Research On Hyperspectral Image Classification Based On Convolutional Neural Network With Small Samples

Optical remote sensing is a passive technology that captures spectral reflection data from surface objects using satellites,aircraft,or drones.Hyperspectral sensors are capable of dividing the visible and infrared spectral ranges into tens to hundreds of narrow bands.As a result,the acquired images contain high-resolution spectral information,enabling more precise characterization of the spectral properties of ground objects.Currently,hyperspectral images have been widely used in agriculture,environmental monitoring,geological exploration,urban planning,and various other fields.The classification task serves as the fundamental prerequisite for various applications involving hyperspectral images.Its primary objective is to assign a distinct class label to every pixel within the image.Traditional shallow-level classification models can only extract low-level or mid-level features from hyperspectral images,and their capacity for representation and fitting is limited,resulting in unsatisfied classification accuracy on large-scale and complex datasets.In the context of computer vision,the ongoing advancements in deep learning have propelled hyperspectral image classification technology into the forefront of contemporary research endeavors.In comparison to traditional methods,deep learning models can autonomously acquire and extract high-level abstract image features via multi-layer neural network architectures.This structural design facilitates the capturing of intricate data structures and inherent associations,resulting in enhanced classification accuracy for the models.While deep learning-based hyperspectral image classification methods have exhibited promising outcomes in numerous instances,the demand for a substantial number of labeled samples has posed hindrance to their extensive development and adoption.Moreover,the process of labeling categories for hyperspectral images is both costly and time-consuming,resulting in a severely limited pool of labeled samples available for practical modeling.In such scenarios,deep learning models may be susceptible to overfitting the noise or randomness present in the training data,thereby giving rise to overfitting concerns.To address this formidable challenge,this dissertation integrates theories of image spatial-spectral feature extraction,ensemble learning,semi-supervised learning,and other methodologies in a comprehensive research endeavor.The goal is to design a deep learning-based hyperspectral image classification method suitable for small sample sizes,enhancing the generalization capabilities of deep learning models and mitigating the impact of the small sample problem on hyperspectral image classification methods.The principal research themes and innovations presented in this dissertation are succinctly outlined as follows:(1)To address the overfitting problem encountered by deep learning models when dealing with small sample sizes,a convolutional neural network that leverages pixel-cluster and spatial-spectral fusion for hyperspectral image classification is proposed.Initially,the model extracts spatial texture features from the original hyperspectral images through gray co-occurrence matrices.These spatial features are subsequently integrated with spectral data to diversify input sample features and provide more discriminative information for the classifier.Moreover,the proposed model designs a pixel-cluster algorithm,which organizes labeled samples into clusters according to specific rules,thereby effectively augmenting the number of training samples and broadening the scope of sample semantics.This,in turn,enables a thorough optimization of the parameters of the convolutional neural network classification model,even in scenarios characterized by a limited number of samples.Subsequently,the model utilizes the constructed pixel-clusters to facilitate the training and fine-tuning of the convolutional neural network’s parameters.Results from experiments conducted on four different hyperspectral datasets demonstrate that the proposed method outperforms traditional CNN approaches,particularly in situations involving constrained sample sizes.(2)Building upon the research content(1)and ensemble learning theory,a deep ensemble classification method based on pixel-cluster and spatial-spectral fusion is proposed.The primary objective of this approach is to address the prevalent overfitting challenge that deep learning models encounter when dealing with limited sample sizes,while concurrently enhancing the model’s generalization capability and robustness.The approach designs a deep ensemble convolutional neural network architecture,which is constructed following the creation of spatial-spectral features and the application of pixel-cluster algorithm to expand the available training samples.More specifically,the Bootstrap sampling method is initially employed to create various training subsets from labeled pixel-clusters with spatial-spectral information.Subsequently,multiple convolutional neural networks are trained utilizing these subsets to establish a diversified set of basic classifiers.In the final step,the outputs of these basic classifiers are integrated through the majority voting strategy to derive the ultimate prediction result.Through the establishment of this ensemble framework,the proposed method comprehensively analyzes data from diverse perspectives and models,effectively mitigating the impact of randomness and overfitting.Consequently,this reinforces the classification performance,generalization capability,and robustness of the ensemble model.Experimental results clearly demonstrate that the proposed approach significantly improves the classification accuracy and robustness of convolutional neural networks for hyperspectral images,especially in the case of small samples.(3)Traditional semi-supervised hyperspectral image classification methods involve training the classifier in two separate stages using both unlabeled and labeled samples,which often leads to increased model complexity and information loss.To tackle this challenge,a deep semi-supervised classification model has been introduced,which utilizes both types of data concurrently.Initially,the pixel-cluster approach is applied to expand the quantity and semantic richness of the limited training samples,enabling the model to better capture the common characteristics within the input data.Subsequently,a limited number of labeled pixel-clusters and a larger set of unlabeled pixel-clusters are employed to train the convolutional neural network,which simplifies the training process and reduces information loss.Furthermore,to enhance the noise resistance of the classification model,random perturbations are added to the input samples,and the consistency regularization method is utilized to minimize differences between output vectors.Ultimately,the parameters of the convolutional neural network classifier are optimized by minimizing the weighted sum of supervised and unsupervised losses.Experimental results validate the effectiveness of this approach in enhancing hyperspectral image classification performance in scenarios with limited samples by concurrently harnessing information from both labeled and unlabeled samples.