3SW-Net: A Feature Fusion Network for Semantic Weed Detection in Precision Agriculture

Early weed detection is crucial for optimizing agricultural productivity and minimizing crop loss. Traditional manual methods of weed identification are labor-intensive and inefficient, particularly in expansive fields. To address this challenge, this study proposes an innovative approach utilizing advanced image processing and deep learning techniques to create an automated weed detection system. We introduce 3SW-Net, a novel deep convolutional neural network specifically designed for weed detection. The method leverages the Simple Linear Iterative Clustering (SLIC) algorithm for efficient segmentation of weed regions and the Histogram of Oriented Gradients (HOG) technique to extract edge and texture features from weed images. By combining the outputs from SLIC, HOG, and grayscale images, a comprehensive feature set is created, significantly enhancing the model’s accuracy. The integrated feature fusion approach demonstrates outstanding performance, achieving a recall of 98.99%, specificity of 99.68%, and an overall accuracy of 99.56% on weed dataset. These results indicate that the combination of SLIC segmentation and HOG feature extraction significantly boosts the effectiveness of the convolutional neural network. The promising outcomes from this model pave the way for developing a robust real-time weed detection system, which can play a crucial role in promoting sustainable agricultural practices and ensuring efficient resource management.