Edge-Optimized Reservoir Computing for Forecasting Complex Ecological Systems: Predicting Fishery Collapse with Incomplete Data

Abstract

Monitoring landings in data-scarce fisheries is hampered by short, noisy series and multi-year reporting gaps. We benchmark Echo State Networks (ESNs) against SARIMA, LightGBM, and LSTM models and against naïve and seasonal-naïve persistence, for one-step-ahead forecasting of the monthly landings (1993–2023) of the shallow-water shrimp industrial trawl fishery of the Colombian Caribbean, using a gap-filled series, equal tuning budgets, and walk-forward validation (h=1, 72 origins) on an embedded NVIDIA Jetson Orin Nano. The ESN obtained the lowest errors (MAE = 1.33 tons, RMSE = 1.82 tons), 28% and 34% below the naïve forecast on the same window (1.85 and 2.77 tons), and Diebold–Mariano tests favored it over every other model; the advantage comes from the test years 2021–2023, and LightGBM is a close competitor. The ESN needed 337 ms per origin (fit plus predict) and 0.08 ms per sample at inference. A bidirectional ESN interpolated 108 unmonitored months; in blind-gap tests, it was no more accurate than linear or Kalman interpolation, and calibrated bands are wide. Because only landings, not fishing effort, are analyzed, the post-2013 decline cannot be separated from fleet contraction.