Semantic Labeling in Remote Sensing Corpora Using Feature Fusion-Based Enhanced Global Convolutional Network with High-Resolution Representations and Depthwise Atrous Convolution

One of the fundamental tasks in remote sensing is the semantic segmentation on the aerial and satellite images. It plays a vital role in applications, such as agriculture planning, map updates, route optimization, and navigation. The state-of-the-art model is the Enhanced Global Convolutional Network (GCN152-TL-A) from our previous work. It composes two main components: (i) the backbone network to extract features and ( ii ) the segmentation network to annotate labels. However, the accuracy can be further improved, since the deep learning network is not designed for recovering low-level features (e.g., river, low vegetation). In this paper, we aim to improve the semantic segmentation network in three aspects, designed explicitly for the remotely sensed domain. First, we propose to employ a modern backbone network called “High-Resolution Representation (HR)” to extract features with higher quality. It repeatedly fuses the representations generated by the high-to-low subnetworks with the restoration of the low-resolution representations to the same depth and level. Second, “Feature Fusion (FF)” is added to our network to capture low-level features (e.g., lines, dots, or gradient orientation). It fuses between the features from the backbone and the segmentation models, which helps to prevent the loss of these low-level features. Finally, “Depthwise Atrous Convolution (DA)” is introduced to refine the extracted features by using four multi-resolution layers in collaboration with a dilated convolution strategy. The experiment was conducted on three data sets: two private corpora from Landsat-8 satellite and one public benchmark from the “ISPRS Vaihingen” challenge. There are two baseline models: the Deep Encoder-Decoder Network (DCED) and our previous model. The results show that the proposed model significantly outperforms all baselines. It is the winner in all data sets and exceeds more than 90% of F1 : 0.9114, 0.9362, and 0.9111 in two Landsat-8 and ISPRS Vaihingen data sets, respectively. Furthermore, it achieves an accuracy beyond 90% on almost all classes.