Fourier-Guided Multi-Scale Vision Transformer for Object-Independent Detection and Severity Grading of Rain-Induced Cracking in Sweet Cherry

Abstract

Abstract Rain-induced cracking is among the most damaging preharvest disorders of sweet cherry because even narrow cuticular fractures reduce fresh-market value and create entry sites for water loss and fungal infection. Conventional inspection is subjective and is least reliable for low-contrast microcracks near the pedicel cavity. This study developed a Fourier-guided multi-scale Vision Transformer (FGM-ViT) for joint fruit-level severity grading and pixel-level crack segmentation. The dataset comprised 960 individual fruit from four commercial cultivars and eight orchard blocks, imaged from four rotational views after natural rainfall exposure or controlled water-immersion challenge. Crack severity was assigned as intact, microcrack, moderate, or severe using stereomicroscopic measurements of cumulative crack length, maximum width, affected surface area, and crack location. All views of an individual fruit and all observations from the same orchard block were constrained to the same fold in a nested five-fold grouped evaluation. FGM-ViT combined a multi-scale spatial encoder with luminance-normalised Fourier residual tokens and frequency-guided cross-scale attention. The complete model achieved 93.8 ± 0.9% fruit-level accuracy, 93.7 ± 1.0% macro-F1, and 0.842 ± 0.018 Dice coefficient for crack segmentation. It exceeded the spatial-only transformer by 2.2 percentage points in accuracy and 2.2 points in macro-F1, with the largest gain observed for microcracks. Classification errors were restricted almost entirely to adjacent severity categories. Under pedicel overlap, glare, brightness shifts, and motion blur, FGM-ViT retained a 2.7–3.5-point advantage over the spatial-only model. Frequency-band occlusion indicated that mid-to-high spatial frequencies carried complementary evidence for narrow fractures, whereas spatial features remained essential for crack location and marketability interpretation. The dual-output framework provides an auditable, low-cost RGB approach for sweet-cherry sorting, cultivar screening, and rain-cracking phenotyping.

Code and data availability

The preprint states that paper-specific assets (crack masks, metadata, out-of-fold predictions, and analysis code) will be deposited in a public repository, but no repository, identifier, or URL is provided and nothing is yet publicly available, so the assets are not currently actionable.

No evidence-backed public reproduction asset is currently recorded.