The challenge of ground subsidence caused by large-scale coal mining activities is becoming increasingly prominent, leading to potential geological risks for sustainable development in mining areas. It is urgently necessary to fully utilize satellite remote sensing technology to conduct landform monitoring research on the dynamic changes of geological environment in mining areas. Digital elevation model-based landform element extraction, or landformc characterization is an important remote sensing technique for representing the geometric shape of landforms. The results of landform characterization can support the digital simulation and terrain on land surfaces, and further extend the extraction and recognition of fine-detailed landform in mining areas. However, the existing technologies still are limited in dealing with landform elements in high-resolution digital elevation models. Based on the new characteristics of surface morphology under high-resolution digital elevation models, this project proposes a spatial-contextual morphological patterns related to aspect and curvature dimension, and establishe a benchmark database of landform characterization. Based on the data enhancement techniques and semantic knowledge in the field of geomorphology, considering the represetnations of spatial scale and spatial distance weights, we develope a multi-scale neural network model for pixel-level landform element extraction by integring convolutional neural networks and deep recurrent neural networks. The results proved that the proposed model support to achieve an accurate characterization of fine-detailed landforms in mining areas with meter-level resolution digital elevation models. We hope our effort could make a contribution on monitoring the dynamic changes of geological environment in mining subsidence areas.