Satellite-derived evapotranspiration gap-filling with spatiotemporal graph neural networks: a masked-reconstruction benchmark for irrigation-scheduling reliability across Sahelian croplands

Reliable evapotranspiration (ET) estimates are essential for irrigation scheduling in the Sahel, yet satellite retrievals are frequently interrupted by cloud cover and sensor gaps, leaving farmers and water managers with incomplete information. The framework represents landscape units as nodes in a graph, encodes spatial adjacency and temporal lagged dependencies, and trains via masked reconstruction so that the model learns to recover missing ET values from surrounding observations. We formalise the gap-filling task as a masked signal recovery problem on a graph and derive conditions under which the reconstruction error is bounded by the spectral properties of the graph Laplacian and the temporal correlation of the signal. Because no site-specific ET data were available for the present analysis, the framework is developed as a methodological proposal and benchmark design; we illustrate the approach using a real reanalysis temperature series for Accra, Ghana (2000–2023, 312 monthly observations, mean 26.521 °C, warming trend 0.1622 °C per decade), which serves as a stand-in for the temporal structure that ET gap-filling must accommodate. We argue that gap-filling accuracy alone is insufficient; reliability must be assessed in terms of the downstream scheduling error it induces, and we propose a benchmark protocol that makes this assessment reproducible.