Data-driven modelling to quantify soil organic carbon in burnt croplands: An integration of remote sensing and machine learning

Stubble burning after harvest is known to degrade soil organic carbon (SOC). However, research on its long-term impacts on SOC is scarce and inconclusive. To address this gap, we introduce a data-driven modeling approach for SOC quantification by integrating remote sensing data with machine learning models to quantify changes in SOC during 2004-2021 across burnt rice areas in Punjab, India. This involved synthesizing literature to obtain SOC values pre- and post-burning, as well as intersecting MODIS burnt areas with rice crop maps to identify stubble burning areas in Punjab from 2004 to 2021. Post synthesis and identification, MODIS satellite band values were extracted for the synthesized experimental plots on pre- and post-burning dates. Further, remote sensing indices, which are sensitive to SOC changes such as NDVI, NBR, RECI, and BSI, were calculated for the pre- and post-burning dates. Using these indices and band values as predictors and literature-derived observed SOC values as response variables, multiple machine learning models were trained, whereby an R² value of 0.3 was obtained. While further efforts are required to improve model accuracy, our study revealed a significant decline in SOC from 2004 to 2018, ranging from 0.1  to 12.5 %,  whereas from 2019 to 2021, SOC increased by 0.7 to 7 % in various districts in Punjab. More specifically, these districts-Sangrur, Ludhiana, and Kapurthala have had the most significant decline from 2014 to 2018, whereas Rupnagar, Patiala, and Fatehgarh Sahib exhibit the highest increase in SOC from 2019 to 2021. The decline in SOC could be attributed to accelerated mineralization driven by combustion and the loss of SOC in the form of CO₂ emissions. Whereas the increase in SOC could be attributed to a reduction in stubble burning and incomplete combustion of residue, leading to the return of unburnt organic matter to the soil. These findings highlight the efficacy of integrating remote sensing frameworks with data-driven machine learning models in monitoring SOC and other aspects of soil health.