Key points are not available for this paper at this time.
The spatial resolution of land cover mapping has been increasing with the evolution of Earth observation technology. However, the higher spatial resolution makes it more laborious to collect training samples for efficient land-cover product updating. Fortunately, the existing historical products with a lower spatial resolution can be used as labels to achieve cross-resolution mapping with the latest images of a higher resolution. Although cross-resolution mapping can generate a large number of low-cost training labels, the labels can be noisy due to resolution mismatch or semantic errors. Furthermore, the existing deep learning based classification models have difficulty in maintaining the output resolution. In this article, a cross-resolution land-cover mapping framework with noisy label learning is proposed to complete high-resolution land-cover mapping based on noisy lower-resolution labels. The proposed method contains three parts: 1) To solve the resolution mismatch problem, the training labels are refined by modeling the spectral similarity and the spatial adjacency between the labels and the images using a conditional random field model. 2) To avoid the problem of resolution loss with the deep fully convolutional neural networks, a high-resolution deep semantic segmentation network is proposed to achieve deep feature extraction while maintaining the output resolution. 3) To eliminate the influence of semantic errors in the generated labels, a class-conditional label correction method is proposed, which detects and corrects the abnormal incorrect labels to facilitate network training. A national-scale land-cover mapping experiment for China was carried out. The 10-m spatial resolution land-cover map of China for 2020 was produced with improved accuracy based on the 30-m spatial resolution historical product, which shows the practicability of the proposed CRLC for rapid land-cover mapping.
Liu et al. (Thu,) studied this question.