13.6 Change Detection

Learning Guide

Remote sensing data have much value in letting us view the state or condition of rangelands at a point in time. However, one of the great strengths of remote sensing is that it permits us to look at changes in systems over time in a way that was not previously available to us.


The idea of remote sensing change detection has been around since the first photographs have been taken. Much has been written on the topic of remote sensing change detection, and a thorough review of the subject is beyond the scope of this module. However, change detection techniques can be broken down into several different types:


Visual Change Detection

Visual change detection is simply looking at two images to identify the changes. You can do a rudimentary change detection on many online GIS or Remote Sensing applications like Google Earth by looking at imagery from different dates


Rango et al. (2005) used visual change detection with historic aerial photographs to document the reestablishment of shrubs following a shrub-removal experiment in the 1930’s in southern New Mexico. From: https://doi.org/10.1016/j.jaridenv.2004.11.001.

Thematic Change Detection

In thematic change detection, images are first classified and then differenes in the classified images are identified. This technique can be sensitive to inaccuracies in the classification or changes in imagery types or classification algorithms over time.


The USGS used thematic change detection to map changes in land cover types across the USA between 1992 and 2001. This example shows urban expansion into rangelands surrounding Las Vegas, NV. Image from https://pubs.usgs.gov/of/2008/1379/

Image Differencing

Image differencing takes two remote sensing images, image products (e.g., NDVI) or models (e.g., models of percent cover, elevation models) and subtracts one image from the other to identify the location and magnitude of changes that have occurred. This technique is less sensitive to changes in imagery or technique over time than thematic change detection, but the question of how much change is meaningful must be addressed. If models are provided with estimates of uncertainty, these can be factored in to the change detection to identify statistically significant changes.


Gillan et al. (2017) used differencing of elevation models in a Chihuahuan Desert rangeland to measure soil erosion and deposition. Image from Gillan et al. (2012), http://dx.doi.org/10.3390/rs9050437


A variation on the image differencing concept is using a long-term series of satellite imagery to define average conditions for each image pixel and then looking for departures or variations from this average condition.

Wylie et al. (2012) used a long-term record of MODIS satellite imagery to define the Expected Ecosystem Performance (EEP) of every pixel in their study area. Differences between Actual Ecosystem Performance and EEP provide a weather-independent way to evaluate the impacts of distrubances (e.g., fire) and management activities. Image from Wylie et al. (2012), http://dx.doi.org/10.2111/REM-D-11-00058.1

Time-series Analysis

In time-series analysis, statistical techniques are used to analyze a time series of remote sensing information to detect trends or changes. This can be simple regression analyses to detect trends over time. However, because vegetation in natural ecosystems exhibits seasonal changes, more sophisticated analyses are generally needed. This is an active area of research. One model that was proposed by Verbesselt et al. (2010, see example below) breaks a time series into a seasonal component and a long-term trend component. The changes in either seasonality or long-term trend can be meaningful for land management.


Verbesselt et al. (2010) developed the Breaks For Additive Season and Trend (BFAST) technique for separating pixels from a remote sensing image time series into seasonal and long-term trend components. This allows for more easy detection and interpretation of changes. This example shows the long-term trend (red lines) component of BFAST overlaid on the original MODIS NDVI time series for a forest plantation planting in 2001 (top), a forest harvest in 2004 (middle), and forest with tree mortality events in 2003 and 2007 (bottom). Image from Verbesselt et al. (2010), http://dx.doi.org/10.1016/j.rse.2009.08.014


Browning et al. (2017) showed that the seasonal component of BFAST analysis corresponded to changes in species composition of a Chihuahuan Desert plant community, whereas the long-term BFAST trend related to overall changes in plant biomass. Image from Browning et al. (2017), http://onlinelibrary.wiley.com/doi/10.1002/eap.1561/full