Hi!

I was thinking of using CE IPCC cards to collect LULC data for my .49 ha plots in Ethiopia. I understand that the operator is supposed to generalize the LULC over the whole plot based on the hierarchy rules developed for the dryland assessment. Which means that each plot of .49 ha is assigned a single LULC class.

What are the limitations regarding the accuracy of the resulting LULC classification to have training plots that have a lower resolution (70m) than the images to be classified (30m)?

Should I consider using the classes of IPCC but estimate their % within the .49 ha? Could I then use these percentages to train the classification algorithm? Could this increase the accuracy?

Maybe there is a methodological paper I can read (and try to understand)? I only have found user manuals so far.

Many thanks for your guidance,

Florence

asked 23 Nov '16, 16:32

flandsberg's gravatar image

flandsberg
4118
accept rate: 0%


Dear Florence,

So the limitations are that as you point out you are using the information collected in the whole plot (70x70 meters) as the training data of the classifier while the geometry used is the central point (which actually represents one single pixel (30x30 meters)

We are going to add an extra column to the Fusion Table export in the coming version of Collect Earth in which you will also get a KML Polygon. This means that in your Fusion table you will be able to choose what type of geometry you want to use in the classification.

My guess is that the classification will improve a bit with that, but clearly the biggest improvement will be achieved by filtering the rows of the Fusion table before using them as training set. For instance, you should remove plots where there is a mixs of Land Covers ( so where there is both trees and water, infrastructure and trees and so on). Also, if you are using the confidence indicators, also use this info to filter out plots where the user has signal that he is not confident of the assessment.

Classifying using Land Use is clearly more challenging than using Land Cover. Good results are usually achieved by using the tree cover percentage. Remember to create classes that are aggregations ( so don't use 2%,4%,6% and so on as classes, instead use classes like 0-6%, 8-20%, 20-40%, 50-70% and 70-100% ). This will give you better results.

We have no paper on this, but you can surely find other sources for that...if you need a GEE script to start with I can find one tomorrow, let me know!

permanent link

answered 23 Nov '16, 23:35

collectearth's gravatar image

collectearth ♦
61515
accept rate: 17%

Hi!

I like your very practical recommendations, thanks!

Just to make sure: the spectral value that will be used for training is the pixel in Landsat that overlays with the red point in the middle of the plot? So the best way is to use only plots who are homogeneous so that the plot value is representative of the central point.

I don't understand your remark on using the tree cover percentage for land use classification. Nor do I understand what the GEE script would be for. Sorry, just too much to understand at once...

Thanks,

Florence

permanent link

answered 29 Nov '16, 20:38

flandsberg's gravatar image

flandsberg
4118
accept rate: 0%

So yes, the spectral value would be that of the pixel that corresponds to the central point (the red point) as this is the geometry used in the FeatureCollection (the fusion table)

As you guess the best solution is to use plots where there are no mixes of land use/land cover. You could record the percentage cover of different land covers in the Collect form and then "clean" out the plots with mixed covers.

Land Use by definition will be more challenging to classify (using the GEE scripts for supervised classification using Collect Earth dat) than Land Cover.

If you use tree cover to classify you get a tree cover map of course. So if you want Land Use you need to use Land Use. My point is that using Land Use the results will not be as good as Tree Cover.

If you come out with a good script or good solution to classify please share the URL in the forum!!!

permanent link

answered 30 Nov '16, 11:29

collectearth's gravatar image

collectearth ♦
61515
accept rate: 17%

So a follow up question here:

If the country's min. area threshold to qualify as forest is 0.50ha and our Collect Earth survey for that country has plot of 0.49ha (0.50ha) size

Does it mean the only plots that would qualify as Forest landuse category are those with 100% tree cover, only? Because the plot size is equal to the threshold.

permanent link

answered 05 Dec '16, 17:04

Tesfay's gravatar image

Tesfay
0226
accept rate: 0%

Hi,

I was wondering if it is now possible to choose what a polygon geometry in the classification in GoogleEarth Engine. If so, how can we do it? Are there any instructions available?

Many thanks,

Florence

permanent link

answered 25 Jan, 13:57

flandsberg's gravatar image

flandsberg
4118
accept rate: 0%

Your answer
toggle preview

Follow this question

By Email:

Once you sign in you will be able to subscribe for any updates here

By RSS:

Answers

Answers and Comments

Markdown Basics

  • *italic* or _italic_
  • **bold** or __bold__
  • link:[text](http://url.com/ "title")
  • image?![alt text](/path/img.jpg "title")
  • numbered list: 1. Foo 2. Bar
  • to add a line break simply add two spaces to where you would like the new line to be.
  • basic HTML tags are also supported

Question tags:

×155

question asked: 23 Nov '16, 16:32

question was seen: 1,517 times

last updated: 25 Jan, 13:57