A machine learning framework for the identification of crops and weeds based on shape curvature and texture properties

Abstract

Identification of weeds in a crop is a challenging problem for designing an automotive weed management system. Due to similar and indistinguishable properties between crop and weed, rather than single, various type of information is required for the successful discrimination of weeds against crops. This paper presents a machine learning-enabled weed classification system based on the fusion of handcrafted shape and texture features of a plant at the feature level. The shape features include several shape factors, region-based (Hu moment invariants), and contour-based descriptors (Fourier descriptors and shape curvature). Texture features include GLCM and LBP. A series of experiments are conducted to investigate the utility of shape curvature in the classification of weeds against crops. Furthermore, the potential of shape features including shape curvature and texture features has been examined by modeling the four classifiers, namely, SVM, KNN, MLP, and Naïve Bayes by using the 10-fold cross-validation approach. The classifiers are compared based on performance metrics- accuracy, error rate, recall, specificity, precision, and F1-score. Experimental results show that the SVM classifier achieved 93.67% overall accuracy with shape curvature features in the classification between crop and weed. In addition, the SVM classifier achieved 99.33% overall accuracy with the integration of shape with shape curvature and texture features.