Satellite Image Classification using Transformers with Map Reduce-Based Pre-processing Framework
Year: 2025
Type: article
Abstract: Objectives: This study aims to develop a deep learning framework with satellite image classification, a framework that is scalable as well as accurate, effectively handling the challenges posed by large-scale remote sensing datasets. The framework has the goal of improving upon classification precision and optimizes through using computational resources. It supports multiple critical applications like agricultural monitoring, land cover mapping, and disaster response with its categorization of satellite images into certain distinct land cover types, including urban, agricultural, forest, desert, and ocean regions. Methods: This study aims to develop a deep learning framework with satellite image classification, a framework that is scalable as well as accurate, effectively handling the challenges posed by large-scale remote sensing datasets. The framework has the goal of improving upon classification precision and optimizes through using computational resources. It supports multiple critical applications like agricultural monitoring, land cover mapping, and disaster response with its categorization of satellite images into certain distinct land cover types, including urban, agricultural, forest, desert, and ocean regions. Findings: The evaluation of that experiment showed that the MapReduce preprocessing pipeline did approximately reduce data handling, along with processing time, by around 68% versus the conventional methods. The Vision Transformer did outperform baseline CNN models (ResNet50 and VGG16) within 4–6% margins and it achieved 95.3% accuracy on EuroSAT and 90.8% on BigEarthNet. In addition, final model size was reduced 37% by pruning techniques, improving deployment readiness with stable classification performance. Therefore, the scalability of the framework and also operational feasibility were validated. Throughput was improved up to 210 images/minute from 50 images/minute. Novelty: This study presents a novel hybrid framework that uniquely combines big data processing via MapReduce with advanced transformer-based deep learning for satellite image classification. While Vision Transformers are emerging in computer vision, their integration with large-scale distributed preprocessing frameworks like MapReduce remains largely unexplored in the context of remote sensing. This combination enables efficient processing of vast satellite image repositories while ensuring high accuracy, making the framework highly suitable for real-time, scalable land cover monitoring applications. Keywords: Satellite Image Classification, Transformer Neural Network (TNN), MapReduce Preprocessing, Land Cover Segmentation, Remote Sensing Automation
Authors: Rondi Pushpa Latha, Persis Voola
Institution: Adikavi Nannaya University
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410732001