Evaluation of Handcrafted Feature Extraction Methods in Cannabis Seed Classification

The growing global acceptance of cannabis is driving research and practical applications, where accurate classification of seed varieties based on morphological characteristics remains a significant challenge. This study presents a comprehensive computer vision and machine learning (ML) framework for classifying cannabis seed varieties by systematically evaluating a diverse range of feature extraction methods. Specifically, we extract a comprehensive suite of features, including basic shape and color descriptors, texture features based on Gray Level Co-occurrence Matrix (GLCM) and Local Binary Patterns (LBP), local descriptors using Scale-Invariant Feature Transform (SIFT), and global image representations with GIST. These features are evaluated using seven ML classifiers. Experimental results reveal that the combination of GIST features with Logistic Regression (LR) achieves the highest classification accuracy of 93.25% on a dataset containing 17 cannabis seed varieties. This outcome underscores the superior performance of GIST features in combination with LR for achieving accurate and scalable cannabis seed classification, offering significant implications for agriculture, quality control, and cannabis-related research.