Multimodal remote sensing and machine learning for sand dune classification in homogeneous environments: A case study from southern Morocco

Abstract

The realistic mapping of dune landscapes is necessary to model dune dynamics, but traditional remote sensing has been generally unimodal, and incapable of guaranteeing a sufficient representation of dune spatial complexity. In this study, we used a multimodal methodology combining Sentinel-1 SAR and Sentinel-2 optical data with five classification algorithms: the Random Forest, the LightGBM, the XGBoost, the support vector machines and the extra trees. We evaluated SAR-only, optical-only, and SAR–optical fusion inputs with spatial cross-validation and morphological post-processing. It is evaluated by accuracy, F1-score, IoU and Matthews correlation coefficient, and other spatial uncertainty analysis. Results demonstrate that fusion strongly boosts performance (37–40% higher F1-scores with respect to the SAR-only and 1–2% with respect to the optical-only inputs). The Random Forest and LightGBM had the highest performance (F1 = 0.725–0.735). A morphological post-processing yielded IoU improvements in 3.4% on average for the purpose of improving spatial coherence. This study proves that SAR–optical fusion is an effective scheme of dune classification and it is also useful for desertification risk assessment and arid landscape engineering.