Integrating Satellite Data and Machine Learning for Estimating Rice Straw Biomass and Emissions in Punjab, India

The current work evaluates the potential of Machine Learning (ML)-based stubble biomass modelling and associated Greenhouse Gases (GHG) and air pollutant emissions using multi-frequency remote sensing observations. The developed model uses Sentinel-1 C-band Synthetic Aperture Radar (SAR) (VV, VH, VH/VV) and Sentinel-2 vegetation indices (NDVI, SAVI, EVI) data, co-located with stratified field observations, as an input stack. The findings reveal that Extra Trees Regressor (ETR) performed better with fused SAR-optical data, with an R 2 = 0.78, RMSE = 0.12 kg/100 m 2 , and MAE = 0.09 kg/100 m 2 at standard 70:30 train-test split and 15-fold cross-validation score. During the post-harvest period of 2023 (October and November), the worked model estimated a staggering fresh (wet) RSB of 4.76 million tonnes of which 1.88 million tonnes (39.6 percent) was open-burnt. Accounting for a standard 60% harvest moisture discount, the related emissions included 1,077,820 tonnes of CO 2 , 4296 tonnes of methane CH 4 , 68 tonnes of N 2 O, while air pollutant emissions comprised 57,283 tonnes of CO, 6180 tonnes of fine particulate matter (PM 2.5 ), and 7918 tonnes of coarse particulate matter (PM 10 ). Regardless of the strong predictive capacity, the analysis is limited by a small sample (n = 109) and data of one growing season. However, the fusion-ETR approach will alleviate the traditional constraints, which will make it easier to have operations, high-resolution monitoring, which can be used to inform air-quality policy and aid in mitigating stubble burning.