Prediction Of Dynamic Evolution Of Land Cover In Remote Sensing Images Based On Deep Learning

Since the industrial age,land use/cover change（LULCC）is crucial to promoting regional and global climate change by promoting the energy cycle and material exchange on the surface of land.Global academia,policymakers and stakeholders believe that space-time LULCC simulation is an effective tool to promote scientific understanding of the response and feedback between land cover change and its socio-economic and natural environment drivers.This complex link and feedback structure is expected to be solved through simulation model to predict future surface cover trajectory and support future land use decision-making.With the development of remote sensing technology and digital image processing technology,more and more people use computer to process remote sensing images to predict land cover change.At present,the application of deep learning in land cover prediction mostly stops at remote sensing change detection.In order to improve the traditional change detection task,we can not only detect which land cover has changed,but also predict the statistical indicators of land cover change and simulate the evolution of land cover in the future.In this paper,we propose two frameworks for land cover prediction of remote sensing images,at the same time of using deep learning models,the spatial transformation module is introduced to make the prediction results more intuitive,so as to realize the prediction of partial boundary evolution of Lake Urmia.In this paper,a series of researches on land cover prediction of remote sensing images are carried out around the deep learning model.The main research contents and innovations of this paper include the following points:（1）A data preprocessing method of gray gradient is proposed.Due to the influence of different factors such as sensor structure,earth rotation,weather,terrain and so on,the consistency of remote sensing data in different time phases can not be fully guaranteed.Therefore,the preprocessing of radiometric correction,geometric registration and so on is generally required before the change detection and other tasks.For the purpose of research,this paper proposes a preprocessing method of gray level linear gradient from the boundary of land cover to the center of land cover,which provides a new preprocessing scheme for land cover prediction.（2）A model of land cover prediction based on U-Net which can be trained end-to-end is proposed.The model uses U-Net network to extract multi-level change features,analyze the trend of land cover change,and predict the future land cover situation.At the same time,in order to optimize the prediction model,the existing model optimization algorithms are used,such as regularization method for over fitting phenomenon,dropout and so on.Compared with traditional prediction models such as Markov model and cellular automata,the model can mine the deep information of surface cover change,automatically fit the trend of time series data change,and compare with the existing depth learning methods,the model can detect and predict the surface cover change at the same time.The prediction results not only include the change statistics,but also show the future space-time information of the cover situation.（3）A land cover prediction model based on U-Net-LSTM which can be trained end-to-end is proposed.Based on（2）,the model introduces LSTM module to realize different memory and forgetting of historical data and the utilization of current input data to different degrees.How to extract the state change of land cover from the time series of remote sensing images and generate the rules of migration change is of great significance for land cover prediction.Here,it is an effective scheme to use LSTM to predict land cover.