3.4.1. Change detection metrics type A (change_A)
- Purpose: Annual forest loss mapping. May only be applicable when annual change data for at least three earlier years exist.
- Reference: Potapov et al., 2019 (https://doi.org/10.1016/j.rse.2019.111278)
- Interval: Annual (January 1 – December 31).
- 16-day interval data: Four years of data is recommended for metric processing (e.g., to generate 2017/2018 change metrics we recommend to download data from 2015 to 2018). See details in the description below.
- Naming convention: Metrics_change_A.xlsx
- Metric generation code: https://glad.umd.edu/gladtools/Tools/compute_metrics_change_A.zip
- Classification code parameters: Use keyword “change_A” to specify this metric set.
- Requires at least 13GB RAM.
- The metrics dataset size for one tile / one year is 9GB.
The change_A metric set was developed and implemented for global (link - http://earthenginepartners.appspot.com/science-2013-global-forest) and regional (Potapov et al., 2019 link - https://doi.org/10.1016/j.rse.2019.111278) annual forest loss mapping. This metric set is sensitive to changes that occurred within the last three years. To map change that occurred in the last year only, the results of change detection must be filtered using the mask of change detected in the preceding years.
Similar to the annual phenological metric algorithm (see Phenological metrics), we utilized several years of data to calculate a metric set for a target year. For the change_A metric set, using four years of data (current and three preceding years) is optimal. The metric set can be generated with less than four years of data, but at least 2 years of data are required. Before processing the phenological metrics, use Landsat ARD Download to download the 16-day data time series.
The process of metrics construction includes two stages: (1) selecting clear-sky observations and building the time series; and (2) extracting reflectance and reflectance change distribution statistics from the time-series.
First, we compile a time-series of annual observations with the lowest atmospheric contamination (Figure A). The per-pixel criterion for 16-day data selection is defined automatically based on the distribution of quality flags within the four years of data. If clear-sky land or water observations are present in the time-series data, only those are used for subsequent analysis. If no such observations are found, the code successively changes the quality threshold for data inclusion, first allowing observations with proximity to clouds and shadows, then allowing all available observations.
To construct the metric set, we used all cloud-free 16-day observations from the current year (the year i), hereafter referred to as time-series C (Fig. A). To create a historical baseline for change detection, we collected an average reflectance from the three preceding years (year i-1 – year i-3) only for those 16-day intervals that have cloud-free observations in the time-series C. If no observations were found for a certain 16-day interval in historic data, we used clear-sky data from the closest observation before/after the missing 16-day composite interval. The compilation of spectral reflectance from the three previous years is referred to as time-series P. For each time-series observation, in addition to normalized reflectance, we calculated normalized ratios, or indices (Band A - Band B)/(Band A + Band B) from selected bands. The per-band and per-index differences were calculated for all 16-day intervals with clear-sky data between time-series C and time-series P and recorded as time-series D.
Similarly to the phenology metrics algorithm, we ranked spectral reflectance and indices values individually for each of the time-series C and P, and extracted selected ranks and averages. We also ranked observation dates by the corresponding NDVI and brightness temperature values and recorded spectral band values for selected ranks. We aggregated time-series P and C into a single time-series, from which we calculated standard deviations and slope of linear regression for each band and index. The time-series D per-16-day difference values were ranked for each spectral band and index and selected statistics were extracted from these ranks.
Metric types and naming conversion
The metrics for each tile is stored in a separate folder as a single-band UInt16 bit GeoTIFF files. The generic naming conversion is the following:
YYYY – corresponding year
B – spectral band or index
T –time-series from which the statistics were extracted. “c” represents the current year (time-series C), “p“ stands for the preceding year (time-series P) and “dif” stands for a time-series of per-16-day interval differences between (time-series D). Regression and standard deviation metrics, which are calculated from the entire time-series, does not have this name section.
S – statistic
C – corresponding band or index used for ranking (only for metrics extracted from ranks defined by a corresponding value)
2018_blue_c_max_RN.tif - The metric represent the value of the normalized surface reflectance of the Landsat blue band for the 16-day interval that has the highest red/NIR normalized ration (also known as NDVI) value during the year 2018.
In addition to spectral metrics, the metric generation software produces a set of technical layers including the number of cloud-free 16-day composites, data quality, and water presence.
The table Metrics_change_A.xlsx has details on the bands, indices, and computed statistics.
The following software should be installed to generate metrics:
- ActivePerl (https://www.activestate.com/products/activeperl/)
- GLAD_1.0 complete package (https://glad.umd.edu/gladtools/Complete_package/GLAD_1.0_master.zip)
- Download all required 16-day composites
- Download and install software
- Make a list of tiles to process (single column, tile names only – see example tiles.txt).
- Use the following command to compute metrics:
> perl C:/GLAD_1.0/metrics_change_A.pl <tile_list> <year> <input folder> <output folder> <threads>
> perl C:/GLAD_1.0/metrics_change_A.pl tiles.txt 2018 D:/Data D:/Metrics 1
The command parameters are:
Input folder: the folder with 16-day composite data. It should contain tile data in subfolders.
Output folder: will be created by the code, tile data will be recorded into subfolders.
Threads: the number of parallel processes. The parameter should be increased only if:
- A computer has a multi-core processor (e.g., Intel Xeon)
- The RAM can hold several processes simultaneously. Each process will use 13GB RAM. To get the total RAM usage, multiply 13GB by the number of processes.