Work

A robust and interpretable hybrid SunAFNet framework for sunflower classification with adversarial resilience and mIoU-validated explainability

Year: 2026
Type: article
Abstract: Accurate and reliable monitoring of sunflower (Helianthus annuus) developmental phases and health status is fundamental for optimizing agricultural yields and ensuring global food security. Manual diagnostics are subjective and time-consuming, and most stand-alone deep learning architectures are limited in their ability to accurately model the fine-grained spatial texture and long-range relationships between the various parts of the capitulum. For this study, we introduce a novel dual-stream hybrid framework: SunAFNet which combines a frozen DenseNet201 backbone with a selectively fine-tuned Vision Transformer (ViT-B/16). The key component of the architecture is a flexible Attention Fusion Module, which dynamically weights local spatial descriptors and global contextual features based on input complexity with an MLP-Softmax mechanism. SunAFNet has been evaluated on a large image dataset of 5,111 images of capitula of five growth and health classes and resulted in a diagnostic accuracy of 97.26%, which is better than the existing CNN and Transformer baselines. The framework showed very high generalization capability, achieving 94.26% accuracy on an independent secondary dataset, and remained very close to the optimal Robustness Index of 0.9899 despite various environmental and digital perturbations. A low value of the Expected Calibration Error (ECE 0.0675) was found, and the model achieved near-perfect selective prediction (accuracy 1.000) on clean data. Furthermore, SunAFNet demonstrated pronounced defensive resilience across a multi-level FGSM adversarial spectrum ( \(\:\epsilon\in\:\left\{0.01,0.05,0.10,0.20,0.30\right\}\) ), maintaining 71.71% accuracy at \(\:ϵ=0.10\:\) and retaining 42.35% diagnostic capability even under extreme perturbation ( \(\epsilon = 0.30\) ), well above stochastic chance. Moreover, a multi-layered Explainable AI (XAI) suite was validated empirically to ensure that the model’s logic is based on biologically relevant floral markers with a statistically significant mIoU of 0.7487 ( p = 0.0098), ensuring grounded diagnostic decisions. SunAFNet can provide a platform to support real-time precision agricultural monitoring, which is both fast and efficient in terms of both computing and inference speed, with an inference latency of merely 38.35 ms.
Language: English
Subfield: Plant Science
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7235309337
FWCI:
Cites: 48
Cited by:
Related to: 10