Answers are grounded in this paper and cite their sources.
Machine learning modeling reveals fluctuating soil salinity patterns over 25 years in an arid basin, highlighting the need for rigorous spatiotemporal validation.
Long-term soil salinity mapping in arid oases is vulnerable to spatial dependence and interannual shifts, yet transfer to unsampled areas and years is rarely tested explicitly. We combined 1107 quality-controlled topsoil observations from the Yanqi Basin, China (2000, 2005, 2015, 2020, and 2025), with 30 environmental covariates. Ten algorithms and five input strategies were compared using site-grouped five-fold, 15-km spatial-block five-fold, and leave-one-year-out (LOYO) validation. Extremely Randomized Trees (ERT) provided the most stable overall performance (R 2 = 0.602, 0.597, and 0.501), whereas Transformer performed best under LOYO (R 2 = 0.523), showing that algorithm rankings depended on the transfer scenario. PCC, VIP, GRA, and GBM feature selection produced no consistent gains; ERT fitted with all 30 predictors was therefore used for mapping. Within the common valid support, the area with SSC ≥ 6 g kg −1 increased from 23.65% in 2000 to 36.41% in 2005, declined to 22.40% in 2015, and rose to 33.88% in 2025, indicating an expansion–relief–reintensification pattern. ERT-SHAP identified SI3, SI1, and Int1 as the leading predictive indicators, while groundwater depth added little independent information after the other covariates were included. ERT systematically underestimated SSC ≥ 10 g kg −1 (spatial-block bias, −4.426 g kg −1 ), requiring caution at the upper end of the observed range. The framework supports reproducible regional comparison and potential hotspot screening by combining spatiotemporal transfer tests, high-value diagnostics, attribution, and long-term mapping.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: