Paper Details

Abstract

Global food security and increasing water scarcity demand more efficient paddy irrigation scheduling, while the limited computational power and memory of edge devices pose challenges for accurate reference crop evapotranspiration (ET₀) prediction. To address this issue, we propose a lightweight GRU variant, termed Tiny-GRU, which integrates temporal attention, unstructured pruning, and post-training dynamic range quantization to enable efficient deployment under resource-constrained edge computing. Nine model configurations were benchmarked on a Raspberry Pi 5 and a resource-constrained Intel i5 platform, with performance evaluated in terms of predictive accuracy, inference latency, memory usage, and thermal stability. Results indicate that the optimized Tiny-GRU configurations achieve high forecasting accuracy (RMSE ≈ 0.02 mm·h⁻¹) while maintaining fast on-device inference (≈ 0.21 ms per sample on Raspberry Pi 5) and stable runtime behavior suitable for continuous edge operation. Statistical analysis confirmed that the observed performance differences among model configurations were significant (p < 0.001). Furthermore, a cross-platform visualization framework was developed to analyze trade-offs between predictive performance and resource consumption, thereby supporting decision-making for edge-oriented benchmarking in smart paddy irrigation. Overall, the proposed Tiny-GRU framework provides a practical and scalable solution for efficient ET₀ forecasting on resource-limited edge devices

To Cite This Article