LabelRS

A toolbox to make deep learning samples from remote sensing images
We strongly recommend using Python scripts instead of ArcGIS plugins

New Features & Updates!

✅ 2023-05-01: Added AutoDraw in LabelRS to to assist segmentation.

• The automatic segmentation algorithms supported by AutoDraw include:
• slic, felzenszwalb, quickshift
• sam (Segment Anything Model)

Note: AutoDraw and LabelRS are independent.

Requirements

• ESRI ArcGIS 10.2 and later versions
• Python Library: tqdm, opencv, pillow

Note: select ArcGIS's python environment as the python interpreter, the location usually is C:\Python27\ArcGIS10.2\python.exe
tqdm, opencv, pillow need to be installed in the ArcGIS python environment

Data preparation

• Make sure your file's coordinate system is WGS84
• Multispectral GeoTif, eg. GaoFen, Sentinel-2, Landsat (Layer Stacking is needed)
• Labeled vector polygon data, which can be obtained from the following sources

1. Public land use vector data
2. Drawn by users in ArcGIS. Build a new shapefile and add a field in attribute table, then start editing and draw polygons, finally set different attribute values for different types of features.
3. Use eCognition. Use multiresolution segmentation in eCognition, and then output the segmentation vectors.
4. Use AutoDraw in LabelRS to to assist segmentation. This is an independent toolkit based on GDAL, OpenCV, skimage and PIL.

Note: Use AutoDraw to assist segmentation, please install opencv, pillow, GDAL and scikit-image libraries, and Python3 environment is recommended.
If you want to use Segment Anything Model, please make sure python>=3.8, and additionally pytorch>=1.7 and torchvision>=0.8. For details, see Segment Anything Model requirements
In short, AutoDraw and LabelRS are independent packages, for example, you can use python 2.7 to execute labelRS, and use python 3 to execute AutoDraw. This may seem strange, but LabelRS depends on ArcPy (Python 2.7), and I'm thinking of fully migrating LabelRS to a Python 3 environment.

Usage

• create semantic segmentation samples

SemanticSegmentation.py --input-image=xxx.tif --input-shpfile=xxx.shp --class-field=xxx --output-path=xxx  

Parameters Description

NAME	TYPE	REQUIRED	DESCRIPTION	DEFAULT	EXAMPLE
--input-image	str	True	the input source imagery	None	C:/GF2_PMS1_ L1A0003131556-MSS1.tif
--input-shpfile	str	True	the labeled vector data	None	C:/Water.shp
--class-field	str	True	The field used to distinguish different features in shpfile	None	class
--tile-size	int	False	The size of the output sample	256	256
--output-path	str	True	output folder	None	C:/output
--output-img-format	str	False	The format of the output sample, including JPEG, PNG and TIFF	TIFF	JPEG
--overlap-size	int	False	tile overlap size	16	16
--band-list	str	False	Bands used to generate samples, default is all bands	None	3,2,1
--write-xml	bool	False	whether to write meta data in xml	True	True
--stretch-method	int	False	Band stretching method 0,Percentage Truncation; 1,Standard Deviation; 2,Maximum and Minimum	0	0
--stretch-parameters	str	False	the input parameters for Percentage Truncation or Standard Deviation	None	0.5,99.5
--gray-level-transformation	int	False	The method of setting output label value, 0,None; 1,Maximum Contrast; 2,Positive Integer; 3,Custom	0	3
--glt-parameters	str	False	The input parameters when Gray Level Transformation is Custom	None	Water:1, building:2
--filter	float	False	Filter out samples with few foreground pixels	0.05	0.05

---

• create object detection samples

ObjectDetection.py --input-image=xxx.tif --input-shpfile=xxx.shp --class-field=xxx --output-path=xxx  

Parameters Description

NAME	TYPE	REQUIRED	DESCRIPTION	DEFAULT	EXAMPLE
--input-image	str	TRUE	the input source imagery	None	C:/GF2_PMS1_L1A0003131556-MSS1.tif
--input-shpfile	str	TRUE	the labeled vector data	None	C:/bridge.shp
--class-field	str	TRUE	The field used to distinguish different categories	None	class
--tile-size	int	FALSE	The size of the output sample	512	512
--output-path	str	TRUE	output folder	None	C:/output
--meta-format	str	FALSE	The format of the output metadata labels, including PASCAL VOC, YOLO and KITTI	PASCAL VOC	PASCAL VOC
--output-img-format	str	FALSE	the output image foramt, including JPEG, PNG and TIFF	TIFF	JPEG
--overlap-size	int	FALSE	The overlap size of the output sample	16	16
--band-list	str	FALSE	output bands list, default is all bands	None	3,2,1
--stretch-method	int	FALSE	Band stretching method. 0,Percentage Truncation; 1,Standard Deviation; 2,Maximum and Minimum	0	0
--stretch-parameters	str	FALSE	the input parameters used for Percentage Truncation or Standard Deviation	None	0.5,99.5
--vision	bool	FALSE	whether to generate visualization results	True	True

---

• create classification samples

Classification.py --input-image=xxx.tif --input-shpfile=xxx.shp --class-field=xxx --output-path=xxx 

Parameters Description

NAME	TYPE	REQUIRED	DESCRIPTION	DEFAULT	EXAMPLE
--input-image	str	TRUE	the input source imagery	None	C:/GF2_PMS1_L1A0003131556-MSS1.tif
--input-shpfile	str	TRUE	the labeled vector data	None	C:/landuse.shp
--class-field	str	TRUE	The field used to distinguish different categories	None	class
--tile-size	int	FALSE	The size of the output sample	128	64
--output-path	str	TRUE	The ouput folder	None	C:/output
--output-img-format	str	FALSE	The format of the output sample, including JPEG, PNG and TIFF	TIFF	TIFF
--overlap-size	int	FALSE	The overlap size of the output sample	16	16
--band-list	str	FALSE	Bands used to generate samples	None	3,2,1
--resampling-type	int	FALSE	Resampling method, including 0,Nearest; 1,Bilinear; 2,Cubic	0	0
--stretch-method	int	FALSE	Band stretching method,including 0,Percentage Truncation; 1,Standard Deviation; 2,Maximum and Minimum	0	0
--stretch-parameters	str	FALSE	the input parameters of band stretching	None	0.5,99.5

Citation

Li J, Meng L, Yang B, Tao C, Li L, Zhang W. LabelRS: An Automated Toolbox to Make Deep Learning Samples from Remote Sensing Images. Remote Sensing. 2021; 13(11):2064. https://doi.org/10.3390/rs13112064