A multi-sensor and multi-temporal GeoAI framework for resolving ecological scale mismatch in Antarctica

Advances in remote sensing (RS) and artificial intelligence (AI) are enabling new approaches to ecological monitoring across spatial and temporal scales. However, vegetation mapping in polar environments remains constrained by a fundamental mismatch between the fine spatial scale of ecological processes and the coarse resolution of satellite observations, as well as by limited availability of biologically reliable training data. In this study, we address these challenges by developing an end-to-end GeoAI framework that integrates field observations, drone imagery, and satellite time-series data to enable cross-scale vegetation mapping. We propose a progressive weak-to-strong labelling workflow that begins with 1) biologically grounded quadrat annotations, to create iteratively expanded training data along with 2) spectral similarity combined with machine learning, followed by 3) training with deep learning (DL) models for pixel-level segmentation. Native Sentinel-2 observations (10–20 m spatial resolution) were processed using DL-based super-resolution to an approximately 1 m output grid. High-resolution drone DL predictions were subsequently transformed into fractional cover labels aligned to the super-resolved satellite grid, supporting cross-scale modelling across centimetre-to-metre resolutions. We evaluate multiple DL architectures using spatially separated UAV evaluation data and assess satellite-scale performance as agreement with UAV-derived fractional cover reference labels. Our results demonstrate that the proposed framework substantially improves the scalability and ecological relevance of vegetation mapping in data-limited environments. The best-performing drone-scale DL model (DeepLabv3+) achieved an overall accuracy of 0.91 and a Macro-F1 score of 0.85 for fine-scale vegetation segmentation. At the satellite scale, the U-Net fractional-cover model achieved strong predictive performance for dominant vegetation classes, including cyanobacteria (R 2 = 0.94) and red Hennediella heimii moss (R 2 = 0.88), with reduced performance for spectrally complex or sparsely represented classes, demonstrating effective cross-scale vegetation mapping despite challenges associated with sub-pixel heterogeneity and spectral mixing. This study provides a generalisable methodological framework for integrating multi-sensor data and AI in ecological RS, demonstrating strong cross-scale performance while illustrating the challenges of capturing fine-scale vegetation heterogeneity at satellite scales.