Sparse-Based Feature Selection For Discriminating Between Crops And Weeds Using Field Images

Control of weed growing in yields is a critical task for reducing crop losses. Recently, image-based systems attempt to discriminate between crops and weeds from a set of features. Although some features have a physiological meaning, most of them are redundant or noisy. Therefore, selecting relevant features must result in interpretable and accurate results while reducing the computational complexity of the system. In this work, we introduce a sparse-based feature selection approach using the Lasso operator that eliminates noisy features aiming to improve the classification of crops. We evaluate our proposal on the Crop/Weed Field Image Dataset, for which we tune the parameters by maximizing the accuracy and minimizing feature dimension. Achieved performance results evidence that our proposed approach improves discrimination in comparison with other feature selection approaches, with the benefit of providing interpretability in weed/crop discrimination tasks.