Edge AI for Climate-Aware ET₀ Forecasting and Autonomous Precision Irrigation: A Hardware-Software Co-Design for Autonomous Precision Irrigation

UR-GAMMA3 Université de Technologie de Troyes Troyes, France
LPMCS Université de Lomé Lomé, TOGO
UR-GAMMA3 Université de Technologie de Troyes Troyes, France
Smart Structural Health Monitoring and Control Laboratory Dongguan University of Technology Dongguan, CHINA

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Precision agriculture is hindered by its dependence on centralized cloud infrastructures, which prevents deployment of advanced deep learning (DL) in disconnected, resource‑constrained rural environments. We propose a globally transferable Edge AI framework capable of executing reference evapotranspiration (ET₀) forecasting directly on low‑power microcomputers. To overcome hardware limitations, we introduce a Distilled Noise‑Invariance training paradigm: offline Empirical Mode Decomposition and Wavelet Denoising (EMD‑WD) are applied only to target variables during training, enabling the network to learn an intrinsic auto‑correction manifold without runtime filtering. We evaluate INT8‑quantized Mamba‑2 and Temporal Fusion Transformer (TFT) models enhanced with Köppen–Geiger climatological latent embeddings across 309 globally distributed sites. The optimized architectures preserve full predictive accuracy while achieving up to 72–80% reductions in memory and 68–75% reductions in latency, all under 5 W of power, enabling cloud‑independent, noise‑resilient precision irrigation .