期刊
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
卷 199, 期 -, 页码 28-39出版社
ELSEVIER
DOI: 10.1016/j.isprsjprs.2023.03.019
关键词
Convolutional neural networks; Very high-resolution images; Smallholder agriculture; Spatiotemporal transferability; Rice-crayfish fields
Accurate identification of crop types at the regional scale is crucial for various purposes in agriculture. This study developed a novel data-driven network, called RAUNet, based on very high-resolution images, which outperformed other methods in mapping rice-crayfish fields in China. The results show that RAUNet achieved high accuracy and had good transferability.
Accurate information on crop type spatial distributions is crucial for yield estimation, agroecological modeling, and agrarian policy development. In comparison with moderate resolution (10-30 m) satellite images, very highresolution (VHR, <= 1 m) images are superior at describing the spatial details and complex texture features of agricultural land. However, VHR images generally suffer from insufficient spatiotemporal coverage and high spectral variation in crop type, making fine-scale crop type mapping at regional or larger scales challenging. In this study, we developed a novel data-driven residual attention U-shape network (RAUNet) to identify crop types at the regional scale based on VHR images. RAUNet was adapted from UNet by introducing a residual module and a gated attention mechanism to strengthen the multilevel representation of crop features from VHR images. We selected rice-crayfish fields (RCFs) in Hubei Province, China, as the case crop type and area to test the performance of RAUNet. Seven models, including RAUNetpan, ResUNet, AttnUNet, UNet, DeepLabV3 + and random forest at the pixel and object levels, were adopted for comparison with RAUNet. Moreover, the model transferability was evaluated over regions with different agricultural landscapes. Our results showed that RAUNet-derived RCF maps obtained an average F1-score and MCC of 0.90 and 0.85 at different cultivation stages, which significantly outperformed other methods. Additionally, RAUNet had good spatiotemporal transferability, with an average F1-score and MCC of 0.93 and 0.88 in four transfer areas. The visualization results of the attention weights indicated RAUNet's prominent capability in combining useful and relevant information from different feature levels to produce refinement predictions for the target crop type. Furthermore, the highlevel features derived from RAUNet can well separate RCFs and non-RCFs over different landscapes, which were also consistent across different cultivation periods. These encouraging results suggest that RAUNet works well for fine-scale crop type mapping in smallholder farming systems, and it is an efficient method for land cover mapping based on VHR images at the regional scale.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据