Global Dynamic World Uncertainty layer

[Geethen]

Contact Details

geethen.singh@gmail.com

Dataset description

Using a modern uncertainty quantification method, conformal prediction, we have quantified the uncertainty for the first non-null dynamic world composite of 2020. Each pixel contains a value from 0 to 9 and represents the length of the prediction set (uncertainty) for a pixel. Each prediction set includes the correct landcover class with, in this case, a 90% probability. The more classes included in a set the greater the uncertainty. If the Dynamic World model was not confident for all 9 candidate classes, an empty set is returned (a value of 0). Therefore, a value of 0 corresponds to the highest uncertainty and a value of 1 corresponds to the lowest uncertainty. A set length value greater one is also associated with high uncertainty.

Research paper: https://arxiv.org/abs/2401.06421

The earth engine app can be used to derive uncertainty for any dynamic world scene
GEE app: https://ee-geethensingh.projects.earthengine.app/view/conformaluq

Earth Engine Snippet if dataset already in GEE

for example

// Prepare Dynamic world
var AfricaDW = ee.ImageCollection('users/grazingresearch/for_geethen').mosaic();
var SAmericaDW = ee.ImageCollection('users/grazingresearch/for_geethen2').mosaic();
//Rest of the world
var Remainder = ee.ImageCollection("projects/ee-geethensingh/assets/UQ/DWsetLength")
var dw = ee.ImageCollection(dw).mosaic();

var dw = ee.ImageCollection.fromImages([dw, AfricaDW, SAmericaDW]).mosaic().rename('setLength');

// Visualise Set length (Uncertainty)
var palettes = require('users/gena/packages:palettes');
Map.addLayer(dw,{min:0, max:9, palette: palettes.matplotlib.viridis[7]}, 'Set lengths');

Sample Code: Add a sample code maybe just adding your datasets in the code editor

Enter license information

CC-BY-4.0

Keywords

Uncertainty, Landcover, Dynamic World

Code of Conduct

• I agree to follow this project's Code of Conduct

[samapriya]

Hi @Geethen
Working on your dataset now can you tell me what each of these collections are? a line a one liner for each maybe. Thanks

[Geethen]

Hi Sam,

They all contain the same type of data/information just for different regions of the world. Ideally, This would be a single image collection.

Specifically, each image contains the set length. i.e., a multiclass prediction based on the calibrated set-valued conformal predictor with alpha (tolerance level) = 0.1. Therefore, the predictions contain the actual class with a 90% probability.

Here is a blog post (https://medium.com/@coachgeets/uncertainty-quantification-in-google-earth-engine-using-conformal-prediction-5e0dd3850564) describing conformal prediction for classification and a demo on how one can do this for any dynamic world scene and area.