In recent years, benefiting from the rapid development of deep learning technology in the field of computer vision, the study of hyperspectral image (HSI) classification has also made great progress. However, compared with ordinary RGB images, HSIs are more like 3D cubes; therefore, it is necessary and beneficial to explore classification methods suitable for the very special data structure of HSIs. In this paper, we propose Multiple Spectral Resolution 3D Convolutional Neural Network (MSR-3DCNN) for HSI classification tasks. In MSR-3DCNN, we expand the idea of multi-scale feature fusion and dilated convolution from the spatial dimension to the spectral dimension, and combine 3D convolution and residual connection; therefore, it can better adapt to the 3D cubic form of hyperspectral data and make efficient use of spectral information in different bands. Experimental results on four benchmark datasets show the effectiveness of the proposed approach and its superiority as compared with some state-of-the-art (SOTA) HSI classification methods.
Compared with general optical images, hyperspectral images (HSIs) contain richer spectral information. On one hand, this provides a sufficient basis for ground object recognition. On the other hand, it results in the intermingling of spatial and spectral information. In order to make better use of the rich spatial and spectral information in HSIs, we resort to Vision Transformer (ViT). To be specific, we propose the Cross Spatial–Spectral Dense Transformer (CS2DT) for spatial-spectral feature extracting and feature fusing. For feature extraction, CS2DT employs the Adaptive Dense Encoder (ADE) module, which enables the extraction of multi-scale semantic information. During the features fusion stage, we use the Cross Spatial–Spectral Attention (CS2A) module based on the cross-attention (CA) operation to better integrate spatial and spectral features. We evaluate the classification performance of the proposed CS2DT on three well-known datasets by conducting extensive experiments. Experimental results demonstrate that CS2DT can achieve higher accuracy and higher stability when compared with the state-of-the-art (SOTA) methods. The source code will be made available at https://github.com/shouhengx/CS2DT.
30 July 2020 in IEEE Geoscience and Remote Sensing Letters
Convolutional neural networks (CNNs) are widely used in the field of remote sensing images. However, the applications of CNNs and related techniques often ignore the properties of remote sensing data. In our study, we focus on the hyperspectral image (HSI) classification problem, and address the issue of including the very rich spectral information present in HSIs in CNN-based models to produce highly accurate classification results. We propose a two-step classification technique, ClusterCNN. The first step divides HSI pixels into different clusters, to form a material map which can be considered as a compressed expression of the original spectral features. The second step trains a CNN that can extract spatial features from the material map, and then exploits these spatial features to classify HSI pixels. The proposed approach follows a strict hierarchy to exploit both the spectral and spatial features in HSIs. Experimental results show the effectiveness of ClusterCNN as compared to the much more complicated state-of-the-art approaches.
In time series prediction tasks, dynamic models are less popular than static models, while they are more suitable for modeling the underlying dynamics of time series. In this paper, a novel architecture and supervised learning principle for recurrent neural networks, namely echo state networks, are adopted to build dynamic time series predictors. Ensemble techniques are employed to overcome the randomness and instability of echo state predictors, and a dynamic ensemble predictor is therefore established. The proposed predictor is tested in numerical experiments and different strategies for training the predictor are also comparatively studied. A case study is then conducted to test the predictor's performance in realistic prediction tasks.
25 June 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Unsupervised Domain Adaptation (UDA) is an effective method for addressing the domain shift issue between the source and target domains in high-resolution (HR) remote sensing images (RSIs) classification. However, existing methods still face significant challenges when dealing with substantial style discrepancies and class imbalance. For instance, the widely used Adaptive Instance Normalization (AdaIN) method reduces the distribution difference by transforming the style of source domain images to match that of the target domain. However, when confronted with large style variations, this method often leads to the loss of fine-grained details such as color, texture, and contrast, negatively impacting the transfer performance of the model. Additionally, class imbalance causes the model to be biased towards dominant classes, resulting in decreased overall classification accuracy. To address these issues, this paper proposes two innovative modules. First, we introduce the SelfAdaIN module, which precisely controls the style transformation process through an adaptive convolutional generation mechanism. This module effectively prevents the loss of fine details found in traditional AdaIN methods, achieving fine-grained style alignment between the source and target domains. Secondly, we propose a Class Balance Loss (CBL) function that adjusts the weight distribution of target domain samples to mitigate the influence of dominant classes and enhance the focus on rare categories, thereby alleviating the issue of class imbalance. Extensive experiments on four cross-domain remote sensing datasets demonstrate that the proposed methods significantly improve classification accuracy. Code is available at https://github.com/nclab-1/FGSA-CB