Hybrid ConvNeXt-Liquid Neural Network for Satellite Image Classification: A Benchmark on EuroSAT Dataset
Gundla Karunasri, Sheshikala Martha, Vishwanath Bijalwan · IEEE Transactions on Geoscience and Remote Sensing · 2025
Recognizing land use and land cover elements in satellite imagery is an important foundational point for environmental tracking applications, as well as urban development and sustainable development projects. This research demonstrates the implementation of a (hybrid ConvNeXt+LNN) deep learning systems that study EuroSAT multispectral optical satellite images to carry out a classification task. The proposed model combination of convolutional feature extraction together with advanced representation learning established superior class separability for detecting the ten different land use and land cover categories. In this, model maintained excellent performance levels were maintained through 50 epochs has achieving a training accuracy of 98.79% with training loss of 0.0350 and validation accuracy 97.41% with validation loss 0.0793, leading to a test accuracy of 97.26% with a test loss of 0.0815. The model received standard classification metrics assessments and confusion matrix analysis to produce evaluation outcomes that examined per-class precision and recall, together with F1-score metrics. ROC curve analysis delivered substantial Area Under the Curve (AUC) results during the complete class examination phase.t-SNE and UMAP were applied to the high-dimensional feature vectors to visualize the learned representations, revealing well-separated clusters for each class. This indicates that the hybrid ConvNeXt+LNN model effectively captures class-specific feature distributions in latent space, enabling robust discrimination across all ten land use and land cover categories. Compared to state-of-the-art baseline models, the proposed ConvNeXt + LNN model achieves a higher test accuracy of 97.26% while maintaining enhanced computational efficiency, requiring only 625.79 seconds per epoch and a reduced parameter count (7.8M vs. 23.5M in ResNet-50). The final research demonstrates validation of the developed approach as an effective and scalable satellite-based LULC classification solution that supports geospatial systems that can be utilized for the National Geospatial Mission and United Nations Sustainable Development Goals.