Detecting crop residue burning in smallholder agricultural areas of Greece using Sentinel-2 and a Random Forest classifier

Crop residue burning is a restricted land management practice, yet its widespread application adds to regional carbon emission budgets and degrades air quality, with environmental and public health costs. Remote sensing offers the only practical means of detecting and monitoring them in a timely fashion. Two cultivated areas in Greece were studied, one during the winter and one during the summer season, to map burned agricultural land and to assess how the time elapsed since the burn affects classification accuracy. Seven spectral indices (NDVI, NBR, NBR+, NBR2, BAI, BAIS2, MIRBI) were computed from multispectral Sentinel-2 imagery. Their separability was then evaluated with the Transformed Divergence (TD) separability metric, to select the index that best discriminates burned from unburned land. Using randomly stratified pixel samples, the Burned Area Index (BAI) proved to be the strongest discriminator (TD = 1.99) and was therefore used for classification. Each case study is assessed only in one season and comprises three consecutive dates within that season, bounded by cloud cover; therefore, there were six dates in total across the two cases. This is due to cloud cover reducing observation availability and to also assess the detectability of the agricultural burn scars since their duration usually lasts a short period of some days. For each of the three dates examined per study area, a Random Forest model was trained to separate burned from unburned areas. The change in burned producer’s accuracy over time was also tracked within the reference burned area boundaries of the first date. Detection accuracy for burned fields consistently reached 95 to 98%. Accuracy over the area that had burned the earliest, by contrast, declined with time, dropping from 93% to 41% within 25 days. The proposed approach is computationally inexpensive, performs well with a limited number of training samples, and could support near-real-time monitoring of burned agricultural land.