The Corine Land Cover (CLC) initiative uses remote sensing data to map land cover (LC) at the European level every six years. CLC is useful for observing continental trends; however higher spatial resolution, improved accuracy, inclusion of local classes and more frequent updates are necessary for informed management of natural resources at the national level. Therefore, in 2015, the Institute for Environmental Solutions (IES) started the SentiSimuLat project where it developed a land cover – land use (LC-LU) algorithm suitable for Latvian end-user needs based on Sentinel-2 data.
European Space Agency’s (ESA’s) Sentinel fleet carrying various instruments and providing regular, as well as free of charge data has enormous potential for nature assessment. The Sentinel-2 satellite provides spatial and temporal data that contributes to scientific research, decision-making and policy planning in many fields. As Sentinel-2A had been successfully launched in 2015, the state-of-the-art airborne hyperspectral laboratory no longer had to simulate satellite data, thus it provided high resolution observations, which alongside in situ data, served reliable reference for the validation of Sentinel-2 observations.
To understand potential end-user needs in Latvia and develop a suitable approach of LC-LU mapping, at the start of the project, IES organised face-to-face interviews with 14 institutions – mostly, representatives of spatial planning and spatial data analysis departments, as well as policy makers, managing administrators, regional and national authorities. The interviews showcased the lack of multi-sectoral LC-LU classification for Latvia, especially, at the regional scale. Most of the respondents expressed their interest in the possibility of acquiring high-resolution satellite information at least 1-2 times per year.
Based on Latvian user needs and Sentinel-2 data, IES identified 12 target classes for annual LC-LU classification: 1) coniferous trees (11110); 2) deciduous trees (11120); 3) shrubland and transnational woodland (11200); 4) grassland (12100); 5) agricultural land (12200); 6) inland marshes (water) (12311); 7) inland marshes (coastal) (12312); 8) peat bogs (12321); 9) peat extraction sites (12322); 10) water (21000); 11) artificial/ urban land (22000); 12) bare land (23000).
To reach the result, the researchers tested various data sets and algorithms. They used all available low-cloud (<10%) Sentinel-2 data from the season of 2015 and 2016 that was combined with reference information. Furthermore, the LC classification result was compared with the information provided by CLC and Copernicus High Resolution (CHR) layers.
The overall accuracy of developed algorithm for all classes in all pilot territories (Cesis, Sigulda and Burtnieki municipalities) is 92%. A weaker performance was detected in wetland classes (78% for peat bogs and 70% of inland marshes (coastal)). The researchers indicated outdated reference data sets as one of the most important causes of the low performance. Nevertheless, the performance of the SentiSimuLat algorithm surpasses the CLC and CHR information layers.
The classification result can be used for cross-checking the accuracy and veracity of available data. For example, potential violations of Common Agriculture Policy (CAP) can be identified by detecting ploughing in territories classed as grasslands. Latvia’s Rural Support Service (RSS) interest in the use of the classification result for control activities under CAP has already resulted in cooperation within the follow-up ESA’s Plan for European Cooperating States (PECS) project SentiGrass.
At the end of the project, IES organised informative seminar that was attended by c.a. 80 potential end-users. During the discussion, they highlighted the interest to cooperate for a further development of the algorithm, adding specific tasks, e.g. detection of invasive species. The team of the SentiSimuLat believes that the developed approach is just in its innovators – early adopters phase and soon, after the release of positive reviews of the first users, it will gain a snowball effect.