Research On Precise Classification Method Of Forest Based On Deep Learning And Sentinel-2A Remote Sensing Images

Remote sensing technology is widely used in forestry survey.Nowadays,machine learning algorithm is mostly used in the task of forest land precise classification based on remote sensing image,which is difficult to extract deep features of data and cannot take advantage of multi-spectral images.To explore the application method of deep learning in the forest land precise classification with multi-spectral remote sensing,this thesis designs a processing method for Sentinel-2A remote sensing data and forest land precise classification task based on convolutional neural network.In the method,the characteristics of data and tasks are fully considered,and the spatial spectrum characteristics of remote sensing data are extracted and utilized efficiently.Through optimization and post-processing,the method has certain generalization in different years,so that it can have better application value in forest survey.In terms of data preprocessing,to comprehensively utilize the spectral,texture,cloud and geological information,this thesis uses SVD to reduce the dimension and clean the multi-temporal remote sensing data,then uses the object-based method with sub-Compartment as the object to process data.After process,superimposing processed data with the raw data,cloud distribution and geological data to generate the multi-temporal composite data set.In terms of neural network model,this thesis takes Mobilenet-v3 as the backbone network,according to the characteristics of composite data and forest land classification standards,Gaussian spatial spectrum data enhancement and multi-output feature sharing classifier are used at the input and output respectively.The experimental results show that the proposed model have average results of80.88%,83.20% and 73.70% in three classification levels of the test area,and the highest classification result of object-based regional is 83.43%,88.95% and 78.45%.In terms of model optimization and post-processing,this thesis explores the portability of model’s transferability between different years.By comparing the results of the multiple proportion of training sets,it is proved that the performance of the transfer model can be improved by an average of 2.73 %compared with that of the non-transfer model under the same conditions,and the improvement effect increases with the decrease of the proportion of training sets.To further compress the model volume and remove the over-fitting phenomenon,this thesis adopts the method of weight pruning to eliminate the redundant parameters from the model.The experimental results show that the effective pruning ratio of the model ranges from 20 % to 60 %,and the relative optimal effect is obtained at 30 %,which improves the classification accuracy of the model by an average of 0.27 %.Through a number of experiments,this thesis obtained a set of forest land precise classification methods based on Sentinel-2A data and convolutional neural network,explored the application value of medium-resolution multi-spectral remote sensing and deep learning in forestry investigation,and provided a new idea for forest land classification task.