改进DeepLabV3+网络的遥感影像农作物分割方法

针对于当前遥感影像农作物提取存在的识别精度较低、边缘识别效果较差、提取速度慢等问题，提出了一种改进DeepLabV3+网络的遥感影像农作物分割方法。将特征提取网络改为更轻量级的MobileNetV2网络，空洞空间金字塔池化模块中的普通卷积改为深度可分离卷积，大幅减少模型计算量，提高模型计算速度；在特征提取模块以及空洞空间金字塔池化模块加入双注意力机制，进一步优化模型边缘识别效果，提升模型分割精度。此外针对农作物数据集类别不平衡问题，引入加权损失函数，给予玉米、薏米与背景类不同的权重，提高模型对农作物区域分割精度。以2019年某地区的无人机遥感影像为研究对象，对玉米、薏米两种农作物进行分割。实验结果表明，改进DeepLabV3+算法像素准确率可达到93.9%，平均召回率可达到90.7%，平均交并比可达到83.3%，优于传统DeepLabV3+、Unet、Segnet等常用于农作物提取的分割方法，对农作物具有更好的分割效果。

Crop Segmentation Method of Remote Sensing Image Based on Improved DeepLabV3+Network

Aiming at the problems of low recognition accuracy, poor edge recognition effect and slow extraction speed in current remote sensing image crop extraction, a remote sensing image crop segmentation method based on improved DeepLabV3+ network is proposed. The feature extraction network is changed to a lighter MobileNetV2 network, and the ordinary convolution in the atrous spatial pyramid pooling module is changed to deep separable convolution, which greatly reduces the amount of model calculation and improves the calculation speed of the model. The double attention mechanism is added to the feature extraction module and the atrous spatial pyramid pooling module to further optimize the effect of model edge recognition and improve the accuracy of model segmentation. In addition, aiming at the imbalance of crop data sets, the weighted loss function is introduced to give different weights to corn, job’s tears and background classes, so as to improve the accuracy of crop region segmentation. Taking the UAV remote sensing image of an area in 2019 as the research object, corn and job’s tears are segmented. The experimental results show that the pixel accuracy of the improved DeepLabV3+ algorithm can reach 93.9%, the average recall can reach 90.7%, and the average intersection and merging ratio can reach 83.3%, which is better than the traditional segmentation methods commonly used for crop extraction, such as DeepLabV3+, Unet, Segnet and it has better segmentation effect on crops.