Deep Spatial Feedback Refined Network With Multilevel Feature Fusion for Hyperspectral Image Subpixel Mapping

The issue of mixed pixels is a prevalent challenge in hyperspectral images (HSIs), largely due to imaging modalities and hardware limitations. Subpixel mapping (SPM) can address this issue by distinguishing between land cover classes at a subpixel scale. In recent years, deep convolutional neural networks have demonstrated their potential and effectiveness for SPM. However, during the SPM process, the unique strengths of features at different network levels are often neglected, resulting in insufficient interaction between features at varying levels. Consequently, the network cannot fully extract and utilize the information from higher and lower level features nor can it thoroughly understand and recognize the input images. This ultimately hampers the model's SPM performance. In this study, we propose a deep spatial feedback refined network with multilevel feature fusion (SFRNet-MLFF) for SPM of hyperspectral remote sensing images. This innovative network incorporates a feedback mechanism and a multilevel feature fusion (MLFF) technique, both of which significantly enhance the interaction and aggregation of information across diverse features. Furthermore, throughout the entire network process, we apply solution space constraints to various spatial resolution contexts. These strategic implementations are expected to substantiate the network's performance in SPM. The experimental results on three hyperspectral datasets demonstrate that SFRNet-MLFF produces results with more distinct feature outlines. These results are better equipped to indicate the precise location and distribution patterns of land cover classes in low-resolution images at higher spatial resolutions. Furthermore, SFRNet-MLFF significantly improves the overall accuracy (OA) compared to the state-of-the-art deep learning SPM networks (DLSPMNets).