Visual Feature Combination Based on Feature Selection for Weed Density Classification

Abstract

Weed removal is crucial in improving agricultural productivity, but current image processing algorithms struggle to accurately differentiate between weeds and crops, leading to suboptimal outcomes. The selection of appropriate feature extraction methods is vital for enhancing object recognition accuracy. This study combines color, shape, and texture features as visual features of leaf to create a comprehensive image representation. Color features offer spectral data, shape features capture object contours, and texture features describe surface patterns. Feature selection using information gain is employed to identify the most relevant features, improving model accuracy by eliminating irrelevant ones. This research used three supervised learning methods. Initial trials using 39 features achieved an accuracy of 68.09% with RBF kernel in Support Vector Machine (SVM), while reducing to 19 features increased accuracy to 95.95%. Additional tests with different kernels (Linear, RBF, and Polynomial) and 19 features showed the RBF kernel was most effective, reaching 97.14% accuracy, due to its ability to handle non-linear data. K-Nearest Neighbors (K-NN) with a k-value of three outperformed other models with an accuracy of 97.78%, and Random Forest with 100 trees achieved 96.82%. Overall, K-NN showed the highest accuracy for weed density classification, as it effectively handled non-linear distributions by considering nearest neighbors with weighted class determination, making it ideal for multi-class data.