Answers are grounded in this paper and cite their sources.
Accurate monitoring of land degradation is critical for achieving land degradation neutrality, yet multitemporal assessments often suffer from low temporal transferability and threshold-related biases. Here, we present a robust machine learning framework to quantify long-term bare soil dynamics (1985–2020) across tropical drylands in Sucre, Colombia. Using Landsat time series, we evaluated Random Forest (RF), Support Vector Machines (SVMs), and Logistic Regression (LR) ML models under three experimental strategies: year-specific assessment, interannual cross-validation, and combined multitemporal training. A total of 11,308 labeled samples were used for model calibration, and vegetation and soil indices were used as ancillary layers for sample support and quality control. SVM maps were additionally validated with 22,000 independent samples. Our results show a marked expansion of bare soil between 1985 and 2001 (+ 25,504 ha; + 53%), followed by a further increase from 2001 to 2010 (+ 53,351 ha; + 71.7%) and a subsequent reduction by 2020 (–28,475 ha; –22%). Although all models achieve high within-year accuracy, SVM showed the most consistent temporal generalization and external validation performance, with accuracy above 98% and Kappa values close to 1 across decades. Spectral contribution analysis identified Red, Green, and shortwave infrared (SWIR2) bands as the most informative predictors. In contrast, NDVI-based reclassification underestimated persistent bare soil by more than 96% relative to the multivariate ML approach. By integrating interannual validation, threshold-sensitivity analysis, and GIS-based persistence mapping, the proposed framework effectively distinguishes transient seasonal exposure from chronic degradation hotspots. These findings provide a scalable and transferable Earth observation workflow to support evidence-based territorial planning and soil conservation in vulnerable tropical landscapes.
No takes yet. Share an insight, caveat, or question.
Lorduy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: