Land cover mapping with random forest using intra-annual sentinel 2 data in central Portugal : a comparative analysis

• Author(s): Lüdtke, Daria
• Date: 2018
• Persistent ID: http://hdl.handle.net/10362/33648
• Origin: Repositório Institucional da UNL
• Subject(s): Data Mining; Land Cover Classification; Multi-temporal classification; Open Access
• Document Type: Master thesis
• Language: English
• Advisor(s): Henriques, Roberto André Pereira; Caetano, Mário Sílvio Rochinha de Andrade; Granell-Canut, Carlos
• Contributor(s): RUN

Description

Dissertation submitted in partial fulfilment of the requirements for the degree of Master of Science in Geospatial Technologies

In recent years, data mining algorithms are increasingly applied to optimise the classification process of remotely sensed imagery. Random Forest algorithms have shown high potential for land cover mapping problems yet have not been sufficiently tested on their ability to process and classify multi-temporal data within one classification process. Additionally, a growing amount of geospatial data is freely available online without having their usability assessed, such as EUROSTAT´s LUCAS land use land cover dataset. This study provides a comparative analysis of two land cover classification approaches using Random Forest on open-access multi-spectral, multi-temporal Sentinel-2A/B data. A classification system composed of six classes (sealed surfaces, non-vegetated unsealed surfaces, water, woody, herbaceous permanent, herbaceous periodic) was designed for this study. Ten images of ten bands plus NDVI each, taken between November 2016 and October 2017 in Central Portugal, were processed in R using a pixel-based approach. Ten maps based on single month data were produced. These were then used as input data for the classifier to create a final map. This map was compared with a map using all 100 bands at once as training for the classifier. This study concluded that the approach using all bands produced maps with 11% higher, yet overall low accuracy of 58%. It was also less time-consuming with about 5 hours to over 15 hours of work for the multi-temporal predictions. The main causes for the low accuracy identified by this thesis are uncertainties with EUROSTAT´s Land Use/Cover Area Statistical Survey (LUCAS) training data and issues with the accompanying nomenclature definition. Additional to the comparison of the classification approaches, the usability of LUCAS (2015) is tested by comparing four different variations of it as training data for the classification based on 100 bands. This research indicates high potential of using Sentinel-2 imagery and multi-temporal stacks of bands to achieve an averaged land cover classification of the investigated time span. Moreover, the research points out lower potential of the multi-map approach and issues regarding the suitability of using LUCAS open-access data as sole input for training a classifier for this study. Issues include inaccurate surveying and a partially long distance between the marked point and the actual observation point reached by the surveyors of up to 1.5 km. Review of the database, additional sampling and ancillary data appears to be necessary for achieving accurate results.