A hybrid 3D CNN-LSTM model with soft spatial attention mechanism for accurate hyperspectral image classification

Hyperspectral imaging (HSI) plays a pivotal role in remote sensing, enabling precise material identification through spectral data across many bands. Despite its advantages, challenges like high dimensionality, spectral mixing, and limited labeled data hinder its classification accuracy. This study introduces a hybrid deep learning model that combines 3D Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks, incorporating residual connections and a soft spatial attention mechanism to overcome these limitations. The model utilizes 3D CNNs for the extraction of joint spatial-spectral features, employs LSTMs to capture sequential dependencies across spectral bands, and incorporates soft spatial attention mechanism to emphasize discriminative spatial regions. The model has been evaluated using benchmark datasets—Indian Pines (16 land cover classes) achieved remarkable overall accuracies of 99.66% and 99.58% on external validation in Salinas (16 agricultural) dataset, respectively, exceeding the performance of current leading methods. Ablation studies confirmed the essential functions of residual connections in enhancing training stability and soft spatial attention mechanism in mitigating redundant features. The findings indicate strong generalization across various datasets, showing only a few misclassifications, even among spectrally similar classes. The proposed model addresses the challenges of spectral feature extraction and enhances classification accuracy by leveraging advanced deep learning techniques. Ablation studies revealed that residual connections improve training stability, while the attention mechanism effectively reduces redundancy. The results suggested that the proposed model could be applied to various fields, including agriculture, environmental monitoring, and land management, with significant improvements in classification accuracy. • Introduces a hybrid deep learning architecture (3D CNN–LSTM with attention) tailored for accurate hyperspectral image classification in remote sensing applications. • Effectively captures spatial, spectral, and sequential dependencies critical for detailed land cover mapping and environmental monitoring using hyperspectral data. • Achieves exceptional classification accuracy on benchmark datasets (Indian Pines: 99.66%, Salinas: 99.58%), outperforming state-of-the-art models in remote sensing contexts. • Incorporates residual connections to ensure model stability and soft spatial attention to enhance feature discrimination in heterogeneous landscapes. • Demonstrates strong generalization across geospatial environments, addressing key remote sensing challenges such as spectral mixing and limited labeled data. • Employs incremental PCA and patch-based training, improving computational efficiency—vital for scaling remote sensing models to large-area earth observation tasks.