Rumo à agricultura inteligente: previsão de produtividade agrícola com dados agrometeorológicos usando machine learning
Carregando...
Data
2021-11-12
Autores
Meneses, Kamila Cunha De [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Sistemas agrícolas baseados em tecnologias digitais podem contribuir para um aumento da segurança alimentar global como parte dos esforços de mitigação e adaptação às mudanças climáticas. Um modelo acurado para previsão da produtividade beneficia muitos aspectos do gerenciamento de áreas produtivas. Portanto, nossa hipótese é a possibilidade do uso de algoritmos de machine learning para a previsão da produtividade em regiões do Brasil. Neste contexto, foram utilizadas séries históricas de produtividade de dois principais cultivos do Brasil. No primeiro trabalho, o objetivo foi prever a produtividade da cana-de-açúcar com seis meses de antecedência da colheita com acurácia em várias regiões produtoras do Brasil, utilizando modelos de machine learning. No segundo trabalho, o objetivo foi prever a produtividade do algodão usando algoritmos de machine learning baseado em elementos climáticos. Para cana-de-açúcar, nós utilizamos dados de produtividade da cana-de-açúcar de 62 localidades do Brasil. Foram utilizados também dados meteorológicos diários de temperatura média do ar, temperatura mínima, temperatura máxima, precipitação, velocidade do vento a 2 metros de altura, umidade relativa, irradiância solar no topo da atmosfera, irradiância solar global coletados na plataforma NASA/POWER. Foi utilizado um método de modelagem tradicional de regressão linear múltipla (RLM) e seis métodos de aprendizado de máquina (ML): support vector machine (SVM), random forest approach (RF), Artificial neural network (ANN), Least Absolute Shrinkage and Selection Operator (LASSO) regression, RIDGE regression e eXtreme Gradient Boosting (XGBoost) para prever a produtividade da cana-de-açúcar com seis meses de antecedência. Para o algodão, foi realizada a previsão da produtividade do algodão em função dos elementos climáticos por meio de algoritmos de aprendizado de máquina com quatro parâmetros ajustados por mínimos quadrados ordinários. Para cana-de-açúcar foram separados 4 grupos de localidades conforme a produtividade a partir de análise de cluster. Nos grupos 1 e 2 ocorrem os maiores valores de deficiência hídrica nas localidades produtoras de cana-de-açúcar. No teste dos modelos de ML, os valores de MAPE foram acima de 20%. No teste dos modelos pelo XGBOOST apresentou MAPEs de 29.71%, 26.79%, 43.5% e 33.36% para os grupos 1, 2, 3 e 4, respectivamente. Os modelos XBOOST e MLP apresentam os melhores desempenhos para a previsão de produtividade da cana-de-açúcar. Os modelos mostram que a produtividade do algodão apresenta tendência sigmóide devido ao acúmulo de precipitação, evapotranspiração potencial, armazenamento de água no solo e excedente hídrico durante o ciclo. É possível prever a produtividade do algodão para as principais regiões produtoras do Brasil usando algoritmos de aprendizado de máquina. Modelos de regressores Extra-trees tiveram melhor desempenho na previsão da produtividade do algodão usando dados climáticos do plantio à floração.
