Extraction Of Cultivated Land From High-resolution Remote Sensing Images Based On MST-DeepLabv3+ Model

Cultivated land is closely related to the economy and people’s livelihoods,and is an indispensable and important resource.However,with the rapid development of urbanization and industrialization in China,cultivated land resources are gradually occupied,threatening the goal of the "red line of cultivated land" policy.Therefore,rapid and accurate extraction of cultivated land information is extremely important to ensure sustainable agricultural development and national food security.Remote sensing images can provide rich information about surface objects,and remote sensing image classification is an important means to extract cultivated land information.With the continuous improvement of remote sensing image quality,high-resolution remote sensing images bring rich feature information.However,extracting its classification with high precision and efficiency is also a significant challenge.The semantic segmentation algorithm in deep learning has gradually become one of the most important tools for processing high-resolution remote sensing images because of its high classification accuracy and strong automatic learning ability.In this paper,the deep learning semantic segmentation model is applied to the extraction of cultivated land information from high-resolution remote sensing images to explore the method’s high precision and high efficiency.Classical semantic segmentation models usually have many training parameters and suffer from inaccurate and inefficient segmentation when performing image segmentation.Most of the methods for cultivated land extraction are only applicable to some specific datasets or specific study areas,and the generalization ability of the models could be better.To address these problems,the main research work and conclusions of this paper are as follows.(1)The DeepLabv3+ model is improved to obtain the MST-DeepLabv3+ model.The feature extraction network Xception of the DeepLabv3+ model is replaced by the lightweight network Mobile Net V2 to reduce the number of model parameters and improve the training speed;comparing different attention mechanism modules,the channel attention mechanism SENet is finally selected to join the model to make up for the accuracy loss caused by the lightweight network and improve the accuracy of semantic segmentation of remote sensing images;introducing transfer learning,the feature extraction network trained on the Image Net dataset is used as a pre-trained model to enhance the model’s ability to acquire features and improve the network segmentation accuracy.The results of several model comparison experiments on the ISPRS Vaihingen dataset and the GID dataset show that the MST-DeepLabv3+ model has excellent segmentation performance and can effectively solve the problems of different degrees of mis-segmentation,under-segmentation,and over-segmentation in the segmentation results of the dataset by the classical model,improve the phenomenon of inaccurate feature boundary and contour segmentation,and provide help for the subsequent cultivated land.In addition,the impact of the improved method on the model is also illustrated by the ablation experiment.(2)Study on cultivated land information extraction.Taking Taikang County of Zhoukou City as the study area,the PMS images of the Gaofen-1 remote sensing satellite were used to produce the Taikang cultivated land dataset.To further explore the methods applicable to cultivated land extraction,two methods,MST-DeepLabv3+(GID)with changing migration model and MST-DeepLabv3+(Nir RG)with a changing image band combination,are proposed on the basis of the MST-DeepLabv3+ model.These two methods were applied to the Taikang cultivated land dataset with MST-DeepLabv3+,PSPNet,UNet,and DeepLabv3+ models.The experimental results show that the three methods MST-DeepLabv3+(GID),MST-DeepLabv3+(Nir RG),and MST-DeepLabv3+ proposed in this paper,which have higher values of all evaluation metrics than PSPNet,UNet,and DeepLabv3+ models,can achieve a complete segmentation of the cultivated land area.It can smooth the segmentation boundary and eliminate the adhesion phenomenon,which can extract information about cultivated land with high efficiency and accuracy.Among them,the MST-DeepLabv3+ model has the highest accuracy and a better overall effect.The experiments confirm that all three models proposed in this paper can effectively extract cultivated land information,which can provide a reference for future high-resolution remote sensing image cultivated land information extraction.Among them,the MST-DeepLabv3+ model not only has the best segmentation performance in the cultivated land extraction task but also gets better segmentation results on other datasets with strong generalization ability.