In the UK, the regeneration and conservation of semi-natural grasslands is important, especially for grasslands protected by legislation such as UK Biodiversity Action Plan (BAP) priority habitats or Sites of Special Scientific Interest (SSSIs). As such, monitoring grassland condition is necessary, but conventional methods of measuring grassland condition are time consuming and limited in their spatial coverage. This thesis tested if remote sensing can provide a more cost- and time-effective solution to measuring grassland condition as defined by the Common Standards Monitoring (CSM).
A field spectroscopy experiment was designed to explore the potential link between grassland spectral reflectance, plus a metric representing a traditional measure of grassland condition referred to as CSM-condition. Partial least squares regressions were used to evaluate the relationship between grassland multi-spectral reflectance and a range of condition-related grassland variables; between the condition related grassland variables and CSM-condition and between the grassland multi-spectral reflectance and CSM-condition. The evaluation tested the relationships across grassland types, seasons and spectral devices used; and between grassland variable observations made in terms of mass or % cover.
When analysing data collected at patch level during the summer; the mass of
bryophytes, dead material and graminoids plus the % cover of forbs can be predicted to a moderate level of accuracy (R2 values of >0.5 and nRMSE <100) when analysing data from all seven grasslands. When analysing data from all Parsonage Down NNR grasslands; the mass of bryophytes, the % cover of live material, % cover-based live:dead ratio and CSM-condition could be predicted to a high level of accuracy (R2values of >0.7 and nRMSE <100). Moisture content plus the % cover of dead material, forbs and gram:forb ratio were all predicted to a moderate level of accuracy as well as CSM-condition predicted by grassland variable values. When using data from all Ingleborough NNR grasslands; the % cover of forbs and biomass plus the mass of bryophytes, dead material and live material could be predicted to a moderate level of accuracy. When using patch level data collected across three seasons; the % cover of dead material, live material and live:dead ratio plus the mass of graminoids could be predicted when using three seasons of data collected on one grassland, or for all three Parsonage grasslands, to at least a moderate level of accuracy although some models trained with % cover data had a high accuracy. Forbs (mass and %
cover) plus the mass of gram:forb ratio, live material and live:dead ratio could be
predicted to at least a moderate level of accuracy for some grasslands.
When using data from all grasslands collected in one season the mass of a range of grassland variables could be predicted to a moderate level of accuracy for the spring and autumn months but not when using % cover data. Using CROPSCAN and SVC data produced similar results, with slightly stronger results from the CROPSCAN, but using data from the Rikola camera produced weaker results. When the results of trained PLSR models were extrapolated to field level, the projected predicted grassland variable values from models trained with CROPSCAN MSR 16R data looked promising but the results have not been externally validated using a separate data set. The most important spectral bands for predicting grassland variables and CSM-condition were NIR and SWIR with the red edge (647nm) and 470nm also having some importance. The most important grassland variables for predicting CSMcondition were depended on whether the grassland variable was mass-based or % cover-based. When using mass data; graminoid:forb ratio mass and live:dead ratio mass were consistently important. When using % cover data; forbs cover, graminoids cover and live:dead ratio cover were consistently important.
Overall, the results suggest that some of the condition-related variables considered in this thesis are predicted with reasonable accuracy and precision at patch level (i.e. R2 values of >0.5 and nRMSE <100), but producing reliable results requires a sufficient quantity of data to train the statistical models (at least 30 quadrats of samples), particularly if the results are to be extrapolated to field level as additional data are required for the external validation of the results. Grassland variable prediction success varied with number of sites considered and with season with no clear consistent pattern. Also, none of the grassland variables could be consistently predicted strongly across all the different grasslands or seasons.
This has implications for any land manager who wishes to emulate the methods in this thesis. The results suggest that this thesis provides a more cost- and timeeffective solution for capturing grassland condition; but anyone emulating these methods would have to carefully choose the variables, grassland types and seasons to collect data and would have to collect a sufficient quantity of data for model training, testing and evaluation. A consequence of adopting, or refining, the approach in this thesis could be more effective monitoring (and therefore timely intervention where necessary) of UK Biodiversity Action Plan (BAP) priority habitats and Sites of Special Scientific Interest (SSSIs). Refining this approach could include testing different modelling approaches and focusing further work on the successful aspects of this research such as key grassland and wavelength variables.
