To the Editor: Korkmaz et al. report high performance for macrofungi identification using convolutional networks with Grad-CAM and Score-CAM visualizations, trained on a 2800-image compilation from natural photographs and GBIF, split 60/20/20 into training, validation, and testing [1]. The topic is timely; nevertheless, some choices can overstate accuracy and make transfer to real-world surveys less certain.
First, independence between training and testing depends on how images relate to specimens, photographers, and sites. When multiple shots of the same individual or collection event are randomly dispersed across the split, near-duplicates and shared backgrounds may leak scene-level cues into evaluation. Grouping by specimen, collector, location, or photographer and then withholding groups—rather than individual images—would better reflect deployment, in line with cross-validation strategies for spatial or hierarchical data [2].
Second, preprocessing, augmentation, and any feature selection should be restricted to the training folds and then applied unchanged to held-out data. Using the full dataset to set parameters, or optimizing on the validation set without nesting, allows target-related structure to seep into assessment, inflating discrimination. A fully nested workflow (split, fit transforms and tuning only on training, freeze, then evaluate) limits selection bias and yields estimates closer to field performance [2].
Third, very high accuracy can arise from shortcuts: models may latch onto background texture, color casts, or photographic conventions instead of morphology, especially with mixed sources. Saliency maps help, but qualitative overlays alone are not decisive; sanity checks and quantitative focus tests can verify that explanations change with the model and that highlighted regions matter for the prediction [3,4]. In addition, calibrated probabilities with reliability assessment—rather than raw scores—let teams choose thresholds appropriate for biodiversity monitoring and curation workflows [5].
For taxonomy and conservation work, what endures is performance on new cameras, unseen habitats, and independent collections. Organizing the split at the level of specimens or sources, nesting all preprocessing and tuning, adding external or site-withheld evaluation, and reporting calibration would make the reported figures a dependable guide for practitioners while preserving the useful interpretability provided by XAI.
Conflicts of Interest
The author declares no conflicts of interest.
References
- Korkmaz, A.F.; Ekinci, F.; Altaş, Ş.; Kumru, E.; Güzel, M.S.; Akata, I. A Deep Learning and Explainable AI-Based Approach for the Classification of Discomycetes Species. Biology 2025, 14, 719. [Google Scholar] [CrossRef] [PubMed]
- Roberts, D.R.; Bahn, V.; Ciuti, S.; Boyce, M.S.; Elith, J.; Guillera-Arroita, G.; Hauenstein, S.; Lahoz-Monfort, J.J.; Schröder, B.; Thuiller, W.; et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 2017, 40, 913–929. [Google Scholar] [CrossRef]
- Geirhos, R.; Jacobsen, J.H.; Michaelis, C.; Zemel, R.; Brendel, W.; Bethge, M.; Wichmann, F.A. Shortcut learning in deep neural networks. Nat. Mach. Intell. 2020, 2, 665–673. [Google Scholar] [CrossRef]
- Adebayo, J.; Gilmer, J.; Muelly, M.; Goodfellow, I.; Hardt, M.; Kim, B. Sanity Checks for Saliency Maps. In Proceedings of the 32nd Conference on Neural Information Processing Systems (NeurIPS 2018), Montreal, QC, Canada, 2–8 December 2018; pp. 9505–9515. [Google Scholar]
- Guo, C.; Pleiss, G.; Sun, Y.; Weinberger, K.Q. On Calibration of Modern Neural Networks. In Proceedings of the 34th International Conference on Machine Learning, Sydney, Australia, 6–11 August 2017; pp. 1321–1330. [Google Scholar]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.