Prediction of Sugarcane Yield Using Remote Sensing and Soil Mineralogy Data

Abstract

Early yield estimation is essential for agricultural planning, localized management and decision-making in commercial areas, especially in view of the high spatial variability of sugarcane fields. However, traditional estimation methods are generally destructive, costly, and inefficient in accounting for this variability. In this context, the integration of remote sensing data, soil attributes and machine learning techniques presents itself as a promising alternative for predicting sugarcane productivity. The objective of this study was to apply predictive models of productivity at a sub-field scale, integrating orbital data, structural information derived from LiDAR (Light Detection and Ranging) and mineralogical attributes of the clay fraction. The study was conducted in a commercial area in the state of São Paulo. The actual productivity data were obtained by sensors embedded in a sugarcane harvester, georeferenced by GNSS with RTK correction. Multispectral images of the constellation Planet Labs were used, from which spectral bands, vegetation indices and texture attributes based on the Gray Level Co-occurrence Matrix (GLCM) were extracted. The height of the plants was obtained by airborne LiDAR surveys, while the soil mineralogy was characterized by magnetic susceptibility measurements, converted into a representative index of the clay fraction. The Random Forest, XGBoost, KNN and Multiple Linear Regression models were evaluated. The results indicated that the XGBoost model presented the best performance among the models tested after selecting the variables, with RMSE and MAPE values of 5.53 t ha⁻¹ and 5.79%, highlighting the texture attributes derived from GLCM as the most important variables for yield prediction, surpassing plant height and soil mineralogy data.