Research On Key Technologies Of Forest Tree Species Classification Based On Hyperspectral Remote Sensing Image

Forest natural resources offer a large number of biological habitats and material bases.The forest has a unique,beautiful natural landscape.It provides the human civilization of material and environment foundation of sustainable development,so the management of forest resources is essential,especially for the classification of forest tree species identification will be a vital research task.In the study of forest tree species identification,forest resource management is gradually transitioning from the means of field investigation,multi-source data,and multi-spectral remote sensing to the analysis of more convenient and quick hyperspectral data.However,there are two critical problems in studying forest tree species hyperspectral data classification: many spectra and difficulty in obtaining the label of sample data.This paper makes an in-depth study of these two problems respectively.In view of the redundancy of the number of bands in hyperspectral data,this paper proposes a band selection scheme based on deep learning interpretability.First,a 1D-CNN model is designed,and a small amount of hyperspectral data of forest species is used for training.Then,the thermal map with discriminant significance is calculated by Grad-CAM to guide the band selection.Then,in order to better display a more delicate heat map,this paper puts forward the Guided backpropagation theory to multiply the Guided backpropagation heat map and Grad CAM to get the Guided-Grad CAM heat map,which can also guide the band selection.Finally,a thermal map is selected,and the selected band is determined according to the brightness of the thermal map.SVM and KNN classification algorithms are used to evaluate the effect of band selection.The experimental results show that this scheme is effective in the classification of tree species of forest resources.It can effectively reduce the hyperspectral data dimension and visualize the critical bands of each tree species.It is not easy to obtain tags,so it is of great significance to choose more valuable tags.This paper proposes an active deep learning scheme.Firstly,a classification model based on Res Net-18 is designed,and model branches are extracted from the middle feature layer for the learning of uncertainty.The framework is composed of two modules and one is classification prediction module,the other is uncertainty prediction module;Then,the combined loss function is designed to have both the function of classification and the function of judging the relative relationship of data loss values.In each round of training,the two modules participate in the training at the same time,and the loss is updated synchronically.Each round of active deep learning processes is evaluated.The experimental results show that this scheme can effectively reduce the number of samples that need to be labeled in the hyperspectral classification of forest tree species,thus improving the classification accuracy of hyperspectral data of forest tree species.The classification effect is better than the traditional scheme in the classification of forest resource tree species.