Research On The Method Of Forest Land Extraction From High-resolution Images Based On Probabilistic Neural Network

For the global terrestrial ecosystems,forest land plays an important role in all aspects of the ecological field,such as regulating the terrestrial climate of the earth and conserving natural water sources.Because of the important ecological and economic value of forest land,the state regards forest land as an important technical and material basis for the precise implementation of sustainable development strategies.The high accuracy of the spatial distribution information of forest land is helpful for studying the regulation effect of forest land on climate change.The application of remote sensing technology in forest land detection provides strong support for forest land ecological monitoring and forest land resource data investigation and data management.By combining traditional remote sensing pre-processing technology with advanced computer image processing technology,researchers have achieved a great improvement in the degree of automation and performance of information extraction.One of the important research directions is the application of neural network technology.Since the application of remote sensing automatic processing technology has realized the popularization of remote sensing image classification,according to the current mainstream neural network learning algorithm,a remote sensing classification model of scene-by-scene training and scene-by-scene classification has gradually been widely adopted.However,problems such as repeated selection of training samples and repeated consumption of model parameters in the classification process still exist.For an image data with a large coverage area and inconsistent climate regions,how to achieve more high-precision forestland information extraction in a large-scale research area through a quickly established neural network classifier model is a problem to be explored.This paper selects the current mainstream three commonly used neural network classifier algorithms including deep belief network,Radial Basis Function neural network,and convolutional neural network,establishing a neural network classifierand compare the classification results of the above-mentioned different classifiers to verify the superiority of the probabilistic neural network classification method mentioned in this article in object-oriented information extraction.First,by extracting the scene-by-view features of the current image data set,and then separately selecting some remote sensing images and their characteristics in the typical woodland area from each land cover type research data set as training samples,the different neural networks selected Learning classifiers were trained separately;Secondly,the optimal model is to use the trained optimal neural network to learn the classifier to achieve the generalized extraction of woodland information in the townships in the entire study area;Finally,according to the land classification standards of the third national land survey,the advantages of each commonly used neural network classifier algorithm were comprehensively analyzed from the aspects of the effect of forest land information extraction,extraction efficiency,configuration and time requirements,and the probabilistic neural network classification was found.The classification accuracy of the filter is up to 90.3%,the kappa coefficient is 0.88,followed by the total classification accuracy of the convolutional neural network classifier reaches 90.2;the total classification accuracy of the Radial Basis Function neural network classifier is 86.4%;finally only the BP neural network The total classification accuracy is the lowest,at 79.04%.According to the classification results,the classification accuracy of the probabilistic neural network is the highest and the classification effect is the best.