RGB Vegetation Index-Based Pixel-Level Classification of Crops and Weeds Using Machine Learning

Abstract

One of the primary causes of increasing poverty among farmers is the uncontrolled and excessive growth of weeds, which directly impacts crop yields. Over time, various surveys and techniques remained advance and implemented to differentiate weeds from crops in order to autonomously eliminate or manage them. Numerous approaches have been employed, including color-based, threshold-based, and machine learning-based methods. This study presents a high-quality, pixel-annotated agricultural dataset comprising of 60 RGB field images utilized for crop–weed–soil segmentation, containing over 75 million labeled pixels. Each image features manually created masks that distinguish soil, crops, and weeds. To address class imbalance, 70,000 samples per class were selected. The dataset is divide into training (60%), validation (20%), and test (20%) sets to enable robust evaluation of machine learning models for pixel-level crop and weed classification, aiming to address this critical issue that contributes to significant mental and physical stress for farmers, as well as substantial financial losses due to wasted cultivated land and inefficient use of labor. In particular, this review focuses on machine learning-based methods applied to evaluating their effectiveness in detecting and classifying weeds through advanced algorithms and model parameters.