Classifying land cover using high-resolution remote-sensing images is challenging. The emergence of deep learning provides improved possibilities, but owing to the limitations of network structures, traditional convolutional neural network methods lose essential information about boundaries and small ground objects. We propose a superpixel-optimized convolutional neural network (SOCNN) framework to overcome this weakness. The SOCNN includes three modules: a semantic segmentation module, a superpixel optimization module, and a fusion module. The performance of the first module was evaluated using several common networks. PSPNet outperformed other networks, obtaining a pixel accuracy of 83.25%, a Kappa coefficient of 0.7862, and a mean intersection over union of 64.19%. Weighted loss was introduced to alleviate the effect of category imbalance, and the class pixel accuracy of category 11 improved by 19.77% with a weight of 20. The subpixel model was evaluated, and the pixel accuracy reached 83.43% with the superpixel-FCN method. Our superpixel optimized module improved the pixel accuracy of the object boundary by 1.37% when the fusion factor was 0.65. These results show that the SOCNN method is effective for recovering boundary information.

中文翻译：

---

基于PSPNet和超像素分割方法的高空间分辨率多光谱遥感影像土地覆盖分类

使用高分辨率遥感图像对土地覆盖进行分类具有挑战性。深度学习的出现提供了改进的可能性，但由于网络结构的限制，传统的卷积神经网络方法丢失了关于边界和小地面物体的基本信息。我们提出了一个超像素优化的卷积神经网络 (SOCNN) 框架来克服这个弱点。SOCNN包括三个模块：语义分割模块、超像素优化模块和融合模块。第一个模块的性能使用几种常见网络进行评估。PSPNet 优于其他网络，获得了 83.25% 的像素精度、0.7862 的 Kappa 系数和 64.19% 的联合平均交集。引入加权损失以减轻类别不平衡的影响，在权重为20的情况下，类别11的类别像素精度提高了19.77%。对亚像素模型进行了评估，使用superpixel-FCN方法的像素精度达到了83.43%。当融合因子为 0.65 时，我们的超像素优化模块将物体边界的像素精度提高了 1.37%。这些结果表明 SOCNN 方法对于恢复边界信息是有效的。