Mapping sugarcane residue burnt areas in smallholder farming systems using machine learning approaches

Satellite remote sensing methods have been proven to quantify the burnt areas resulting from different fire activities reliably. However, they were reported to perform poorly to identify crop residue burnt areas, particularly in smallholder systems due to lack of frequent and high spatial resolution satellite data. In this study, we used Harmonized Landsat Sentinel-2 (HLS) observations and evaluated two machine learning classifiers (i.e., Support Vector Machines (SVM) and Artificial Neural Networks (ANN)) to map sugarcane burnt areas for the 2019–20 season in a smallholder farming region in Thailand.Results showed that both classifiers performed well in identifying the spatial patterns of sugarcane burnt areas in the region. The ANN outperformed SVM at both pixel and regional scales. At pixel level, ANN accuracy was 93.4 % while SVM's best-performing Polynomial kernel accuracy was 82.7 %. The ANN estimated average percent burnt area (51.1 %) in the region was closer to reported value (48.7 %) by Thailand's Office of Cane and Sugar Board (OCSB), compared to the SVM estimate (62.9 %). The total estimated burnt areas by ANN and SVM (315 and 418 thousand ha, respectively) deviated more from OCSB's area (240 thousand ha) compared to percent burnt area. However, area estimates from classifiers had significantly better accuracy than the estimates of MODIS burnt products.Overall, this study demonstrated that HLS observations provided required spectral information to build promising models to map burnt areas in smallholder systems with higher accuracy than global products. Our mapping algorithm using the ANN classifier showed the potential to monitor sugarcane burnt areas reliably, and contribute to the successful implementation of regulatory policies in Thailand.