Deep learning based weed classification in corn using improved attention mechanism empowered by Explainable AI techniques

Abstract

The agricultural crops, like corn, suffer from the presence of undesirable plants known as weeds, which compete for sunlight and water, leading to lower crop yields. Recognizing weeds during their early growth stage is vital for minimizing their impact on crop growth and maximizing yield. By leveraging a lightweight deep neural network, this research endeavours to classify corn and the weeds that often grow alongside it. To achieve this, the Enhanced Convolutional Block Attention Module (CBAM) embedded EfficientNet model (ECENet) is proposed by integrating the enhanced CBAM with EfficientNetB0 model and the inclusion of extra layers. The Enhanced CBAM has been created by modifying the original CBAM through the parallel arrangement of the Channel Attention Module (CAM) and Spatial Attention Module (SAM). The simultaneous use of attention modules eradicates the need for CAM and SAM to be dependent on each other, resulting in the independent extraction of attention feature maps. The ECENet model was trained and tested on the corn weed dataset to understand the discriminative features of corn and weed. The proposed system yielded 99.92% overall recognition accuracy, with 4,772,010 parameter footprints, a model size of 57.4 megabytes, and 0.796 giga floating-point operations per second (GFLOPs). The proposed ECENet takes 37%, 91%, 80%, and 78% fewer parameters than DenseNet121, InceptionResNetV2, ResNet50V2, and XceptionNet respectively. The proposed model excels in diagnosing weed and crop differentiation, outperforming previous studies and state-of-the-art models. Finally, interpretability of the proposed model has been provided using explainable AI techniques (XAI) such as GradCAM and LIME. Due to its small memory requirement and high accuracy, the ECENet is ideal for real-time corn and weed classification on handy and mobile devices with minimal computational capabilities. The system can also be expanded to be included in agricultural robots for real-world weeding in large farmlands.