Pixel-Level Weed Classification Using Evolutionary Selection of Local Binary Pattern in a Stochastic Optimised Ensemble

Abstract

In the early growth stages of young crops, weeds compete for living resources while often grow close together and with overlaps. Conventional weeding techniques such as mechanical weeding and herbicide application are not precision methods. To allow for high precision target and spraying of weeds, autonomous field robots equipped with computer vision systems could be deployed to manage and control the weeds in the crop line. The robots are therefore required to classify weeds at pixel-level accuracy with high recall rates. This paper presents a pixel-level weed classification model using rotation-invariant uniform local binary pattern (LBP) features. The model design is based on an ensemble with a two-level optimisation structure. The first-level leverages on Genetic Algorithm (GA) optimisation to select the best rotation-invariant uniform LBP configurations. The second level utilises covariance matrix adaptation evolution strategy (CMA-ES) in the Neural Network (NN) ensemble to select the best combinations of voting weights of the predicted outcome for each classifier. This model design allows for different feature inputs selected by GA for each of the NN classifiers in the ensemble, unlike classical ensembles which share the same input data. For ensemble weight optimisation, we compared differential evolution, particle swarm optimisation, and CMA-ES. A crop/weed field image dataset public dataset consisting of 60 images of weeds and carrots at early true leaf stage was used for validation with existing works. The results demonstrated that our LBP-based pixel-level classifier exhibits excellent predictive capability with high recall and F1 scores when compared to the literature. It was found that CMA-ES optimised ensemble gave the best configuration and demonstrated a classification accuracy of 87.9%, which exceeded the existing statistical feature classification of 85.9%. This highlights some potentialities for chip-based application-specific integrated circuit LBP pixel-level weed classification in smart agriculture applications.