Wildland-urban Extraction From High Resolution Remote Sensing Image Which Based On Deep Learning

As an important part of urban ecological construction,Wildland-Urban plays an irreplaceable role in people’s health and sustainable development of cities and towns.It is not only limited to the scope of forestry survey,but also includes areas with high fragmentation of trees in cities and towns.Based on this,this paper puts forward the concept of Wildland-Urban and extracts Wildland-Urban resources in the study area.Distribution and growth status of Wildland-Urban can be judged by naked eyes from high-resolution remote sensing image.It brings complex visual information,but also mixes with the distortion of mixed spectral information and spatial information.The accuracy,information richness and algorithm stability of ground object recognition based on spectral information and hue are far from meeting the application requirements.Spatial texture information is the main feature of terrain type recognition.Due to the complexity and diversity of spatial texture features,many algorithms have been developed for quantitative description and object segmentation of spatial texture information(such as gray level co-occurrence matrix,multi-scale segmentation,etc.),they have certain effects in the application of specific data,specific objects and specific regions,but their universality is low,It has poor generalization.In recent years,the deep learning technology represented by neural network has achieved great success in the application of image pattern recognition such as face recognition and handwriting recognition,and the recognition accuracy has reached or even surpassed the level of human eyes.Theoretically,Neural network has the advantage of automatic and comprehensive application of all spectral and texture features of ground objects for recognition.The application of this technology in remote sensing image recognition is also an inevitable trend.However,there are some special problems in remote sensing applications,such as face recognition,handwriting recognition and so on.The main problem is how to construct a simple and functional deep learning network based on remote sensing texture features without reducing the recognition accuracy,so as to improve the universality and stability of the network to meet the needs of remote sensing image application in large areas,while avoiding the huge network(up to 1000 layers in depth)has dramatically increased the demand for sample(up to millions of samples).Therefore,exploring information extraction methods that conform to the characteristics and inherent laws of high-resolution remote sensing images has become a research hotspot in the field of high-resolution remote sensing information processing.This research shows the precision of Wildland-Urban extraction from high resolution remote sensing image which based on deep learning is higher than traditional object-oriented method.It could sufficiently improve the classification efficiency while ensuring the accuracy.This study has a certain reference value for more intelligent classification of remote sensing images.