Automatic Desert Off-Road Track Mapping Using The Fusion Of Sentinel-2 And Prisma Data

The ever-changing characteristics of desert tracks require consistent detection, and leveraging the advantages offered by remote sensing can address these challenges. This research focuses on mitigating spectral indistinctiveness exacerbated by dynamic landscapes. It tackles spatial limitations inherent to Synthetic Aperture Radar (SAR) by utilizing the spatial resolution of Sentinel-2 data and the heightened spectral sensitivity and spatial resolution of PRISMA data. This study introduces an automated framework that integrates Sentinel- 2 and PRISMA data to automatically map off-road tracks in desert environments. The proposed approach involves fusing Sentinel-2 RGB and Infrared bands with the panchromatic band of PRISMA data, thereby augmenting the number of samples for narrow tracks. Edge detection techniques (Laplacian and Canny) and the NDWI index are employed to distinguish drainage networks, refining the data for machine learning classifiers (Multilayer Perceptron and Random Forest). Validated against data from the INCT (Institut National de Cartographie et de Télédétection), the proposed system demonstrates good generalizability across new areas of interest, highlighting its potential in desert track mapping. The high accuracy achieved with the MLP (0.97) and RF (0.98) classifiers is noteworthy. However, false positives occur due to homogeneous desert landscapes, prompting the need for further refinement.