Improving global land cover fraction change detection using a Markov chain model

Organised by Laboratory of Geo-information Science and Remote Sensing

Thu 19 May 2022 08:30 to 09:00

Venue Gaia, building number 101
Room 1

By Rob Burger

Global land cover mapping has aided monitoring of the complex Earth’s surface and provided vital information to understand the interactions between human activities and the natural environment. Most of the global land cover products are provided with discrete classes, indicating the most dominant land cover class in each pixel. Fraction mapping, which expresses the proportion of each land cover class in each pixel, is able to characterize heterogeneous areas covered by multiple land cover classes. However, land cover fraction maps have shown unrealistic inconsistent year-to-year changes, which makes it difficult to detect robust trends. To improve the detection of land cover fraction change, three models were compared in this thesis: a basic Random Forest (RF) model, a recurrent RF model and a Markov chain model. Based on Landsat 8 derived features, initial fraction estimations were predicted on a global scale for the years 2015 until 2018. The RF regression models were trained on over 150.000 reference points, provided by the Copernicus Global Land Service Land Cover project (CGLS-LC100). The recurrent RF model incorporated predicted fractions of its previous year to account for temporal information. The Markov chain model, which incorporated temporal information from its adjacent years and class co-occurrence dependency, was applied on the initial fraction estimations as a postprocessing step. An accuracy assessment was performed with a subpixel confusion-uncertainty matrix to compare the performance of the models. All fraction estimations are validated on over 30.000 reference points that contain multitemporal fraction data, and also account for possible change areas. The results indicate that the Markov chain model achieved better overall accuracies compared to the other models. Compared to the basic RF model, the Markov chain postprocessing approach consistently improved the overall and per-class accuracies. The recurrent RF model obtained considerable poorer results in the later years, especially in areas that experienced land cover change. These results confirm that Markov postprocessing has the potential to reduce spurious and inconsistent multitemporal land cover change, which is in line with other studies. The findings highlight the importance of the addition of Markov models for future land cover mapping studies, so that multitemporal predictions provide realistic and consistent land cover change, essential for global trend estimations.