Dominant Tree Species Classification Based On Multi-Source Remote Sensing Images With Deep Learning Method

It is helpful for people to better understand,manage and utilize forests from tree species classification map quickly and accurately.There is a large amount of spectral information in Hyperspectral remote sensing image(HSI),and spatial information for multispectral remote sensing data(MSI).The combined use of the two remote sensing data can play their respective advantages,which is helpful to the task of tree species classification.In recent years,deep learning methods have made great progress in tree species classification based on multi-source remote sensing data,and their classification results are generally better than other traditional algorithms.However,existing deep learning methods still have some problems in tree species classification.For example,the ability of feature extraction and learning is lower,and the potential information of multi-source data has not been deeply mined,which will lead to low accuracy of tree species classification.Computational redundancy leads to low classification efficiency.Classification depends on large-scale annotated information.In view of this,this paper studied remote sensing image data of forest farms in Tahe area taken by HJ-1A(small Environmental satellite)and Sentinel-2(Sentinel 2),and carried out the following work to solve the problems or defects of deep learning methods in tree species classification task:(1)Image preprocessing of HJ-1A HSI,Sentinel-2 MSI and class II small class survey data in the study area,including destriping,atmospheric correction,geometric correction,vector attribute data and raster data conversion,cropping and stacking,etc.For the HJ-1A image data,the dual-line interpolation algorithm is used to resampling,so that the resolution of the two remote sensing image data in pixel level is consistent.Three representative plots were selected in the study area and three multi-source tree species datasets were made.(2)In order to solve the problems of high redundancy,loss of depth features and low tree species recognition rate of existing tree species classification algorithms,A multi-source feature fusion tree species classification network based on Convolutional neural network(CNN)and Long short-term memory network(LSTM)is proposed.The aim is to improve the accuracy of network for tree species classification.Sentinel-2 data as Multispectral Image(MSI)input has high spatial resolution(HR),while HJ-1A data as hyperspectral Image(HSI)has low spatial resolution(LR).In order to mine the correlation between HSI and MSI,this paper uses the spectrum of LR-HSI and the corresponding spatial neighborhood in HR-MSI as the input of the network.The features of corresponding neighborhoods with two CNN branches are extracted in LR-HSI and HR-MSI,respectively.Then,the dual branches are connected and fed back to the activation layer,and finally the fusion layer outputs the spectrum classification map.The classification accuracy of the proposed feature fusion network model is 6% higher than that of single data source and 8% higher than that of other advanced classification methods,which achieves good performance and provides a new idea for the application of deep learning methods in forestry.(3)In order to solve the problem that the nonlinear activation function in the deep learning algorithm of tree species classification has gradient disappearance and negative values directly return to zero,a new activation function named Smish is designed in this paper.This function can not only guarantee negative activation and derivative value,but also maintain partial sparsity and regularization effect of negative input.The experimental results show that Smish improves the classification accuracy by 2% compared with Logish on CIFAR10,MNIST and SVHN.The classification accuracy of tree species using Smish’s lightweight model EfficientNetB3-Smish was improved by 4% compared with Logish.In addition,the ESDNet tree species classification model is designed based on EfficientNetB3-Smish,and the deep cross-attention mechanism module is introduced to enhance effective features and reduce redundancy.The experimental results show that the training duration of ESDNet network is reduced by about 50 minutes,and the classification is more efficient.(4)According to the problem of manual annotation difficulty,limited samples of half a supervised classification were more seriously,but there’s not tag samples in half a supervised classification method is easy to wrongfully convicted of defects,so this paper presents a multisource fusion hypergraph convolution half a supervision tree species classification model of neural network.The network head of HSI and MSI for canonical correlation analysis,the highlevel view features are extracted from the associated features,and the two data sources are projected into the same dimension to reduce redundancy.In addition,the weight matrix of HSI and MSI is fused to calculate the adjacency matrix of multi-modal graph,and the complementary and related information between HSI and MSI is deeply mined to enhance the global association ability of graph structure model.From the perspective of graph embedding,a graph-based loss function is introduced to constrain the fusion features extracted from the network and accurately guide the learning performance of the model.Through the experiment,the classification accuracy of the method reached 83.01%,and which achieved better effect of identification for the spruce which is difficult to be recognized.(5)Based on the problem of semi-supervised model which still need artificial mark for tree species classification,this paper puts forward the multiple source tree species classification based on pixel level since the supervised learning algorithm,this algorithm integrated the advantages of the typical comparative study and the supervised learning,with two different encoder multi-source data extracted different characteristics to enhance processing information respectively,The two auto-encoder networks proposed in this paper have good feature learning ability,and the designed comparative learning network can further learn better representative features from the two modules.From the process of feature fusion,the low-dimensional spatial structure information of data can be learned simultaneously in the unified network.The classification accuracy of the tree species dataset rises to 78%.The proposed method can obtain high-quality features and is more suitable for the task of tree species classification without labels.