Soil Order Classification Using Deep Learning Architecture and Multimodal Covariate Integration

Digital soil mapping (DSM) and its related studies are gaining momentum considering the need for an assorted database on soil resources required for several policy decisions. With numerous machine learning (ML) based algorithms being standardized for operational use, the inclusion of the deep learning architectures for DSM is still under the research phase. The study explores the utility of multiclass deep learning approach to characterize the soil order classes using multimodal covariate datasets. Six covariate information, including Mean annual air temperature, Elevation, Normalized Difference Vegetation Index (NDVI), Ferrous Minerals Difference Ratio (FMDR), VV, and VH representing the SCORPAN model, were downloaded and derived for the Madurai district, Tamil Nadu, India. Legacy soil order information of the study area was used as a labelled dataset to train the model using a standard U-Net deep learning architecture. Six soil order classes were found representing the study area, which was then classified besides other landform features with a mean IoU score of 0.62, obtained from evaluating the validation image patches.