FRF-Net: Land Cover Classification From Large-Scale VHR Optical Remote Sensing Images

Improved Land Cover Classification of VHR Optical Remote Sensing Imagery Based Upon Detail Injection Procedure

“…Next, efficient spatial pyramid (ESP) [36] refinement module combined with point-wise convolution is proposed for preserving large receptive field with fewer parameters and memory footprints for the detail feature description. In addition, the refinement module of FRF-Net [34] is designed for unifying channel numbers of the different feature layers and adjusting the feature maps slightly to improve classification performance. Thus, similar to depth wise separable convolution, a 1×1 convolution is used to extract the channel-wise dependency, while 3×3 convolutions focused on the spatial-wise dependency extraction.…”

“…DeepLab v3+ [14] utilizes an encoding-to-decoding structure to encode abundant contextual information, and a simple yet effective decoder is adopted to recover the detail information. In FRF-Net [34], an encoding-to-decoding process is designed based on two types of attention mechanism. Self-attention is set up to build the encoder for capturing long-range dependence.…”

“…For the encoding process, a great amount of research [22]- [34] focuses on exploring receptive field and capturing long-range dependency to improve pixel-level interpretation ability. The self-attention mechanism [37] is adopted as an efficient way to capture the long-range dependency in the feature domain as shown in Fig.…”

“…Recently, to improve the land cover classification performance of CNN-based methods in VHR optical remote sensing images, encoding-to-decoding frameworks and skip connections were widely employed in [25]- [34]. Liu et al [27] proposed the multi-scale full convolution network (FCN) to implement maritime semantic segmentation from optical remote sensing images.…”

“…Then, Yuan et al [33] also incorporated a wide-range attention unit for feature selection into a densely-connected U-Net (WRAU-Net) to achieve the road segmentation from remote sensing images. In our previous work of full receptive field network (FRF-Net) [34], feature-wise attention mechanisms were proposed to design a novel encoding-to-decoding process, which captured the globally consistent information efficiently. These abovementioned methods utilized attention mechanisms in the channel or feature domain to select more significant features or enhance their effectiveness, while the detail features are easy to be ignored.…”