Combining Random Forests and object-oriented analysis for landslide mapping from very high resolution imagery

The increasing availability of very high resolution (VHR) remote sensing images has been leading to new opportunities for the cartography of landslides in risk management and disaster response. Object-oriented image analysis has become one of the key-concepts to better exploit additional spatial, spectral and contextual information. The multitude of additional object attributes calls for the use of advanced data mining and machine learning tools to identify the most suitable features and handle the non-linear classification task. In this study we used the Random Forest algorithm for the selection of useful features and object classification in the context of landslide mapping. A workflow for image segmentation, feature extraction, feature selection and classification was developed and tested with multi-sensor optical imagery from four different test sites. Due to class imbalance and class overlap between landslide and non-landslide areas the classifier can be heavily biased towards over- and under-prediction of the affected areas. This is a common issue for many real-world applications and a procedure to estimate a well-adjusted class ratio from the training samples was designed and tested. A number of potentially useful object metrics was evaluated and it was demonstrated that topographically guided texture measures provide significant enhancements. Employing 20% of the image objects for training accuracies between 73.3% and 87.1% were achieved at four different test sites.