Mobile System with a Hybrid Hierarchical Model for Estimating the Foliar Severity of Downy Mildew in Quinoa under Controlled Capture Conditions

Abstract

Visual quantification of downy mildew foliar severity in quinoa is a labor-intensive process and may vary across observers, thereby limiting its standardization in phenotypic assessment and phytopathological monitoring. The present study aimed to experimentally evaluate a mobile system designed to support the automated quantification of downy mildew foliar severity in quinoa. A system named QuinuApp was developed, comprising a mobile application, a standardized image acquisition device (BlackBox), and a hierarchical hybrid estimator based on interpretable visual features. The study followed a quantitative approach under an experimental design, using 772 RGB images of individual adaxial leaves from 92 accessions. Severity, expressed on a continuous scale from 0% to 100%, was assessed by an expert plant pathologist and used as the operational reference. The pipeline included photometric normalization, main leaf segmentation, extraction of 55 image-derived clinical-textural features, auxiliary classification by biological phases, and hybrid regression with isotonic calibration, incorporation of sporulation as a contextual variable, accession-based contextual adjustment, and biologically informed rules. On the test set, the model achieved Pearson’s r = 0.699, CCC = 0.694, MAE = 17.17%, and RMSE = 24.75%, attaining the best overall performance compared with linear regression, Random Forest, XGBoost, and SVR. Bland–Altman analysis showed a mean bias of −0.58%, with 50.78% of predictions within ±10% and 69.43% within ±20%. The proposed system demonstrated technical feasibility as a supportive tool for the automated estimation of downy mildew foliar severity in quinoa under controlled experimental conditions, with potential applicability in assisted digital phenotyping.