Accurate monitoring and classification of tillage practices, such as Zero Tillage (ZT) and Conventional Tillage (CT), are essential for sustainable agricultural management due to their significant environmental, economic, and agricultural impacts. Traditional methods for monitoring tillage practices are often costly and limited in scope, whereas remote sensing provides a more efficient, and scalable solution. Building upon a previous work that achieved a 75% accuracy using random forest, this study on the Indo-Gangetic Plains (IGP) uses Sentinel-2 data and advanced machine learning (ML) techniques such as XGBoost, Logistic Regression (LR), and LightGBM, to improve the classification accuracy while optimizing the number of features. Different feature engineering and selection techniques are employed here to reduce the number of features from 171 to an optimal 35, while maintaining the classification accuracy, or even slightly improving it like for the XGBoost and LightGBM models (76%), demonstrating their potential for effective tillage monitoring.
TILLAGE MONITORING: DETERMINING THE OPTIMAL NUMBER OF FEATURES IN MULTI-SPECTRAL IMAGES: A CASE STUDY IN THE INDO-GANGETIC PLAINS
Mukherjee J.;Dell'Acqua F.
2024-01-01
Abstract
Accurate monitoring and classification of tillage practices, such as Zero Tillage (ZT) and Conventional Tillage (CT), are essential for sustainable agricultural management due to their significant environmental, economic, and agricultural impacts. Traditional methods for monitoring tillage practices are often costly and limited in scope, whereas remote sensing provides a more efficient, and scalable solution. Building upon a previous work that achieved a 75% accuracy using random forest, this study on the Indo-Gangetic Plains (IGP) uses Sentinel-2 data and advanced machine learning (ML) techniques such as XGBoost, Logistic Regression (LR), and LightGBM, to improve the classification accuracy while optimizing the number of features. Different feature engineering and selection techniques are employed here to reduce the number of features from 171 to an optimal 35, while maintaining the classification accuracy, or even slightly improving it like for the XGBoost and LightGBM models (76%), demonstrating their potential for effective tillage monitoring.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
