Digital mapping of soil fauna abundance and diversity using Sentinel-1/2 and environmental attributes

Soil fauna abundance and diversity serve as crucial indicators for assessing ecosystem health, making their spatial predictions vital for understanding soil faunal community ecology and maintaining biodiversity. While digital soil mapping has proven effective for quantitatively predicting various soil properties, its application to soil faunal communities remains limited. To evaluate soil fauna abundance and diversity, we measured the number of macro-fauna, the number of meso-fauna, the total number of soil macro-fauna and meso-fauna, the Shannon-Wiener diversity index, the Simpson index, the Taxonomic richness, and the Pielou evenness index. High-resolution Sentinel-1 (S1) synthetic aperture radar (SAR), Sentinel-2 (S2) multispectral imagery, and digital elevation model (DEM) derivatives were employed in conjunction with machine learning (ML) models to map those soil faunal indices. Using a dataset of 84 soil samples (0–20 cm) gathered through hand-picking method, the Tullgren funnel approach, and Pitfall trap in the Danjiangkou Reservoir area, we compared the performance of random forest (RF), support vector machine (SVM), and Cubist models based on the coefficient of determination (R 2 ), the root mean square error (RMSE) and mean absolute error (MAE). Our analysis revealed that backscatter coefficients derived from S1 radar data played a more significant role in predicting soil properties compared to optical features from S2 imagery and DEM derivatives. However, a combined approach utilizing S1 data, S2 data, and DEM derivatives yielded satisfactory predictions. Furthermore, we found that RF and SVM models consistently outperformed the Cubist model, with RF achieving the most favorable results (R 2 =0.406, RMSE=0.821, and MAE=0.572) in predicting the abundance of meso-fauna. Additionally, uncertainty analysis demonstrated the superior stability of the RF model in predicting soil fauna abundance and diversity indices. These findings emphasize the potential of high-resolution S1 and S2 imagery, coupled with ML techniques, for local-scale soil property mapping and underscore the reliability of RF for ecological assessments. This highlights its potential for future research in soil faunal community ecology. • High-resolution optical and radar data were firstly used to map soil fauna feature. • The most important factor was S1 derivative for soil faunal density and diversity. • S1, S2, terrain factors, and machine learning helped to map soil fauna. • RF and SVM outperformed Cubist in spatially predicting soil fauna community.