Research On Land Cover Classification Algorithms With Spectral Data Based On Attention Mechanism

Spectral information implies the material composition and state of the target,and the spectral radiation differences basically reflect the inherent material differences of the target.For multitemporal multisensor crop classification and hyperspectral classification,the high dimension,strong correlation and nonlinearity of spectral data make it difficult to distinguish the objects with different radiation characteristics,especially the similar categories with little spectral differences.Attention mechanism in cognitive psychology and cognitive neuroscience is analyzed in this dissertation.The attention model in cognitive psychology explains how attention focuses on some high-priority information and ignores the other input.And in neuroscience,The electrophysiological studies,stimulation techniques,imaging studies,and individual data on neurological injuries all reflect the neural processing mechanisms of attention.In deep learning,the attention mechanism has been successfully introduced into many fields.Inspired by the research paradigm of sequence model,attention mechanism can be combined with hyperspectral classification and multitemporal crop classification,and then attention mechanism can be utilized to focus on the subtle spectral differences between similar categories.The main contributions of this dissertation are as follows:In order to distinguish the subtle spectral differences,the SA-CNN-GRU approach is proposed to perform hyperspectral classification.For the hyperspectral dataset,the pixelbased sample can be seen as spectral sequence.The CNN layers are used to extract the spatial features and GRU networks are utilized to model the spectral sequence.With using the proposed attention module,the different weights of bands which indicate different contributions can be obtained for the classification.The experiments show that the attention module can not only help to improve the classification accuracy of hyperspectral,but also perform quantitative analysis of spectral separability in a certain sense.For the case that the differences between similar crops are mainly subtle phenological differences,this dissertation proposes a classification model which focuses on the subtle phenological differences of crops.In the model,temporal attention mechanism can be described as reinforcing the phenological differences between crops and restraining their phenological similarities.So,benefitting from temporal attention mechanism,the fine phenological differences of crops can be distinguished.For the experiment with Sentinel-2 and Landsat-8 data,firstly,convolutional neural networks are used to unify the spatial-spectral scale of Sentinel-2 and Landsat-8 on pixel-based multiband images,and then a query module is introduced to search ”what is the important temporal information” over the whole sequence,through computing the attention weights the importance of each sequence time can be obtained.And the experimental results show that this method can achieve better results in distinguishing the categories with similar phenology.By borrowing the successful Transformer architecture in natural language processing(NLP),this dissertation also proposes a CNN-Transformer model for multitemporal crop classification.Multitemporal data can be viewed as a sequence of features,to deal with the sequence information,the Transformer architecture is borrowed from the knowledge of NLP to handle the correlations of sequence features,which has powerful modeling capability of sequence information.In the model,after first obtaining the unified multitemporal features,the feature embeddings and position embeddings of the sequence information can be easily calculated.Secondly,the encoder module derived from Transformer is used to express the correlation of the sequence,and by stacking encoder modules with some layers,the depth pattern characteristics of the sequence can be acquired.The experimental results finally illustrate that the framework can achieve the better classifications.Inspired by Squeeze-Excitation Networks(SENet)and model redundancy elimination strategy based on Interleaved Group Convolutions(IGC),this dissertation further studies the computational efficient multitemporal crop classification model based on group attention and interleaved group convolutions.The model utilizes Group-Squeeze-Excitation Networks to obtain the attention weights for each temporal group feature,and then uses“group feature recalibration” module to highlight phenological differences over the whole sequence.In addition,a regular convolution module with dense matrix operation can be replaced by an IGC module with two sparse kernels,so,the networks can be simplified and the number of parameters is greatly reduced.