Mitchell Bonney Profile picture
Jun 10 32 tweets 10 min read
With the release of #DynamicWorld and other #global maps from the last few years, I have some (rambling) thoughts on global products, the challenge of accuracy assessment, and why I think there is still a lot of room for high-quality local/regional predictive mapping (1)
First, these global products are super cool! I played around with Google's DynamicWorld 10 m land cover today and being able to see the probabilities for each class for each pixel is great and being able to do this for every #Sentinel image through time is really unique (2)
However, at #local scales, they can be hit-or-miss, and none are ideal for every need. Some fail to pass a visual inspection (the "eye test") and even the best products require (in my opinion) local quantitative accuracy assessment before use for decision making or in models. (3)
In this thread, I begin by visually inspecting some maps over a complex suburban landscape...

For land cover, I considered four global products #Copernicus 100m, @Esri 10 m, @ESA_EO 10 m, @googleearth 10 m at local scale.

Simple #EarthEngine code
code.earthengine.google.com/507cbbb6f1d7dd… (4)
The reference: very high resolution Google Earth basemap over Brampton, ON, Canada (-70.80, 43.74). (5)
Copernicus 100 m (2019). Large # of classes (nice to see attempts at mapping different forest types). Low spatial resolution means it misses smaller patches and linear features. (6)
ESRI 10 m (2020). Map building process leads to “splotchy” patches. I find it overestimates built-up areas and misses many smaller patches/linear features despite the 10 m spatial resolution. Seems like there is aggregation under the hood that makes it 10 m in name only. (7)
ESA 10 m (2020). I consider ESA’s effort the gold standard (so far) in global land cover maps. Lots of local detail and mostly passes the eye test vs. VHR basemap. My only notable complaint would be the slight overestimation of grassland. (8)
Google DW 10 m (continuous since 2015, summer 2021 composite shown). The temporal aspect is the innovation, but at a single date it does not match ESA and feels similar to ESRI (lack of detail, splotchy patches). Appears to overestimate tree patches (ESRI underestimates). (9)
As you can see, ESA’s product is (in my opinion) the strongest offering for point-in-time global land cover. It may be useful for general needs at local-global scales. (10)
However, accuracy should be quantified locally before extensive use. Beyond this, specialized local products are still useful. For example, users may be interested in more/finer classes (e.g., forest types), products through time, and likely both! (11)
There is also the case of % cover products (e.g., % tree canopy cover). These are continuous products, compared to categorical land cover products. For truly understanding landscapes and land change, I think both are required. (12)
For example, a “Residential” or “Built up” land cover class may contain wide % differences in canopy cover, impervious cover etc. and those covers may/will change through time without a land cover transition occurring. (13)
There have been efforts to build global % cover products for a variety of types. Here I will focus on two % tree canopy products that are on Earth Engine: Copernicus 100m, GFCC 30 m. Note that images shown after this are scaled 0 (black) to 100% (white). (14)
First, very high resolution basemap over Mississauga, ON, Canada (-79.70, 43.53). (15/X)
Copernicus 100 m (2019). Low spatial resolution. 0% tree canopy in suburban residential areas. 15-30% canopy in tree-less agricultural fields. Close to 50% in some natural grasslands. Does not pass the eye test. (16)
GFCC 30 m (2015, but 2000/2005/2010 also available). Passes initial eye test. Reasonable % tree canopy in residential areas (can even see roads vs. yards). Only minor 5-10% overestimation in most agricultural fields. (17)
However, it strongly underestimates % canopy in forest patches. The highest pixel value I could find in this area was 65% (I can assure you there are denser forests than that here). (18)
In my PhD I explored building % tree canopy maps using Landsat. Although not a global product (see the edge), using dense local training data... it does (I think) a good job. It does not have the granularity of GFCC due to 3x3 averaging to fit scale of training plots... (19)
But is more accurate across a wider range of %/land covers. It has < 1% tree canopy in many agricultural fields and over most impervious surfaces, there is reasonable 80-90% canopy in forest patches, and variation between lower and high canopy neighborhoods is clear. (21)
I was also able to make these maps annually since 1972, to meet the needs of my project. There was no global product providing that, so I built it myself with local/dense training data. This led to better maps over my AOI than are likely achievable with global products. (22)
But they are not perfect maps! Image: Region of Peel, Canada – RGB image (R: 1972 tree canopy, G: 1996, B: 2020)... you can see areas of tree canopy loss/gain over nearly 50 years. (23)
Overestimation of tree canopy in some agricultural fields in 1972 propagates through to this RGB map. This is a struggle when training data, even if randomly/densely sampled across space, are used for prediction based on data (usually sat imagery) from different times. (24)
The challenges of setting up robust training data building and accuracy assessment for predicted map time-series requires more study. (25)
So even local products are not error-free and need to be used with caution. But accuracies are often higher than global products (if training data is a dense random probably sample!) and may pass the eye test. (26)
For prediction maps at all scales, robust accuracy assessments should be conducted, preferably with quantitative assessments for different subsets (e.g., cover classes, climate zones etc.). (26)
As noted by Meyer and Pebesma 2022 (doi.org/10.1038/s41467…) this can be a challenge for global products especially. (28)
They argue that “showing predicted values on global maps without reliable indication of global and local prediction errors or the limits of area applicability, and distributing these for reuse, is not congruent with basic scientific integrity”. (29)
And for “global maps of ecological variables be published only when they are accompanied by properly derived local and global accuracy measures.” (30)
Another interesting comment on this subject is Wyborn and Evans 2021 (nature.com/articles/s4155…): “Conservation needs to break free from global priority mapping”. (31)
Is yet another global map really needed? Or would our time and effort be better spent at local-regional-national scales where most conservation/environmental decisions are actually made? (32).

Sorry this was so long, but hope it made you think! 🛰️🤓
@NamratShresth Saw your tweet to
@jonnoruppert @YuhongHe4 earlier, got me thinking about this subject that has been on my mind!

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with Mitchell Bonney

Mitchell Bonney Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us on Twitter!

:(