• Home  
  • Coconut Disease AI Reached 97% Accuracy, but Optimizer Gains Were Unclear
- Research Briefs

Coconut Disease AI Reached 97% Accuracy, but Optimizer Gains Were Unclear

A coconut disease image model achieved 97.13% average test accuracy, but its optimiser did not clearly beat random search.

A coconut disease image model achieved 97.13% mean accuracy across repeated experiments. Its specialised optimisation procedure, however, did not clearly outperform random search under a matched evaluation budget.

What researchers studied

Researchers combined features from three pretrained image networks: InceptionV3, EfficientNetB0 and MobileNetV2. They evaluated 5,798 field images spanning five coconut disease categories. Exact-image hashing, perceptual similarity checks and deep-feature clustering helped identify related photographs. The researchers separated data by plant to limit leakage, and restricted augmentation and class-balancing to the training partition. Multiple runs and group-aware validation were used to assess reliability.

What they found

Mean test accuracy across five runs was 97.13%, with a reported 95% confidence interval of 96.56%–97.70%. Macro-F1 was 96.32%, while nested group-aware cross-validation reached 97.99% accuracy. Not grouping images from the same plant led to an approximately 1.6-percentage-point overestimate. The proposed dual-optimizer performed comparably to random search under equal budgets, and the additional gains were not statistically significant.

What it means

The work shows why agricultural AI evaluation should separate genuinely different plants rather than relying on random image splits. Closely related photographs can make performance look better than it would be on unseen crops. The model offers a basis for further testing, and the leakage audit is useful beyond coconut farming. A high experimental score is not evidence that farmers can safely use the system without expert support.

Important limitation

Independent testing across farms and geographic locations remains necessary. Lighting, plant varieties, camera quality and unfamiliar disease conditions could change field performance. The study did not demonstrate operational diagnostic reliability or a statistically superior optimiser. Its findings should be interpreted as controlled image-classification results rather than confirmed agricultural outcomes.

Source Information

Study: Dual-optimizer driven hybrid ensemble model for automated coconut disease recognition

Authors: B. Arulgnanaselvi and S. Prasanna

Journal / Source: Frontiers in Artificial Intelligence

Published: 20 November 2026

DOI: 10.3389/frai.2026.1869799

Contact Us

Research Today is a South African digital publication that makes credible research easier to understand.

TERMS OF USE & PRIVACY POLICY

follow us