Time series change detection

Available with Image Analyst license.

A time series of imagery or rasters is made up of data collected over time, usually at regular time intervals, often for the purpose of analyzing changes at the earth's surface. In ArcGIS AllSource, a time series of raster data can be organized in a multidimensional raster dataset or multidimensional mosaic dataset, and tools can be used to extract information about a pixel's history over time.

The purpose of modeling a pixel's history over tens or hundreds of images is typically to find the date at which some type of change occurred.

The CCDC Analysis raster function and the LandTrendr Analysis raster function can be chained together with the Detect Change Using Change Analysis raster function to extract date of change information from a multidimensional raster.

The Analyze Changes Using LandTrendr tool or the Analyze Changes Using CCDC tool, in conjunction with the Detect Change Using Change Analysis Raster tool, can be used to identify changes in pixel values over time to indicate land use or land cover changes.

The Change Detection Wizard combines the available tools and functions to guide you through the process of extracting date of change information from a time series of imagery or rasters. The output from the wizard is a raster in which each pixel has a date value corresponding to the time of a particular type of change.

The following section provides details on each pane in the Change Detection Wizard when performing time series change detection.

Change Detection Wizard

The Change Detection Wizard is launched from the Change Detection drop-down button on the Imagery tab, in the Analysis group. The button is unavailable if you are not working in a 2D map scene, or if you do not have the Image Analyst extension.

Configure

The first pane in the Change Detection Wizard is the Configure pane, where you can select the Change Detection Method option you want to use. To extract date of change information from a multidimensional raster, set the Change Detection Method to Time series change detection.

Parameter	Description
Input Raster	The input multidimensional raster dataset to be analyzed. Supported inputs include multidimensional Cloud Raster Format (.crf) files, multidimensional mosaic datasets, or multidimensional image services. This tool extracts changes in an observed feature, so the ideal input multidimensional imagery should capture a consistent observation throughout time and should not include atmospheric or sensor interference, clouds, or cloud shadow. The best practice is to use data that has been normalized and can be masked using a QA band—for example, Landsat Collection 1 Surface Reflectance products with a cloud mask. If you have already generated a change analysis raster using the Analyze Changes Using LandTrendr or the Analyze Changes Using CCDC tools, you can provide the result as the input raster in the wizard, and the next pane will be skipped.
Processing Extent	The processing extent for the output change raster.

Analyze Time Series

The Analyze Time Series pane allows you to specify the type of model to run to perform time series analysis, and to configure the model. This pane will not appear if you entered an existing change analysis raster in the Configure pane.

The parameters visible in this pane depend on the modeling option selected in the Change Analysis Method parameter:

• CCDC—The Continuous Change Detection and Classification (CCDC) algorithm will be used to evaluate changes in pixel values over time. To use this option, the input multidimensional raster must contain at least 12 slices, spanning at least 1 year. For information on the algorithm and parameters, see How Analyze Changes Using CCDC works.
• LandTrendr—The Landsat-based detection of trends in disturbance and recovery (LandTrendr) algorithm will be used to evaluate changes in pixel values over time. For more information on the algorithm and parameters, see How the Analyze Changes Using LandTrendr tool works.

CCDC Change Analysis Parameters

Parameter	Description
Bands for Detecting Change	The spectral bands to analyze for change detection. The default is to use all bands.
Bands for Temporal Masking	The bands to use for cloud, cloud shadow, and snow masking. Because cloud shadow and snow are very dark in the shortwave infrared (SWIR) band, and clouds and snow are very bright in the green band, it is recommended that you use the SWIR and green bands for masking. If no bands are selected, no masking will occur.
Chi-squared Threshold for Detect Change	The chi-square statistic change probability threshold. If an observation has a calculated change probability that is above this threshold, it is flagged as an anomaly, which is a potential change event. The default value is 0.99.
Minimum Consecutive Anomaly Observations	The minimum number of consecutive anomaly observations that must occur before an event is considered a change. A pixel must be flagged as an anomaly for the specified number of consecutive time slices before it is considered a true change. The default value is 6.
Updating Fitting Frequency (in years)	The frequency, in years, at which to update the time series model with new observations. Updating a model frequently can be computationally costly and the benefit can be minimal. For example, if there are 365 slices or clear observations per year in the multidimensional raster, and the updating frequency is for every observation, the processing will be 365 times more computationally expensive compared to updating once per year, but the accuracy may not be higher. The default value is 1.

