Application of machine learning methods for gap filling of satellite data

Garcia, M. (Main supervisor), Baum, A. (Supervisor)

Activity: Examinations and supervisionSupervisor activities

Description

Remote Sensing (RS) geophysical model products from optical and thermal satellites, can efficiently monitor the spatiotemporal distribution of water cycle components such as evapotranspiration (ET), in very large areas. Although products could assist in the mitigation of climate change problems especially in arid regions, they suffer from data loss mainly because of cloudiness. The main data loss problem, is that clouds are blocking the incoming solar radiation, decreasing the ET values of the surface, compared to the ET in a sunny day. This study, aims to assess the performance of different imputation methods in a spatiotemporal ET dataset from Spain. The used imputation method, Historical Average which is based on sunny days, was overestimating the imputed ET values. The tested imputation algorithms, ranged from naïve methods to time-series (TS), machine learning (ML), matrix completion (MC) and tensor decomposition (TD). The study area was the Majadas Oak Savanna in Extremadura.
Period1 Feb 2020
Examinee
Examination held atTechnical University of Denmark