基于可变形全卷积神经网络的冬小麦自动解译研究

Automatic Interpretation of Winter Wheat Based on Deformable Full Convolution Neural Network

以高分二号遥感影像为研究对象进行冬小麦多元特征的提取,在U-Net模型基础上进行改进,将一种可变形全卷积神经网络(DFCNN)模型引入到遥感影像自动解译领域。为提高网络模型对几何变化特征的提取能力,引入可变形卷积的思想,将可训练的二维偏移量加入到网络中的每个卷积层前,使卷积产生形变,并获得对象级语义信息,从而增强了模型对不同尺寸及空间分布的冬小麦特征的表达。使用DFCNN模型对数据集进行训练及微调,得到最优的网络模型,其像素精度为98.1%,解译时间为0.630 s。采用FCNN模型、U-Net模型及RF算法得到的冬小麦自动解译像素精度分别为89.3%、93.9%、90.0%,说明基于DFCNN模型的冬小麦自动解译精度相对较高,且对复杂的几何变化特征有较好的表达,具有较好的泛化能力。

China is a big producer of winter wheat.Obtaining the growth and distribution of winter wheat in a timely and accurate manner can provide a strong basis for China's agricultural policy and distribution of agricultural products.Complex geometric changes and foreign body phenomena in high-resolution remote sensing images limited the recognition ability of ground objects.The multivariate features of winter wheat were extracted from GF2.Based on the U-Net model,a deformable full convolutional neural network(DFCNN)model was introduced into the field of automatic interpretation of remote sensing images.In order to improve the ability of the network model to extract geometric features,the idea of deformable convolution was introduced.A trainable two-dimensional offset was added to the front of each convolutional layer in the network to deform the convolution and obtain object-level semantic information.Thus,the expression of winter wheat features with different sizes and spatial distribution was enhanced,and the interference of foreign bodies in high-resolution remote sensing images was eliminated.A deformable convolution module was added to the improved full convolutional neural network model,and the data set was trained and fine-tuned to obtain the optimal network model with an accuracy rate of 98.1%and a time cost of 0.630 s.Based on FCNN model,U-Net model and random forest(RF)algorithm,the accuracy of automatic interpretation was 89.3%,93.9%and 90.0%,respectively.The results showed that the winter wheat based on DFCNN model had the highest accuracy.Moreover,it can express complex geometric change characteristics well and had good generalization ability.