FewShotCropNet: Real-Time Detection of Emerging Crop Diseases with Limited Labels Using Spectral-Temporal Attention Prototypical Networks

The acquisition of labelled data for new or emerging plant diseases is difficult due to the high cost and logistical complexity of ground-truth collection, challenges that are further compounded by the limited infrastructure available to smallholder farmers in sub-Saharan Africa. This study presents FewShotCropNet, a few-shot learning model based on Spectral-Temporal Attention Mechanisms and Prototypical Networks that utilises multispectral time-series data from Sentinel-2 to classify crop diseases. Two principal innovations are incorporated in the proposed model: (1) a spectral attention mechanism based on the Squeeze-and-Excitation approach to learn disease-relevant spectral band weights; and (2) a temporal attention pooling mechanism to identify the most discriminative growth stages for disease classification. The model employs a two-phase training strategy comprising supervised pre-training followed by episodic meta-learning, enabling the generation of optimal feature representations under extreme label scarcity. Crop disease detection experiments were conducted in Edo State, Nigeria on cassava and maize using monthly Sentinel-2 composites from 2024 (10 spectral bands and five vegetation indices across twelve temporal steps). Under a 4-way 5-shot classification scenario with 100 GPS-validated labelled samples (25 per-class), FewShotCropNet achieved a mean accuracy of 98.15% with a 95% confidence interval of ±0.58%. An equitable comparison was enabled by introducing a Pre-trained Simple Prototypical Network, a variant sharing the same two-phase training strategy as FewShotCropNet but without the attention modules—which achieved 97.75% (±0.63%). FewShotCropNet exceeded the Pre-trained Simple ProtoNet by +0.40 percentage points (t = 1.52, p = 0.13), with the attention module contribution verified as positive though not statistically significant on the current dataset. Statistically significant improvements over models trained without pre-training were observed: FewShotCropNet outperformed the Relation Network (94.40%), Matching Network (94.40%), and the Optimised Baseline Convolutional neural networks (CNN) (86.65%, pre-trained backbone with 5-shot linear probe), with p < 0.0001 for both episodic meta-learning comparisons. The spatial extraction buffer was reduced from 500 m to 250 m to minimise mixed-pixel contamination. Ablation studies confirmed that both spectral and temporal attention modules contribute positively to classification performance when pre-training is controlled. These findings suggest that attention-enhanced meta-learning, combined with supervised pre-training and narrower spatial extraction buffers, represents a viable pathway for scalable, real-time crop disease monitoring in data-scarce agricultural environments.