- Landscape change assessment
, change detection
, landscape structure
r.change.info [-ac] input=name[,name,...] output=name[,name,...] [method=string[,string,...]] [size=integer] [step=integer] [alpha=float] [--overwrite] [--help] [--verbose] [--quiet] [--ui]
- Do not align input region with input maps
- Use circular mask
- Allow output files to overwrite existing files
- Print usage summary
- Verbose module output
- Quiet module output
- Force launching GUI dialog
- input=name[,name,...] [required]
- Name of input raster map(s)
- output=name[,name,...] [required]
- Name for output raster map(s)
- Change assessment
- Options: pc, gain1, gain2, gain3, ratio1, ratio2, ratio3, gini1, gini2, gini3, dist1, dist2, dist3, chisq1, chisq2, chisq3
- Default: ratio3
- pc: proportion of changes
- gain1: Information gain for category distributions
- gain2: Information gain for size distributions
- gain3: Information gain for category and size distributions
- ratio1: Information gain ratio for category distributions
- ratio2: Information gain ratio for size distributions
- ratio3: Information gain ratio for category and size distributions
- gini1: Gini impurity for category distributions
- gini2: Gini impurity for size distributions
- gini3: Gini impurity for category and size distributions
- dist1: Statistical distance for category distributions
- dist2: Statistical distance for size distributions
- dist3: Statistical distance for category and size distributions
- chisq1: CHI-square for category distributions
- chisq2: CHI-square for size distributions
- chisq3: CHI-square for category and size distributions
- Window size (cells)
- Default: 40
- Processing step (cells)
- Default: 40
- Alpha for general entropy
- Default = 1 for Shannon Entropy
- Default: 1
calculates landscape change assessment
for a series of categorical maps, e.g. land cover/land use, with
different measures based on information theory and machine learning.
More than two input
maps can be specified.
r.change.info moves a processing window over the
input maps. This processing window is the current landscape
under consideration. The size of the window is defined with
size. Change assessment is done for each processing window
(landscape) separately. The centers of the processing windows are
step cells apart and the output maps will have a
resolution of step input cells. step should not be larger
than size, otherwise some cells will be skipped. If step
is half of size , the moving windows will overlap by 50%. The
overlap increases when step becomes smaller. A smaller
step and/or a larger size will require longer processing
The measures information gain, information gain
ratio, CHI-square and Gini-impurity are commonly
used in decision tree modelling (Quinlan 1986) to compare
distributions. These measures as well as the statistical distance are
based on landscape structure and are calculated for the distributions
of patch categories and/or patch sizes. A patch is a contiguous block
of cells with the same category (class), for example a forest fragment.
The proportion of changes is based on cell changes in the current
The method pc
calculates the proportion of changes
the actual number of cell changes in the current landscape divided by
the theoretical maximum number of changes (number of cells in the
processing window x (number of input maps - 1)).
For each processing window, the number of cells per category are
counted and patches are identified.
The size and category of each patch is recorded. From these cell and
patch statistics, three kinds of patterns (distributions) are
- 1. Distributions over categories (e.g land cover class)
- This provides information about changes in categories (e.g land
cover class), e.g. if one category becomes more prominent. This detects
changes in category composition.
- 2. Distributions over size classes
- This provides information about fragmentation, e.g. if a few large
fragments are broken up into many small fragments. This detects changes
- 3. Distributions over categories and size classes.
- This provides information about whether particular combinations of
category and size class changed between input maps. This detects
changes in the general landscape structure.
The latter is obtained from the category and size of each patch, i.e.
each unique category - size class combination becomes a separate class.
The numbers indicate which distribtution will be used for the selected
method (see below).
A low change in category distributions and a high change in
size distributions means that the frequency of categories did not
change much whereas the size of patches did change.
The methods gain1, gain2 and gain3
calculate the information
after Quinlan (1986). The information gain is the difference
between the entropy of the combined distribution and the average
entropy of the observed distributions (conditional entropy). A larger
value means larger differences between input maps.
