| dc.description.abstract | Wetlands provide essential ecosystem services for the world, but in recent years, due to direct human activities as well as global warming and other climatic variations, there has been significant ongoing reductions in both wetlands' quantity and quality. Vegetation monitoring is key to assessing the overall health and dynamics of a wetland. Multiple vegetation surveys such as the national bog survey, national fen survey, etc. have been carried out in previous years to map and understand the extent of vegetation on these wetlands. These manual surveys, however, are often time-consuming and require a large number of people. Also, once surveyed, it is unlikely that such an effort will be repeated regularly to update the map. The remoteness and inaccessibility of many wetlands are also limiting factors. Hence, there is a growing recognition of remote sensing (RS) techniques as a cost-effective and viable alternative to field-based ecosystem monitoring. This study aims to identify and monitor the environmental conditions of the wetlands using machine learning (ML) techniques.
The study begins with a pixel-based approach to map vegetation communities across raised bogs and fens using ensemble classifiers such as bagged tree (BT). Due to turf-cutting and other practices, the boundary of such wetlands is often ill-defined, and therefore the study initially develops a boundary delineation algorithm. Using edge detection techniques such as entropy filtering, canny edge detection, and Lazy snapping, the wetlands' boundaries were successfully delineated. The pixel-based approach applied initially only takes into account the spectral properties of the area. Therefore, the study was further extended to segment-based learning using graph cut maximum a-posteriori (MAP) segmentation. This takes into account the contextual information on top of the spectral information. This segmentation process acted as a post-classification smoothing for the wetland maps. The algorithm was tailored for land-cover, especially vegetation identification and hence, was termed the Mapping Vegetation Communities (MVC) algorithm. A total of up to 18 classes were mapped, using multi-date satellite data, inside 13 wetlands using the MVC algorithm with an average accuracy of 84% for the years 2017 and 2018. The algorithm works very well for larger vegetation communities, but some small communities were not adequately identified due to the restriction of spatial resolution (10 m) of Sentinel-2 data. Therefore, drones or unmanned aerial vehicles (UAVs), otherwise known as drones, were employed to gain high spatial resolution.
Drones provide very high spatial resolution and flexibility in temporal resolution. In order to confirm the applicability of these methods for mapping vegetation inside wetlands, a comprehensive comparison was made between multiple DL and ML algorithms. The study reveals that DL provides higher accuracy compared to ML by ~2%, but also is time and cost-intensive. Hence, the choice of the algorithm should be application dependant.
The study then extends the ability of remote sensing-based monitoring of wetlands by combining the high spatial resolution of drones with the global coverage of the satellite data to create seasonal maps of vegetation communities within the wetlands. This nested methodology incorporates geo-referenced land-cover maps, scaled at the drone resolution level and up-sampled to S2 imagery level through interpolation. A colour correction technique was introduced in the pipeline to improve consistency between drone image capture sessions. The proposed framework has been evaluated on various wetlands across Ireland, and results are presented herein for an ombrotrophic peatland, Clara Bog. Additionally, to automate the process, a majority voting technique was applied to seasonal images. The obtained monthly maps were united to produce a more precise annual vegetation map of the wetland for the year 2019. The application of this method thereby reduces the number of field surveys typically required to assess long-term ecological change on wetland habitats.
Wetlands are known to be the largest natural source of methane (CH4), and the atmospheric lifetime of CH4 is about 9 +/- two years which makes it a good target for climate change mitigation. In this study, the use of the methane data along with the retrieved surface albedo (SA) from the recently launched European Space Agency's Sentinel-5 Precursor (S-5p) mission has been explored over Canada. The sensitivity of the methane emissions from wetlands over Canada was seen in the total column methane measured by S-5p, which was then used for the land type classification. The data from 2018 and 2019 were used separately to create individual maps for the two years and compared to the reference ground truth. It was seen that spring, and autumn-time CH4 measurements are high, and the lowest values are during the summer. As the area covered by bog and fen are small and mixed with different land types (e.g., marsh, swamp, forest), the low variation in sensitivity of the CH4 data makes it difficult to determine these land types. However, six of the ten land types were identified with great confidence. Amongst them are two major wetland types in Canada (marsh and swamp), covering a significant area of the country. Also, identification of forest land type is significant in monitoring the area covered by forest and its change over time. This is an entirely new use of the S-5p CH4 product, and the study showed the high potential of the data with applications in land type identification. The S-5p CH4 data can be applied to multiple areas which are yet to be explored.
Finally, this study has also investigated a 28-year hydrological record and water levels on four turloughs in the west of Ireland (1989 to 2017) with respect to their ecohydrological metrics using statistical analysis. For each vegetation community, the metric was defined using the flood duration and flood depth, as well as global radiation and temperature as a proxy for the time of the year when the floodwaters first recede and the vegetation can emerge. These trends were compared with the latest RS map produced using the MVC algorithm, and it was seen that the key communities stayed intact despite some extreme flood events over the past 20 years. The metrics were further refined using hierarchical clustering for the range of parameters. Such an ecohydrological metric is beneficial for forming policies and defining pressures associated with drainage and other land use activities that may take place. | en |
