基于U-Net的葡萄种植区遥感识别方法

Remote Sensing Recognition Method of Grape Planting Regions Based on U-Net

为提高葡萄种植区遥感识别精度,基于高分二号卫星遥感影像,对U-Net网络进行改进:从空间和通道维度自适应校准特征映射,以增强有意义的特征,抑制不相关的特征,提升地物边缘分割精度;减少下采样次数,使用混合扩张卷积代替常规卷积操作,以增大卷积核感受野,降低图像分辨率的损失,提高对不同尺寸地物的识别能力。实验结果表明,本文模型在测试集上的像素准确率、平均交并比和频权交并比分别为96.56%、93.11%、93.35%,比FCN-8s网络分别提高了5.17、9.57、9.17个百分点,比U-Net网络提高了2.39、4.59、4.39个百分点。此外,本文通过消融实验和特征可视化证明了注意力模块和混合扩张卷积在精度提升上的可行性。本文模型结构简单、参数量少,能够识别不同面积的葡萄种植区,边缘分割效果良好。

The accurate acquisition of the spatial distribution of grape planting regions from remote sensing imagery is of great significance for optimizing the layout of grape planting regions and promoting the structural adjustment of grape industry.Due to the problems of the large differences in the size,unfixed spectral characteristics and complex background environment of the objects,it brings many challenges to accurate crop remote sensing recognition.In order to improve the accuracy of crop remote sensing recognition,a pixel-level accurate recognition method was proposed for grape planting regions based on the GF-2 satellite remote sensing imagery and the U-Net model was taken as the basic skeleton.The main improvements to U-Net were recalibrating the feature maps separately along channel and space adaptively,to boost meaningful features and improve the accuracy of edge segmentation,while suppressing weak ones,and reducing the number of downsampling and using hybrid dilated convolution instead of conventional convolution operation,to cut down the loss of image resolution and improve the recognition of objects of different shapes and sizes.The experiments showed that the pixel accuracy,mean intersection over union(MIoU),and frequency weighted intersection over union(FWIoU)of the model on the test set were 96.56%,93.11%and 93.35%,respectively,which were 5.17 percentage points,9.57 percentage points and 9.17 percentage points higher than those of the FCN8s model,and 2.39 percentage points,4.59 percentage points and 4.39 percentage points better than those of the original U-Net model.In addition,the impacts of the attention modules and hybrid dilated convolution on this model were analyzed through ablation experiments.The proposed model was simple with few parameters,capable of identifying different sizes of grape planting regions with fine edge segmentation effect,and it can provide an effective way to improve the accuracy of crop remote sensing recognition.