Technical Summary: Few-Shot Hyperspectral Aphid Detection via FastGAN, Transformer-Based Classification, and Explainable AI

Problem Statement

Early detection of aphid infestation (Aphis fabae) in crops, specifically faba bean (Vicia faba), is critical for preventing yield loss and minimizing unnecessary pesticide application. While hyperspectral imaging (HSI) combined with Spectral Information Divergence (SID) analysis offers a non-destructive method for monitoring plant health, the application of deep learning to this domain is severely constrained by the scarcity of labeled datasets. Deep convolutional neural networks (CNNs) and Vision Transformers (ViTs) typically require large volumes of data to generalize effectively; small datasets restrict model complexity and increase the risk of overfitting. Furthermore, infested leaves exhibit high heterogeneity in spectral and structural patterns (e.g., variable lesion patterns, discoloration), making them more difficult to model than healthy leaves.

Methodology

The study proposes a data-efficient framework integrating synthetic data generation, deep learning classification, and explainable AI (XAI) to address data scarcity in hyperspectral aphid detection.

1. Dataset and Preprocessing

The study utilized hyperspectral SID images of faba bean leaves collected under controlled greenhouse conditions, covering both healthy and aphid-infested samples. Preprocessing included reflectance calibration, region-of-interest (ROI) extraction, Standard Normal Variate (SNV) normalization, and noise removal. To adapt hyperspectral cubes for 2D architectures, two strategies were employed: (i) generating three-channel representations via Principal Component Analysis (PCA-RGB) or selected VIS/NIR bands, and (ii) maintaining a narrowband stack (710–825 nm) identified as discriminative for aphid stress. The dataset was partitioned at the plant level (70% training, 15% validation, 15% testing), with the test set remaining entirely real to ensure objective evaluation.

2. Synthetic Data Generation (FastGAN)

To augment the limited SID dataset, the authors employed FastGAN, a lightweight generative framework designed for high-resolution synthesis from few-shot data.

• Architecture: The generator utilizes hierarchical up-sampling and residual blocks linked via Skip-Layer Channel-wise Excitation (SLE) modules to stabilize training and preserve multi-scale context. The discriminator employs multi-scale convolutions with spectral normalization.
• Training: The model was trained using non-saturating logistic loss with an R1 gradient penalty. Differentiable augmentations (translation, cutout, brightness/contrast) were applied to both real and synthetic samples to regularize training.
• Output: A converged generator produced 10,000 synthetic SID images that preserved the spatial and spectral structure of real samples.

3. Image Quality Evaluation

The fidelity of the generated images was quantified using the Fréchet Inception Distance (FID). Two variants were used:

• HSI-Encoder FID: Features derived from a self-supervised 3D-CNN trained on real SID cubes.
• PCA-RGB FID: Hyperspectral cubes reduced to three principal components and processed by an Inception-V3 model.
  Results indicated that while healthy leaves converged rapidly to low FID scores (75–90), infested leaves required more iterations to stabilize (100 FID) due to the greater heterogeneity of infection patterns.

4. Classification Architectures

Four deep learning models were trained and evaluated on both real-only and Real+GAN augmented datasets:

• VGG16: A classical deep CNN with uniform 3×3 convolutional filters.
• ResNet-50: Utilizes residual skip connections to mitigate vanishing gradients.
• EfficientNet: Applies compound scaling to balance network depth, width, and resolution.
• Vision Transformer (ViT-B/16): Replaces convolutions with self-attention mechanisms to capture long-range spatial dependencies.
  All models were initialized with ImageNet weights and fine-tuned for binary classification (healthy vs. infested).

5. Explainable AI (XAI)

To interpret model decisions, the study compared Grad-CAM and LayerCAM. LayerCAM was selected for its ability to generate higher-resolution localization maps by aggregating gradients from multiple convolutional layers, preserving fine-grained spatial details crucial for detecting small-scale aphid infestations.

Key Results

1. Synthetic Data Quality

FastGAN successfully generated realistic synthetic images. Visual progression showed that while healthy leaf generation stabilized quickly (FID < 100 by ~70 iterations), infested leaf generation exhibited oscillations and required ~120 iterations to stabilize, reflecting the complexity of modeling heterogeneous disease symptoms.

2. Classification Performance

• Impact of Augmentation: Dataset augmentation significantly improved classification robustness across all architectures, reducing false negatives and improving generalization, particularly for models trained on smaller datasets.
• Model Comparison:
  • ViT: Achieved the highest performance, reaching perfect accuracy (1.00) and F1-scores (1.00) with 2,000–10,000 samples and maintaining high performance (0.97) even with 100,000 samples. The self-attention mechanism proved superior in capturing complex spectral-spatial patterns.
  • EfficientNet: Demonstrated strong, balanced performance (F1-score 0.95 with 20,000 samples) and high sensitivity to disease characteristics.
  • ResNet-50: Showed moderate improvements over VGG16 with stable behavior across dataset scales.
  • VGG16: Exhibited signs of overfitting on small datasets and degraded performance as dataset complexity increased, showing reduced adaptability to diverse data.
• Confusion Matrices: Advanced architectures (ViT and EfficientNet) significantly reduced false negatives compared to VGG16, which struggled to detect infested leaves (high false-negative rate).

3. Explainability

LayerCAM visualizations revealed that mid-level feature layers (Feature 5) in EfficientNet provided the most accurate localization of aphid infestations, highlighting specific texture and structural alterations. Deeper layers (Features 7–8) focused on global plant structures, while shallow layers lacked semantic context. This confirmed that intermediate layers retain the critical fine-grained information necessary for disease detection.

Significance and Contributions

The paper claims four primary contributions:

1. Framework Development: The creation of a FastGAN-based framework capable of generating synthetic SID images from limited hyperspectral datasets of aphid-infested faba bean leaves.
2. Architectural Comparison: A comprehensive evaluation of four distinct deep learning families (VGG16, ResNet-50, EfficientNet, ViT) under both real and GAN-augmented conditions.
3. Few-Shot Robustness: Demonstration that synthetic data augmentation effectively improves the robustness of aphid detection under few-shot conditions, mitigating overfitting in deep learning models.
4. Interpretability: The integration of LayerCAM to identify specific image regions contributing to classification, offering more biologically meaningful insights than traditional Grad-CAM.

Conclusion

The study concludes that combining hyperspectral imaging with FastGAN-based data augmentation and transformer-based learning provides a promising, data-efficient approach for early, non-destructive aphid detection. The Vision Transformer emerged as the most reliable model for discriminating between healthy and infested leaves, attributed to its ability to capture global context and long-range dependencies. While the framework showed significant potential, the authors note limitations regarding generalizability to other crop-pest systems, the need for spectral fidelity validation, and the requirement for future testing under uncontrolled field conditions. The findings suggest that such integrated approaches could contribute to improved crop monitoring and more sustainable pest management strategies.