Remote Sensing Images Change Detection Of Grape Plantation Based On U-Net

Remote sensing is an important data source for crop species identification and change detection.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 and change detection.China is one of the world’s largest grape producing countries.The accurate acquisition of the spatial distribution and change information 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.Based on GF-2 satellite remote sensing imagery and the U-Net model,this paper carried out research on grape planting regions remote sensing recognition and change detection methods.The main research contents and results are as follows:(1)Generation method of dataset for grape planting regions.Firstly,this paper used orthorectification,image registration,image sharpening and cropping operations to preprocess multi-temporal remote sensing imagery.Secondly,the grape planting regions in the multitemporal remote sensing imagery was manually labeled,and labeled images with the same size as the corresponding original imagery were generated.The original imagery and the labeled image were cropped.Finally,the data augmentation technology was used to generate pixel level dataset of grape planting regions,which laid the foundation for remote sensing recognition and change detection of grape planting regions.(2)Remote sensing recognition method of grape planting regions.In order to improve the accuracy of crop remote sensing recognition,the main improvements to U-Net were:(1)recalibrated the feature maps separately along channel and space adaptively by attention modules,to boost meaningful features and improve the accuracy of edge segmentation.(2)reduced the number of downsampling and used 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,MIo U(Mean Intersection over Union),and FWIo U(Frequency Weighted Intersection over Union)of the model on the test set were96.56%,93.11%,93.35%,respectively,which were 5.17,9.57 and 9.17 percentage points higher than those of the FCN-8s model,and 2.39,4.59 and 4.39 percentage points better than those of the original U-Net model,and 2.44,4.57 and 4.43 percentage points higher than those of the Seg Net model.In addition,this paper analyzed the impacts of the attention modules and hybrid dilated convolution on this model through ablation experiments.(3)Change detection analysis of grape planting regions.The trained model in(2)was used to predict the multi-temporal remote sensing imagery that had been preprocessed separately.Morphological filtering methods were used to post-process the prediction results.By comparing the prediction results of multi-temporal remote sensing imagery,the change detection results of grape planting regions were obtained.The experiments showed that the accuracy,F1-score,and Kappa Coefficient of the model on the test se t were 95.43%,86.39%,81.27%,respectively,which were 7.89,11.99 and 18.42 percentage points higher than those of the FCN-8s model,and 2.39,5.64 and 8.18 percentage points better than those of the original U-Net model.The proposed model is simple with few parameters,capable of identifying different sizes of grape planting regions with fine edge segmentation effect,and provides an effective way to improve the accuracy of crop remote sensing change detection.