Domain-aligned feature extraction with Mobilenetv2-GRL and KNN for wheat leaf disease detection

• Introduces a novel pipeline combining MobileNetV2 with Gradient Reversal Layer (GRL) for robust domain-invariant feature extraction, followed by K-Nearest Neighbors (KNN) for interpretable classification of 12 wheat leaf disease classes. • Achieves 98.0% accuracy on clean images and 94.0% on domain-shifted (augmented) data—outperforming the baseline MobileNetV2 model by 10% under real-world simulation conditions. • Implements the model in a lightweight, browser-based Streamlit interface for real-time wheat disease detection, enabling practical use by farmers and agricultural field workers in resource-constrained settings. Leaf diseases of wheat can significantly impact the global food supply, making it critical to detect them as soon as they occur and provide accurate estimates for monitoring their spread and evolution. The commonly used manual detection approaches are prone to errors, expensive, and inefficient, spurring a growing interest in automated alternatives. While CNN-based solutions have demonstrated promising results, they often fall short in terms of generalization and interpretability. We present a novel, hybrid, domain adaptive pipeline for wheat leaf disease identification that combines semantic segmentation and classification. Our approach employs a pretrained MobileNetV2 model augmented with a Gradient Reversal Layer (GRL) for extracting domain invariant feature embeddings. These embeddings, in turn, are passed through a K-Nearest Neighbors (KNN) classifier for final prediction to boost interpretability and domain invariance. The system design separates feature extraction from the classification, resulting in improved generalization performance. The feature extractor is fine-tuned using a multi-task loss, enabling it to learn domain agnostic embeddings. The GRL enforces feature consistency between clean and augmented images by reversing the gradient direction during backpropagation. Finally, the model outputs a 256-D feature embedding to be classified using KNN. The approach achieves 98.0% and 94.0% accuracy on the clean test set and the combined clean-augmented dataset, respectively, outperforming standard MobileNetV2 baselines. We also provide a lightweight, browser-based, Streamlit app for fast and easy disease diagnosis by simply uploading an image of the leaf to enable low-resource field-level solutions for real-world monitoring of these plant diseases.