Spectral indices to predict drought stress in native plant species

Vegetation indices have been shown to be good predictors of drought stress-related physiological variables for many plant species. These indices can be built using different sections of the light spectrum (known as spectral bands) depending on the plant species and on the physiological variable that is being predicted. In order to optimize these vegetation indices, we want to create predictive models for each species and variable studied (10 species and 4 variables) using every single possible combination of two 1 nm spectral bands available in our data sets, in the form of simple ratio indices and normalized different indices, and across different regions of the collected light spectra.
Person

Principal investigator

Jaume Ruscalleda Alvarez jaume.ruscalledaalvarez@research.uwa.edu.au
Magnifying glass

Area of science

Plant biology
CPU

Systems used

Zeus
Computer

Applications used

Measurement of drought stress on native plants
Partner Institution: The Universite of Western Australia| Project Code: director2125

The Challenge

Physiological monitoring of vegetation at large scales remains a challenge, particularly for diverse plant communities. This research project occurs in the context of mine site restoration monitoring, and in particular, is focused in developing novel methods to monitor plant physiological performance in faster and more efficient ways when compared to traditional ecophysiology methods.
Hyperspectral reflectance measures have been shown to provide reliable information on plant physiological status in the form, among other, of spectral vegetation indices. These indices can vary across species and depending on the physiological variable that they are estimating, and thus need to be optimized for each species-variable combination. Ground-level hyperspectral data collected in this project provides us with information of to what extent is a plant (or one of its leaves) reflecting light at different electromagnetic levels, at a very fine scale (1 nm steps, for a total of 2500 steps).
Our aim is to optimize different existing forms of spectral vegetation indices that use a combination of 2 spectral bands in the short-wave infra-red part of the spectrum, as well as creating all possible 2 spectral band combination across the whole collected spectrum, exploring new indices that might be able to predict more accurately any of our species-variable combinations.
Once the optimum spectral bands are identified, we will be able to determine if hyperspectral data derived indices are good proxy indicators of drought stress, as well as consider if this method can be up-scaled to remote platforms such as drones or satellites that could collect hyperspectral data in the form of images, covering large areas in shorter time periods.

The Solution

The method consists in creating all possible predictive models for each vegetation index -physiological variable combination, in the form of linear, quadratic and cubic models, as well as a log-transformed linear model. For each one of these species-variable-regression type combination the script delivers a set of 4 different model performance metrics that are stored in different tables (coefficient of determination, Root Mean Square Error, Standard Error and p-values). This happens for every single 2 wavelength combination being studied in each vegetation index. For example, the vegetation index called Normalized Difference Water Index (NDWI) is obtained through the following calculation: NDWI = (RIR– RSWIR)/(RIR + RSWIR), where RIR and RSWIR represent the reflectance value at a 1 nm spectral band in the near infra-red region (for RIR), and a reflectance value in the short-wave infra-red region (for RSWIR), respectively. The near infra-red region is composed of 250 different 1 nm bands, and the short-wave infra-red region is composed of 1950 bands. That provides a total of 487,500 different wavelength combinations for each species-variable combination (ten species and four variables). That is, we are building 19.5 million models for the NDWI index. We are exploring 2 more indices that use these same spectral regions (the Moisture Stress Index and the Simple Ratio Water index). Additionally, we want to explore all possible combinations of all wavelengths in the spectrum (a total of 2500) for each species-variable combination, which adds up to 250 million models for each data set (we have a total of 3 data sets).

The Outcome

Results from our overarching study have actually shown that optimised spectral indices obtained through our calculations with Pawsey are the best method (among a group of 4 different methods) to predict physiologically significant variables , such as leaf water potential, in a wide range of native species, corresponding to different plant functional types. Therefore, access to Pawsey computing power has been key in order to squeeze out as much information as possible from very large hyperspectral data sets, and find a suitable method to predict drought stress in native plants, which is the main goal of my PhD.