Semantic Segmentation Of Poppy Field In High-resolution Remote Sensing Image Based On Deep Learning

Banning the cultivation of drug progeny such as poppy at source is one of the important tools to combat the drug and drug problem.Rapid and accurate detection of drug plant offenses,such as illicit poppy cultivation,is an important basis for accurately assessing the scale and dynamics of illicit cultivation,and for carrying out source and flow elimination efforts.The use of satellite or low-altitude remote sensing platforms has become the main means for national anti-drug departments or organizations to detect illicit cultivation activities of drug progeny plants,and its successful application has yielded great social benefits.In current anti-drug practice,target detection and area extraction for poppy illicit cultivation plots still face problems of accuracy and efficiency,which restrict rapid eradication and combat operations.Among them,the traditional methods of extracting poppy illicit cultivation plots using satellite images,such as visual interpretation and object-oriented,have certain limitations.The rapid development of target detection techniques in deep learning provides an effective technical approach for accurate and rapid identification of drug progeny plants,and has been initially applied.The existing applied research content focuses on the rapid detection of target locations of illicit drug plant plots,but neglects to focus on plot area information.However,the scale of illicit cultivation revealed by area information is also of very high operational value in anti-drug work and should be given sufficient attention.The target detection model can quickly and accurately obtain information about the location of poppy plots.However,target detection methods that want to obtain specific area information for each plot require further use of object-oriented or semantic segmentation methods.Therefore,this study was conducted to quickly and accurately obtain location information and area information of poppy illicit cultivation plots.This study was carried out to extract poppy plots using semantic segmentation models such as PSPNet on GF-1 satellite images.And this study was conducted to improve the PSPNet model for the characteristics of poppy illicit cultivation plots.The dissertation research work and results include the following:(1)This study was carried out to extract poppy plots using semantic segmentation models such as PSPNet.In this study,we train a poppy illicit cultivation plot extraction model based on UNet,PSPNet and Deep Labv3+.In this study,MPA(mean pixel accuracy)and MIo U(mean intersection ratio)were used as evaluation metrics to assess the accuracy of the model.And in this study,the efficiency of the model is evaluated using FPS(frames per second transmitted,which in this paper refers to the number of images detected per second)as an evaluation metric.The experimental results show that the accuracy and speed of poppy illicit cultivation plot extraction by PSPNet model are optimal.The PSPNet model achieves an MIo U of80.25%,an MPA of 87.59%,and a detection speed of 44.4 FPS.Therefore,it is optimal to choose PSPNet model as the base model for further improvement study in this study.And it is feasible to obtain information on the location and area of poppy illicit cultivation plots simultaneously and quickly using semantic segmentation.(2)Performance study of the improved I-PSPNet model based on PSPNet.In this study,model improvement is performed to address the problems that arise during the training and testing of PSPNet models.The improvements used in this study include replacing the backbone network,incorporating the SE module(Squeeze and Excitation Module)and the encoder-decoder structure and improving the loss function.Experimental results showing an increase in prediction speed from 44.4 FPS to 96 FPS by using the Mobile Netv2 backbone.This study solves the problem of segmentation results producing holes when extracting large plots of illicit poppy cultivation by introducing the channel attention SE module,with a 5% and 4.3%increase in MPA and MIo U.This study improves the accuracy of extracting the contours of poppy illicit cultivation plots by introducing an encoder-decoder structure,with MPA and MIo U improves by 0.8% and 0.45%.This study improves the positive and negative sample imbalance by using an improved loss function,and MPA and MIo U improves by 1.4% and 1%.The improved model improves the prediction speed by 90% over the PSPNet model,and MPA and MIo U by 6.2% and 4.3%.The IPSPNet model has good performance for RGB band datasets and GF-2 images.The main innovations of this study are the following:(1)In this study,the semantic segmentation technique is applied to quickly and accurately extract poppy illicit cultivation plots from high-definition satellite images.And in this study,the best algorithm based on PSPNet model was preferentially determined.(2)In this study,the PSPNet model is optimized and improved in four aspects:backbone network,SE module,encoder-decoder structure and loss function.The improved model improves the accuracy and speed of plot detection and can provide technical support for anti-drug authorities to quickly detect poppy illicit cultivation plots.The I-PSPNet model constructed in this study can quickly and accurately obtain information on the location and area of poppy illicit cultivation plots.And it can provide technical support for the anti-drug department to quickly discover poppy illicit cultivation plots,objectively assess the scale of illicit cultivation,and implement precise crackdowns on illegal drug criminal activities.