基于CBAM-U-HRNet模型和Sentinel-2数据的棉花种植地块提取

Classification of Cotton Planting Area Using CBAM U HRNet Model and Sentinel 2 Data

棉花是我国重要的经济作物和战略储备物资,及时、准确地获取棉花空间分布信息对于棉花产量预测、农业政策的制定与调整具有重要意义。针对高分辨率遥感影像获取难度大以及传统机器学习对特征信息利用不足的问题,本文以新疆南部地区图木舒克市为目标区域,提出一种以U-HRNet为基本框架,融合CBAM注意力机制的CBAM-U-HRNet棉花种植地块提取模型。选择U-Net、HRNet和U-HRNet作为对比模型,评估CBAM-U-HRNet模型在Sentinel-2(10 m)和GF-2(1 m)2种空间分辨率数据集上的表现以及在棉花地块提取的优势。结果表明,基于Sentinel-2遥感影像的CBAM-U-HRNet组合模型对棉花地块的提取精度最优,mIoU和mPA分别达到92.78%和95.32%。与Sentinel-2数据集相比,空间分辨率更高的GF-2数据在HRNet、U-Net和U-HRNet网络上取得了更高的精度。对于两种不同空间分辨率的数据集,基于CBAM-U-HRNet模型的棉花地块提取精度较为接近,表明CBAM-U-HRNet模型能够减少由于数据集空间分辨率不同导致的错分。与随机森林算法相比,CBAM-U-HRNet模型对棉花地块提取的准确率更高。研究结果可以为干旱地区棉花识别与种植地块快速提取提供技术支撑。

Cotton is an important economic crop and strategic reserve material in China,timely and accurate acquisition of cotton spatial distribution information is of great significance for cotton yield prediction and agricultural policy development and adjustment.In order to address the problems of the difficult availability of high⁃resolution remote sensing data and insufficient usability of feature information by traditional machine learning,a CBAM U HRNet classification model was established to extract cotton planted area,where U HRNet and CBAM attention mechanism were combined,and Tumxuk City in the southern Xinjiang was taken as an study area.Firstly,the Sentinel 2 remote sensing data were pre⁃processed and annotated.Secondly,the attention mechanism CBAM was introduced into U HRNet to enhance the important features for cotton classification,suppress the relatively unimportant features,and reduce the interference caused by complex background information.Finally,U Net,HRNet and U HRNet were selected to compare with CBAM U HRNet model to test their performance in the classification of cotton planted area.During this process,two different spatial resolution datasets such as Sentinel 2(10 m)and GF 2(1 m)were used,and the advantages of CBAM U HRNet model were evaluated by using the best feature subset.The results showed the CBAM U HRNet model that using Sentinel 2 remote sensing data had the best classification accuracy for cotton planted area,with mIoU and mPA reaching 92.78%and 95.32%,respectively.Comparing with the Sentinel 2 dataset,the GF 2 data had higher spatial resolution and achieved higher accuracy by using HRNet,U Net and U HRNet networks.For the two datasets with different spatial resolutions,the classification accuracies of cotton planted area using the CBAM U HRNet model was comparable to each other.The CBAM U HRNet model can reduce the misclassification induced by the difference in spatial resolution of the two datasets.Comparing with the random forest algorithm,the CBAM U HRNet model had higher accuracy in the classification of cotton.The research results can provide technical support for the classification of cotton,and the fast and objective extraction of vegetation planted area in arid regions.