Estimativa de indicadores de qualidade do solo por meio de modelos de "machine learning"
Carregando...
Data
2022-11-09
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O monitoramento de indicadores de qualidade do solo contribui para mitigar as mudanças climáticas globais sem diminuir a produtividade dos sistemas, além de avaliar a eficiência de sistemas sustentáveis e complexos. O uso de modelos de “machine learning” no atual cenário do aumento dos dados é necessário para que as informações dos sistemas agrícolas sejam mais colaborativas, de maneira que, diminua custos e se traduza em otimização dos sistemas, tornando-os mais sustentáveis. Com este trabalho objetivou-se testar modelos de machine learning para estimar indicadores de qualidade do solo, a fim de melhorar o monitoramento dos solos agrícolas em sistema de monocultivo de milho, de pecuária e de Integração Lavoura-Pecuária, avaliando: 1) as combinações de variáveis para estimar o carbono e nitrogênio orgânico solo; e 2) variáveis que impactam a disponibilidade do nitrogênio potencialmente mineralizável no solo. Os modelos de machine learning testados foram Multilayer Perceptron Regressor (MLP), Random Forest Regressor (RF), K Neighbors Regressor (KNN), Support Vector Regression (SVR), Multiple Linear Regression (MLR), Adaptive Boosting Regressor (AdaBoost) e eXtreme Gradient Boosting Regressor (XGBoost), usando a linguagem de programação Python. Os modelos utilizados se mostraram eficientes para estimar carbono e nitrogênio do solo. As variáveis combinadas de plantas-animais-solo-clima estimaram com boa acurácia e precisão o carbono e nitrogênio orgânico do solo. No sistema de monocultivo de Milho, os modelos AdaBoost e MLP obtiveram alta precisão e MAPE<1%. No sistema monocultivo de Pecuária, os modelos MLP, SVR e Adaboost foram mais precisos e acurados (MAPE<3%). No sistema de integração Lavoura-Pecuária, todos os modelos tiveram elevada acurácia (MAPE<5%), no entanto, o modelo SVR obteve a menor precisão para estimar carbono orgânico do solo (R2 < 2%), e o SVR e KNN obtiveram a menor precisão para estimar o nitrogênio orgânico do solo (R2 < 50%). Para estimar o nitrogênio potencialmente mineralizável no solo, o modelo XGBoost foi o mais preciso, com o menor acurácia e viés (R2 = 0.97, MAPE = 3% e MBE = 0.10 mg kg-1), superando o AdaBoost, RF e MLR, nessa ordem. As diferentes combinações de variáveis estimadoras podem indicar processos importantes que influenciam na liberação de nitrogênio potencialmente mineralizável no solo, por meio de métodos de explicações aditivas de Shapley (SHAP). No geral, esses resultados fornecem uma nova perspectiva com o uso de aplicação de machine learning para estimar importantes nutrientes do solo, aproveitando os mais diversos históricos de dados de sistemas agrícolas, que podem ser úteis para os tomadores de decisão na produção de alimentos.
Monitoring soil quality indicators contribute to mitigating global climate change without reducing the influence of systems, in addition to measuring the efficiency of systems and complexes. The use of "machine learning" models in the current scenario of the data age is necessary for information from agricultural systems to be more collaborative, in a way that reduces costs and translates into the optimization of systems, making them more intelligent. The objective of this work was to test machine learning models to estimate soil quality indicators, to improve the monitoring of agricultural soils in a corn and pasture monoculture system and CropLivestock Integration system, evaluating: 1) a combination of variables to estimate carbon and organic only; and 2) variables that impact the availability of mineralizable power in the soil. The machine learning models tested were Multilayer Perceptron Regressor (MLP), Random Forest Regressor (RF), K Neighbors Regressor (KNN), Support Vector Regression (SVR), Multiple Linear Regression (MLR), Adaptive Boosting Regressor (AdaBoost) and eXtreme Gradient Boosting Regressor (XGBoost), using the Python programming language. The models used are efficient to estimate soil carbon and tolerance. The combined plant-animal-soil-climate variables accurately and accurately estimated soil carbon and organics. In the corn monoculture system, the AdaBoost and MLP models achieved high accuracy and MAPE<1%. In the livestock monoculture system, the MLP, SVR, and Adaboost models were more precise and accurate (MAPE<3%). In the Crop-Livestock integration system, all models had high accuracy (MAPE<5%), however, the SVR model obtained a lower accuracy for estimating soil organic carbon (R2 < 2%), and the SVR and KNN obtained the lowest precision to estimate the soil organic (R2 < 50%). To estimate the electrically mineralizable in the soil, the XGBoost model was the most accurate, with the lowest accuracy and bias (R2 = 0.97, MAPE = 3% and MBE = 0.10 mg kg-1), surpassing AdaBoost, RF, and MLR, in that order. The different combinations of estimated variables may indicate important processes that influenced the release of potentially mineralizable material in the soil, using Shapley's induced additive methods (SHAP). Overall, these results provide a new perspective with the use of machine learning applications to estimate important soil nutrients, taking advantage of the most diverse historical data of agricultural systems, which can be useful for decision-makers in food production.
Monitoring soil quality indicators contribute to mitigating global climate change without reducing the influence of systems, in addition to measuring the efficiency of systems and complexes. The use of "machine learning" models in the current scenario of the data age is necessary for information from agricultural systems to be more collaborative, in a way that reduces costs and translates into the optimization of systems, making them more intelligent. The objective of this work was to test machine learning models to estimate soil quality indicators, to improve the monitoring of agricultural soils in a corn and pasture monoculture system and CropLivestock Integration system, evaluating: 1) a combination of variables to estimate carbon and organic only; and 2) variables that impact the availability of mineralizable power in the soil. The machine learning models tested were Multilayer Perceptron Regressor (MLP), Random Forest Regressor (RF), K Neighbors Regressor (KNN), Support Vector Regression (SVR), Multiple Linear Regression (MLR), Adaptive Boosting Regressor (AdaBoost) and eXtreme Gradient Boosting Regressor (XGBoost), using the Python programming language. The models used are efficient to estimate soil carbon and tolerance. The combined plant-animal-soil-climate variables accurately and accurately estimated soil carbon and organics. In the corn monoculture system, the AdaBoost and MLP models achieved high accuracy and MAPE<1%. In the livestock monoculture system, the MLP, SVR, and Adaboost models were more precise and accurate (MAPE<3%). In the Crop-Livestock integration system, all models had high accuracy (MAPE<5%), however, the SVR model obtained a lower accuracy for estimating soil organic carbon (R2 < 2%), and the SVR and KNN obtained the lowest precision to estimate the soil organic (R2 < 50%). To estimate the electrically mineralizable in the soil, the XGBoost model was the most accurate, with the lowest accuracy and bias (R2 = 0.97, MAPE = 3% and MBE = 0.10 mg kg-1), surpassing AdaBoost, RF, and MLR, in that order. The different combinations of estimated variables may indicate important processes that influenced the release of potentially mineralizable material in the soil, using Shapley's induced additive methods (SHAP). Overall, these results provide a new perspective with the use of machine learning applications to estimate important soil nutrients, taking advantage of the most diverse historical data of agricultural systems, which can be useful for decision-makers in food production.
Descrição
Palavras-chave
Combinação de variáveis, Inteligência artificial, Nitrogênio, Carbono, Agricultura