基于改进CBAM-DeepLab V3+的苹果种植面积提取

Apple Planting Area Extraction Based on Improved DeepLab V3+

为提高苹果种植区域的提取精度,提出了一种基于Sentinel-2和MODIS融合影像的CBAM-DeepLab V3+模型。影响苹果种植区域提取精度的主要因素包括遥感影像的质量以及语义分割模型的性能。从影像质量角度来看,采用基于时序的时空融合算法ESTARFM,通过融合Sentinel-2和MODIS的遥感影像数据,实现更高空间分辨率和时间分辨率数据的获取。与此同时,将训练样本从原始的800幅扩充至2400幅,为后续语义分割模型提供更为充足的样本容量。在语义分割模型优化方面,为了进一步提高苹果种植面积的提取精度,以DeepLab V3+网络结构模型为基础,引入基于通道和空间的CBAM注意力机制,进而发展出CBAM-DeepLab V3+模型。与原始DeepLab V3+模型相比,加入CBAM注意力机制的CBAM-DeepLab V3+模型在拟合速度较慢、边缘目标分割不精确、大尺度目标分割内部不一致和存在孔洞等缺陷方面实现了突破,这些改进提高了模型的训练与预测性能。本研究采用原始Sentinel-2影像及时空融合后的影像数据集,结合烟台市牟平区王格庄镇的数据集和观水镇苹果数据集对U-Net、FCN以及DeepLab V3+模型和CBAM-DeepLab V3+模型进行对比,研究发现在苹果种植面积提取方面,CBAM-DeepLab V3+优化模型所取得的MIoU为84.6%,苹果种植面积提取准确率达90.4%。U-Net、FCN和DeepLab V3+模型的MIoU分别为79.2%、75%、81.2%。此外,该模型预测的烟台市牟平区王格庄镇苹果种植面积为3433.33 hm^(2),与烟台市国民经济和社会发展统计公报公布的3666.66 hm^(2)相比,误差为233.33 hm^(2),预测准确率高达93.64%。

To improve the accuracy of apple cultivation area extraction,a CBAM-DeepLab V3+model based on the fusion of Sentinel-2 and MODIS satellite images was proposed.The main factors affecting the accuracy of apple cultivation area extraction included the quality of remote sensing images and the performance of semantic segmentation models.From the perspective of image quality,a time-series spatiotemporal fusion algorithm called ESTARFM was employed to fuse Sentinel-2 and MODIS remote sensing image data,achieving higher spatial and temporal resolution data.Simultaneously,the training samples were increased from the original 800 to 2400,providing more abundant sample capacity for the subsequent semantic segmentation model.In terms of optimizing the semantic segmentation model,in order to further improve the accuracy of apple cultivation area extraction,a CBAM attention mechanism based on channel and spatial information was introduced into the DeepLab V3+network,resulting in the development of the CBAM-DeepLab V3+model.Compared with the original DeepLab V3+model,the CBAM-DeepLab V3+model with the addition of CBAM attention mechanism achieved significant breakthroughs in terms of slower fitting speed,less accurate edge target segmentation,inconsistency in segmenting large-scale targets,and existence of holes.These improvements enhanced the training and prediction performance of the model.The original Sentinel-2 images and the spatiotemporal fusion images were used,combined with the datasets of Wanggezhuang Town in Muping District and the apple dataset of Guanshui Town to compare the U-Net,FCN,DeepLab V3+models,and the CBAM-DeepLab V3+model.The research findings indicated that in terms of apple cultivation area extraction,the overall accuracy(MIoU)achieved by the optimized CBAM-DeepLab V3+model was 84.6%,and the accuracy of apple cultivation area extraction reached 90.4%.In comparison,the MIoU of U-Net,FCN,and DeepLab V3+models were 79.2%,75%,and 81.2%,respectively.Additionally,the predicted apple cultivation area of Wanggezhuang Town in Muping District was 3433.33 hm^(2),with only 233.33 hm^(2) deviation compared with the data of 3666.66 hm^(2) published in the Yantai City National Economic and Social Development Statistics Report,resulting in a high prediction accuracy of 93.64%.