How LandTrendr Works

Most points on the Earth's surface have a "life history" captured in images taken by the Landsat Thematic Mapper instruments. With LandTrendr, we use statistical algorithms to separate trends from noise, and thereby identify periods of stability and of change for every pixel (30 by 30 meters in size). This process of trajectory segmentation is the core of all our mapping. The trajectory segmentation phase is preceded by substantial "pre-processing" needed to ensure that the data are clean enough to discern trends, and is followed by several different types of "post-processing" that make the maps ultimately of interest to most users.  

Trajectory-based change detection.

Above, we show trajectories for three pixels. Light brown lines show the original data from Landsat, while colored lines show how LandTrendr segmentation simplified those data. Sharp jumps correspond to forest harvest, while slow declines correspond to gradual regrowth after harvest. The features of the segments can be mapped to show when events happened, or how fast processes such as regrowth are occurring. 
Back Home

How LandTrendr works (this page)

Overall workflow  Pre-processing Segmentation Mapping Change Mapping Landcover


LandTrendr Workflow

LandTrendr processing occurs in a series of steps shown below. 


The schematic above outlines the overall workflow of LandTrendr. Images must be acquired and "cleaned" before temporal segmentation occurs, resulting in a "Landsat data stack".  Temporal segmentation distills each pixel's history into straightline segments. The primary product of segmentation is a set of files (vertex files) that compactly capture the endpoints of those straightline segments. These can then be queried to make maps of change, or they can be used to temporally-smooth other bands of the same image. Temporally-smoothed images can then be used to build yearly maps of land cover. 
Back Home

How LandTrendr works

Overall workflow (this page)  Pre-processing Segmentation Mapping Change Mapping Landcover



The core step in LandTrendr processing is the simplification of a time-series of Landsat spectral values into their salient "shape", as represented by successive connected straightline segments.   

Above, we show flow of steps needed to identify straightline segments in a single pixel's spectral trajectory.  Details are given in Kennedy et al. (2010). Briefly, we first eliminate spikes (a), identify possible transition points (vertices) in the time series (b), remove extraneous vertices (c), identify the best path through the vertices (d), then sequentially simplify the time series (e) and finally select the best model using a simple fitting statistic (f).  
Back Home

How LandTrendr works

Overall workflow  Pre-processing Segmentation (this page) Mapping Change Mapping Landcover


Making maps from trajectories


Once segments have been identified, they can be summarized in maps in different ways.  One approach labels the change that has occurred in each pixel according to characteristic sequences of change over time. 


Above, we show maps of LandTrendr Change Labels for areas near Yosemite and Sequoia and Kings Canyon National Parks in the Sierras of California. Greens indicate where vegetation has been increasing steadily for years, while brown indicates where vegetation has been declining steadily.  Pinks and purples identify areas of disturbance, in some cases caused by fires and others by harvest, with different colors corresponding to different sequences of change before and after the major disturbance. 

Temporal smoothing

Temporal segmentation of the Landsat archive can also be used to "clean up" noise in sequential series of imagery, a process we call "LandTrendr temporal smoothing."   

The segmentation algorithms are run on a single index (in the figure below, on the NBR, or "normalized burn ratio").  The rules determined using that index for each pixel are then used to smooth the noise for that pixel in other spectral bands.   



What's different about LandTrendr?

There are many change detection efforts in the remote sensing community.  

All change detection efforts infer change on the ground from changes in spectral properties of images (i.e. the color of the images in different parts of the visible and infrared spectrum).  

The problem is that many spectral changes are associated with uninteresting change (see below). By fitting imagery across years, the uninteresting change largely becomes noise around a longer trend, allowing better separation of real from false change, and allowing capture of trends (such as growth of vegetation or chronic vegetation mortality).