Ensemble Learning Based on Feature Selection and Distance Normalization for Enhancing Corn and Weed Classification

Weeds need to be removed from the immediate areas surrounding crops as they compete for soil nutrients. Farmers currently clear weeds manually, which is both tiring and imprecise. Therefore, researchers have developed artificial intelligence (AI) using deep learning or non-handcrafted methods to facilitate precise detection. However, these methods have yet to achieve real-time inference speeds. Consequently, this study adopts a handcrafted approach that employs visual leaf features for classification via ensemble learning. The objective is to employ feature selection and data normalization to create an accurate and efficient machine-learning model. The experimental findings obtained in this work demonstrate that Information Gain effectively reduces features by 50%, from 22 to 11, while maintaining accuracy. Chebyshev normalization emerges as the most suitable normalization technique, as it significantly enhances classification accuracy in ensemble learning. The accuracy obtained when using histogram gradient boosting is found to be 0.92 with an inference time of 5.955 ms per image. These outcomes illustrate that handcrafted features achieve higher accuracy than non-handcrafted methods, ultimately improving efficiency and enabling real-time implementation.