Choosing an Algorithm¶

Zeit implements more than a dozen methods. Most projects need one or two. Start from the question you are trying to answer, and let it pick the method.

Start from your question¶

Your question	Use	Needs
When was this area disturbed (cleared, burned) and how badly?	LandTrendr	One image per year
Did anything change, at any time of year? What did it change into?	CCDC	Every clear image (dense, multi-band)
Is something changing right now, compared with a stable past?	BFAST Monitor	Regular series (e.g. 16-day composites)
How many changes happened, and when, in a regular series?	BFAST Lite	Regular series
Did the trend change, the seasonal cycle, or both?	BFAST	Regular series
Is this area getting greener or browner over the decades?	Mann-Kendall	One value per year (or `method="seasonal"`)
When does the growing season start, peak and end? Is this year late?	Phenology	Dense series, several per month
Which crop or land-cover type is this pixel, given a few examples?	TWDTW	Reference time series per class
Which land-cover types exist here, with no labels at all?	SOM	Any cube
Can I work with fields and patches instead of pixels?	SNIC	Any cube
I have many labelled samples and want the best classifier.	TempCNN / LightTAE	Labelled time series
I have two dates and labelled change masks.	Siamese Change Detector	Image pairs + masks

A quick decision guide¶

flowchart TD
    Q{"What do you want to know?"}
    Q -->|"When / where did it change?"| C{"How often is it observed?"}
    Q -->|"Is there a long-term trend?"| MK["Mann-Kendall"]
    Q -->|"What is it?"| L{"Do you have labels?"}
    Q -->|"When is the growing season?"| PH["Phenology"]

    C -->|"Once a year"| LT["LandTrendr"]
    C -->|"Every clear image"| CC["CCDC"]
    C -->|"Regular composites"| B{"Past or present?"}
    B -->|"Whole history"| BL["BFAST Lite / BFAST"]
    B -->|"New observations only"| BM["BFAST Monitor"]

    L -->|"A few examples per class"| TW["TWDTW"]
    L -->|"Many labelled samples"| DL["TempCNN / LightTAE / U-TAE"]
    L -->|"None"| SO["SOM clustering"]

Change detection methods compared¶

The four change-detection families answer related but different questions. The figures below all come from their tutorials.

Annual · retrospectiveLandTrendrStraight segments through yearly values. Best for forest disturbance and recovery over decades.

Dense · retrospectiveCCDCHarmonic models of the seasonal cycle, multi-band. Dates changes within the year and describes the new state.

Regular · retrospectiveBFAST Lite / BFASTStatistically optimal number of breaks in trend and season. Single index.

Regular · near real timeBFAST MonitorIs the newest data consistent with the stable past? The basis of many alert systems.

	LandTrendr	CCDC	BFAST Lite / BFAST	BFAST Monitor
Input	1 index, 1 value per year	Up to 7 bands, every clear observation	1 index, regular series	1 index, regular series
Models the season?	No (annual values)	Yes (harmonics)	Yes (harmonics)	Yes (harmonics)
Detects	Abrupt and gradual change, recovery	Abrupt change, any time of year	Breaks in trend (and season, for BFAST)	The first break after a date
Typical use	Forest disturbance history	Land-cover change mapping and classification	Retrospective break analysis	Alerts
Speed per pixel	Very fast	Moderate	Moderate	Very fast

Combine methods

The methods complement each other. A common pattern is Mann-Kendall to find where a region is trending, then LandTrendr or CCDC to find when and how abruptly it changed there. Another is Tmask to clean the data before CCDC. See the User Guide for every workflow.