Logistic regression product-unit neural networks for mapping Ridolfia segetum infestations in sunflower crop using multitemporal remote sensed data

Abstract

Remote sensing (RS), geographic information systems (GIS), and global positioning systems (GPS) may provide the technologies needed for farmers to maximize the economic and environmental benefits of precision farming. Site-specific weed management (SSWM) is able to minimize the impact of herbicide on environmental quality and arises the necessity of more precise approaches for weed patches determination. Ridolfia segetum is one of the most dominant, competitive and persistent weed in sunflower crops in southern Spain. In this paper, we used aerial imagery taken in mid-May, mid-June and mid-July according to different phenological stages of R. segetum and sunflower to evaluate the potential of evolutionary product-unit neural networks (EPUNNs), logistic regression (LR) and two different combinations of both (logistic regression using product units (LRPU) and logistic regression using initial covariates and product units (LRIPU)) for discriminating R. segetum patches and mapping R. segetum probabilities in sunflower crops on two naturally infested fields. Afterwards, we compared the performance of these methods in every date to two recent classification models (support vector machines (SVM) and logistic model trees (LMT)). The results obtained present the models proposed as powerful tools for weed discrimination, the best performing model (LRIPU) obtaining generalization accuracies of 99.2% and 98.7% in mid-June. Our results suggest that a strategy to implement SSWM is feasible with minima omission and commission errors, and therefore, with a very low probability of not detecting R. segetum patches. The paper proposes the application of a new methodology that, to the best of our knowledge, has not been previously applied in RS, and which obtains better accuracy than more traditional RS classification techniques, such as vegetation indices or spectral angle mapper.