There is a substantial risk that agricultural diseases pose to the safety of food around the world. These diseases can cause large decreases in productivity and lead to greater expenses for farmers. Rapid and accurate detection of these diseases is necessary to improve agricultural output and ensure the industry's continued viability. Traditional techniques of visual examination, on the other hand, can be expensive, subjective, and imprecise at times, particularly in large-scale agricultural operations. This study presents AgriHAFNet, a dual-branch deep learning methodology for scalable multi-crop disease identification to address this challenge. The framework employs EfficientNet-B0 to encode specific lesion appearance features and ConvNeXt-V2 for its capacity to collect long-range, contextual data. An attention-based method connects these networks to increase robustness and accuracy. Training on 18,592 images from six crops with thirty disease categories yielded 99.77% accuracy on training images, 96.96% accuracy when validating with other images, 95.65% accuracy for unseen testing images and a macro F1-score of 0.94. The results suggest that AgriHAFNet has excellent generalization ability and potential for implementation in precision agriculture.