[Paper] From Code to Field: Evaluating the Robustness of Convolutional Neural Networks for Disease Diagnosis in Mango Leaves

Source: arXiv - 2512.13641v1

Overview

This paper investigates how well popular convolutional neural networks (CNNs) can diagnose mango leaf diseases when the input images are degraded by realistic noise, blur, and weather effects. By creating a corrupted version of the MangoLeafDB dataset and benchmarking several models, the authors show that a lightweight, purpose‑built network (LCNN) can be more robust than heavyweight architectures like ResNet‑101—an insight that matters for deploying AI on edge devices in agriculture.

Key Contributions

• Robustness‑focused dataset: Extended the public MangoLeafDB into MangoLeafDB‑C, injecting 19 artificial corruptions (e.g., Gaussian noise, defocus blur, motion blur) at five severity levels.
• Comprehensive benchmark: Evaluated five CNN families—ResNet‑50, ResNet‑101, VGG‑16, Xception, and a custom lightweight CNN (LCNN)—using F1, Corruption Error (CE) and relative mean Corruption Error (mCE).
• Empirical finding: LCNN consistently outperformed larger models on corrupted images, achieving the lowest mCE while maintaining competitive accuracy on clean data.
• Practical recommendation: Highlights the importance of robustness testing for AI‑driven plant disease diagnostics, especially for low‑resource, edge‑computing scenarios.
• Open research direction: Calls for robustness to become a first‑class evaluation metric in agricultural AI pipelines.

Methodology

1. Dataset preparation
   • Started from the original MangoLeafDB (healthy vs. diseased leaf images).
   • Applied the ImageNet‑C style corruption pipeline to synthesize MangoLeafDB‑C, covering 19 distortion types (e.g., Gaussian noise, snow, fog, JPEG compression) at five increasing severity levels.
2. Model selection
   • Chose four off‑the‑shelf CNNs (ResNet‑50/101, VGG‑16, Xception) representing modern deep architectures.
   • Designed LCNN, a shallow, parameter‑efficient network tailored to mango leaf features (e.g., texture‑focused filters, reduced depth).
3. Training & evaluation
   • Trained each model on the clean training split of MangoLeafDB.
   • Tested on both clean and corrupted test sets.
   • Computed F1 score (balanced precision/recall), Corruption Error (CE) per distortion, and relative mean CE (mCE) to aggregate robustness across all corruptions.
4. Analysis
   • Compared performance drop from clean to corrupted conditions and ranked models by overall robustness (lowest mCE) and by specific realistic corruptions (defocus/motion blur).

Results & Findings

• LCNN achieved the lowest mCE (0.97), meaning it degraded the least across all corruptions.
• On defocus blur and motion blur—common in field photography—LCNN’s F1 dropped <5 %, whereas ResNet‑101 fell >20 %.
• Larger models retained high accuracy on pristine images but were significantly more sensitive to noise, compression artifacts, and weather‑related distortions.
• The gap widened with higher severity levels, confirming that sheer depth does not guarantee robustness.

Practical Implications

• Edge deployment: LCNN’s small footprint (≈1.2 M parameters) and robustness make it ideal for smartphones, Raspberry Pi‑class devices, or custom IoT cameras used by smallholder farmers.
• Cost‑effective disease monitoring: Robust models reduce the need for expensive image‑preprocessing pipelines (e.g., denoising, deblurring) in the field, lowering latency and power consumption.
• Model selection guidelines: When building AI tools for agriculture, prioritize robustness metrics (CE/mCE) alongside accuracy, especially if the target environment includes variable lighting, motion, or compression.
• Data collection strategy: The MangoLeafDB‑C pipeline can be reused for other crops, encouraging developers to simulate realistic field conditions before model release.
• Integration with decision support: Reliable leaf‑disease predictions can feed directly into automated spraying systems or advisory apps, improving yield and reducing pesticide overuse.

Limitations & Future Work

• Synthetic vs. real-world corruptions: The study relies on artificially generated distortions; real field images may exhibit compound effects (e.g., simultaneous blur and illumination changes) not fully captured.
• Single‑crop focus: Findings are specific to mango leaves; transferability to other plant species remains to be validated.
• Model diversity: Only five architectures were tested; newer vision transformers or self‑supervised models could behave differently.
• Hardware evaluation: The paper reports inference speed only indirectly; a thorough benchmark on actual edge hardware would strengthen deployment claims.
• Future directions: The authors suggest extending robustness testing to multi‑modal data (e.g., hyperspectral), incorporating domain‑adaptation techniques, and exploring automated robustness‑aware neural architecture search for agricultural tasks.

Authors

• Gabriel Vitorino de Andrade
• Saulo Roberto dos Santos
• Itallo Patrick Castro Alves da Silva
• Emanuel Adler Medeiros Pereira
• Erick de Andrade Barboza

Paper Information

• arXiv ID: 2512.13641v1
• Categories: cs.LG, cs.AI, cs.CV
• Published: December 15, 2025
• PDF: Download PDF