Few-Shot Learning for Land Cover Classification Using Sentinel-2 Imagery: A Lightweight GeoAI Framework

Abstract

Recent advances in satellite remote sensing have led to the widespread availability of high-resolution Earth observation data. However, the performance of deep learning models in land cover classification often depends on large volumes of annotated data, which are costly and time-consuming to obtain. This study applies a lightweight GeoAI approach for land cover classification using Sentinel-2 imagery under limited training data conditions. We implement a lightweight GeoAI framework utilizing prototypical networks with pre-trained convolutional backbones to classify land cover types under low-data regimes. Experiments on a subset of the EuroSAT dataset demonstrate that our approach achieves competitive accuracy, even with as few as one to five labeled examples per class. This work highlights the potential of FSL techniques in reducing labeling costs and enabling rapid deployment of
land cover models in data-scarce environments.