Benchmarking of the Symbolic Machine Learning classifier with state of the art image classification methods - application to remote sensing imagery

Abstract

A new method for satellite data classification is presented. The method is based on symbolic machine learning (SML) techniques and is designed for working in complex and information-abundant environments, where it is important to assess relationships between different data layers in model-free and computational-effective modalities. In particular, the method is tailored for operating in earth observation data scenarios connoted by the following characteristics: i) they are made by a large number of data granules (scenes), ii) they are made by heterogeneous sensors and iii) they are mapping a large variety of different geographical areas in different data collection conditions. The volume, variety and partially unstructured nature of these scenarios can be associated with the characteristics of Big Data. The results of an experiment observing the behavior of the SML classifier by injecting increasing levels of noise in the training set are discussed. Spatial generalization, random thematic noise and spatial displacement noise are tested. Seven supervised classification algorithms have been considered for comparison: Maximum Likelihood, Logistic Regression, Linear Discriminant Analysis, Naive Bayes, Decision Tree, Random Forest and Support Vector Machine. According to the results of the experiment, the SML classifier performed very well providing outputs with comparable or better quality than the other classifiers. Furthermore, the better performances were released with a much less expensive computational cost. Consequently, the SML classifier was evaluated as the best available solution in the specific data scenario under consideration. Few applicative examples of the new SML classifier using Spot5, Sentinel1, and Sentinel2 data inputs are provided.