Classification of water deficit in blueberry plants using low-cost thermography and artificial intelligence

Abstract

Introduction. Water deficit stress is a limiting factor in blueberry (Vaccinium corymbosum L. ‘Biloxi’) production. Infrared thermography detects this stress, but its high cost limits its adoption. Objective. To develop and validate a low-cost system, based on low-resolution thermography and a multilayer perceptron architecture, for the automatic classification of water status in blueberry plants under controlled conditions. Materials and methods. A longitudinal single-case design, structured as a computational proof of concept, was used with a ‘Biloxi’ plant in a micro-greenhouse in Cundinamarca, Colombia, for 53 days, between June and August 2025. A microcontroller managed a matrix thermal sensor (32 × 24 pixels), environmental sensors, and an RGB camera, and logged data every 20 min. A labeled dataset (“healthy” or “stressed”) was generated using an empirical Crop Water Stress Index (CWSI-E). A multilayer perceptron with a 768-30-1 architecture was trained. Results. The model reached the target mean squared error in eleven epochs. The final validation accuracy was 99.34 %. The confusion matrix showed four false negatives and one false positive, which evidenced the discrimination capability. Conclusions. Low-resolution thermal patterns contain sufficient information for a multilayer perceptron to classify water status in blueberry plants under controlled conditions. This methodology constitutes a low-cost and accessible alternative for precision agriculture.