A novel hybrid feature method for weeds identification in the agriculture sector

Weed identification and controlling systems are gaining great attention and are very effective for large productivity in the agriculture sector. Currently, farmers are facing a weed control and management problem, and to tackle this challenge precision agriculture in the form of selective spraying is much-needed practice. In this article, we introduce a novel framework for a weed identification system that leverages (hybrid) the robust and relevant features of deep learning models, such as convolutional neural network (CNN) and handcrafted features. First, we apply the image pre-processing and augmentation techniques for image quality and dataset size enhancement. Then, we apply handcrafted feature extraction techniques, such as local binary pattern (LBP) and histogram of oriented gradients (HOG) to extract texture and shape features from the input. We also apply the deep learning model, such as CNN, to capture the relevant semantic features. Lastly, we concatenate the features extracted from a different domain and explore the performance using different classifiers. We achieved better performance and classification accuracy in the presence of the extreme gradient boosting (XGBoost) classifier. The achieved results witnessed the effectiveness and applicability of the given method and the importance of concatenated features.