Introduction and Workshop Overview

Ujaval Gandhi

ujaval@spatialthoughts.com

GEE Dynamic World Workshop

CC BY 4.0

Introduction to Dynamic World

What is Dynamic World?

CC BY 4.0

Dynamic World: Near Real Time land cover data

Global Land Cover “Bands”

10m resolution based on ESA Sentinel-2

Near Real Time: 2-5 day globally�for seasonal and recent events

Per-pixel probabilities across 9 classes

Free, Open License model & dataset

​

01

02

03

​

04

05

�Dataset and AI Model

Dynamic World

• A freely available and openly licensed global Landcover dataset by Google and WRI, launched June 2022.
• Based on Sentinel-2 dataset
• Landcover predictions are generated by a neural network trained on large amounts of global training data.
• The model classifies input images into 9 landcover classes.
• The output is per-pixel probabilities in each of the 9 classes
• Entire Sentinel-2 archive from 2015 onwards has been classified.
• Model runs in near real-time on new scenes and results are available in the data catalog.

Near Real-time Data Production

Dynamic World data available from June 23, 2015 to….2-5 days ago.

Snow/Ice

​

​

Dynamic World Dataset

Per-pixel probabilities across 9 classes

Time-Series of Class Probabilities

“map with you, not for you”

A good mental model to use for Dynamic World is to NOT think of it as landcover product, but as a dataset that provides 9 additional bands of landcover related information for each Sentinel-2 image that can be refined to build a locally relevant landcover map or change detection model.

Dynamic World Training Data

Spatial Context

• The receptive field of the Dynamic World model is ~21 pixels across.
• Reflectance values from up to 100m away from a target 10m pixel are used for prediction.
• Pros:
• This greatly reduces the “salt-and-pepper” effect typically seen when pixels are classified in isolation.
• Captures phenomena such as urbanization
• Cons:
• Certain classes lose fine-grained details (a.k.a. they appear blobby)

​

​

​

Sentinel-2 Composite (2019)

Sentinel-2 Composite (2020)

ESA WorldCover 2020 (Bare/Sparse Vegetation)

Sentinel-2 Composite (2019)

Sentinel-2 Composite (2020)

Dynamic World (built)

Temporal Context

• The Dynamic World model needs to generate near-realtime predictions.
• Each image is classified without knowledge of previous or subsequent images.
• It does NOT have any temporal context.
• Cons:
• ‘crops’ are often labeled as ‘bare’ or ‘grass’ or vice-versa.
• Users need to add their own temporal context by aggregating multiple dynamic world images over time. [Xu, P. et al. (2024)]

Summary

• Dynamic World is different from other pixel-based classification datasets.
• ‘urban’ not ‘built’
• ‘crops’ not ‘cropland’
• …
• Use Dynamic World probability bands as inputs to your models.
• Class probabilities have much closer correlation with landcover than spectral bands or spectral indices.
• You can process and aggregate dynamic world probability time-series to derive products.

Module Overview

Javascript Basics

​

Creating Composites

​

Importing Data

​

Computation with Images

​

Raster to Vector Conversion

​

Export

​

Module 1

Change Detection

Introduction to Machine Learning

​

Collecting training data

​

Classifying images

​

Accuracy Assessment

​

Exporting Results

​

​

Module 2

Supervised Classification

Processing Time-Series

Creating Charts

Building User Interfaces in Earth Engine

Publishing your first Earth Engine App

Module 3 and 4

Time-Series and Earth Engine Apps

Module 1: Change Detection

2020

2022

New Urban Areas

Module 2: Supervised Classification

Sentinel-2 Image

Training Samples

Classified Image

Module 3: Exploring Time Series

Module 4: Earth Engine Apps

Let’s get coding

What is your favorite programming language?

Image Source: Reddit

Javascript vs. Python

• You are learning the Earth Engine API
• The API is exactly the same regardless of the language you choose
• With a few small caveats
• Javascript API is the most mature and easiest to get started.
• No installation required
• No need to worry about authentication
• Very easy to share scripts, ask for help
• Building and deploying apps is very easy
• Python API is much more powerful
• Integrate with other data science libraries for data processing and plotting
• Automate launching and managing Exports
• You can easily convert any Javascript code to Python when needed