Backwarding: An Overfitting Control for Genetic Programming in a Remote Sensing Application

Abstract

Overfitting the training data is a common problem in supervised machine learning. When dealing with a remote sensing inverse problem, the PAR, overfitting prevents GP evolved models to be successfully applied to real data. We propose to use a classic method of overfitting control by the way of a validation set. This allows to go backward in the evolution process in order to retrieve previous, not yet overfitted models. Although this “backwarding” method performs well on academic benchmarks, there is not enough improvement to deal with the PAR. A new backwarding criterion is then derived using real satellite data and the knowledge of plausible physical bounds for the PAR coefficient in the geographical area that is monitored. This leads to satisfactory GP models and drastically improved images.