Research On Semantic Segmentation Of Remotely Sensed Land Parcels Based On Convolutional Neural Networks

The plot is the basic unit of agricultural production.The precise delineation and location of agricultural plots is important for the dynamic monitoring of land resources and the efficient management of agricultural production.Traditional methods are based on edge detection models and area segmentation algorithms to extract plots with regular shapes,or to segment plots with prominent features based on texture and colour characteristics.However,these methods rely too much on the appropriate selection of parameters,require a lot of manual effort and long data processing cycles,and have low accuracy rates.With the rapid development of remote sensing technology in China and the in-depth application of convolutional neural networks in the field of image segmentation,a plot segmentation method combining remote sensing images and convolutional neural networks has been gradually developed.However,publicly available domestic remote sensing plots datasets are scarce,and there are few cases of segmenting plots in remote sensing images based on convolutional neural networks.In order to improve the efficiency and accuracy of remote sensing plots segmentation,this thesis collects remote sensing satellite images to establishes the dataset,builds and improves Trans UNet network framework,designs the vectorised post-processing steps,and conducts research on the semantic segmentation method of remote sensing plots:(1)A remote sensing plot dataset is established for common arable land characteristics in China.The GF-2 satellite data was collected and the geographic coordinate system of Arc GIS software was used to adjust the colour channels to highlight feature characteristics and to mark farmland in the images by manual visual interpretation.Once labelling is complete,crop and pre-process the image using mirroring,rotation,brightness enhancement etc.Finally,the entire data is divided into a training set,a test set and a validation set.(2)To explore the performance of different convolutional neural networks in remote sensing plot segmentation.The same experimental conditions were set and the UNet,Seg Net,Deeplab V3+ and Trans UNet networks were trained with the datasets in this thesis and the segmentation results were compared.The experimentally best performing Trans UNet network was then analysed and improved in terms of both residual modules and skip connections respectively,making it more suitable for segmenting high-resolution remote sensing plot images.Experiments show that the improved Trans UNet network can take into account both global semantic information and plot segmentation details,and performs better in segmentation of high-resolution remote sensing images.(3)In order to obtain better segmentation results at the lowest possible cost,the multi-resolution properties of remote sensing images and spatial location consistency are used to design the plot segmentation method in combination with vectorised post-processing steps.Two segmentation models are trained using both high and low resolution data,and these two models are used to predict both resolutions of the same region.Experiments have shown that using low-resolution data for training and segmentation can to reduce data costs and time consumption to one fifteenth.Due to the presence of fine porosity and uneven edge lines in the segmentation results,a combination of vectorisation operations and morphological processing is used as a post-processing solution to optimise the segmentation results and implement same-region data reuse.