Improved strawberry disease classification under class imbalance through in-backbone latent diffusion

Strawberry production suffers major yield losses from fungal diseases such as Powdery mildew, Fusarium wilt, and Gray mold. These yield losses make early and accurate classification essential for effective crop management. Achieving accurate classification, however, is challenging in greenhouse conditions due to class imbalances and data scarcity. To mitigate these limitations, diffusion-based approaches, including pixel-space diffusion, have been proposed for data synthesis, but they often fail to preserve the fine-grained pathological features necessary for reliable classification. To overcome this challenge, we introduce an in-backbone latent diffusion framework that integrates a stochastic feature refinement process directly into the intermediate feature representations of a deep image classifier. By bypassing pixel reconstruction and integrating diffusion between conv3 and conv4 of a ResNet-50, the method enhances disease-discriminative features while preserving structural consistency within the feature space. On a four-class strawberry dataset (three fungal diseases and healthy) collected under real greenhouse conditions, our method achieved a macro-averaged F1-score of 0.90. This considerably outperformed the baseline without augmentation (0.34), pixel-level diffusion (0.50), and state-of-the-art architectures including EfficientNet-B4 (0.31) and ViT-B/16 (0.19). Our approach was particularly effective for the minority classes, improving recall for Powdery mildew from 0.08 to 0.90, and for Fusarium wilt from 0.13 to 1.00. Compared to pixel-level diffusion, the proposed method achieved a 6.5× reduction in inference latency with only 3% memory overhead relative to the baseline ResNet-50. Further analysis showed that in-backbone latent diffusion promotes cleaner class separation and preserves subtle lesion cues. Our study demonstrates that in-backbone latent diffusion provides a robust, efficient, and domain-adapted feature purification alternative to pixel-level diffusion synthesis and conventional classification architectures.