Abstract/Summary

With 40% of the world’s plants estimated to be under threat of extinction and ever lowering levels of ecological intactness of biological systems, the requirement to effectively monitor plant species and diversity has never been more pressing. Globally, natural, and semi-natural grassland ecosystems are at particular risk of degradation and conversion. Semi-natural grasslands in the UK currently make up about 1-2% of the permanent lowland grassland cover. Once degraded due to agricultural additions or inappropriate management, they can be difficult and costly to restore. As these systems display high levels of plant and invertebrate diversity, there is a need to safeguard their decline. However, there are currently significant challenges to providing the data needed to assess the condition of these systems. Remote sensing could contribute by providing information on herbaceous plant diversity and vegetation state across a wide range of spatial scales and time. Optical traits are a subset of plant traits that are detectable using reflectance data from leaf to canopy scales, dependent on the configuration of the sensor employed and can be linked to taxonomic diversity and condition of vegetation. Very high spatial resolution hyperspectral imaging technologies are, for the first time, enabling in-situ grassland plant phenotyping at the leaf, individual and high-resolution canopy scale. Analyses of these spectra have demonstrated promising results in application of mapping of taxonomic units and diversity metrics. However there is little evidence of the temporal stability of these observations. At the landscape scale, openly available, higher spatial resolution satellite data is also enabling examination of smaller field parcels, which are typical of UK fragmented landscapes. In this context, spectral time-series have the potential to be used to predict the condition of vegetation communities of conservation interest. In this thesis, the use of optical remote sensing data to further our understanding of semi-natural grasslands and to safeguard their decline, is examined, with a particular focus on the exploitation of multi-temporal sampling. Firstly, spectral variation in space, as a surrogate measure for species or community type diversity (also known as the spectral variation hypothesis), is assessed via a meta-analysis of existing studies. The results of the synthesis reveal some promise for the approach, but a large amount of variation between study outcomes is observed, suggesting that methodological approaches are important in the effectiveness of the proxy. Secondly, spectral data is collected alongside botanical and phenological diversity data at high spatial resolution over a growing season to test the stability of the spectral variation hypothesis over time. The results of these experiments show that the ability to detect biodiversity using this method is seasonally, and possibly, site dependent. Next, the suitability of hyperspectral leaf reflectance for distinguishing 17 herbaceous species growing within a calcareous grassland is examined. The application of machine learning classification models to multi-temporal leaf spectra show that although species are distinguishable at most sampling times within the year, the transferability of these models is very limited between sampling dates. Finally satellite time-series of vegetation indices are used to predict favourable or unfavourable vegetation condition criteria in calcareous fields across two years. A number of indices were successful in distinguishing between the different condition criteria but there was variation in results found between the two years sampled, due to differences in intra-annual vegetation phenology. Overall the results of this thesis, show promise for remote sensing of grassland biodiversity and condition. Both high spatial resolution hyperspectral data, as well as coarser resolution multi-spectral data sets, can be useful in evaluation of these systems. However, the dynamic nature of leaves and canopies over time, will require a multi-temporal approach to model building, which should be an integral part of developing these methods in the future.