vapor pressure (mmHg), koc, ton of straw per hectare, rainfall simulation (mm) and herbicide dose in m kg ha-1. The database was prepared with 1872 data that were randomly divided into two sets: 70% (1310) were used for model training, while the remaining 30% (562) were used for model validation. For the ANN soil model, the total sample used consisted of 124 soil types and 15 herbicides, totaling 1860 observations. We also split the data into training set, with 1302 Examples, and validation set, with 558 Examples. The inputs used for the model were characteristics such as physical-chemical characteristics of the herbicides, molar mass, vapor pressure, solubility in water, dissociation constant (pka), octanol-water coefficient (kow), half-life, sorption coefficient normalized for MO (koc), and soil area, clay, silt, pH, organic matter, sum of bases (SB), cation exchange capacity (CEC) and percentage of CEC with SB (V%). For the development of the ANN, the statistical software R with the H2O package was used. Training was done by varying the number of championships and the number of neurons in each layer, as well as the number of times of application. As the weights are elements started with random values, different values were obtained for each training session. In this way, each network was trained 10 times with the same architecture to select the best result. In evaluating the performance of the ANN, 30% of the data were used for validation. The parameters adopted to evaluate the ANN model were the Sample Mean Squared Error (EQM), the Root Mean Squared Error (REQM) and the Root Mean Squared Error Percentage. The smaller this error in the validation set, the better the architecture of the ANN associated with it. Furthermore, two additional performance indices were used: the coefficient of performance (r2) and the confidence index (c) to evaluate the results obtained with the developed ANNs. The study concluded that all neural network architectures performed optimally, with a positive positive (r2) above 0.95 against the validation data. The architecture with 1 hidden layer and 9 neurons stood out, which provided precision with a low REQM of 8.9% for the transposition of the herbicide in the straw. The ANN model for estimating herbicide availability in soil also had promising results, with the threemode architecture hidden and more neurons allowing for more accurate results. The percentage of importance in the analysis indicated that "herbicide dose" and "vapor pressure" were crucial for predictions of herbicide dynamics in straw, while in soil, pH and organic matter were the most relevant factors. The strong psychology between the ANN estimates and the validation data demonstrates the model's ability to generalize to new data. These findings are important for the practical application of neural network technology in estimating herbicide transposition in straw and herbicide availability in soil.
Understanding the environmental behavior of a herbicide is extremely important to ensure its correct application, allowing for the right moment and the appropriate modality. Therefore, the development of mathematical models capable of estimating the transposition of the herbicide in the straw and its availability in the soil has become increasingly crucial. In this context, artificial neural networks (ANNs) have proven to be a successful approach for modeling relationships, especially involving time series and non-linear behavior. One of the main advantages of neural networks is their ability to learn through a set of examples (patterns), allowing them to later provide accurate answers. Thus, the objective of this work is to develop a model using artificial neural networks to estimate the transposition of different herbicides in the straw, as well as the availability of these herbicides in the soil. For the RNA straw model, the following variables were used as model data inputs: molecular mass, water solubility (mg L-1), pka, kow, vapor pressure (mmHg), koc, ton of straw per hectare, rainfall simulation (mm) and herbicide dose in m kg ha-1. The database was prepared with 1872 data that were randomly divided into two sets: 70% (1310) were used for model training, while the remaining 30% (562) were used for model validation. For the ANN soil model, the total sample used consisted of 124 soil types and 15 herbicides, totaling 1860 observations. We also split the data into training set, with 1302 Examples, and validation set, with 558 Examples. The inputs used for the model were characteristics such as physical-chemical characteristics of the herbicides, molar mass, vapor pressure, solubility in water, dissociation constant (pka), octanol-water coefficient (kow), half-life, sorption coefficient normalized for MO (koc), and soil area, clay, silt, pH, organic matter, sum of bases (SB), cation exchange capacity (CEC) and percentage of CEC with SB (V%). For the development of the ANN, the statistical software R with the H2O package was used. Training was done by varying the number of championships and the number of neurons in each layer, as well as the number of times of application. As the weights are elements started with random values, different values were obtained for each training session. In this way, each network was trained 10 times with the same architecture to select the best result. In evaluating the performance of the ANN, 30% of the data were used for validation. The parameters adopted to evaluate the ANN model were the Sample Mean Squared Error (EQM), the Root Mean Squared Error (REQM) and the Root Mean Squared Error Percentage. The smaller this error in the validation set, the better the architecture of the ANN associated with it. Furthermore, two additional performance indices were used: the coefficient of performance (r2) and the confidence index (c) to evaluate the results obtained with the developed ANNs. The study concluded that all neural network architectures performed optimally, with a positive positive (r2) above 0.95 against the validation data. The architecture with 1 hidden layer and 9 neurons stood out, which provided precision with a low REQM of 8.9% for the transposition of the herbicide in the straw. The ANN model for estimating herbicide availability in soil also had promising results, with the threemode architecture hidden and more neurons allowing for more accurate results. The percentage of importance in the analysis indicated that "herbicide dose" and "vapor pressure" were crucial for predictions of herbicide dynamics in straw, while in soil, pH and organic matter were the most relevant factors. The strong psychology between the ANN estimates and the validation data demonstrates the model's ability to generalize to new data. These findings are important for the practical application of neural network technology in estimating herbicide transposition in straw and herbicide availability in soil.
Descrição
Palavras-chave
Dinâmica de herbicidas, Palhada, Sorção, Aprendizado de máquina, Inteligência artificial, Herbicide dynamics, Straw, Sorption, Machine learning, Artificial intelligence