A Novel VGG-16 Adaptation for Multi-band Satellite Image Classification: Optimized Preprocessing and Class-specific Augmentation

Recent innovations in satellite imaging have significantly improved the effectiveness of Earth observation through high-resolution imagery used in environmental analysis, urban ecosystem, city planner mapping, and disaster management. Nevertheless, with high-dimensional Landuse Landcover multispectral data available, the classification of satellite images remains a challenging task due to the variability of the data. The conventional approaches to machine learning do well but fail at managing the intricacies of multispectral data without shredding and feature extraction. This research presents a novel technique by applying the VGG-16 structure, a deep convolutional neural network for image recognition, to classify the EuroSAT multispectral satellite image dataset, consisting of various European terrains captured across multispectral bands, including visible and near-infrared, which are essential for environmental analysis. The proposed technique presents a VGG-16 model enhanced with new convolutional blocks and dropout layers, specifically designed to classify multiband satellite imagery data. The first phase of convolutions has been made to work with four spectral bands, more specifically, RGB and near-infrared, to increase the capacity of distinguishing the type of ground cover. In addition, the output layer has improved to provide a ten-class scene classification (forest, residential, industrial, highway, pasture, river, sea lake, herbaceous vegetation, annual crop, permanent crop) for various landcover types, enhancing the model's applicability. Augmentation techniques, including rotation, flipping, and shifting, have been used to increase the diversity of the training dataset; additionally, transfer learning leveraged the resultant augmented datasets. This adaptation not only enhances the mechanism for classifying the satellite images but also decreases the time and computation resources needed, thereby making it applicable to big data. The proposed modifications to the VGG-16 allow obtaining a higher classification accuracy rate of ±96.7%, which ensures the effectiveness of an automated system for analyzing satellite imagery. Received: 15 November 2024 | Revised: 10 March 2025 | Accepted: 14 April 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The EuroSat data that support the findings of this study are openly available in GitHub at https://github.com/phelber/eurosat. Author Contribution Statement Ranu Sewada: Conceptualization, Methodology, Software, Validation, Investigation, Writing – original draft, Writing – review & editing, Visualization. Hemlata Goyal: Conceptualization, Formal analysis, Writing – review & editing, Supervision, Project administration.