Prior Knowledge-Based Deep Convolutional Neural Networks for Fine Classification of Land Covers in Surface Mining Landscapes

Abstract

Land cover classification is critical for urban sustainability applications. Although deep convolutional neural networks (DCNNs) have been widely utilized, they have rarely been used for land cover classification of complex landscapes. This study proposed the prior knowledge-based pretrained DCNNs (i.e., VGG and Xception) for fine land cover classifications of complex surface mining landscapes. ZiYuan-3 data collected over an area of Wuhan City, China, in 2012 and 2020 were used. The ZiYuan-3 imagery consisted of multispectral imagery with four bands and digital terrain model data. Based on prior knowledge, the inputs of true and false color images were initially used. Then, a combination of the first and second principal components of the four bands and the digital terrain model data (PD) was examined. In addition, the combination of red and near-infrared bands and digital terrain model data (43D) was evaluated (i.e., VGG-43D and Xcep-43D). The results indicate that: (1) the input of 43D performed better than the others; (2) VGG-43D achieved the best overall accuracy values; (3) although the use of PD did not produce the best models, it also provides a strategy for integrating DCNNs and multi-band and multimodal data. These findings are valuable for future applications of DCNNs to determine fine land cover classifications in complex landscapes.