Extracting Farmland From UAV Images Based On Deep Learning Semantic Segmentation Model

In the process of intelligent agricultural development,extracting spatial distribution information of farmland and spatial distribution elements of crops from massive remote sensing farmland images captured by drones based on deep learning technology is the foundation and prerequisite for precise operation of drone AI.However,the current unmanned aerial vehicle(UAV)high-altitude operations still rely on manual assistance,mainly due to the low accuracy of extracting spatial distribution information of farmland from UAV aerial remote sensing images,especially at the edges of farmland and small farmland,which leads to low efficiency and high cost of UAV high-altitude operations(such as sowing,fertilizing,and spraying).In response to this problem,this thesis conducts specific analysis and research,The main research work and innovation of the thesis are as follows:(1)Aiming at the problem of low precision of spatial distribution information of farmland extraction from UAV images,this thesis proposes a method of improved DeepLabv3+ model with GhostNet as the backbone network.Firstly,the feature extracted from the backbone network is enhanced through the feature pyramid network;In order to further integrate multi-scale features,the 1×1 convolution of the atrous spatial pyramid pooling module(ASPP)in the encoder and decoder is replaced by the spatially aware standalone self-attention layer,and the dilation rate in the ASPP module is adjusted to improve the extraction accuracy of the farmland edge;Finally,without reducing the performance of the model,the feature concatenate is replaced by feature fusion(Add)to reduce the training parameters of the model.The experimental results show that the mean intersection over union(mIOU)of the improved DeepLabv3+ model can reach 94.57%,and the mean pixel accuracy(mPA)can reach 97.16%,which is 4.53% and 2.93% higher than that of the DeepLabv3+ model respectively,effectively improving the accuracy of farmland edge and small spatial distribution information of farmland extraction from UAV images.(2)In order to further improve the accuracy of extracting farmland spatial distribution information,there is an urgent need for a method that can quickly extract semantically complete and accurate farmland spatial distribution information.Due to the efficiency and robustness of the SegFormer model in the field of semantic segmentation,this thesis further studies and proposes a method for extracting farmland spatial distribution information from unmanned aerial vehicle images based on the improved SegFormer model.This method first optimizes the Transformer Block module in the encoder and introduces efficient channel attention(ECA).Then,the output results of each optimized block are inputted into the introduced bidirectional feature pyramid network(BiFPN)layer for enhanced feature extraction.Finally,the weighted fused multi-level features in the encoder are input into the lightweight decoder,and the multi-level features are aggregated through the multi-layer perceptron(MLP)layer to combine local and global attention,Then,the squeeze-andexcitation(SE)module is used to enhance channel features and improve model performance.The experimental results show that the mIOU and mPA of the improved SegFormer model are 96.73% and 98.45%,respectively;Compared to the improved DeepLabv3+,it is 2.16%and 1.29% higher,respectively.The IOU for farmland recognition reached 98.33%,which is 1.12% higher than the improved DeepLabv3+,further improving the accuracy of extracting spatial distribution information of farmland in UAV images.