Research On Forest Remote Sensing Data Analysismethodology Based On Machine Learning

With the rapid development of remote sensing technology,technologies and means for processing and analyzing multi-source remote sensing data are gradually increasing.Hyperspectral imagery（HSI）,Light detection and ranging（LIDAR）,and Unmanned aerial systems derived multi-spectral imagery（UAS_MSI）Multi-source data systems provide a way to capture forest-related information.In detail,hyperspectral images contain rich spectral information that can be used to classify forests.Laser detection and ranging data is used to provide structural information on the forest（eg,tree heights）.Based on the high-resolution multispectral data collected by UAV,it is possible to analyze the relative coverage area of vegetation in the forest.However,the distribution of land features in the forest is complex,and the available monitoring information for analyzing forests is more difficult to obtain.In order to detect the scope of the forest and draw the relative coverage of the plants in the forest,this paper proposes that use the two advanced machine learning algorithms（Multi-task learning（MTL）and Deep Learning（DL））for quantitative analysis of forests.The content of this paper is as follows:First,sparse representation（SR）uses the l₀,l₁-norm to process tasks.This algorithm does not consider specific domain information between various tasks.In order to supervise the case with fewer samples and make full use of the specific domain information contained in the surveillance information to improve the detection effect of the algorithm,this paper proposes a multi-task learning target detection algorithm based on spatial-spectral flow support（ICRTD_MTL）.The ICRTD_MTL algorithm consists of two parts:（1）Multi-feature learning（MFL）,which extracts tensors derived from different features of hyperspectral images（eg,spectral-value features（SVF）,spectral gradient features（SGF）and spectral texture features（TF）,etc.）.（2）Multi-task learning（MTL）.Assuming that each feature of the hyperspectral image has a specific contribution to subsequent task processing,the collaborate representation vector is solved by using an objective function that simultaneously optimizes the weights.Then,intermediate forest pixels are detected using the positions of the weight vectors in the collaborate representation vector.Finally,the spatial-spectral correlation of the pixels is used to improve the detection effect of the detector.The advantages of ICRTD_MTL are as follows:（1）Sparsity stability of different features of unknown pixels is guaranteed;（2）Pixels in a small area can share a common low-rank subspace.The experimental results of hyperspectral images in AVIRIS and HyMap are used to prove the performance of the algorithm,and ICRTD_MTL is used to achieve the detection of range of intermediate forest.Second,the flow-based multitasking learning target detection algorithm（CRTD_MTL）looks for associations between different tasks through l_2,1 regular terms and uses the sharing mechanism of associated information to achieve multi-task learning.The disadvantages of this flow-based sharing approach are:（1）The parameter selection process of the l_2,1regularization term is more complex.（2）The prior knowledge provided by the overcomplete dictionary is not used to optimize the sparse and low-rank processes of the joint representation coefficient matrix.Aiming at the above problems,a multi-task learning target detection algorithm based on probability graph（MTL_NFF）is proposed in the paper,that is,in the framework of multi-task learning（MTL）,according to Bayesian rule,using the Maximum A-posteriority（MAP）and Singular Value Decomposition（SVD）constitute a probability map to optimize the process of solving the collaborate representation coefficients of each node in multitasking learning.MTL_NFF establishes the connection between the input data and the output task through the probability map,improving the detector’s effectiveness.Then,the first short-wave infrared（SWIR1）region of the hyperspectral image was used to detect forest biomass changes,and biomass change maps and tree height information provided by LIDAR was used to quantitatively analyze forest changes.Finally,MTL_NFF is compared with the deep neural network algorithm results to prove the performance of the MTL_NFF algorithm.Finally,unmanned aerial systems derived multispectral imagery（UAS_MSI）have an extremely high spatial resolution（usually 0.07m²）.The single-layer SR algorithm based on l₀,l₁ norm can not effectively use UAS_MSI derived features to classify targets.In order to make full use of the tensor formed by UAS_MSI derived features to improve the classification accuracy under the condition of limited sample size,this paper proposes a deep neural network classification algorithm based on spatial-spectral sparse tensor（SDFS-DNN）.The SDFS_DNN contains multiple feature learning（MFL）,sparse autoencoders（AEs）and logistic regression classifiers.MFL is used to provide complementary,uncorrelated UAS_MSI derivative spectral feature tensors（DFS）.According to the similarity of the spatial neighborhood spectra in the image,the spatial-spectral tensor stack（SDFS）is obtained by using the information in the spatial neighborhood of the pixel to smooth DFS.Afterward,the sparsely constrained AEs multi-level sparse SDFS are used to provide features to the logistic regression classifier for classification.The advantages of SDFS_DNN are:（1）To make full use of the complementary polymorphic sparse feature tensors provided by the SDFS-based spatial-spectral-based joint structure（matrix-tensor）.（2）Using AEs layer-by-layer sparse SDFS to remove information redundancy in SDFS and extract spectral,geometric and texture features in SDFS.The relative coverage area of dead trees,lianas and living trees included in the study area was plotted by SDFS_DNN.Finally,the performance of the SDFS_DNN algorithm is demonstrated by comparing it with the SVM results.