AgroInsect: A hybrid AI-driven edge-cloud architecture for precision pest detection and geostatistical interpolation in agriculture

de Almeida, Guilherme Pires Silva and dos Santos, Leonardo Nazário Silva and de Oliveira, Ruy and Teixeira, Marconi Batista and De Oliveira, Mario and Gontijo, Pablo da Costa and de Oliveira Bailão, Adriano Soares and do Carmo França, Heyde Francielle and Novak, Sergio Souza and Stirle, Jéssica Lauanda (2026) AgroInsect: A hybrid AI-driven edge-cloud architecture for precision pest detection and geostatistical interpolation in agriculture. Computers and Electronics in Agriculture, 256. p. 112386. ISSN 0168-1699

Preview

Text 1-s2.0-S0168169926009841-main.pdf - Published Version Available under License Creative Commons Attribution. Download (4MB)

Abstract

This study presents AgroInsect, a hybrid edge-cloud framework for agricultural pest monitoring that integrates on-device deep learning inference with server-side geostatistical mapping into a unified, field-deployable smartphone workflow. Three lightweight YOLO variants (YOLOv5n, YOLOv8n, and YOLOv11n) were trained and benchmarked under three TensorFlow Lite quantization schemes (Float32, Float16, and INT8), evaluating detection accuracy (mAP@0.5, mAP@0.5:0.95), inference latency, energy consumption, and thermal behavior on a mid-range Android device. YOLOv11n-INT8 was selected for deployment, achieving mAP@0.5 of 79.3% with a model size of only 2.98 MB and an estimated battery drain of 8.0% h−1. The system detects four pest species of agronomic relevance to soybean and maize crops in Brazil (Diabrotica speciosa, Dalbulus maidis, Diceraeus spp., and Spodoptera frugiperda) directly on the device, while geolocation is automatically extracted from image EXIF metadata and validated against farm boundaries using the Haversine formula. Detection records are synchronized to a cloud database, where Ordinary Kriging (PyKrige) generates continuous pest density surfaces that are returned to the mobile interface for visualization. Field evaluation achieved 95.1% overall accuracy and F1-scores of 0.94 or higher for all species. Kriging validation under dense synthetic sampling yielded R2 up to 0.9656 (species-dependent), with performance degrading under sparse conditions (R2=0.41–0.68 at 200 × 200 grids). Kriging was further validated against Inverse Distance Weighting (IDW) using leave-one-out cross-validation on real field detections (n=79), which showed comparable point-prediction accuracy between the two methods, while Kriging preserved more spatial variance and captured a larger share of the observed infestation hotspot. Our approach integrates on-device pest detection, automated georeferenced validation, and server-side geostatistical mapping within a single operational pipeline, offering a practical tool for Integrated Pest Management in low-connectivity rural environments.

Item Type:	Article
Identification Number:	10.1016/j.compag.2026.112386
Dates:	Date Event 31 August 2026 Accepted 9 September 2026 Published Online
Uncontrolled Keywords:	YOLOv11, TensorFlow lite, Deep learning, Pest detection, Ordinary kriging
Subjects:	CAH10 - engineering and technology > CAH10-01 - engineering > CAH10-01-01 - engineering (non-specific)
Divisions:	Architecture, Built Environment, Computing and Engineering > Engineering
Depositing User:	Gemma Tonks
Date Deposited:	10 Sep 2026 12:43
Last Modified:	10 Sep 2026 12:43
URI:	https://www.open-access.bcu.ac.uk/id/eprint/17236

Actions (login required)

View Item