LandTrendr Change Analysis Parameters

Parameter	Description
Processing Band	The image band name to use for segmenting the pixel value trajectories over time. Choose the band name that will best capture the changes in the feature you want to observe. The default is the first band.
Snapping Date	The date used to identify a slice for each year in the input multidimensional dataset. The slice with the date closest to the snapping date will be used. This parameter is required if the input dataset contains sub-yearly data. The default is 06-30, or June 30, which is approximately midway through a calendar year.
Maximum Number of Segments	The maximum number of segments to be fitted to the time series for each pixel. The default is 5.
Vertex Count Overshoot	The number of additional vertices beyond maximum number of segments + 1 that can be used to fit the model during the initial stage of identifying vertices. Later in the modeling process, the number of additional vertices will be reduced to maximum number of segments + 1. The default is 2.
Spike Threshold	The threshold to use for dampening spikes or anomalies in the pixel value trajectory. The value must range between 0 and 1 in which 1 means no dampening. The default is 0.9.
Recovery Threshold	The recovery threshold value, in years. A feature in a landscape will often take time to recover from a nonpermanent change such as a forest fire or insect infestation. Use this parameter to control the rate of recovery recognized by the model. If a segment has a recovery rate that is faster than 1/recovery threshold, the segment is discarded and not included in the time series model. The value must range between 0 and 1. The default is 0.25.
Minimum Number of Observations	The minimum number of valid observations required to perform fitting. The number of years in the input multidimensional dataset must be equal to or greater than this value. The default is 6.
P-Value Threshold	The p-value threshold for a model to be selected. After the vertices are detected in the initial stage of the model fitting, the tool will fit each segment and calculate the p-value to determine the significance of the model. On the next iteration, the model will decrease the number of segments by one and recalculate the p-value. This will continue and, if the p-value is smaller than the value specified in this parameter, the model will be selected and the tool will stop searching for a better model. If no such model is selected, the tool will select a model with a p-value smaller than the lowest p-value × best model proportion value. The default is 0.01.
Best Model Proportion	The best model proportion value. During the model selection process, the tool will calculate the p-value for each model and identify a model that has the most vertices while maintaining the smallest (most significant) p-value based on this proportion value. A value of 1 means the model has the lowest p-value but may not have a high number of vertices. The default is 1.25.
Prevent One Year Recovery	Specifies whether segments that exhibit a one year recovery will be excluded. Checked—Segments that exhibit a one year recovery will be excluded. This is the default. Unchecked—Segments that exhibit a one year recovery will be not be excluded.
Recovery Has Increasing Trend	Specifies whether the recovery has an increasing (positive) trend. Checked—The recovery has an increasing trend. This is the default. Unchecked—The recovery has a decreasing trend. The recovery from a change in landscape can occur in the positive or negative direction. For example, when a landscape experiences forest loss, a time series of vegetation index values shows a drop in index values, and the recovery shows a gradual increase in vegetation index values, or a positive recovery trend.
Output Other Bands	Specifies whether other bands will be included in the results. Checked—Other bands will be included in the results. The segmentation and vertices information from the initial segmentation band specified in the Processing Band parameter will also be fitted to the remaining bands in the multiband images. The model results will include the segmentation band first, then the remaining bands. Unchecked—Other bands will not be included in the results. This is the default.

Detect Date of Change

The Detect Date of Change pane provides the parameters for you to specify the date of change information you want to extract from the model.

