Modelling Wild-Oat Density in Terms of Soil Factors: A Machine Learning Approach BEATRIZ DIAZ ANGELA RIBEIRO RICARDO BUENO AND DOMINGO GUINEA bdiaz@iai.csic.es angela@iai.csic.es rbueno@iai.csic.es domingo@iai.csic.es Industrial Automation Institute-CSIC, Ctra.Campo Real. Km 0.2, 28500 Madrid, Spain JUDIT BARROSO DAVID RUIZ AND CESAR FERNADEZ-QUINTANILLA CCMA-CSIC, Serrano 115 B, 28006 Madrid, Spain judith@ccma.csic.es david.ruiz@ccma.csic.es cesar@ccma.csic.es Abstract. In crop fields, weed density varies spatially in non-random patterns. Initial knowledge of weed distribution would greatly improve weed management for Precision Agriculture operations. Site properties could be correlated to weed distribution, since the former vary among crop fields and also certain factors such as soil texture or nitrogen may condition the weed growth. This paper presents a method, based on artificial intelligence techniques, for inducing a model that appropriately predicts the heterogeneous dis- tribution of wild-oat (Avena sterilis L.) in terms of some environmental variables. From several experi- ments, distinct rule sets have been found by applying a genetic algorithm to carry out the automatic learning process. The best rule set extracted was able to explain about 88% of weed variability. Keywords: artificial intelligence, data mining, genetic algorithms, machine learning, rules, weed density Introduction Weed infestations in crops are still a challenge that has to be met in agriculture. Usually, weeds are heterogeneously distributed in agricultural fields (Cardina et al., 1997). Thus, different sampling procedures have been used to detect and describe the spatial distribution of weeds within a field (Rew and Cousens, 2001). However, weed discrimination is often a difficult task, particularly when weeds and crops have similar morphological and/or spectral characteristics. In spite of this, the spatial variability of weed abundance constitutes the basis for site specific weed manage- ment systems. Using these systems, farmers could spray selectively to reduce the amount of herbicide usage thereby diminishing environmental impact as well as economic cost (Earl et al., 1996). The persistence of high-density weed areas in fields over time suggests a non- random distribution that probably depends on environmental variability in the field. Moreover, soil characteristics as well as the properties of plant species, have a strong influence on the growth and reproduction of both crop and weed. Some studies Precision Agriculture, 6, 213–228, 2005 Ó 2005 Springer Science+Business Media, Manufactured in The Netherlands.