Implementation of Ensemble Deep Learning Coupled with Remote Sensing for the Quantitative Analysis of Changes in Arable Land Use in a Mining Area

Abstract

The high-intensity exploitation of mineral resources in mining cities in China has created new environmental challenges and serious environmental situations in recent years. The land use patterns in mining areas change rapidly for natural and anthropogenic reasons. Land-use/land-cover (LU/LC) change is extremely important in the sustainable development of mining cities. Large-scale opencast coal mining results in the destruction of built-up land, industrial land and mining land as well as arable land resources. Therefore, the analysis of changes in mineral resources and arable land use has attracted increasing attention. With the development of remote sensing (RS) and deep learning (DL) technology, many forms of data for detecting land-use changes are available. The goal of this paper is to promote coordinated land-use development. In this study, an ensemble feature pyramid convolutional neural network model (EFPCNNM) approach, which combines the theory of convolutional neural network (CNN), feature pyramid (FP) and ensemble learning (EL), was used to analyze land-use change in the Longmen Mountain opencast coal mine area in Anqing City, Huaining County, Anhui Province, China. Unlike the CNN, and ensemble convolutional neural network model (ECNNM), which only used the high-level features to predict the ground objects, and the feature information of other layers was not fully considered, EFPCNNM enhanced the small target information in high-level features and improved the detection performance of significant ground object. EFPCNNM was highly portable and could be embedded in many models to further boost performance. Comparative experiments on support vector machine (SVM), extreme learning machine (ELM), artificial neural network (ANN), CNN and ECNNM methods demonstrated that the EFPCNNM could exceed the other models in terms of detection accuracy and inference speed, e.g., the overall accuracy was 93.5036%, kappa coefficient was 0.9423, and time was 2.7569 s. Multitemporal RS images from eight periods starting in 2005 and ending in 2019 were used as the land-use data. With a typical classification system and RS quantitative analysis, this paper also evaluated the temporal and spatial changes in arable land use in the mining area. Of all the land-use types, the area of arable land decreased the most over the study period, and the area of mining land and construction land increased the most. In addition, changes in the seepage area associated with collapse sites in the study area were analyzed. The results of this change detection process could provide decision support for the coordinated development of land use and mineral resources in mining areas.