Abstract
Soil drainage conditions are highly important to farmers and the environment. To map drainage classes efficiently, several analytical approaches, such as decision tree classification, can be used. Decision tree classification can be improved by combining the predictions of several trees with boosting and bagging techniques. This study tested the relative performance of boosting and bagging for the prediction of drainage classes. Furthermore, as drainage classes form an ordered series rather than unrelated classes, differential costs for misclassification were tested in combination with each technique. Decision tree models were trained from 1135 observations of soil drainage classes and validated using leave-one-out cross validation and a hold-out validation sample with 567 observations. The best model was achieved using bagging combined with differential costs for misclassification (overall accuracy=52.0%). On the other hand, differential costs for misclassification reduced the overall accuracy of boosted decision trees from 50.8% to 49.2%. The best models obtained with boosting and bagging were used to produce maps of drainage classes on a national extent. The maps predicted the same drainage class in 81% of the study area. Finally, with boosting as well as bagging, the models had a high usage of the predictor variables wetlands, slope to channel network, clay content, land use and geology.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have