Anthropogenic activities have great impacts on ecological environmental protection. Landscape patterns are important indicators of ecological risks. Therefore, effectively and accurately predicting landscape pattern variations is vital for qualitatively identifying the ecological risks faced by an area, especially in ecologically fragile underground coal mining areas with intense anthropogenic mining activities. Conventional landscape pattern prediction models mainly consider the driving factors of climatic and topographical conditions, whereas geological conditions are rarely considered. To overcome this limitation, seven representative driving factors of landscape pattern variations including two geological property-related factors (the thickness of coal seams and groundwater level), two climatic factors (precipitation and evaporation), and three topographical factors (elevation, slope, and NDVI which is abbreviated for the normalized difference vegetation index) were selected for statistical correlation with six major landscape pattern indices. Landscape pattern variation prediction models were subsequently developed using three different machine learning approaches and applied in the Yushenfu mining area of China. The model validation results showed that the model using the particle swarm optimization–extreme learning machine (PSO–ELM) method outperformed that using an extreme learning machine (ELM) and an artificial neural network (ANN). Furthermore, the spatial distribution law of the six landscape pattern indices under coal mining conditions was predicted using the PSO–ELM-based prediction model. The results show that the average increase rates of landscape pattern indices LSI, DP, LPI, and PSA are 7.6%, 26.1%, 44.9%, and 135.8%, respectively; whereas, the average decrease rates of LDI and SHDI are 19.1% and 12.5%, respectively, after mining in the #3 and #4 planning areas. Mining activities reduced the diversity of landscape patterns and increased regional ecological risks, as two high-risk areas were identified. The proposed prediction models were proven to be useful for planning mining areas and protecting local ecological environments.