Research On UAV Image Vegetation Recognition Method Based On Deep Learning

With the development of precision forestry technology,the demand of rapid and accurate vegetation recognition is increasing.However,the vegetation recognition for satellites and unmanned aerial vehicle(UAV)remote sensing images is still challenging due to complex handcraft feature extraction low recognition accuracy,slow processing speed and so on.Therefore,a method of UAV image vegetation recognition based on deep learning was proposed in this study.Anji County of Zhejiang Province and Jiufeng Forest Park of Beijing were selected as research areas for vegetation recognition.Five variant models of fully convolutional network(FCN)were built by multi-scale feature fusion.In order to further reduce the model size and improve the recognition speed,five variant models of MobileNet fully convolutional network(MFCN)were designed based on the model compression algorithm and multi-scale feature fusion.The method can automatically extract and learn the features of images during end-to-end training.The FCN and MFCN models were compared with the pixel-based classification method of ENVI software and the object-oriented classification method of eCognition software.The results demonstrated that the FCN-8s was the best model on the recognition of vegetation and the average overall accuracy for forest grassland and farmland was 88.95%,the average overall accuracy for Platycladus orientalis,Chinese pine and oak was 83.40%.The recognition accuracy of MFCN-2s model was slightly lower than that of FCN-8s,but the 3.35s runtime was the shortest,which was 23.62%of FCN-8s model,0.22%of the object-oriented classification method,and 0.002%of the pixel-based classification method.Therefore,the proposed method is feasible and effective,and it has high recognition accuracy and speed.This study facilitates the development of forest resources investigation and ecological monitoring.