Maize crop residue cover mapping using Sentinel-2 MSI data and random forest algorithms

The return of crop residues to cultivated fields has numerous agronomic and soil quality benefits and, therefore, the areal extent of crop residue cover (CRC) could provide a rapid measure of the sustainability of agricultural production systems in a region. Recognizing the limitations of traditional CRC methods, a new method is proposed for estimating the spatial and temporal distribution of maize residue cover (MRC) in the Jilin Province, NE China. The method used random forest (RF) algorithms, 13 tillage indices and 9 textural feature indicators derived from Sentinel-2 data. The tillage indices with the best predictive performance were STI and NDTI (R 2 of 0.85 and 0.84, respectively). Among the texture features, the best-fitting was Band8AMean-5∗5 (R 2 of 0.56 and 0.54 for the line-transect and photographic methods, respectively). Based on MSE and InNodePurity, the optimal combination of RF algorithm for the line-transect method was STI, NDTI, NDI7, NDRI5, SRNDI, NDRI6, NDRI7 and Band3Mean-3∗3. Likewise, the optimal combination of RF algorithm for the photographic method was STI, NDTI, NDI7, SRNDI, NDRI6, NDRI5, NDRI9 and Band3Mean-3∗3. Regional distribution of MRC in the Jilin Province, estimated using the RF prediction model, was higher in the central and southeast sections than in the northwest. That distribution was in line with the spatial heterogeneity of maize yield in the region. These findings showed that the RF algorithm can be used to map regional MRC and, therefore, represents a useful tool for monitoring regional-scale adoption of conservation agricultural practices.