Weed detection using machine learning: A systematic literature review

Abstract

Recently, many researchers and practitioners used Machine Learning (ML) algorithms in digital agriculture to help farmers in decision making. This study aims to identify, assess and synthesize research papers that applied ML algorithms in weed detection using the Systematic Literature Review (SLR) Protocol. Based on our defined search string, we retrieved a total of 439 research papers from three electronic databases, of which 20 papers were selected based on the selection criteria and thus, were synthesized and analyzed in detail. The most applied ML algorithm is Neural Networks in these models. Thirteen evaluation parameters were identified, of which accuracy is the most used parameter. 75% of the selected papers used cross-validation as the evaluation approaches, while the rest used holdout. The challenges most encountered were insufficient data and manual labeling of the pixel during image segmentation. Based on the ML algorithms identified, we concluded that supervised learning techniques are the most used techniques in weed detection.