NDVI Change Analysis at Coorabulka Station
ProRaster Scientific 2022 – 2023
I used ProRaster Scientific to quantify the spatial and temporal changes in vegetation over a year at a large outback cattle station in Australia called Coorabulka. Coorabulka station is located in Western Queensland in the “Channel Country” between the Georgina and Diamantina Rivers. The channel country is covered with numerous intertwined rivulets that are generally dry but flood when rain falls in the catchments to the east. Wikipedia reports that Coorabulka station has an area of 6370 square kilometres and carries up to 8000 head of cattle. I do not have any relationship with Coorabulka Station or its owners.
I presented this study in a series of six videos, presented below, and also viewable on YouTube.
In Part 1, above, I introduce Coorabulka station and summarise the results that I will present over the series of videos. In the background, I play an animation of Landsat data acquired covering the station over a year and also an animation of NDVI (Normalised Difference Vegetation Index) computed over the same period. These animations were produced by exporting images from ProRaster Scientific and then importing those images into OpenShot.
For this study, I downloaded 98 individual Landsat 8 and 9 scenes. A mosaic of two scenes is required to provide spatial coverage of Coorabulka, so I acquired 49 revisits in total. This represents every available revisit over this time period. All data was Level 1, top of atmosphere.
In Part 2, above, I show the reader how to build mosaic products, and mosaic sequence products, and introduce the concept of collation before building a collated mosaic sequence product.
In ProRaster Scientific, a “Product” is created by combining one or more scenes together either by mosaicing, collating, sequencing, or some combination of all three. The inputs to a product are all virtual rasters, and the product itself will also be a virtual raster. Because it is a raster, it can be operated upon as a single dataset, even though it may be comprised of many scenes spatially and temporally.
I explain how to order your scenes when you build a mosaic product or a collated product. When building a mosaic, you will want low-quality scenes (generally affected by high levels of cloud) at the bottom of the stack, and high-quality scenes (generally cloud free) at the top of the stack. When collating, you might choose to order by acquisition time. In this study, I order by proximity to the preferred acquisition time for each temporal event. This means that each event contains data that is primarily derived from that event time, then drawn from revisits before and after the event time as needed.
When you create a sequence from two or more existing products, you do not need to specify the order of the products as this is automatic and based on the acquisition time of the product. Therefore, it is important to make sure your products have a valid acquisition time.
In Part 3, above, I build the final collated mosaic sequence and render this sequence. Firstly, I show you how to use the “Order by proximity” options to order the scenes in a collated mosaic or collated scene. This is a critical part of the pre-processing which ensures that the statistics for each temporal event (revisit) represent the state of the Earth at that time, as closely as possible. At the same time, by bringing in high quality data from adjacent revisits to replace low quality data at the target revisit, we improve the overall quality of our statistics.
In this study, I collated five revisits into one mosaic at each temporal event. This includes the two revisits prior to the target, and the two revisits after the target. At the beginning and end of the sequence, where revisits either before or after may not be available, I simply use what is available.
A sequence product can be built progressively as new scenes are acquired by the satellite constellation. You can always add new scenes to the product database, and build new products. You can edit any existing products and change the scenes or products that are used in that product.
At the end of the video I show the mosaic sequence and the final collated mosaic sequence rendered as videos.
In Part 4, above, I demonstrate how to use the Product Editor to build a branched processing chain. This is the crux of multispectral imagery analysis in ProRaster Scientific.
A product is a virtual raster that is derived from one or more scenes or products that are, themselves, virtual rasters. You can execute raster processing operations on virtual rasters, and the output from this will be a new virtual raster. The exceptions to this rule are the Export and Statistics operations. The “Render” and “Display” operations produce rendering algorithms as output, but these are actually a form of virtual raster as well.
So, you can build a chain of raster processing operations in the Product Editor. The input to each link in the chain is the virtual raster output from the previous operation. You can also branch the chain to build multiple chains. This is variously described as either a tree, cascade, or waterfall. When you edit any operation in the tree, all of the operations (following all branches) that are subordinate to that operation will be updated. In most cases, this will just trigger the regeneration of a virtual raster. For Export and Statistics operations, it triggers the computation of the raster cell values. This is when your computer finally has to do some work – maybe a bucket load of it!
Because these final computations can take a long time to execute, you can control whether editing an operation automatically triggers a regeneration of the tree. If you turn off automatic execution, you have to remember to manually trigger execution when you have finished editing and are ready to compute.
In Part 5, above, I play the videos showing the full sequence of 49 events, from July 2022 to July 2023. You can now clearly see three floods come down the Georgina River – two major floods and a smaller flood in between. The first flood generates a flush of vegetation growth which then gets covered and degraded by the subsequent floods. But, in good news for the grazier, the vegetation recovers after the second major flood.
The distribution histogram for the NDVI index data clearly tracks the changes in vegetation over the study period. It shows the vegetation in the rangeland country slowly reducing over time, and the vegetation in the channel country declining before the first flood arrives, and rejuvenates the system.
In Part 6, above, I import the temporal statistics into a spreadsheet and graph the results. I am now looking at the NDVI index statistics for the data clipped to the station boundary. Note that the boundary polygon I acquired for the station seems to be too small by about 1500 square kilometres. But, bear in mind that this study was only undertaken to demonstrate and test the capabilities of ProRaster Scientific.
One of the most useful (and simple) outputs is the “mean” value statistic. This responds well to changes in the vegetation cover and, if you multiply the mean by the number of valid cells, you acquire a proxy for the “volume” of vegetation. In other words, it can be used as a biomass proxy.
A more accurate biomass proxy would exploit a known (or assumed) relationship between NDVI and vegetation biomass. In this case, you would use the distribution statistics and, for each “bin” in the histogram, compute vegetation biomass which would then be summed for all bins to acquire the total vegetation biomass on the station at that time.