AquaSustain-Bench: A Standardised Multi-Farm Benchmark Dataset and Evaluation Framework for Aquaculture Water Quality Prediction, Disease Risk Detection, and Autonomous Sustainability Control

Abstract

AquaSustain-Bench is the first open benchmark dataset and evaluation framework for commercial aquaculture precision intelligence. It provides 47.8 million validated sensor-timestep observations from 12 commercial partner farms across Egypt, Saudi Arabia, and Bangladesh, covering three aquaculture system types (Nile tilapia earthen ponds, whiteleg shrimp biofloc RAS, catla/rohu freshwater polyculture ponds) and 551 IoT sensor nodes. The dataset includes 228 documented disease and anomaly events with ground-truth labels at 15-minute resolution, complete actuator command histories, and digital twin state vectors. Eight standardised benchmark tasks are defined: (1) 72-hour water quality forecasting, (2) event detection and early warning, (3) autonomous sustainability control, (4) cross-farm transfer learning, (5) new-farm onboarding speed, (6) digital twin state estimation, (7) CPS pipeline latency, and (8) federated multi-farm learning. Seven published baselines and five AquaFarm stack model implementations are included. GitHub repository: https://github.com/Drwae/AquaSustain-Bench