The availability of new multispectral sensors capable of capturing high resolution images through low altitude flights using drones, provides access to large amounts of information of the Earth Surface at a much lower cost than images captured by other devices. These images have a limited number of spectral bands but cover a very large area of landscape with resolutions that can reach several centimeters per pixel. They are adequate for application to different fields such as for mapping natural ecosystems such as river ecosystems through classification. Solving this problem for a variety of images requires to acquire knowledge from some of them to be applied to other. The images required to learn need to be labeled using information provided by experts, repositories or other sources. In most of the cases costly field visits are required for labeling. As a consequence the number of labeled images is usually scarce. The process of extracting the information available in the labeled images (source domain) to classify other different but related ones, that contain similar elements and that are not labeled (target domain) is known as Domain Adaptation (DA). Transfer Component Analysis (TCA) is a kernel-based feature extraction technique especially designed for DA. TCA tries to learn a transformation matrix across domains by minimizing the distribution distance measure. TCANet is an scheme por unsupervised DA that simulates the behavior of a convolutional network but for which the computation of the filter coefficients is performed directly through TCA instead of being computed through a back-propagation algorithm. The high computational cost of TCA together with the large size of the high resolution datasets makes the use of both parallelization techniques and the application of spatial information extraction algorithms indispensable to solve the problem. In this paper we propose an optimized superpixel-based DA technique for river ecosystem classification using high-resolution multispectral images. The proposal combines the extraction of spatial information using superpixel-based segmentation algorithms for establishing the input patches to the DA technique with the parallelization of the algorithm by using a distributed memory architecture.

This work was supported in part by the Ministerio de Ciencia e Innovación, Government of Spain (grant numbers PID2019-104834GB-I00), the Consellería de Educación, Universidade e Formación Profesional (grant number 2019--2022 ED431G-2019/04), and by the Junta de Castilla y León (project VA226P20 (PROPHET II Project)). All are co-funded by the European Regional Development Fund (ERDF). Alberto S. Garea acknowledges USC for its "Convocatoria de Recualificación do Sistema Universitario Español - Margarita Salas" postdoctoral fellowship under the "Plan de Recuperación y Transformación" program funded by the Spanish Ministry of Universities with the European Union's NextGeneration funds and to Prof. Begüm Demir from TU Berlin.