Parameter	Description
Change Type	Specifies the change information to calculate for each pixel. When using the CCDC change analysis method, you can choose from the following options: Time of latest change—The most recent date and time at which a pixel was flagged as being changed. Time of earliest change—The earliest date and time at which a pixel was flagged as being changed. Time of largest change—The date and time at which the calculated change was the most significant for a pixel. Number of changes—The total number of times the pixel changed. When using the LandTrendr change analysis method, the following additional options are available: Time of longest change—The date at which a pixel was flagged as being changed, at the beginning or end of the longest period of change. Time of shortest change—The date at which a pixel was flagged as being changed, at the beginning or end of the shortest period of change. Time of fastest change—The date at which a pixel was flagged as being changed, at the beginning or end of the fastest period of change. Time of slowest change—The date at which a pixel was flagged as being changed, at the beginning or end of the slowest period of change. Filter by Start Value—Filter results by start value so that only the changes of a given start value are included in the output. Filter by End Value—Filter results by end value so that only the changes of a given end value are included in the output.
Maximum Number of Changes	The maximum number of changes per pixel that will be calculated. This number corresponds to the number of bands in the output raster. The default is 1, meaning only one change date will be calculated, and the output raster will contain only one band. This parameter is not applied when the Change Type parameter is set to Number of changes.
Segment Date	Specifies whether the date at the beginning of a change segment will be extracted or at the end. This parameter is available only when using the LandTrendr change analysis method.
Change Direction	Specifies the direction of change to be included in the analysis. All Directions—All change directions are included in the output. This is the default. Increasing—Only change in the positive or increasing direction is included in the output. Decreasing—Only change in the negative or decreasing direction is included in the output. This parameter is available only when using the LandTrendr change analysis method.
Filter by Year	Specifies whether the output will be filtered by a range of years. Checked—Results will be filtered so that only changes that occurred within a specific range of years will be included in the output. Unchecked—Results will not be filtered by year. This is the default. This parameter is available only when using the LandTrendr change analysis method. Use this parameter to identify changes that occurred within a specific time period, for example, if you are looking for changes that occurred in a landscape during five years of drought. If checked, you must enter the minimum and maximum years to use for filtering results.
Filter by Duration	Specifies whether results will be filtered by the change duration. Checked—Results will be filtered by duration so that only the changes that lasted a given amount of time will be included in the output. Unchecked—Results will not be filtered by duration. This is the default. This parameter is available only when using the LandTrendr change analysis method. Use this parameter to identify changes that occurred over a given range of years, for example, if you are only interested in abrupt changes over 1 or 2 years. You can calculate the duration you are interested in using the formula end year - start year +1. Gaps in the time series will be included. If checked, you must enter the minimum and maximum duration values to use for filtering results.
Filter by Magnitude	Specifies whether results will be filtered by change magnitude. Checked—Results will be filtered by magnitude so that only the changes of a given magnitude will be included in the output. Unchecked—Results will not be filtered by magnitude. This is the default. This parameter is available only when using the LandTrendr change analysis method. Use this parameter to identify changes of a given magnitude, for example, if you are only looking for large changes in the vegetation index NDVI. Magnitude is an absolute value, so the minimum and maximum values cannot be negative. To specify directional change, use the Change Direction parameter. If checked, you must enter the minimum and maximum magnitude values to use for filtering results.
Output Date of Change Raster	The output dataset. The output is a multiband raster in which each band contains change information depending on the change type selected and the maximum number of changes specified. For example, if the Change Type parameter is set to Time of earliest change and the Maximum Number of Changes parameter is set to 2, the tool calculates the two earliest dates when change occurred throughout the time series for every pixel. The result is a raster in which the first band contains the dates of the earliest change per pixel, and the second band contains the dates of the second-earliest change per pixel.

Example

The following example extracts the date of the most rapid change from a time series of yearly NDVI rasters from 2000 to 2019.

1. Add the NDVI multidimensional raster dataset to your map.
2. With the layer selected in the Contents pane, open the Change Detection Wizard from the Imagery tab in the Analysis group.
3. In the Configure pane, set the Change Detection Method to Time Series Change and ensure the Input Raster is set to the NDVI multidimensional raster. Click Next.
4. In the Analyze Time Series pane, configure the parameters to perform LandTrendr modeling.
   1. Set the Change Analysis Method parameter to LandTrendr.
   2. Set the Maximum Number of Segments parameter to 10.
   3. Leave all other defaults.
   4. Click Next.
5. In the Detect Change pane, configure the parameters to extract the beginning of the most rapid, highly negative (loss of NDVI) change in the series.
   1. Set the Change Type to Time of Fastest Change.
   2. Set the Change Direction to Decreasing.
   3. Check the Filter by Magnitude check box.
   4. Set the Minimum magnitude to 0.5 and the Maximum magnitude to 2.
   5. For the Output Date of Change Raster, type FastestNDVILoss.crf.
6. Click Run.