Research On Vegetation Classification Of UAV Remote Sensing Image Based On Semantic Segmentation Model

The segmentation of UAV Vegetation-cover image was complex and not precisely.To this end,a segmentation model with good accuracy was developed,which is suitable for UAV visible-light bands image segmentation to improve the accuracy of vegetation classification.The main research contents and conclusions of the paper are as follows :(1)In this study,the UAV was used to acquire Vegetation-cover images,and the Vegetation-cover image data set was established by the UAV visible-light bands images which were annotated at pixel level.Firstly,based on original Deep Lab V3+ model,several improvements were made,Then,the image feature extraction algorithm in digital image processing technology was combined into the improved Deeplab V3+ model,and several different ways of combination were discussed.(2)The research uses four basic models of Random Forest,Seg Net,U-Net,and Deeplab V3+ for experimental comparison at first.Among them,three semantic segmentation models performed better than the traditional image segmentation method Random forest algorithm,and Deeplab V3+ performed best.(3)The improved Deeplab V3+ model can achieve better recognition results on UAV vegetation coverage images.The amount of presented model parameters is only 6.74% of the original model.It performed well with pixel accuracy of 94.97% and mean intersection-over-union of 74.08% on the validation set,which was 1.33% and 7.65% higher than that of the original Deep Lab V3+ model,it can obtain Vegetation-cover data faster and more accurately,(4)It didn't improve the model's indicators by using image features as input alone,and the image feature combined with the original image input method can make the model get better segmentation results.Compared with the improved Deeplab V3 model,the Accuracy of the original image combined with HSV color feature as input is increased by 2.18%,while Precision,Recall increased by 0.57%,2.24%,and MIOU increased by 2.40%,reaching 76.48%.In summary,this paper uses the vegetation-cover image dataset as the research data,and uses the semantic segmentation method as the research method to realize the semantic segmentation of the vegetation image,the method proposed in this paper has certain validity and applicability,it can obtain vegetation coverage data more accurately,which provide a reference for the extraction of vegetation coverage from UAV visible light band images.