EUVIMEDEUVIMED entdecken
Crossref

A Simulation-First Edge AI Framework for Agricultural Decision Support: Separating Observability Limits from Quantization Error

John Nyamu Kibe, Morris Kimathi Mwangi, Samuel Macharia Kang'iri

Journal of Applied Science, Engineering and Technology for Development · 16.11.2026 · 6 (1)

DOI: 10.33803/y448fy.2 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Applied Science, Engineering and Technology for Development, online erschienen am 16.11.2026. Verfasst von John Nyamu Kibe und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Embedded agricultural decision support is often built by training a classifier small enough for a microcontroller on the outputs of a physically based model (policy distillation), quantizing it, and reporting its agreement with that model. Such evaluations report footprint and quantization error but rarely test whether the decision can be recovered from what the node senses. This study tests it against a simulated reference: a dual crop coefficient water balance and a thermal time blight model, driven by a weather generator fitted to central Kenyan highland climatology, so the latent state behind each decision is known exactly. For a node that senses temperature, humidity, wind and pressure, each decision is scored from the calendar alone, from the sensed variables with and without the calendar, and with the latent state supplied. Irrigation timing agrees with the reference on 97.0% of days, equal to the majority-class rate of 97.0%, and its Matthews correlation is -0.001. Supplying the latent depletion raises the same linear model to +0.858, whereas rainfall inputs, including rainfall minus evapotranspiration accumulated since planting, leave it near zero. Fertilizer timing scores +0.875, but day of year alone scores +0.892 and removing it leaves +0.266 (+0.032 to +0.266 over six generator seeds), so the classifier is predicting the calendar. Pest risk is partly observable, +0.497 against +0.984 with the latent severity. Running the water balance on the node, with Penman-Monteith evapotranspiration from radiation estimated from the temperature range as FAO-56 recommends, reproduces the reference schedule to a median of 1 day, with an event-level F1 of 0.757 at a three day tolerance against 0.299 when scored per day; calibrated Hargreaves scores lower at every seed. At Q15 no decision changes in 5,200 samples. For such nodes, decision quality is limited by what the sensors observe, not by quantization.

Zitieren

APA

Kibe, J. N., Mwangi, M. K., & Kang'iri, S. M. (2026). A Simulation-First Edge AI Framework for Agricultural Decision Support: Separating Observability Limits from Quantization Error. Journal of Applied Science, Engineering and Technology for Development, 6(1). https://doi.org/10.33803/y448fy.2

Vancouver

Kibe JN, Mwangi MK, Kang'iri SM. A Simulation-First Edge AI Framework for Agricultural Decision Support: Separating Observability Limits from Quantization Error. Journal of Applied Science, Engineering and Technology for Development. 2026;6(1). doi:10.33803/y448fy.2

BibTeX
@article{Kibe2026,
  author = {Kibe, John Nyamu and Mwangi, Morris Kimathi and Kang'iri, Samuel Macharia},
  title = {A Simulation-First Edge AI Framework for Agricultural Decision Support: Separating Observability Limits from Quantization Error},
  journal = {Journal of Applied Science, Engineering and Technology for Development},
  year = {2026},
  volume = {6},
  number = {1},
  doi = {10.33803/y448fy.2}
}
Kostenlos starten →