AgroMetLLM: An evapotranspiration and agro-advisory system using localized large language models in resource-constrained edge
Year: 2025
Type: article
Abstract: We introduce AgroMetLLM, an on-device agrometeorological advisory system that combines five validated evapotranspiration (ET) models with various quantized Large Language Models (LLMs) on a Raspberry Pi 4B. The users specify a location, 3-7-day horizon by using Gradio interface and LLM; Open-Meteo APIs then supply daily inputs Tmax, Tmin, Tmean, RHmean precipitation, Food and Agriculture Organization (FAO) reference evapotranspiration (ET0), and Rs for multiple Indian sites. Computed ET ranges (mm day⁻¹) across locations were: FAO ET₀ 2.84-6.21; Hargreaves-Samani 6.28-13.74; Turc 0.17-0.21; Priestley-Taylor 5.64-9.06; Makkink 2.73-4.38. A few-shot prompting strategy, based on curated examples for 3-, 5-, and 7-day forecasts, is used to guide the Qwen LLM under Ollama to produce structured, five-point advisories in 1-2 s. One-way ANOVA (F = 3.30-6.71, p = 0.016-0.0002) and Kruskal-Wallis tests (χ2 = 9.61-15.48, p < 0.05 except Turc p = 0.088) confirm significant ET differences among models and LLM sizes. All outputs and metadata persist in SQLite, and Matplotlib renders comparative bar charts in the dashboard. These results demonstrate that compact, quantized LLMs can reliably deliver actionable irrigation guidance-matching cloud-based accuracy-while operating offline, with minimal latency and energy use, thus empowering resource-constrained smallholder farmers.
Source: Journal of Agrometeorology
Authors: PARTHA PRATIM RAY, MOHAN PRATAP PRADHAN
Institution: Sikkim University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDGs: Zero hunger, Climate action
Open Access status: diamond
OpenAlex ID: W4413865735