Desert classification is the fundamental for preventing and/or controlling desertification. Topographical features of desert remote sensing images change constantly due to the uncertainty of desert terrain, illumination, and other properties. Therefore, it is a very challenging task to accurately classify desert areas. In order to quickly and accurately classify desert from remote sensing images, this paper proposed a multi-scale residual network based on an attention mechanism. The network used conventional convolutions to perform preliminary feature extraction on images, and subsequently adopted a multi-scale residual module to further process the feature maps. Based on the idea of fusing multi-scale features, the multi-scale residual module effectively reduced information loss and possible gradient disappearance because of using skip connections. By introducing the attention mechanism, dependencies between feature channels were established, as a result, the network could recalibrate channel characteristic responses adaptively. Experimental results showed that the proposed network had better generalization ability and a higher accuracy on classification of multispectral desert remote sensing images compared with other methods.

沙漠分类是预防和/或控制荒漠化的基础。由于沙漠地形，光照和其他属性的不确定性，沙漠遥感图像的地形特征不断变化。因此，准确地对沙漠地区进行分类是一项非常艰巨的任务。为了快速，准确地从遥感影像中进行沙漠分类，提出了一种基于注意力机制的多尺度残差网络。该网络使用常规卷积对图像执行初步特征提取，随后采用多尺度残差模块进一步处理特征图。基于融合多尺度特征的想法，由于使用跳过连接，多尺度残差模块有效地减少了信息丢失和可能的梯度消失。通过引入注意力机制，建立了特征通道之间的依赖关系，从而网络可以自适应地重新校准通道特征响应。实验结果表明，与其他方法相比，该网络在多光谱沙漠遥感图像分类中具有较好的泛化能力和较高的分类精度。