Ecological context supports image-based fungal species recognition in citizen-science observations
Year: 2026
Type: article
Abstract: Citizen-science initiatives generate large numbers of image-supported fungal observations that can contribute to biodiversity recording, but their scientific value depends partly on reliable species identification. Recognition from field photographs is challenged by variable imaging conditions, a long-tailed species-frequency distribution, and visual similarity, while ecological descriptors recorded with the same observations provide context unavailable to image-only models. We examined whether habitat, substrate, and month improve fungal species recognition and compared matched strategies for incorporating these fields. Using Danish Fungi 2024-Mini (36,393 images from 20,618 observations of 182 species), we developed the Multi-scale Ecological Fusion Network (MEF-Net), which combines pooled multi-scale ConvNeXt descriptors with ecological descriptors through direct fusion and bounded channel recalibration. At an input resolution of 224 × 224 pixels, MEF-Net achieved mean Top-1 accuracy, Top-3 accuracy, and Macro-F1 values of 72.21%, 88.35%, and 61.68%, respectively, across five independent runs, a 5.51-percentage-point increase in Macro-F1 over the matched image-only ConvNeXt-Base model. At 384 × 384 pixels, mean Macro-F1 was 63.72%, 4.21 percentage points higher than the matched image-only model. These overall gains reflected the combined contributions of the multi-scale visual representation and ecological-context integration. Habitat produced the largest mean gain among the individual fields, whereas using all three fields jointly produced the best metadata-aware result. In the controlled strategy comparison, ecological conditioning of the pooled visual descriptors achieved the highest mean performance, while direct concatenation and calibrated Bayesian Late Fusion also improved upon the matched multi-scale image-only reference. MEF-Net achieved higher Macro-F1 across rare, medium-frequency, and frequent species, with the largest difference in the rare group. These findings indicate that ecological observation context can complement image evidence and support candidate ranking within human-validated fungal biodiversity-recording workflows.
Authors: Chong Liu, Jianxin Wang, Yeru Wang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,077
OpenAlex ID: W7225752784