Accurate recognition of potato leaf diseases is crucial for timely intervention and yield preservation in precision agriculture. Conventional deep learning-based approaches often rely solely on probabilistic supervision, which may fail to capture the fine-grained structural and spectral cues essential for distinguishing visually similar disease symptoms, especially under class imbalance. This paper proposes a spectral-topological integrated loss (STIL) that unifies probabilistic, frequency-domain, and topology-domain supervision within a single training objective. The spectral term aligns the frequency statistics of class activation maps with class-specific prototypes, enhancing edge and texture fidelity, whilst the topological term constrains global shape consistency by preserving connectedness and hole patterns. An adaptive difficulty weighting strategy further directs structural supervision towards ambiguous and minority-class samples. We integrate STIL into diverse backbone architectures, including CNN, VGG19, MLP, ResNet18, MobileNetV2, ShuffleNetV2 and Vision Transformer, and evaluate its performance on the imbalanced PlantVillage potato leaf dataset. Experimental results demonstrate consistent improvements across all metrics, with gains in accuracy of 0.9–1.2% and F1-score of 0.9–1.9% over standard cross-entropy loss. Ablation studies confirm the complementary benefits of spectral and topological consistency, with notable advantages in recovering hard positives and reducing fragmented predictions. Owing to its plug-and-play nature and architecture-agnostic design, STIL offers a generalizable solution for structure-aware visual recognition, with significant potential for deployment in agricultural disease monitoring systems.
Full publication URL
