Research Of Unmanned Aerial Vehicle Landcover Imagery Segmentation Method

As a vital part of global change research,landcover type identification can provide key guidance information for many projects.Landcover classification data's timely and accurate acquisition is of great practical significance.Traditional acquisition methods mostly use manual field investigation,which is expensive and impossible to obtain an accurate geographical distribution of different landcover areas.In order to obtain landcover classification data quickly and accurately,landcover images are collected through the lens carried by drone and automatic region segmentation and classification based on semantic segmentation is performed in this paper.Firstly,a high-precision heavyweight semantic segmentation model is proposed to achieve accurate region segmentation and classification of landcover imagery taken by unmanned aerial vehicles(UAVs).Based on Deep Lab V3+,targeted improvements have been made to constitute this model.Considering the characteristics of the UAV landcover imagery dataset,this segmentation model replaces backbone to accelerate model convergence,and add an adaptive joint upsampling module after the backbone to enhance the information communication between feature maps.And beyond that,this model adjusts the original atrous spatial pyramid pooling module to satisfy the dataset and upgrades the decoder to enhance the weight of position information.Experimental results show that,compared with original Deep Lab V3+,the optimized model improved pixel accuracy and mean intersection-over-union on the test dataset by 14.55 and 25.49 percentage points respectively,reaching 95.06% and 81.22%,which fully meets the requirements for high-precision landcover classification data acquisition.Secondly,based on the heavyweight segmentation model,the lightweight optimization of semantic segmentation model is explored and a new training method based on knowledge distillation is adopted in this paper.In this training method,the heavyweight model can guide the training process of student network LEDNet to improve its generalization ability through pixel-wise knowledge distillation,pairwise knowledge distillation and holistic knowledge distillation.Experimental results show that,with this training method,the pixel accuracy and mean intersection-over-union of student network LEDNet on the test dataset reached 82.09% and 70.81%,increased by 6.8 and 5.35 percentage points respectively.And student network can handle 30 images per second on NVIDIA 1050 ti GPU with model parameters being only 0.91 MB.In conclusion,this model can basically meet the needs of realtime processing and even on-end processing of UAV imagery.