Evolving a cloud-robust water index with genetic programming
João E. Batista, Sara Silva · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Over the years, remote sensing (RS) experts created many indices to help them study satellite imagery by highlighting characteristics like vegetation, water, or burnt areas, among others. In this work, we study water indices. Although there is a large number of water indices that work perfectly in unclouded imagery, clouds and shadows cast by clouds are often mistaken for water. This work is focused on the automatic feature construction using genetic programming (GP), in an attempt to make features that are more robust to these issues. To do this, we use a dataset containing pixels from areas where we could find these issues to evolve models that learn how to classify those pixels correctly. The results indicate improvements when comparing evolved features with indices, but further improvements are required to tackle other issues found.