Behind the Paper: Testing a new satellite-derived vegetation index in a new biome

The last chapter of my PhD dissertation was published earlier this year in the International Journal of Applied Earth Observation and Geoinformation. From conception to publication, this paper took about two-and-a-half years of work. It signifies the end of my PhD era, so to speak. The idea came from my supervisor, who suggested that testing the relatively new Plant Phenology Index (PPI) in semi-arid biomes would be a worthwhile cause because it was only evaluated in the boreal biome. PPI was developed in the department where I did my PhD by Hongxiao Jin.

PPI was shown to better detect phenology transition periods in northern latitudes than the more commonly used indices such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI). This, combined with its ability to track the dynamics of green foliage in the vegetation canopy, was the motivation to test it in my area of interest: African drylands. I was particularly interested in whether PPI can be used to estimate carbon uptake, i.e. gross primary productivity (GPP), in the semi-arid savanna biomes of Africa. The reason is that African drylands have a disproportionately low, and spatially clustered, number of studies on the dynamics of the carbon cycle.

I used an open-access global network of eddy covariance flux towers called FLUXNET as my primary source of in-situ data. The advantage of using FLUXNET is that the database is an amalgamation of different sites that is standardized to a common format. This is a huge deal because often different sites processes data and produce the main variables (like GPP) differently, which makes combining data from different sites to create a model problematic. Anyway, I ended up with four sites, two in the Sahel and two in Southern Africa. It’s not ideal, I would have preferred to have at least twice that amount of sites as there are more flux towers in Africa. Unfortunately, those sites are not open-access and the people in charge of them have not put the data online, nor have they standardized the data according to FLUXNET specifications. Had this data been available, the results of this study would have been even better (indeed this was one reviewer’s critique), but I cannot use data I do not have access to (which was my response to the critique). The good news was the four sites had a total of 14 site-years of data, which was sufficient to create a robust model and partly offset the spatial setback.


In order to see whether PPI does a good job in estimating GPP, I needed to do two things: (1) develop a GPP model using PPI as a foundation, and (2) test that model against both the in-situ flux tower data and some already established satellite-derived GPP models. For the first part, I used satellite-derived land surface temperature (LST) from NASA’s MODIS sensor as the basis to estimate vapor pressure deficit (VPD), which is a fancy way of measuring how dry the air is. VPD is an important parameter because it regulates tiny pores on leaves called stomata, which open to absorb CO2 for photosynthesis and let out water for transpiration. But they close when the VPD is too high because water will leave the leaf and into the atmosphere in large amounts if they remain open, thus causing problems for the plant. The human equivalent to transpiration is perspiration (sweating). So, you can imagine if someone is continuously sweating in hot weather that s/he will soon suffer from dehydration. The same is true of plants, they risk wilting if they do not control the exit of water from their leaves in hot and dry temperatures. The combination of PPI and VPD enables the model to account for amount of green leaf area available for carbon uptake and the environmental factor that controls carbon uptake in the hot and dry ecosystems that are the focus of the study.


After developing the PPI-based GPP model, I tested it against the flux tower data and three other GPP models. As is mostly the case, these were models that were developed for more temperate ecosystems, so I was curious how they themselves would perform in dryland settings. The results (a) were promising because the PPI-based GPP model outperformed all other GPP models and was closely linked to the flux tower data. Because the method derives VPD from LST data, the uncertainties in the LST data can be transferred over to the resultant dataset. I tested whether using the VPD data from flux towers themselves would improve my result . And of course, replacing the LST-derived VPD with observed VPD from the flux tower markedly improved the model (b). You’re probably wondering why I didn’t include the observed data in the first place? The reason is that the flux tower data is from a single point while the LST data is a gridded product that spatially continuous, i.e. it’s a map. And I wanted to develop a model that is applicable not only in areas where flux towers are present, but also where they are absent.

Download the paper here.

Citation: Abdi, A. M., Boke-Olén, N., Jin, H., Eklundh, L., Tagesson, T., Lehsten, V., & Ardö, J. (2019). First assessment of the plant phenology index (PPI) for estimating gross primary productivity in African semi-arid ecosystems. International Journal of Applied Earth Observation and Geoinformation, 78, 249-260.