Deep learning semantic segmentation supported risk monitoring of tailings reservoir basin

Abstract

Tailings reservoir is a necessary facility for mining activity, and it also causes danger to surrounding environment.. Watershed risk of the tailings reservoir in Chicheng was monitored and analyzed using GF-1 high-resolution remotely sensed data with the help of multiscale fusion and deep learning method, as well as the support of Remote Sensing (RS) and Geographic Information System (GIS) technology for a comprehensive and detailed identification and extraction of the risk information of the tailings reservoir and to study the dam-break path of the tailings reservoir in watershed and the risk to land surface over mining area.In this research, a sample set library for target detection was constructed by analyzing texture, hue, shape, and size of the tailings reservoir on the remotely sensed data. Subsequently, an improved multi-scale fusion algorithm (e.g., Multi_Scale Feature Map_SSD (MSF_SSD)) was constructed by adding a deconvolution module and a connection module to the original single shot multiBox detector (SSD). Next, the Pyramid Scene Parsing network (PSPnet) algorithm was selected to achieve the structure of the tailings reservoir on the basis of the target detection results. The internal structure of the tailings reservoir was obtained. With the help of RS and GIS technology, the surface of upstream catchment and the possible danger runoff were extracted, and dam-break path of the tailings reservoir is simulated on the basis of the arc hydro model. Finally, the range area affected by the dam-break were obtained by constructing the buffer zone of the dam-break path.The research results shown that the dam-break path of the tailings reservoir in Chicheng is generally from west to east and from north to south, and the total area affected by the dam-break was 480 km2. The combination analysis with land use/ cover classification indicated that forest land was 176.52 km2, farm land was 175.52 km2, urban land was 43.74 km2, rural construction land was 2.47 km2, water body was 17.72 km2, grassland was 3.60 km2, and pasture was 1.22 km2.The sample library constructed using GF-1 remotely sensed data and Google Earth 16 level image can provide the basis for the automatic recognition of tailings reservoir with deep learning framework. With the help of improved MSF_SSD and PSPnet algorithm, the semantic segmentation accuracy of the test area for pixel accuracy, mean IoU, F1 score, and mean F1 score is 0.98, 0.97, 0.99, 0.98, respectively. A comprehensive analysis of tailings reservoir dam-break range and possible damage to land surface types are performed with the help of hydrological analysis method and random forest classification results. Outcomes of this research can be used to analyze the impact area caused by dam-break, promote the capabilities of risk management and emergency response of tailings reservoir, and provide fundamental theories for decisions making in relevant departments.