Study On Forest Land Classification And Dynamic Change Simulation Based On Deep Learning

As the main carrier of forest,forest land is the foundation of long-term survival and development of forest and an important part of ecological environment.At present,China has successfully drawn a national forest land "one map" based on 3S technology(remote sensing technology,geographic information system,global positioning system).Its formation can greatly improve the level of fine,automatic and intelligent management of forest resources.it is of great significance to study the changing law of forest resources and environment and promote the construction of ecological civilization in our country.At present,with the continuous development of human economy and society and the continuous change of global climate,the forest environment is also changing,but due to the complexity and diversity of forest resource system and the untimely census data,the update of "one map" lags behind.It is difficult to meet the needs of forest resources change monitoring.Therefore,there is an urgent need to study automatic forest land monitoring techniques and methods to achieve accurate extraction of forest land information.With the rapid development of satellite sensor technology and artificial intelligence technology,all kinds of remote sensing images are becoming more and more abundant,and deep learning technology is becoming more and more mature.The use of remote sensing data and deep learning technology to extract forest land information has become a hot topic in the field of forestry.Taking Maiji District of Tianshui City,Gansu Province as the main research object,this paper studies the classification of complex forest land by using conditional random field model and deep convolution neural network,and puts forward a double-pool convolution neural network architecture.to provide a strong guarantee for the renewal and construction of "one map" of forest land.At the same time,based on the simulation method of forest land dynamic change based on stack sparse self-coding network,the spatio-temporal evolution law and driving mechanism of forest land are analyzed to realize the dynamic monitoring of forest land resources.The main research contents are as follows:(1)Forest land classification method based on random forest-conditional random field modelForest land resources have complex and diverse characteristics,and the classification method based on a single pixel is often difficult to achieve a good classification effect.Integrate multi-temporal remote sensing data and spatial information into the classification process,use the random forest method to calculate the classification probability,and use the classification probability as the correlation potential function of the conditional random field model,while considering the spatial relationship between the pixels,and use the Gaussian function to construct the conditions The second-order potential function of the random field model realizes the accurate classification of forest land.Finally,it is compared with traditional machine learning classification methods,such as support vector machine,maximum likelihood classification,and decision tree.The results show that the classification method combined with random forest and conditional random field model has the highest classification result accuracy.For Landsat and Sentinel2,the overall classification accuracy(OA)is 86% and 87.9%,respectively,compared with the original random forest The classification accuracy has been improved by about 3.5%;the user accuracy(UA)is 86.6% and 88.4%.(2)Forest land classification method based on the combination of dual-pooling convolutional network and conditional random fieldDue to the diversity of forest land types in different periods,the spectral values of the same ground objects on remote sensing images cannot obey the same probability distributionondifferent time and space.Based on the deep convolutional neural network,this paper uses a multi-spectral remote sensing image classification methodindual-channel convolutional structure,which extracts the joint features of spatio-temporal spectra from multispectral remote sensing images.At the same time,this paper proposes a dual-pooling convolutional neural networkby analyzing the advantages and disadvantages of different pooling methods.Then,a dual-pool convolutional neural network coupled with the conditional random field model is further proposed,in whichthe output of the convolutional neural network is used as the first-order potential function of the conditional random field model,and the Gaussian function is used as the second-order potential function.The experimental results show that the overall classification accuracy(OA)of the dual-pooled deep convolutional neural network structure combined with the conditional random field model on Landsat8 and Sentinel2 data is92.5% and 95.3%,and the user accuracy(UA)is 93% and 95.5%.The classification accuracy on the forest land is higher than the one-dimensional,two-dimensional,two-channel convolutional neural network and dual-pooled deep convolutional neural network methods,and good classification results are obtained.(3)Forest land distribution simulation based on stacked sparse Stacked Auto-encoders and cellular automata modelIn order to analyze the temporal and spatial variation characteristics of different forest land types,the time-phase classification and comparison method is used to analyze the changes from the aspects of forest land typemore accurately,distribution,area,etc.,and we study the conversion relationship between various types of forest land over the years,and analyze the forest land in different periods.In this study,the relationship between each driving factor was established with stacked auto-encoder and logistic regression,and the analysis and study of the spatial-temporal pattern change of forest land resources were carried out based on the cellular automata method of spatio-temporal simulation of forest land change trend.the research results show that: The accuracy of SAE-CA simulation is 87%,which is 3% higher than that of LogisticCA.By simulating the change of forest land in 2018-2028,it is found that construction land shows an increasing trend,and the forest land area decreases slightly,The main reason is that human activities are frequent,which speeds up the development and utilization of land resources.In summary,based on multi-temporal feature images,this paper studies the methods of forest land classification from two theoretical perspectives of statistical learning and deep learning.From the perspective of statistical learning classification,fully considering the spatial correlation between pixels,a method of forest land classification based on random forestconditional random field model is proposed.From the perspective of deep learning theory,a dual-pooling convolutional neural network architecture is proposed,which improves the ability of network feature extraction.At the same time,the conditional random field model is used to optimize the output of the network to make the classification results more reasonable in spatial distribution.Finally,this paper applies the proposed classification method to the simulation of forest land,and proposes a method for simulating the dynamic changes of forest land in a stacking sparse autoencoder network by taking full advantage of the nonlinear fitting advantages of deep neural networks based on the cellular automaton model,which provides a theoretical reference for updating and constructing a map of forest land.