Monitoring Crop Residue Burning in Smallholder Farms at Sub-Field Scale Using High-resolution Satellite Imagery

Crop residue burning in smallholder farming systems represents a critical source of atmospheric pollution and greenhouse gas emissions. However, current operational active fire monitoring products from coarse-resolution MODIS/VIIRS, restrict their application to mapping and monitoring crop stubble burning in smallholder farms. These smallholder farming systems have field sizes that may vary between 0.5 and 2 hectares, resulting in an underestimation across ~40% of the global croplands. This current limitation necessitates the need for high-resolution alternatives that can help track and monitor crop burn practices. This enables the accurate quantification of GHG emissions and the implementation of regulations in densely populated areas. To address this limitation, we developed a machine-learning approach for high-resolution mapping and monitoring of stubble burning using PlanetScope (3-5 m resolution) and Sentinel-2. Our results demonstrate that the burn detection model applied to PlanetScope achieved an accuracy of 81%, outperforming the Sentinel-2-based detection model, which had an accuracy of 69%. We attribute this to the finer resolution of Planetscope, which even compensated for the spectral limitation in detecting the burn events. The predicted PlanetScope burn detection product further enabled the delineation of burn patterns within individual farm boundaries, allowing us to classify whether a farm is entirely burned or partially burned based on the percentage of burnt area per field. Random Forest feature importance indicated that Global Environmental Monitoring Index (GEMI) consistently outperformed as the optimal spectral predictor, compared to the traditional indices, including the Normalized Burn Ratio and the Normalized Difference Vegetation Index. We also found that GEMI can effectively discriminate between burnt signatures and spectrally similar agricultural activities, such as post-harvest tillage and crop residue management operations.  Our results demonstrate that high-resolution commercial imagery can significantly enhance operational agricultural monitoring. Moreover inspiring confidence in policymakers and researchers by enabling the accurate quantification of emissions, effective policy enforcement, and environmental health protection across smallholder regions globally. However, a significant challenge persists in the scalability of research-grade studies to operations due to extremely higher costs associated with PlanetScope's commercial data acquisition (exceeding $200,000 annually for district-level continuous monitoring). These costs present significant barriers for resource-constrained governmental agencies and research institutions in developing countries, despite their demonstrable technical superiority. Future studies should address these challenges by developing data fusion-based hybrid frameworks that offer a scalable solution, striking a balance between technical needs and fiscal realities, while supporting climate mitigation and sustainable agricultural practices that strategically leverage complementary sensor capabilities. 

 Keywords: remote sensing, machine learning, PlanetScope, Sentinel-2, crop residue burning, burnt area detection