Benchmarking Deep Learning for Agricultural Potential Estimation from Multispectral Satellite Time Series

Agricultural potential estimation -assessing land suitability for crop growth-is a critical yet challenging task for food security and sustainable land use planning. While traditional approaches rely on labor-intensive field surveys or coarse-resolution data, deep learning offers a scalable alternative. In this paper, we present the first systematic benchmark of deep learning approaches for agricultural potential estimation by leveraging time series data from multispectral satellite imagery. We benchmark three prediction strategies -continuous space discretization, categorical classification, and K-rank ordinal classification-across four state-of-the-art deep learning architectures (2D CNN, 3D CNN, ViViT, and UTAE) on the public AgriPotential dataset. Our experiments make several key findings. Continuous space discretization consistently achieved the best performance, achieving up to 87.97% accuracy±1 and a mean absolute error below 0.7. We show that temporal modeling is essential: architectures leveraging time series data significantly outperform their static counterparts, while multi-year data improves robustness against inter-annual variability. Through spectral ablation, we show that RGB and multispectral bands provide complementary information, and error analysis confirms that regression yields the most reliable predictions, with misclassifications predominantly occurring between adjacent classes -which is critical for real-world applications. Our results establish a strong baseline for agricultural potential estimation and provide practical guidance for integrating machine learning into agricultural planning applications. Our findings highlight the importance of temporal modeling, multispectral feature fusion, and robust loss functions-key considerations for future research in remote sensing and applied machine learning. Code at github.com