Thai University RankingsRESEARCH RADAR
Evidence of global relevance

Leakage-Aware Transfer Learning with Explainable AI and CPU-Efficient Deployment for Mango Leaf Disease Classification on the MangoLeafBD Benchmark

After detecting duplicate and near-duplicate MangoLeafBD images before group-aware splitting, EfficientNetB0 achieved 99.50% accuracy and weighted F1 on a 599-image test set, with grouped cross-validation averaging 99.63%. CPU latency averaged 25.3 ms per image. Performance is strong on the benchmark but unproven in real orchards and new regions.

01

Key findings

  • EfficientNetB0 achieved 99.50% accuracy and weighted F1 on 599 test images; the ensemble matched but did not exceed it. Grouped cross-validation yielded 99.63±0.20% accuracy and 99.62±0.20% F1. Mean CPU latency was 25.3 ms, p95 27.6 ms, with throughput up to 166.5 images/s.
02

Why this matters globally

Mango disease threatens production globally. The leakage-aware protocol offers a reproducible evaluation standard for agricultural AI beyond headline accuracy.

03

Thai researcher contribution

Khon Kaen, Kasetsart and Sakon Nakhon Rajabhat University researchers combined accuracy, interpretability and low-cost deployment.

04

Limitations to consider

MangoLeafBD is a public benchmark, not a representative Thai orchard sample. Lighting, backgrounds, cultivars, severity and co-disease may shift. Grad-CAM is not causal explanation, and external validation, calibration, abstention and farmer-impact evidence are absent.

05

Verify the original sources

Applied SciencesRead the original article

DOI: 10.3390/app16146989

KEEP EXPLORING

More Thai research to explore