Research And System Implementation Of UAV Grassland Pasture Segmentation Based On MDCD-DeeplabV3+ Network

Grasslands are an important part of the earth’s ecosystem and play an irreplaceable role in the development of livestock farming.With the intensification of global climate change and unreasonable human exploitation,grassland ecosystems have been destroyed,so effective management and restoration of grasslands is an important foundation for achieving national sustainable development.At present,the digital management platform of grassland uses more macro-scale satellite remote sensing,but its overall resolution is low,high cost,long cycle time and vulnerable to weather,so its application is somewhat limited.As a new type of low-altitude remote sensing platform,UAV makes up for the shortage of satellite remote sensing with many advantages such as lightness,high resolution and low cost.UAV pasture segmentation is an important part of the digital management of grassland,providing data support for grassland research.However,due to the complex grassland environment,the scattered distribution of forage and the inconspicuous color difference between different growth periods and background,the UAV forage image segmentation needs further improvement.Therefore,this thesis conducts an in-depth study on the UAV grass image segmentation method using deep learning model.The main research contents are as follows:(1)Establish UAV pasture dataset and compare and study the UAV pasture segmentation method based on full convolutional neural network.The current mainstream image semantic segmentation algorithms are studied,and Seg Net,UNet,PSPNet,HRNet and DeeplabV3+ semantic segmentation networks are used to segment the dataset established in this thesis,and the average intersection ratio(m IOU),average pixel accuracy(m PA),average precision(m Precision),average recall(m Recall),Flops,Param,and Inference times are compared and analyzed to select the best segmentation network.The comparison concludes that DeeplabV3+ has the best segmentation performance with m IOU,m PA,m Precision and m Recall reaching 86.46%,94.76%,90.11% and 90.76%,respectively,with only 102.589 G and 54.709 M in the number of parameters and computation,and only1.04 s in the single test set image segmentation time.(2)Proposed MDCD-DeeplabV3+ based UAV grazing segmentation network.The network uses DeeplabV3+ as the base network,firstly,the lightweight Mobilenet V2 network is selected as the backbone network for initial feature extraction,and the network configuration is adjusted to suit the forage segmentation task;secondly,the depth separable convolution is used instead of the normal convolution to lighten the network;in addition,the DASPP module is used to capture a larger perceptual field to enhance the interaction between the features;meanwhile The hybrid attention mechanism(CBAM)is introduced to reassign weights to enhance feature extraction.The experiments demonstrate that compared with DeeplabV3+ network,the proposed MDCD-DeeplabV3+ network m IOU,m PA,m Precision and m Recall improve by 5.26%,3.36%,5.45% and 4.27% respectively,and the computational and parametric quantities decrease by 87.437 G and 50.294 M,and the segmentation speed are also improved.The results prove the effectiveness of the model in this thesis.(3)Design and implement the UAV forage segmentation system based on MDCDDeeplabV3+.The system uses Py QT5 for front-end UI interface design and Python language for back-end operation implementation,and the MDCD-DeeplabV3+ network proposed in this thesis is embedded in the system to realize forage segmentation.The main functions of the system include picture and video upload,picture and video segmentation,pause or continue of video segmentation,end of video segmentation,etc.