Multi-year Land Cover Dynamics of the Dried Aral Sea Bed (Aralkum), 1990–2024: Random Forest and U-Net Classification of Satellite Imagery

This dataset documents the multi-year land-cover dynamics of the Aralkum — the desert that formed on the dried bed of the Aral Sea (58–63°E, 43–46°N, Uzbekistan and Kazakhstan). Land cover was classified into four classes (water, saltflat, dry playa, vegetation) for eight time points: 1990, 1995, 2000, 2010, 2020, 2022, 2023 and 2024. A single Random Forest model was trained on five sensor-independent spectral indices (NDVI, NDWI, BSI, SI, SAVI), allowing one consistent model to be applied across Landsat-5/7/8 and Sentinel-2 imagery spanning 1990–2024 (overall accuracy 91.1%, Cohen's Kappa 0.877). The main finding is a decline of the water surface from approximately 31,265 km² in 1990 to about 2,776 km² in 2024 — a reduction of roughly 91% over 34 years — while the dry playa expanded correspondingly. In addition, a U-Net deep-learning model was applied to the area for the first time and compared with Random Forest, reaching an overall accuracy of 93.0% and producing spatially more coherent maps. Contents:- Classified land-cover maps (GeoTIFF, EPSG:4326) for each year- All figures at 600 DPI (publication quality)- Training features, ground-truth points and per-year area tables (CSV)- Trained Random Forest and U-Net models- Python scripts and Colab notebook for full reproducibility- Bilingual documentation (English and Uzbek) Note: the water trend is highly reliable; saltflat and vegetation values fluctuate between years due to sensor differences and weakly-labelled training data and should be read as trends. The U-Net model was trained on Random Forest labels, so independent field validation remains future work. License: CC BY 4.0

Not yet translated