【摘要】 针对遥感影像在时域上缺失或特征不对齐影响作物识别效果这一问题，在条件对抗域适应^[1]模型（CDAN）基础上提出一种基于可学习样本权重CDAN模型的作物分类方法。一方面，使用ResNet^[2]提出的并联卷积结构组成特征提取模块，对于低分辨率地块对象提取出丰富的特征；同时为解决困难样本给模型带来的负迁移问题，使用可学习的样本加权网络代替原模型直接使用熵计算的方式，来更好地度量样本的可迁移性。通过采集到的不同年份多月影像数据，在水稻分类任务上进行跨时域实验。结果表明，直接使用跨时域遥感影像进行预测会显著降低水稻分类精度，使用改进CDAN模型在多种迁移数据场景下的指标均有较大提升，最终分类精度达97%。更多还原

【Abstract】 To address the problem that remote sensing images are missing in the time domain or not aligned and thus reducing the accuracy of crop recognition, a crop classification method based on the conditional domain adaptation^[1]（CDAN） model is proposed with a learnable sample weight sub-network. On the one hand, the parallel convolution structure proposed by ResNeXt^[2] was used to extract abundant features for low-resolution field objects. At the same time, in order to reduce the negative transfer of difficult samples, the learnable sample weighted network was used to better measure the transferability of samples rather than entropy. Multiple months of images were collected to perform cross-domain experiment on rice classification. The results show that the cross-domain prediction of remote sensing images would significantly reduce the classification accuracy of the original model, and the improved CDAN model proposed in this paper has great improvement in several metrics in a variety of data transfer scenarios, with a final classification accuracy of 97.00%.更多还原