Study on extraction of small to micro artificial wetlands in check-dam gullies on the Loess Plateau based on the GEE platform and DNN network: a case study of Suide County

To address dynamic monitoring needs for small-scale artificial wetlands in Loess Plateau check-dam gullies, this study integrates Google Earth Engine (GEE) with a Deep Neural Network (DNN) to develop a multi-feature classification approach for high-resolution remote sensing. Using PlanetScope 3B imagery, 18 spectral, textural, and topographic features were extracted and reduced to 10 key predictors via Recursive Feature Elimination (RFE). The DNN model outperformed Random Forest (RF) and CART in accuracy, achieving a Kappa coefficient of 93.31% and overall accuracy (OA) of 95.00%. Results show 390,882.2 m² of wetlands in Suide County, primarily along gullies and river valleys, with DNN demonstrating superior boundary continuity in complex terrain. This framework validates the feasibility of deep learning for precise wetland mapping, offering practical solutions for Loess Plateau soil conservation and ecological restoration.