Description

Land cover data serve as a primary input for monitoring various environmental conditions including global carbon cycle or biodiversity change. These data can be generated from remote sensing imagery using advanced machine learning techniques. Nevertheless, land cover mapping at the national and regional level is challenged by the following factors: (1) lack of reliable training and validation samples used by the machine learning classifiers to ‘learn’ the characteristics of the target land cover classes; and (2) lack of transferability of the trained classifiers across different areas

Objectives and Methodology

The overall goal of this MSc thesis is to develop a scalable machine learning solution for mapping land cover classes in Africa. For this purpose, we will use the LandCoverNet training samples shared on the Radiant Earth Foundation ML hub: https://www.radiant.earth/mlhub/ and the following multi-source data: multi-temporal PlanetScope, Sentinel-2 or harmonized Landsat-Sentinel-2 remote sensing data, Global Human Settlement Layer (GHSL), Night Time Light (NTL) etc. The analysis can be carried out using one of the following machine learning techniques: e.g. deep learning, Random Forests, Support Vector Machine. The following sub-objectives will be investigated
(1) to assess the importance of multi-source data for land cover mapping in the target area
(2) to reduce the noise in the available training and validation sample set using spectral-spatial information
(3) to assess the generalization capability of the training machine learning classification model from one area to another area
The training samples are available for the whole continent. Nevertheless, the thesis research will be focused only on one particular country.
The classification results will be compared with other land cover data available for the study areas such as those recently published by Li et al. (2020) or Midekisa et al. (2017).

Further reading

Gómez, C., White, J.C., & Wulder, M.A. (2016). Optical remotely sensed time series data for land cover classification: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 116, 55-72