Hyperspectral imaging-based land use classification using a hybrid Convolutional Neural network-vision transformer model

Effective land use and land cover (LULC) classification is critical for sustainable environmental management, particularly in the context of rapid urbanization. This study was motivated by the need to move beyond conventional, labor-intensive field survey methods by adopting advanced remote sensing technologies and artificial intelligence (AI)-based approaches. Traditional machine learning methods, such as Support Vector Machines (SVM), have shown limitations in capturing the complex spectral and spatial characteristics of hyperspectral data. To address these challenges, a hybrid deep learning framework is proposed, combining Convolutional Neural Networks (CNNs) with Vision Transformers (ViTs), leveraging the local feature extraction strengths of CNNs and the global contextual modeling capabilities of Transformers. The proposed methodology involves the collection of hyperspectral aerial imagery and ground-based spectral reflectance data from the Geum River Basin in South Korea. A hyperspectral reflectance library was constructed using over 5,700 field measurements obtained via a FieldSpec-4 spectroradiometer. These data were used to train and validate a CNN-ViT model, which integrates a series of 1D convolutional layers for spatial-spectral feature extraction with a Transformer encoder for capturing long-range dependencies. To ensure robust training and generalization, preprocessing steps such as spectral band alignment and reflectance normalization were applied, and hyperparameter tuning was employed using Optuna. The classification performance of the CNN-ViT model was evaluated against the traditional SVM model and CNN model, showing significantly improved accuracy, particularly in areas where spectral mixing occurs due to the 2-meter spatial resolution of airborne hyperspectral imagery. In particular, traditional methods struggled to distinguish between mixed surfaces such as soil and vegetation, whereas the CNN-ViT model successfully resolved these ambiguities.The CNN-ViT model achieved an overall classification accuracy of 0.98, outperforming the SVM model’s 0.72 and CNN model’s 0.86. Furthermore, our approach was implemented in a scalable geospatial processing framework that enables efficient batch classification and raster map generation. These results highlight the effectiveness of the CNN-ViT hybrid approach in advancing LULC classification using hyperspectral data, with implications for enhancing land monitoring systems, environmental planning, and resource management. • Development of a CNN-ViT Hybrid Model. To overcome the limitations of traditional machine learning models (SVM, RF), a hybrid model integrating Convolutional Neural Networks (CNN) and Vision Transformers (ViT) was developed. • Superior Classification Accuracy of the CNN-ViT Model . This model demonstrated higher classification accuracy and improved boundary delineation, utilizing airborne hyperspectral imagery and field-measured reflectance data for land use and land cover (LULC) classification. • Real-World Applicability. The CNN-ViT model successfully generated high-resolution LULC maps based on hyperspectral imagery, demonstrating its potential applications in environmental monitoring, urban planning, agriculture, and water resource management.