Abstract
Anthracnose of banana caused by Colletotrichum species is one of the most serious post-harvest diseases, leading to significant yield losses. Clarifying the infection mechanism using non-destructive methods is crucial for timely discrimination and control. This study presents an approach for tracking growth and identifying different infection stages of C. musae in bananas using Vis/NIR spectroscopy. A total of 330 banana reflectance spectra were collected over ten consecutive days. Models were built using traditional feature extraction (PCA, CARS, SPA) combined with machine learning (PLSDA, SVM), and compared against a one-dimensional convolutional neural network (1D-CNN). The 1D-CNN models performed best, achieving accuracies of 95.18% for identifying infection levels and 97.37% for early-stage daily identification. These results indicate the feasibility of identifying infected banana fruit with daily precision.
Introduction
Anthracnose, caused by the fungus Colletotrichum musae, is a major post-harvest challenge for the banana industry. It causes sunken black lesions that lead to significant yield losses during storage and transport. The primary difficulty lies in the fact that early-stage infection is often "asymptomatic" to human inspectors. Traditional detection relies on slow, destructive laboratory techniques or manual inspection, which often catches the disease only after it has already spread. This study seeks a fast, non-invasive way to detect the chemical and structural changes occurring inside banana tissue immediately after fungal contact.
Methods
The researchers compared traditional machine learning pipelines against a modern deep learning approach using a custom dataset of 330 reflectance spectra (339–1019 nm) sampled every 24 hours for 10 days.
- Classification Tasks:
- Four-class: Control, Acceptable (Days 1-4), Moldy (Days 5-7), and Highly Moldy (Days 8-10).
- Five-class (Early Stage): Control vs. Day 1 vs. Day 2 vs. Day 3 vs. Day 4.
- Traditional ML Pipeline: Preprocessing with Standard Normal Variate (SNV) was followed by feature selection (PCA, CARS, SPA) and classification using SVM or PLSDA.
- Deep Learning: A 1D-CNN was implemented for "end-to-end" learning, automatically extracting high-level features from raw spectral curves without manual wavelength selection.
Results
The 1D-CNN model demonstrated superior performance, particularly in identifying the specific day of early-stage infection.
- 1D-CNN Performance: Achieved 95.18% accuracy for identifying infection levels and 97.37% accuracy for early-stage daily identification.
- Traditional ML Performance: The PCA-SVM model was the best non-deep-learning performer, with accuracies of 93.98% and 94.74% for the four- and five-class patterns, respectively.
- Spectral Fingerprints: The study identified specific regions (around 600–680 nm for pigments and 740/842/899 nm for chemical bonds) as critical indicators of fungal activity. Feature selection (SPA) reduced variables to just 11–14 while maintaining over 89% accuracy.
Discussion
The results indicate that Vis/NIR spectroscopy, combined with deep learning, can identify banana fruit infected with C. musae with a resolution accurate to a single day. The model's ability to distinguish between Day 1 and Day 2 of infection suggests that Vis/NIR sensors could be used in ripening rooms to flag infected fruit before they are shipped to retailers. While the study was conducted under controlled laboratory conditions, it paves the way for cheaper, specialized sensors for non-destructive quality control in the banana industry.
Summary updated