Generative AI-based Land Cover Classification via Federated Learning CNNs: Sustainable Insights from UAV Imagery

This paper introduces a novel approach method for decentralized land cover and land use classification, utilizing federated learning in conjunction with Convolutional Neural Networks (CNNs) on imagery obtained from Unmanned Aerial Vehicles (UAVs). Integration of UAV imagery provides high-resolution spatial data, facilitating precise classification of land cover types. Federated learning (FL) ensures data privacy and reduces communication bandwidth usage by enabling model training on local devices (e.g., UAVs) without the need to share data with a centralized server. However, these UAVs have limited resources and often struggle to capture a sufficient number of images for training a model without encountering overfitting. We address the challenges of scarce data samples on UAVs by leveraging DCGAN (Deep Convolutional Generative Adversarial Networks) to generate synthetic images, promoting sustainability by minimizing the need for extensive data collection. These synthetic images are used to train local models of UAVs, offering enhanced results while combating overfitting. Our developed FL model achieved an accuracy exceeding 97%, indicating a 7% improvement over Vanilla FL. The proposed sustainable approach exhibits promising outcomes in achieving accurate and scalable land cover and land use classification, showcasing its applicability in environmental monitoring, urban planning, and precision agriculture.