| In recent twenty years,the area of Plastic-Mulched Landcover(PML)in China has been increasing,and PML has become an important agricultural landscape.Plastic mulched planting has greatly improved the yield and the quality of crops in the region and our whole country,but it has induced the problem that the residual plastic mulches has changed the soil structure and threatened the sustainable development of agriculture.Monitoring the spatio-temporal distribution of PML can provide data basis for agricultural water resources allocation,residual plastic mulches recovery and precision agriculture development in PML regions.Some progress has been made in remote sensing information extraction of PML,but the accurate and rapid extraction of PML is still facing great challenges because its spectral characteristics are easily affected by crop types and growth conditions,soil types and moisture,and farming methods.Convolutional Neural Network(CNN)models can automatically extract the deep features of target ground objects and improve the classification accuracy of remote sensing images.Therefore,this paper,selecting Guyuan,Ningxia(where the PML is concentrated)as study area,focused on applying deep learning methods to extract PML from self-made PML data sets of high-resolution remote sensing images.The main research works include:(1)Nine mainstream CNNs and four fine-tuned AlexNet models are trained and tested,and the accuracies of scene classification results from these 13 CNNs are compared.(2)A Two-Step Ensemble Convolutional Neural Network(TSE-CNN)algorithm is proposed for further improving the scene classification accuracies.Combining the ensemble learning idea,the algorithm integrates13 CNNs as the first step to generate the base classifiers for each land use type;in the second step,the weighted integrated base classifier obtains the final scene classification result.(3)By analyzing 7 Atrous Spatial Pyramid Pooling(ASPP)schemes and choosing [6,12,18],the DeeplabV3+ semantic segmentation models based on 4 mainstream backbone networks were trained and applied to extract pixel-based PML information,and the influence of data enhancement on semantic segmentation accuracies were analyzed.(4)A Threshold-Scene-Classification-Based Semantic Segmentation(TSCB-SS)approach,which applied DeeplabV3+ models for detecting PML from the results of scene classification combining threshold method,was proposed and adopted for extracting pixel-level PML.The research results mainly include:(1)The overall accuracies(OAs)of scene classification of 13 CNN models are from 70% to95%,10 CNNs achieve OAs higher than 90%,and RestNet34 attains the highest OA(94.34%)and F1 of PML(97.61%).(2)Compared with the single CNN model obtaining the highest classification OA,the proposed TSE-CNN algorithm has improved OA by 1.22%,which indicates that TSE-CNN is effective for detecting PML from high resolution remote sensing imagery.(3)The backbone networks acquiring the highest Intersection-over-Union(Io U)of plastic-mulched farmland and greenhouse farmland are different,namely ResNet18(0.8190)and Xception65(0.8418).Having experimented with datasets of different spatial resolutions,it is suggested that the trained semantic segmentation models have certain generalization abilities.(4)Data enhancement can improve F1 values of PML scene classification with the mainstream CNNs,while it can not significantly improve the semantic segmentation accuracies of the validation images.(5)The TSCB-SS approach can greatly reduce the computational time for identifying pixel-level PML information,and can effectively improve the Io U of pixel-based classification of greenhouse farmland from validation images by 0.2059. |
