Study On Research On The Classification Technology Of Hyperspectral Remote Sensing Images Based On Few-shot

Hyperspectral remote sensing images are extremely rich in spectral bands and contain a great deal of spectral and spatial information,which can enable more accurate feature identification.Hyperspectral remote sensing technology has been widely used in several fields,including food quality and safety,precision agriculture,industrial inspection,geological disaster warning and monitoring,criminal investigation,civilian inspection,environmental monitoring,and biomedicine.Since each pixel point in a hyperspectral remote sensing image contains multiple bands of information,and the annotation of each pixel point requires considering multiple bands of information,the annotation work demands significant manpower and time.Even the labeling of certain ground features necessitates the expertise and experience of professionals.These two factors contribute to a limited number of labeled samples in hyperspectral remote sensing images.Therefore,achieving high-precision classification of hyperspectral remote sensing images with limited labeled samples is of great value.In recent years,deep learning methods have achieved promising results in the classification of hyperspectral remote sensing images.However,the classification accuracy is limited due to problems such as the limited number of labeled samples and the high correlation of information between spectral layers.To address these issues,this thesis focuses on utilizing deep learning networks to achieve higher classification accuracy with a small number of labeled samples.The main research is as follows:(1)To address the issue of few labeled samples,a hybrid relational network model is proposed.The network adopts the concept of metric learning to perform similarity learning between pairs of samples,thereby achieving improved classification results with a small number of labeled samples.The utilization of a hybrid 3D/2D convolutional neural network allows for comprehensive extraction of spectral-spatial features,reducing network complexity and enhancing computational efficiency.Moreover,the relational learning module incorporates a sample pairing method,effectively mitigating the problem of limited labeled samples and facilitating more accurate learning of sample correlations.Additionally,transfer learning is introduced during training to minimize the required number of labeled samples for hyperspectral remote sensing image classification.Experimental comparisons are conducted with other methods using three public datasets,demonstrating the superior effectiveness of the proposed approach in classifying hyperspectral remote sensing images with a reduced training sample size.(2)To address the issue of high correlation between spectral layers in hyperspectral remote sensing images,a method combining vegetation indices and hybrid relational networks is proposed on the basis of Research Component 1.Due to similar physical or chemical properties among different ground objects or low spatial resolution,there may be high similarity in spectral curves,leading to potential misclassification.By incorporating the fusion of vegetation indices,the differences in spectral curves between vegetation types can be enhanced,resulting in more discriminative spectral-spatial features and improved classification accuracy.The vegetation index data is overlaid onto the original hyperspectral remote sensing data and then fed into the hybrid relational network model for classification.Experimental results on multiple datasets have demonstrated that the fusion of vegetation indices achieves better classification performance,especially when dealing with limited labeled samples.