TITLE:
Intelligent System to Aid in the Detection of Cocoa Diseases by Processing Images of Infected Leaves
AUTHORS:
Guié Toa Bi, Vangah Wognin, Alico Nango, Kouassi Francis Yao, Kadja Anicet, Sié Ouattara, Alain Clement
KEYWORDS:
Foliar Diseases, Digital Images, Random Forest, Transfer Learning (SqueezeNet), Tropical Crop (Cocoa)
JOURNAL NAME:
Advances in Molecular Imaging,
Vol.15 No.3,
September
28,
2026
ABSTRACT: Cocoa leaf diseases cause 40% losses in the Tonkpi region due to a lack of accessible diagnostic tools. Existing automated solutions do not cover this diagnosis in this specific geographic context. This work hypotheses that a system combining image segmentation, quantitative descriptors, and machine learning can detect these diseases under real-world conditions in Tonkpi. A corpus of 805 images of cocoa leaves, divided into six classes, was compiled in January 2026 in four locations using a Tecno Spark 20C smartphone, following a standardized protocol. Two segmentation methods, Otsu and K-means (K = 3), were compared. Fifty-three descriptors selected by ANOVA were then used to populate the k-NN and Random Forest classifiers, while SqueezeNet processed the raw images. The implementation in MATLAB R2023b employs five-fold cross-validation and a 60/20/20% distribution. The pct_necrose_R descriptor is the most discriminating pathological signal for this crop. Random Forest achieves 100% accuracy in binary detection, surpassing k-NN (97.52%). SqueezeNet identifies all six classes with 99.38% accuracy. SIADMA is the first system to compare two segmentations, 53 descriptors, and three classifiers on a local tropical corpus dedicated to cocoa cultivation in the Tonkpi region. The optimal architecture of Otsu, Random Forest, and SqueezeNet, deployable on smartphones, provides a concrete foundation for a decision-support tool accessible to farmers and field agents.