Research On Forest Parameter Extraction Of High Resolution Remote Sensing Images Based On Deep Learning

Forests are the second largest biologically suitable environment in the earth ecosystem,an important part of the global carbon cycle,and an important line of defense against global warming.Accurate real-time monitoring of forest parameters in a large area is helpful for future forestry development planning based on forest growth status.With the rapid development of China’s space industry,remote sensing technology,with its advantages of low cost,high efficiency and strong operability,has shown a strong application prospect in forest parameter observation.Optical remote sensing data contains abundant horizontal texture information.Synthetic Aperture Radar(SAR)data is more sensitive to forest canopy height information.In this paper,optical and SAR remote sensing data are used as data sources,and Jingyuetan National Forest Park in Changchun City,Jilin Province is taken as the experimental area to explore the forest parameter acquisition algorithm suitable for Northeast China.The specific research contents and achievements are as follows:(1)Study on forest type classification based on multi-temporal remote sensing data.The normalized vegetation index(NDVI),vertical vegetation index(PVI),bare land index(BI)and shadow index(SI)were used as characteristic parameters to establish the estimation model of forest canopy density grade by weight allocation method,and the canopy density grade mapping of the study area was obtained.Spring,summer,autumn selected one scene data,set up for remote sensing of forest types classification data sets,convolution neural network is adopted to establish the classifier based on pixels,overall forest types classification accuracy can reach 85.58%,joined the crown density information in the control experiment,the experimental results show that the join crown density information can significantly improve the classification accuracy is about 5%.(2)Forest Aboveground Biomass Retrieval Based on Multi-source Remote Sensing Data.By combining the spectral data of Sentinel-2 with the radar data of GF-3,a number of vegetation indices,such as ratio vegetation index(PVI),normalized vegetation index(NDVI)and soil-regulated vegetation index(SAVI),were extracted according to the spectral characteristics.For GF-3 radar data,the backscattering coefficients of four polarization modes were extracted in this paper,and the variables of high importance to the model were selected to participate in the construction of the inversion model of forest biomass and forest stock based on random forest.The prediction accuracy of the model is R~2=0.78,RMSE=50.56t/hm~2(3)A study based on improved normalized vegetation index.this paper,the trend distribution of 16 common vegetation indices in one year was analyzed,and the suggestion for selecting the superposition feature indices of multi-temporal data was given,and the feature indices which contributed to the classification of forest types were screened out.The improved red edge normalized vegetation index(RE-NDVI)was innovatively proposed,and the experiment proved that it could improve the classification accuracy of forest types when multi-temporal data were superimposed.Based on the perspectives of multi-source data fusion and multi-temporal data fusion,this paper explores the advantages of satellite remote sensing data for extracting forest parameters,and improves the existing algorithm.The experimental results show that the accuracy of forest parameters obtained by remote sensing technology in this paper can reach more than 84%on average.