AMFFNet: attention-guided multi-level feature fusion network for land cover classification of remote sensing images

Abstract

In the field of remote sensing, the classification of land cover is a pivotal and challenging issue. Standard models fail to capture global and semantic information in remote sensing images despite the fact that a convolutional neural network provides robust support for semantic segmentation. In addition, owing to disparities in semantic levels and spatial resolution, the simple fusion of low-level and high-level features may diminish the efficiency. To address these deficiencies, an attention-guided multi-level feature fusion network (AMFFNet) is proposed in this study. The proposed AMFFNet approach is designed as an encoder-decoder network with the inclusion of a multi-level feature fusion module (MFF) and a dual attention map module (DAM). A DAM models the semantic association of features from a spatial and channel perspective, and an MFF bridges the semantic and resolution gaps between high-level and low-level features. Furthermore, we propose a residual-based boundary refinement upsample module to further optimize the object boundaries. The experimental results indicate that the proposed strategy can considerably enhance the accuracy of land cover classification, achieving a mean intersection over union of 90.39% on the LandCover.ai dataset and 63.14% on the Gaofen Image Dataset with 15 categories (GID-15).

SABNet: Self-Attention Bilateral Network for Land Cover Classification

“…In order to merge global and local features extracted from both sides smoothly, we propose a decoder similar to a feature pyramid structure that fuses features progressively at different levels from bottom to top. It also enhances the representation of local features at the first layer by connecting it to the contextual path [48] using a double attention module consisting of parallel spatial attention and channel attention. Specifically, the SFF module has five layers, with the first layer using channel attention, spatial attention and two 3x3 standard convolution blocks, the structure of which can be seen in Figure 4, and the remaining layers consisting of two 3x3 standard convolution blocks and one 2x upsampling block.…”

“…To demonstrate the advanced nature of our method, we conducted a series of comparison experiments on landcover.ai between SABNet and nine other state-of-the-art land cover classification methods. To demonstrate the fairness of the experiments, in addition to the benchmark network BiseNet, we selected network structures with similar parametric numbers as SABNet, namely: UNet [16], Deeplabv3+ [26], PSPNet [17], SegNet [25], DIResUNet [51], FCN-8s [15], AMFFNet [48], HFENet [29], DEANet [52] and Segformer [41], and analysed each network. UNet, DeepLabv3+, SegNet, DIRe-sUNet and AMFFNet represent encoder-decoder networks.…”

“…To further validate the generalisation of the SABNet network, we conducted a series of comparison experiments using the GID-15 dataset with 15 classes and seven other advanced land cover classification methods, namely: BiseNet [22], PSPNet [17], Segformer [41], UNet [16], Deeplabv3+ [26], AMFFNet [48], HPSNet [53] and CAFHNet [54], and calculated three quantitative metrics as well as the number of parameters and computation of the model for the different network experimental results.…”