Measuring Ecological Vulnerability Status of Chongqing Using Deep Learning Algorithms

Abstract

Abstract The ecological environment is the welfare of human survival. With the global climate change in recent years, the deterioration of the ecological environment has exceeded any time. The purpose of this study is to evaluate the ecological vulnerability of Chongqing, China, and draw an ecological vulnerability map. The study screened the impact factors by information gain ratio, and finally considered 16 ecological vulnerability impact factors, using multilayer perception (DNN) and convolutional neural network (CNN) methods to model vulnerability. A total of 1200 ecological points were recorded by remote sensing images, field survey and local data, and divided into training set and test set according to the ratio of 7: 3. Finally, two ecological vulnerability status maps were produced. The DNN and CNN models were evaluated by receiver operating characteristic curve (ROC), area under curve (AUC), mean absolute error (MAE) and root mean square error (RMSE). The results showed that the ecological vulnerability evaluation effect of CNN (AUC = 0.926) was better than that of DNN (AUC = 0.888). Calculate the contribution rate of vulnerability impact factors to different model results by random forests. The results show that the highest contribution rate of the two models are geological factors. It can be understood that the relative ecological vulnerability of Chongqing is mainly affected by its own karst landform. The areas with high vulnerability in the study area are the northeast and southeast regions, as well as the mountain valleys and urban in the central and western regions. The main ecological problems are low forest quality and unreasonable structure caused by its own geological factors, and serious rocky desertification and soil erosion. Human activities, including deforestation, over-reclamation and stone mining, are also important factors contributing to the ecological vulnerability of the study area. The machine learning method of this study creates an appropriate and accurate ecological vulnerability status map, which can support the future ecological environment protection and governance decisions in the study area.