Open cnakalembe opened 1 year ago
@cnakalembe what year?
Preference is 2022 since the season is done
On Nov 12, 2022, at 10:45 AM, Ivan Zvonkov @.***> wrote:
@cnakalembe what year?
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.
Context: When we randomly sample points for evaluation on CEO, 1% or less of the sample may be crop
Issue: Low crop sample size make evaluation difficult
Potential solution:
From the literature review on stratification techniques, I summarized some of the methods used in the papers listed in the Google Doc -- https://docs.google.com/document/d/1QfBemFjJtRUJ3C8Z70tszGlNZo5iB50knRK0y6iusqk/edit
Happy to hear your comments!
Read through the report in some more detail, great summaries!
Context: NDVI by stratification is the easiest and most intuitive way of sampling random points
Issue: Low crop sample size makes evaluation difficult
Potential solution:
Changing year to 2019
@ivanzvonkov, @hannah-rae, I tried the NDVI by quartiles. What do you think?
Also adding the GEE code - https://code.earthengine.google.com/0d6758d51ef68a4bbdb881d11edd1eb3
@MsPixels okay took a look at this in some more detail couple questions:
[10,25,50,75,90]
in ee.Reducer.percentiles
, shouldn't it be [25, 50, 75]
?percentiles.get("NDVI_p90"))
shouldn't it be p100 or something like that?After clarification of these questions the next step is to figure out how many points to sample. The Olofsson paper: https://www.sciencedirect.com/science/article/abs/pii/S0034425714000704 is one resource for this. Once you have that number we can figure out how to create a Collect Earth Online labeling set from these points.
@ivanzvonkov, I was curious about the difference between the 10th and 25th quartiles, that's why I calculated it. However, the 10th quartile doesn't reflect in the strata map. Also, the 25th, 50th, and 75th percentiles only exclude the woody vegetation zone of Mali.
Again, changing the 90th percentile to the 100th generalizes the map.
Okay I see. So given:
NDVI_p25: 0.09578940770143446
NDVI_p50: 0.1190916155868571
NDVI_p75: 0.19721850558714377
NDVI_p90: 0.2832040283154737
As I understand your suggestion is:
But these are not true quartiles, right? Wouldn't quartiles be:
This can be plotted with:
var NDVI_threshold = ee.Image(1)
.where(clip.gte(ee.Number(percentiles.get("NDVI_p25"))), 2)
.where(clip.gte(ee.Number(percentiles.get("NDVI_p50"))), 3)
.where(clip.gte(ee.Number(percentiles.get("NDVI_p75"))), 4)
;
Would this make more sense as a stratification? What are the pros and cons of your suggested ranges?
I chose the 25th, 50th, 75th, and 90th percentiles based on this article I found Long-Term_Land_UseLand_Cover_Change_Assessment_of_the_Kilombero_Catchment_in_Tanzania. I will go ahead and use the quartiles you suggested.
Also, after going through Olofsson's paper, I came up with this sampling design for three scenarios based on the standard error of the overall accuracy. Will be on standby for your comments. Sampling Design for Mali
Stratification by LULC. This code combines 11 LULC layers to get a majority vote of crop and noncrop zones. Based on the strata, I sampled the crop and noncrop points.
Month: Feb Year: 2019