A Regularized Version of Adaboost for Pattern Classification in Historic Air Photographs

In this work, we present a novel classification method for geoinformatics tasks, based on a regularized version of the AdaBoost algorithm implemented in the GIS GRASS. AdaBoost is a machine learning classification technique based on a weighted combination of different realizations of a same base model. AdaBoost calls a given base learning algorithm iteratively in a series of runs: at each run, the algorithm modify a set of weights for the contribute of the training samples. Initially the weights are set equally; at each step of the algorithm, the weights of incorrectly classified examples are increased so that the base learner is forced to focus on the hard examples in the training set. The final classifier is obtained by a weighted majority vote of the base classifiers. The AdaBoost algorithm is one of the most successful classification methods in use. While the algorithm largely preserves its general and practical applicability, theoretical and experimental works show that AdaBoost can overfit when it is applied to noisy data. As presented in (Merler et al, 2003), the procedure to regularize AdaBoost consists in sorting data points by hardness, as emerging from analysis of the evolution of AdaBoost weights, and in progressively eliminating the hardest from the data set. In the noisy case, it appears consistently able to reduce both the bias and net variance components of the model error with respect to the standard algorithm. The regularized AdaBoost algorithm was tested on synthetic noisy data. Finally, it was employed for the identification of craters of exploded bombs from geolocated reconnaissance images. Historical aerial material from World War Two was considered (data from the UXB-Trentino project – Autonomous Province of Trento). The final result is a map of the spatial density of craters that can be used as an indicator of the risk of presence of unexploded aerial bombs.