GeoCrop-Bench: Cross-Regional Evaluation of Models for Crop Recognition and Ecological Monitoring

Abstract

The accurate recognition of crop types and the continuous monitoring of agricultural ecology are foundational to global food security, climate change mitigation, and sustainable land management. While deep learning models have demonstrated exceptional capabilities in localized remote sensing tasks, their performance frequently degrades when deployed across diverse geographical regions due to spatial, temporal, and spectral domain shifts. To address this critical gap, this paper introduces GeoCrop-Bench, a comprehensive cross-regional evaluation framework and benchmark dataset designed to rigorously assess the generalization capabilities of crop recognition and ecological monitoring models. The dataset integrates multimodal satellite imagery, including synthetic aperture radar and multispectral optical data, spanning multiple continents, climate zones, and agricultural paradigms. By evaluating various baseline architectures, from convolutional neural networks to vision transformers, we reveal significant performance disparities when models trained in data-rich regions are transferred to data-scarce environments. The analysis extends beyond simple classification accuracy to include ecological indicators such as vegetation health, soil moisture proxies, and carbon sequestration potential. Our findings indicate that while multi-modal fusion architectures exhibit superior robustness, substantial algorithmic innovations are still required to achieve true global scalability. This study provides a vital stepping stone toward developing resilient, globally applicable agricultural monitoring systems that can support international ecological policy and sustainable farming practices.