Spatiotemporal Graph Neural Network Ensembles for Artificial Intelligence Agriculture Locust Swarm Trajectory Prediction Under Radar Surveillance Gaps: A Computer Coverage-Calibration Evaluation
Year: 2026
Type: article
Abstract: Locust swarm trajectory prediction under radar surveillance gaps is a coverage problem before it is an accuracy problem: a forecaster that reports a nominal 95% interval but covers the true swarm position in only four of five cases will misdirect aerial control assets even when its point forecasts are unbiased. The estimator is a linear model of the form intercept, exposure, group and severity, fitted by ordinary least squares to 600 data sets of N = 500 drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2, seed 714928413). The evaluation characterises the sampling behaviour of the estimator that any graph ensemble's readout layer must inherit. Across the 600 replications the intercept was estimated with a bias of -0.07778 and an RMSE of 0.87843, the exposure coefficient with a bias of 0.00173 and an RMSE of 0.01625, and the nominal 95% confidence intervals achieved empirical coverage of 1.0 for both parameters. Coverage calibration, not point accuracy, is the binding constraint on operational deployment of graph-based swarm predictors, and it specifies the calibration diagnostics that a radar-gap-aware ensemble must satisfy before its intervals can be trusted for asset allocation.
Author: Barasa, Otieno Wekesa
Institution: University of Nairobi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223093319