High-Precision Extraction, Spatiotemporal Variations and Driving Factors of Small and Micro Wetlands (SMWs) in the Ulungur Lake Basin of China Using Sentinel-2 Imagery

Small and micro wetlands (SMWs) constitute critical ecological units sustaining ecological security and biodiversity in arid regions. Nevertheless, their small patch size, fragmented distribution, and prominent inter-annual variability hinder accurate extraction, and the mechanisms governing their dynamics remain poorly understood. Taking the Ulungur Lake Basin (ULB) in Xinjiang as the study area, this study integrated multidimensional features acquired from 2019–2024 Sentinel-2 time-series imagery, including spectral bands, remote-sensing indices, tasseled-cap transformation components, textural features, and topographic variables. The Boruta algorithm and ant colony optimization (ACO) were adopted for feature selection. Random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) models were then implemented under both year-by-year independent and multi-year integrated modeling schemes to extract SMWs across the study area. Sen’s slope estimator, the Mann–Kendall trend test, and Geographical Detector (GD) were further applied to investigate the spatiotemporal dynamics and driving forces of SMWs. The results revealed that (1) the ACO+XGBoost model achieved optimal overall performance among the tested configurations, with a six-year mean overall accuracy (OA) of 0.923, Kappa coefficient of 0.903, and Macro F1-score of 0.919, together with low interannual variability; (2) the August-observed SMWs area showed substantial interannual fluctuation, with an endpoint difference from 2841.94 ha in 2019 to 3065.49 ha in 2024, but the monotonic trend was not statistically significant; these wetlands were mainly concentrated in lakeshore zones, along river channels and irrigation ditches, and within low-lying depressions; and (3) at the 10 km analytical scale, DEM was the dominant factor of SMWs spatial variation, whereas climate and land cover were more important at 1–5 km scales, indicating scale-dependent controls. These findings offer a technical reference for the accurate identification and dynamic monitoring of SMWs in arid regions.