NAME
i.segment.uspo - Unsupervised segmentation parameter optimizationKEYWORDS
imagery, variance, segmentation, threshold, parallelSYNOPSIS
Flags:
- -k
- Keep all segmented maps
- -h
- Use hierarchical segmentation
- -a
- Use adaptive spectral bandwidth (with mean shift)
- --overwrite
- Allow output files to overwrite existing files
- --help
- Print usage summary
- --verbose
- Verbose module output
- --quiet
- Quiet module output
- --ui
- Force launching GUI dialog
Parameters:
- group=name [required]
- Group to use for segmentation
- maps=name[,name,...]
- Raster band(s) for which to calculate variance (default: all group members)
- seeds=name
- Seeds for segmentation
- output=name
- Name for output file (- for standard output)
- segment_map=name
- Prefix for "best" output segmentation map per region
- regions=name[,name,...] [required]
- Regions in which to analyze the variance
- segmentation_method=string [required]
- Segmentation method to use
- Options: region_growing, mean_shift
- Default: region_growing
- thresholds=float[,float,...]
- Thresholds to test
- threshold_start=float
- Lowest threshold to test
- threshold_stop=float
- Threshold at which to stop (not included)
- threshold_step=float
- Step to use between thresholds
- minsizes=integer[,integer,...]
- Minimum number of cells in a segment to test
- minsize_start=integer
- Lowest minimum segment size to test
- minsize_stop=integer
- Value for minimum segment size at which to stop (not included)
- minsize_step=integer
- Step to use between minimum segment sizes
- radiuses=float[,float,...]
- Radiuses to test
- radius_start=float
- Lowest radius to test
- radius_stop=float
- Radius at which to stop (not included)
- radius_step=float
- Step to use between radiuses
- hrs=float[,float,...]
- Spectral bandwidths to test
- hr_start=float
- Lowest spectral bandwith to test
- hr_stop=float
- Spectral bandwith at which to stop (not included)
- hr_step=float
- Step to use between spectral bandwidths
- autocorrelation_indicator=string
- Indicator for measuring inter-segment heterogeneity
- Options: morans, geary
- Default: morans
- optimization_function=string
- Optimization function used to determine "best" parameters
- Options: sum, f
- Default: sum
- f_function_alpha=float
- Alpha value used for F-measure optimization function
- Default: 1
- number_best=integer
- Number of desired best parameter values and maps
- Default: 1
- memory=integer
- Total memory (in MB) to allocate (will be divided by processes)
- Default: 300
- processes=integer
- Number of processes to run in parallel
- Default: 1
Table of contents
DESCRIPTION
i.segment.uspo provides unsupervised segmentation parameter optimization for i.segment determined by the compromise between intra-segment variance and inter-segment spatial autocorrelation.
The module runs segmentation across a user defined set of thresholds and minimum segment sizes, as well, for the mean shift algorithm, a set of spectral and spatial bandwiths. For the mean shift algorithm, you can also activate adaptive bandwidth using the -a flag.
The user provides an imagery group and the name of an output text file where parameter and optimization values for all tested segmentations are stored. The user can either give a list of thresholds and minimum sizes, or provides start, stop and step values for each. In addition, the user can provide a list of named regions for which to test the segmentation. This allows to not test the entire image, but rather to test specific areas in the image that might be characterstic for specific types of land cover.
The module then selects the parameters providing the highest values of a given optimization function. The number of "best" parameter combinations to provide to the user per region is defined by number_best.
Two optimization functions are available via the optimization_function parameter: A simple sum of the normalized criteria values as defined by Espindola et al (2006), or the F-function as defined by Johnson et al (2015). When using the F-function, the user can determine the f_function_alpha value which determines the relative weight of the intra-segment variance as compared to the inter-segment spatial autocorrelation. A value of 0.5 gives the former half weight of the latter, A value of 2 gives the former double weight than the latter.
The optimization functions use intra-segment variance and inter-segment spatial autocorrelation. For the latter, the user can chose to use either Moran's I or Geary's C.
The user can chose between non-hierarchical (default) and hierarchical segmentation using the h flag. The latter uses each segmentation at a given threshold level as seed for the segmentation at the next threshold level within a given minimum segment size. Note that this leads to less optimal parallelization as for a given minsize, all segmentations have to be done sequentially (see below).
The segment_map parameter allows to provide a basename for keeping the number_best best segmentations for each given region according to the optimization function. The resulting map names will be a combination of this basename, the region, the threshold, the minsize and the rank of the map within the region according to its optimization criteria value.
The module uses high-level parallelization (running different segmentations in parallel and then running the collection of parameter values in parallel). The parameter processes allows to define how many processes should be run in parallel. Note that when using hierarchical segmentation the number of parallel processes is limited to the number of different mininum segment sizes to test.
The k flag allows to keep all segmentation maps created during the process.
NOTES
The module depends on the addon r.neighborhoodmatrix which needs to be installed.
Any unsupervised optimization can at best be a support to the user. Visual and other types of validation of the results, possibly comparing several of the "best" solutions, remain necessary.
Even though the module allows the user to test different minsizes, it is probably better to run the module with minsizes=1 and then adapt the minsize in the final run of i.segment depending on the desired minimum mapping unit.
In hierarchical segmentation mode, each segmentation is used as seed for the next hierarchical level. This means that these segmentations have to be run sequentially. Currently, parallelization is thus only used if more than one value was given for minsize. In a future version, parallelization should optionally be run by region if the number of regions is larger than the number of different minsize values.
EXAMPLE
g.region -au n=220767 s=220392 w=638129 e=638501 res=1 save=region1 g.region -au n=222063 s=221667 w=637659 e=638058 res=1 save=region2 i.group ortho input=ortho_2001_t792_1m i.segment.uspo group=ortho regions=region1,region2 \ output=ortho_parameters.csv segment_map=ortho_uspo \ threshold_start=0.02 threshold_stop=0.21 threshold_step=0.02 \ minsizes=5,10,15 number_best=5 processes=4 memory=4000
REFERENCES
G. M. Espindola , G. Camara , I. A. Reis , L. S. Bins , A. M. Monteiroi (2006), Parameter selection for region-growing image segmentation algorithms using spatial autocorrelation, International Journal of Remote Sensing, Vol. 27, Iss. 14, pp. 3035-3040, https://doi.org/10.1080/01431160600617194B. A. Johnson, M. Bragais, I. Endo, D. B. Magcale-Macandog, P. B. M. Macandog (2015), Image Segmentation Parameter Optimization Considering Within- and Between-Segment Heterogeneity at Multiple Scale Levels: Test Case for Mapping Residential Areas Using Landsat Imagery, ISPRS International Journal of Geo-Information, 4(4), pp. 2292-2305, https://doi.org/10.3390/ijgi4042292
SEE ALSO
i.segment, i.group, i.segment.hierarchical, r.neighborhoodmatrixAUTHOR
Moritz LennertSOURCE CODE
Available at: i.segment.uspo source code (history)
Latest change: Monday Nov 11 18:04:48 2024 in commit: 59e289fdb093de6dd98d5827973e41128196887d
Main index | Imagery index | Topics index | Keywords index | Graphical index | Full index
© 2003-2024 GRASS Development Team, GRASS GIS 8.4.1dev Reference Manual