Environmental space similarity model maps dry red soil under limited samples in the Yuanmou Dry Hot Valley.

Obtaining sufficient field observations for natural resources in complex mountainous environments is often constrained by limited soil samples. Geographic environmental similarity may provide a useful basis for spatial inference under small-sample conditions. Here we infer the spatial distribution of dry-red soil in the Yuanmou Dry-Hot Valley, China. Field surveys produced a training set (n = 72) and an independent validation set (n = 46; dry-red soil/non-dry-red soil = 1:1). After screening covariates related to soil, terrain, climate, and vegetation, we quantified environmental similarity among samples and computed a composite similarity score. Using stratified repeated random splitting on the independent validation set, we calibrated the classification threshold, built an Environmental Similarity Model (ESM), and generated an uncertainty layer for interpretation and extrapolation-risk identification. Results show that (1) dry-red soil is mainly distributed in basins and valley areas; (2) the ESM showed relatively strong performance on the independent validation set within the present study area (mean AUC = 0.946, Accuracy = 0.867, Recall = 0.870, F1-score = 0.862), exceeding the benchmark machine-learning models under the current validation design (mean AUC = 0.654-0.809, Accuracy = 0.589-0.707); and (3) the uncertainty layer identifies environmentally under-represented areas with higher extrapolation risk, which may support prioritized supplementary sampling and iterative updating. Overall, these results suggest that the ESM provides an interpretable option for dry-red soil inference under the present small-sample and heterogeneous environmental setting.