Agricultural systems based on digital technologies can contribute to an increase in global food security as part of climate change mitigation and adaptation efforts. An accurate model for forecasting productivity benefits many aspects of productive area management. Therefore, our hypothesis is the possibility of using machine learning algorithms to predict productivity in regions of Brazil. In this context, historical series of productivity of two main crops in Brazil were used. In the first work, the objective was to accurately predict the productivity of sugarcane six months before harvesting in several producing regions in Brazil, using machine learning models. In the second work, the objective was to predict cotton productivity using machine learning algorithms based on climatic elements. For sugarcane, we used sugarcane productivity data from 62 locations in Brazil. Daily meteorological data of mean air temperature, minimum temperature, maximum temperature, precipitation, wind speed at 2 meters high, relative humidity, solar irradiance at the top of the atmosphere, global solar irradiance collected on the NASA/POWER platform were also used. A traditional multiple linear regression (RLM) modeling method and six machine learning (ML) methods were used: support vector machine (SVM), random forest approach (RF), Artificial neural network (ANN), Least Absolute Shrinkage and Selection Operator (LASSO) regression, RIDGE regression and eXtreme Gradient Boosting (XGBoost) to predict sugarcane productivity six months in advance. For cotton, cotton yield prediction was performed as a function of climatic elements through machine learning algorithms with four parameters adjusted by ordinary least squares. For sugarcane, 4 groups of locations were separated according to productivity from cluster analysis. In groups 1 and 2, the highest values of water deficit occur in sugarcaneproducing localities. In the test of the ML models, the MAPE values were above 20%. In the test of the models by XGBOOST they presented MAPEs of 29.71%, 26.79%, 43.5% and 33.36% for groups 1, 2, 3 and 4, respectively. The XBOOST and MLP models present the best performances for predicting sugarcane productivity. The models show that cotton yield has a sigmoid trend due to the accumulation of P, PET, STO and EXC throughout the cycle. It is possible to predict cotton productivity for the main producing regions in Brazil using machine learning algorithms. Extra-trees regressor models performed better in predicting cotton yield using climatic data from planting to flowering. With this, it is possible to have an average anticipation of around 80 days, allowing the producer time to plan their activities such as harvesting and sales strategies.
Agricultural systems based on digital technologies can contribute to an increase in global food security as part of climate change mitigation and adaptation efforts. An accurate model for forecasting productivity benefits many aspects of productive area management. Therefore, our hypothesis is the possibility of using machine learning algorithms to predict productivity in regions of Brazil. In this context, historical series of productivity of two main crops in Brazil were used. In the first work, the objective was to accurately predict the productivity of sugarcane six months before harvesting in several producing regions in Brazil, using machine learning models. In the second work, the objective was to predict cotton productivity using machine learning algorithms based on climatic elements. For sugarcane, we used sugarcane productivity data from 62 locations in Brazil. Daily meteorological data of mean air temperature, minimum temperature, maximum temperature, precipitation, wind speed at 2 meters high, relative humidity, solar irradiance at the top of the atmosphere, global solar irradiance collected on the NASA/POWER platform were also used. A traditional multiple linear regression (RLM) modeling method and six machine learning (ML) methods were used: support vector machine (SVM), random forest approach (RF), Artificial neural network (ANN), Least Absolute Shrinkage and Selection Operator (LASSO) regression, RIDGE regression and eXtreme Gradient Boosting (XGBoost) to predict sugarcane productivity six months in advance. For cotton, cotton yield prediction was performed as a function of climatic elements through machine learning algorithms with four parameters adjusted by ordinary least squares. For sugarcane, 4 groups of locations were separated according to productivity from cluster analysis. In groups 1 and 2, the highest values of water deficit occur in sugarcaneproducing localities. In the test of the ML models, the MAPE values were above 20%. In the test of the models by XGBOOST they presented MAPEs of 29.71%, 26.79%, 43.5% and 33.36% for groups 1, 2, 3 and 4, respectively. The XBOOST and MLP models present the best performances for predicting sugarcane productivity. The models show that cotton yield has a sigmoid trend due to the accumulation of P, PET, STO and EXC throughout the cycle. It is possible to predict cotton productivity for the main producing regions in Brazil using machine learning algorithms. Extra-trees regressor models performed better in predicting cotton yield using climatic data from planting to flowering. With this, it is possible to have an average anticipation of around 80 days, allowing the producer time to plan their activities such as harvesting and sales strategies.
Descrição
Palavras-chave
Inteligência artificial, Agricultura, Produtividade agrícola, Artificial intelligence, Agriculture