基于深度语义分割的无人机多光谱遥感作物分类方法

Crop Classification Method of UVA Multispectral Remote Sensing Based on Deep Semantic Segmentation

摘要:

为精准获取农田作物种植分布信息以满足农业精细化管理需求，基于DeepLab V3+深度语义分割网络提出了一种面向无人机多光谱遥感影像的农田作物分类方法。通过修改输入层结构、融合多光谱信息和植被指数先验信息、并采用Swish激活函数优化模型，使网络在响应值为负时仍能反向传播。基于2018—2019年连续2年内蒙古自治区河套灌区沙壕渠灌域的无人机多光谱遥感影像，在2018年数据集上构建并训练模型，在2019年数据集上测试模型的泛化性能。结果表明，改进的DeepLab V3+模型平均像素精度和平均交并比分别为93.06%和87.12%，比基于人工特征的支持向量机（Support vector machine, SVM）方法分别提高了17.75、20.8个百分点，比DeepLab V3+模型分别提高了2.56、2.85个百分点，获得了最佳的分类性能，且具有较快的预测速度。采用本文方法能够从农田作物遥感影像中学习到表达力更强的语义特征，从而获得准确的作物分类结果，为利用无人机遥感影像解译农田类型提供了一种新的方法。

Abstract:

In order to accurately obtain the field crop planting distribution information to satisfy the needs of refining the management of agriculture, a field crop classification method was proposed for unmanned aerial vehicle (UAV) multispectral remote sensing images based on DeepLab V3+ network. In which the structure of the input layer was modified to fuse multispectral information with the prior features of vegetation indexes, and the activation function of Swish was adopted to maintain the backpropagation capability of the model when the response was a negative value. The research region was Shahaoqu irrigation field in the Hetao Irrigation District, Inner Mongolia Autonomous Region, whose UAV multispectral remote sensing images collected in 2018 and 2019 were taken as samples. The classification model was constructed and trained on the data of 2018, and the generalization performance of the model was tested on the data of 2019. The experimental results showed that the improved DeepLab V3+ model got excellent classification with fast speed. Its mean pixel accuracy and mean intersection over union were 93.06% and 87.12%, respectively, which were 17.75 percentage points and 20.8 percentage points higher than those of the traditional support vector machine (SVM) method using artificial features, and 2.56 percentage points and 2.85 percentage points higher than those of the original DeepLab V3+ model. Therefore, this method can learn more expressive semantic features from the field crop remote sensing images, thus obtaining accurate crop classification. The research result provided a new technical basis for the interpretation of farmland types using UAV remote sensing images.