An Explainable Deep Learning Framework for Sorghum Weed Classification Using Multi-Scale Feature Enhanced DenseNet

Sorghum is a vital crop known for its resilience in harsh environmental conditions, playing a key role in food security and bioenergy production across the globe. Accurate sorghum weed classification is crucial for efficient agricultural management, yet traditional methods are labour-intensive and resource-heavy. This research presents the first work on the sorghum weed classification dataset, introducing a novel deep-learning model using a customised DenseNet-169, enhanced with two innovative feature extraction blocks: Dimensional Convolutional Feature Fusion (DCFF) and Deformable Multi-Kernel Channel Fusion (DMKCF). These blocks improve multi-scale feature extraction and adaptively capture complex patterns from the weed images. The model also integrates Triplet Attention (TA) to refine feature representations and employs Grad-CAM and LIME for visual interpretability, offering insights into the classification process. The proposed network, when tested on the SorghumWeedDataset_Classification, achieved a testing accuracy of 99.07%, surpassing several state-of-the-art CNN models. The proposed model provides a robust solution for sorghum weed classification, advancing agricultural monitoring and management practices. Additionally, the model achieved 100% accuracy for the broad leaf weed class, further underscoring its effectiveness in identifying and classifying sorghum weeds with absolute precision.