Machine learning model ensemble for predicting sugarcane yield through synergy of optical and SAR remote sensing

Abstract

Pre-harvest estimate of sugarcane production is required by sugar mill officials for proper planning about intra or inter-regional trading of sugarcane if expected production is more or less than mill's crushable capacity. Integration of optical and synthetic aperture radar (SAR) remote sensing has shown to improve biomass prediction accuracy of a perennial crop like sugarcane, particularly when optical data is unavailable due to presence of clouds. This study aims at estimating sugarcane yield using optical data from Sentinel-2 and SAR data from Sentinel-1 at mill catchment level of four sugar mills in Gujarat and Maharashtra, India. A variety of machine learning (ML) algorithms, including those based on Bayesian inference, as well as ensemble methods like bagging or boosting, were utilized to predict biomass. Additionally, a specific type of ensemble technique known as model stacking was also employed for predicting cane biomass. Fusing optical and SAR based yield driving variables from different active growth phases of sugarcane in an ensemble modeling framework explained about 63–70% variations of pixel-level above ground biomass during model training and 44–60% during testing stage. Aggregation of pixel level sugarcane yield at micro-zone level (cluster of villages) showed good prediction accuracies in Gujarat (NRMSE of 18%) and Maharashtra (NRMSE of 32%) at least 1–2 months before harvesting.