改进D-LinkNet模型在国产卫星影像云检测中的应用

大部分国产卫星波段少，利用传统方法难以进行云检测。本文提出了基于改进D-LinkNet模型的云检测算法，并应用于国产卫星遥感影像的云检测。首先，利用自然资源部国土卫星遥感应用中心提供的人工勾云样本生成二值图标签；其次，对D-LinkNet50的编码器进行改进，使用带有通道注意力机制的ResNeSt50-Block替换原始的ResNet50-Block；然后，对损失函数进行加权，测试分析发现只用交叉熵损失作为损失函数时，检测精度更高；最后，使用条件随机场（CRF）对预测结果进行后处理。试验结果表明，改进D-LinkNet模型在测试集上的IoU提升了1.93%，精度提升了2.45%，保持了较好的云边缘信息，云检测效果高于原D-LinkNet模型。

Application of improved D-LinkNet model in cloud detection of domestic satellite image

Due to the lack of band of most domestic satellites, cloud detection is difficult using traditional methods. In this paper, a cloud detection algorithm based on improved D-LinkNet model is proposed and applied to domestic satellite remote sensing image cloud detection. Firstly, the binary image label is generated by using ResNeSt50-Block with channel attention mechanism. Secondly, the encoder of D-LinkNet50 is improved by using ResNeSt50-Block with channel attention mechanism to replace the original ResNet50-Block. And then the loss function is weighted, and it is found that the cross-entropy loss is the only loss function with higher precision. Finally, the conditional random field (CRF) is used to post-process the predicted results. The experimental results show that the MIoU and precision of the improved D-LinkNet model are improved by 1.93% and 2.45% respectively, and the cloud edge information is kept well. It can be used in cloud detection, and the effect is obviously higher than that of the original D-LinkNet model.