Information gain indicates the absolute amount of information gained
(to be precise, reduced uncertainty) when considering the individual
input maps instead of their combination. When cells and patches are
distributed over a large number of categories and a large number of size
classes, information gain tends to over-estimate changes.
The information gain can be zero even if all cells changed, but the
distributions (frequencies of occurrence) remained identical. The square
root of the information gain is sometimes used as a distance measure
and it is closely related to Fisher's information metric.
Information gain ratio
The methods ratio1, ratio2 and ratio3
calculate the information
that changes occurred, estimated with the ratio of the
average entropy of the observed distributions to the entropy of the
combined distribution. In other words, the ratio is equivalent to the
ratio of actual change to maximum possible change (in uncertainty). The
gain ratio is better suited than absolute information gain when the
cells are distributed over a large number of categories and a large number
of size classes. The gain ratio here follows the same rationale as
the gain ratio of Quinlan (1986), but is calculated differently.
The gain ratio is always in the range (0, 1). A larger value means
larger differences between input maps.
The methods chisq1, chisq2 and chisq3
calculate CHI square
after Quinlan (1986) to estimate the relevance of the different input
maps. If the input maps are identical, the relevance measured as
CHI-square is zero, i.e. no change occurred. If the input maps differ from
each other substantially, major changes occurred and the relevance
measured as CHI-square is large.
The methods gini1, gini2 and gini3
calculate the Gini
, which is 1 - Simpson's index, or 1 - 1 / diversity, or 1
- 1 / 2^entropy for alpha = 1. The Gini impurity can thus be regarded
as a modified measure of the diversity of a distribution. Changes
occurred when the diversity of the combined distribution is larger than
the average diversity of the observed distributions, thus a larger
value means larger differences between input maps.
The Gini impurity is always in the range (0, 1) and calculated with
G = 1 - ∑ pi2
The methods information gain and CHI square are the
most sensitive measures, but also the most susceptible to noise. The
information gain ratio is less sensitive, but more robust
against noise. The Gini impurity is the least sensitive and
detects only drastic changes.
The methods dist1, dist2 and dist3
calculate the statistical
from the absolute differences between the average
distribution and the observed distributions. The distance is always in
the range (0, 1). A larger value means larger differences between input
Methods using the category or size class distributions (gain1,
gain2, ratio1, ratio2 gini1,
gini2, dist1, dist2) are less sensitive than
methods using the combined category and size class distributions
(gain3, ratio3, gini3, dist3).
For a thorough change assessment it is recommended to calculate
different change assessment measures (at least information gain and
information gain ratio) and investigate the differences between these
Entropies for information gain and its ratio are by default Shannon
, calculated with
H = ∑
The entropies are here calculated with base 2 logarithms. The upper
bound of information gain is thus log2(number of classes).
Classes can be categories, size classes, or unique combinations of
categories and size classes.
option can be used to calculate general entropies
after Rényi (1961) with the formula
= 1 / (1 - α) * log2
An alpha < 1 gives higher weight to small frequencies,
whereas an alpha > 1 gives higher weight to large
frequencies. This is useful for noisy input data such as the MODIS land
cover/land use products MCD12*. These data often differ only in
single-cell patches. These differences can be due to the applied
classification procedure. Moreover, the probabilities that a cell has
been assigned to class A or class B are often very similar, i.e.
different classes are confused by the applied classification procedure.
In such cases an alpha > 1, e.g. 2, will give less weight to
small changes and more weight to large changes, to a degree alleviating
the problem of class confusion.
Assuming there is a time series of the MODIS land cover/land use
product MCD12Q1, land cover type 1, available, and the raster maps have
then a change assessment can be done with
r.change.info in=`g.list type=rast pat=MCD12Q1.A*.Land_Cover_Type_1 sep=,` \
radius=20 step=40 alpha=2
Available at: r.change.info source code (history)
Main index |
Raster index |
Topics index |
Keywords index |
Graphical index |
GRASS Development Team,
GRASS GIS 7.8.4dev Reference Manual