Explainable medicinal plant leaf disease diagnosis using linear transformer enhanced InceptionResNetV2 under adversarial evaluation

Abstract Traditional manual inspection of medicinal plants is subjective, tedious, and labor-intensive. It is essential for guaranteeing the purity of pharmaceuticals and their therapeutic efficacy; medicinal plant diagnostics play an important role. This research proposes IRNet-3, a high-precision hybrid diagnostic framework that combines the localized hierarchical feature extraction of Convolutional Neural Networks (CNNs) with the scalable long-range global dependency modeling of Linear Transformers (Linformers). The study has targeted three species of medicinal plants of strategic importance, namely Centella asiatica , Kalanchoe , and Mikania micrantha . IRNet-3 architecture uses a 70% frozen InceptionResNetV2 backbone, along with a Multi-Head Linformer block (Heads = 4, $$k=64$$ ) for computational efficiency, and uses a balanced dataset of 9000 images, composed of samples from primary sources validated by experts refined using a multi-stage preprocessing pipeline (Contrast limited Adaptive Histogram Equalization [CLAHE], Gamma correction, and Gaussian denoising. The experimental results confirm that IRNet-3’s overall accuracy rate is superior (97.11%) and the micro-average (Area Under the Curve, approaching 1.00) is close to the ideal performance, significantly outperforming five high-performing baseline architectures such as InceptionV3, ResNet152V2, EfficientNetB3, EfficientNetV2S and standalone InceptionResNetV2. The model’s decision-making process was carefully explained using post-hoc feature attributions in the multi-perspective Explainable AI (XAI) toolbox, covering LIME, Grad-CAM, Grad-CAM++, and Occlusion Sensitivity, to reduce the black-box effect and establish its trustworthiness in diagnostics. These interpretations validated the model’s emphasis on biologically realistic pathological features (necrotic lesions and structural warping) as opposed to aberrational background features. In addition, a comprehensive adversarial security evaluation with Fast Gradient Sign Method (FGSM), Fast Gradient Method (FGM) and Basic Iterative Method (BIM) attacks (with $$\varepsilon $$ varying in the range 0.00 to 0.10) demonstrated that IRNet-3 is well protected against single-step attacks but fail under iterative gradient manipulation, which can be effectively addressed by adversarial fine-tuning. This research proposes a framework that delivers a reliable, upstream, non-destructive, interpretable and computationally efficient diagnostic solution for quality assurance and biomass triage in botanical agriculture.