Mapping wetlands in Sweden using multi-source satellite data and random forest algorithm

Abstract

Wetlands are valuable ecosystems, and assets for human life, that must be regularly monitored, starting with accurately mapping their location and extent. However, an updated national inventory of wetlands is needed. The availability of multi-source data, and advanced machine learning algorithms in Google Earth Engine (GEE) offers excellent opportunities to map wetlands on a country-wide scale. This study mapped wetlands in Sweden using optical, radar, and topographical data, with the Random Forest algorithm, and labels from digitized polygons within the boundaries of the latest national wetlands inventory of Sweden (VMI), completed in 2005. This study discriminates between three classes (non-wetlands, wetlands, and water). From the digitized polygons, 30,000 points were sampled per class in each county (1,890,000 in total). A single RF classifier was trained for each county of Sweden, and a new Swedish national wetlands inventory (RFWI) was generated. The accuracy assessment with testing samples showed that the country-wide overall accuracy (OA) of the classified validation set of points is 98.97%, with a kappa value of 0.985, where the counties with the best, and worst OA are Kronobergs (99.84%), and Norrbottens (97.40%), respectively. RFWI agrees with VMI to a large extent, thus, there are new wetlands mapped, and wetlands surveyed in VMI disappeared. The countrywide area classified as wetlands in RFWI is 30.8% bigger than VMI, as VMI does not incorporate many small wetlands. Nevertheless, the results between counties are mixed. Six out of the 21 counties are estimated to have suffered an overall loss of wetlands area, as big as 73%. RFWI wetlands coverage is higher than VMI’s in the remaining counties, lesser than 30% (small) in seven counties, between 38-64% (mild) in four, between 202-245% (high) in other three, while Stockholm presents a huge difference (450%). The decrease observed in some counties was corroborated with Google Earth imagery. The small, and mild differences are due to the incorporation of wetlands not present in VMI. While high, and huge differences found in four counties are overestimated due to label-related issues. Approximately 70% of the areal difference between RFWI & VMI is contained in Jämtlands county. Besides VMI, the developed RFWI was compared to other continental, and global wetlands products specifically LUCAS, ESA WorldCover, the Ramsar Database, MODIS LC1, and GLWD-3. LUCAS dataset was reclassified to match this study’s classes, and for extracting the class from RFWI, this procedure indicated good agreement between the products (87.88% OA). All the other large-scale inventories mentioned underestimated wetlands occurrence in Sweden to a large extent. The area of wetlands in RFWI could be considered as a realistic maximum in most counties except Stockholms, and Jämtlands. Limitations of this study are discussed and recommendations for future studies are given.

Popular Abstract

This study mapped wetlands in Sweden by combining optical, radar, and topographical data, with high-quality data containing the whereabouts of wetlands, discriminating wetlands from water bodies, and from all other types of land covers (such as forests, urban, croplands, among others), creating a new national wetlands inventory (RFWI). RFWI was compared and validated with several products that mapped or surveyed wetlands in Sweden.
Wetlands are valuable ecosystems, and assets for human life. They act as ‘living filters’, capturing carbon from the atmosphere, and filtering pollutants from water bodies, alleviating the effects of human economic activities. Wetlands can also diminish the effects of flooding, protecting humans living in coastal zones and close to water courses. Therefore, wetlands must be regularly monitored, starting with accurately mapping their location and extent. However, an updated national inventory of wetlands is needed because the Swedish national wetlands inventory (VMI) is potentially outdated. Additionally, continental, and global inventories were found to underestimate the wetlands occurrence in Sweden.
For mapping wetlands, we employed data from multiple sources, for training a machine learning classifier in each county, more specifically the Random Forests algorithm. The whereabouts of wetlands are called labels, which summed with the rest of the data served as input for the classifier. The labeling, crucial for training the algorithm, was performed twice with different methods. The availability of multi-source data, and advanced machine learning algorithms in the novel Google Earth Engine (GEE) planetary-scale tool played a huge role in the development of this study.
The accuracy assessment showed that the country-wide overall accuracy of the classified validation set is 98.97%. RFWI agrees with VMI to a large extent, thus, there are new wetlands mapped, and wetlands surveyed in the old inventory that disappeared. The countrywide area classified as wetlands in RFWI is 30.8% bigger than VMI, expected as VMI does not incorporate many small wetlands, and has a coarser spatial resolution. Nevertheless, the results between counties are mixed. Despite that the accuracy is very high, spatial autocorrelation between the training and validation datasets suggests that the accuracy reported is over-optimistic, and that the results are not entirely trustworthy. Specially in Stockholms and Jämtlands counties, where an overestimation of wetlands occurrence is one of the limitations of RFWI. RFWI’s uncertainty is related to issues in the labeling process, required for training the classifier algorithm, it is possible that the label of some samples was correctly set because of the difficulty attached to visually interpreting poor quality satellite imagery.
Recommendations for future studies are given, and the scripts used during this study are open-source and hosted in the cloud for further development in the subject.

• author: Hernandez Malave, Salvador
• supervisor: Zheng Duan
• organization: Dept of Physical Geography and Ecosystem Science
• course: NGEM01 20221
• year: 2022
• type: H2 - Master's Degree (Two Years)
• subject: Earth and Environmental Sciences
• keywords: physical geography and ecosystem analysis, multi-temporal, Google Earth Engine, machine learning, classification, country-wide, geomatics
• publication/series: Student thesis series INES
• report number: 574
• language: English
• id: 9092982
• date added to LUP: 2022-06-23 15:55:22
• date last changed: 2022-06-23 15:55:22