CLASSIFICATION OF LAND BASED ON ANALYSIS OF REMOTELY-SENSED EARTH IMAGES

• US 20150071528A1
• Filed: 09/11/2013
• Published: 03/12/2015
• Est. Priority Date: 09/11/2013
• Status: Active Grant

• Current Assignee: Maxar Intelligence Incorporated (Galileo Parent, Inc.)
• Original Assignee: DigitalGlobe, Inc. (Galileo Parent, Inc.)
• Inventors: Marchisio, Giovanni B., Tusk, Carsten, Koperski, Krzysztof, Tabb, Mark D., Shafer, Jeffrey D.
• Granted Patent: US 9,147,132 B2
• Time in Patent Office: Days
• US Class Current: 382/159
• CPC Class Codes: G06F 16/29 Geographical information databases
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06V 20/13 Satellite images
• CPC Class Codes: G06V 20/194 using hyperspectral data, i.e. more or other wavelengths than RGB

Abstract

Land classification based on analysis of image data. Feature extraction techniques may be used to generate a feature stack corresponding to the image data to be classified. A user may identify training data from the image data from which a classification model may be generated using one or more machine learning techniques applied to one or more features of the image. In this regard, the classification module may in turn be used to classify pixels from the image data other than the training data. Additionally, quantifiable metrics regarding the accuracy and/or precision of the models may be provided for model evaluation and/or comparison. Additionally, the generation of models may be performed in a distributed system such that model creation and/or application may be distributed in a multi-user environment for collaborative and/or iterative approaches.

Claims

1. A land classification system operable for analysis of very high resolution (VHR) remotely-sensed multispectral Earth imagery, comprising:

an image store containing image data corresponding to VHR remotely-sensed multispectral Earth images;

at least one feature extraction module in operative communication with the image store, the feature extraction module being operable to produce feature data regarding at least a portion of image data;

a feature stack comprising the image data and the feature data;

a client interface operable to receive training data regarding at least one or more pixels of the image data from a user regarding at least one class to which the one or more pixels belong; and

a classification compute module operable to generate a classification model at least in part based on a portion of the feature stack corresponding to the one or more pixels and the training data, wherein the classification model relates to classification of pixels of image data into one or more classes.

2. The system of claim 1, wherein the classification compute module is operable to apply the classification model to the image data to classify pixels of the image data into the one or more classes.

3. The system of claim 2, wherein the classification compute module classifies pixels of the image data based on an analysis of the feature stack in relation to the classification model.

4. The system of claim 3, further comprising:

a post classification analysis module operable to reclassify at least one pixel of the image data based on at least one post classification rule.

5. The system of claim 4, wherein the post classification rule comprises at least one of a minimum mapping unit rule, a surrounded by rule, a topological relation rule, a majority analysis rule, or a smoothing rule.

6. The system of claim 1, wherein the VHR remotely-sensed multispectral Earth image data comprises spectral band data corresponding to at least 8 multispectral bands.

7. The system of claim 6, wherein the multispectral bands collectively range from at least about 300 nanometers in wavelength to at least about 1100 nanometers in wavelength.

8. The system of claim 7, wherein the multispectral bands collectively range from at least about 300 nanometers in wavelength to at least about 2400 nanometers in wavelength.

9. The system of claim 8, wherein the multispectral bands comprise a plurality of short wave infrared (SWIR) bands.

10. The system of claim 9, wherein the plurality of SWIR bands range from at least about 1100 nanometers in wavelength to at least about 2400 nanometers in wavelength.

11. The system of claim 1, wherein the feature stack comprises a plurality of data layers, wherein each data layer includes different feature data for each pixel of the image data.

12. The system of claim 11, wherein the feature stack includes at least one spectral feature layer, at least one morphological feature layer, or at least one textural feature layer.

13. The system of claim 12, wherein the at least one spectral feature layer comprises data values for each pixel of the image data based on spectral band data collected by a remote image acquisition platform.

14. The system of claim 13, wherein the remote image acquisition platform comprises a satellite imaging system.

15. The system of claim 13, wherein at least one spectral feature layer comprises a relative measure between at least two spectral band data layers.

16. The system of claim 12, wherein the at least one morphological feature layer comprises for each given pixel of the image data information based on the arrangement of adjacent pixels relative to the given pixel.

17. The system of claim 12, wherein the textural feature layer comprises information for a given pixel based on a spatial distribution of tonal variations within one or more spectral band data layers relative to a given pixel.

18. The system of claim 1, wherein the client interface comprises an internet browser-based client operable to provide a user interface for receiving the training data from a user.

19. The system of claim 18, wherein the client interface comprises a plurality of distributed client interfaces operable to receive training data from a plurality of users, wherein the classification model is at least partially based on training data received from a plurality of users.

20. The system of claim 1, further comprising a plurality of feature extraction modules each operable to generate a feature data layer of the feature stack.

21. The system of claim 20, wherein one or more of the feature extraction modules are selectively applicable to image data by a user to generate feature data layers.

23. The system of claim 21, wherein at least one of the feature extraction modules comprises a fractal dimension feature extraction module;

and

wherein the fractal dimension feature extraction module is operative to generate feature data by;

identifying an image of a predetermined size of pixels;

dividing the image into a plurality of abutting square windows having a side dimension less than the predetermined size of the image;

constructing for each plurality of abutting square windows a column having a height as represented in increments of the side dimension of the abutting window, wherein the height is determined such that the ratio of the predetermined size of the image to the side dimension of the abutting windows equals the ratio of a maximum pixel value for a given multispectral band of the image to the height, wherein the squared side dimension and the height represent a plurality of boxes whose volume is defined by increments of the side dimension;

determining, for each abutting window a box from the plurality of boxes in which a maximum pixel value and a minimum pixel value occur;

calculating, for each of the plurality of abutting windows, the number of boxes separating the box in which the maximum pixel value is disposed and the box in which the minimum pixel value is disposed; and

summing the number of boxes separating maximum and minimum pixel levels over all of the plurality of abutting windows to generate a summed value;

repeating each of the constructing, determining, calculating, and summing operations for all possible values of the side dimension of the plurality of abutting windows such that each side dimension is less than or equal to half of the predetermined size of the image and at least greater than three pixels in length;

plotting each summed value generated in the repeating step to determine a slope of the summed values, wherein the slope comprises the fractal dimension of the image.

29. The system of claim 21, wherein at least one of the feature extraction modules comprises a total variational (TV) feature extraction module;

and

wherein the TV feature extraction module comprises a multispectral filtering module using a plurality of spectral bands of an image to generate a filtered image corresponding to the image.

39. The system of claim 21, wherein at least one of the feature extraction modules comprises a rotational invariant histogram of gradients (HOG) Mature extraction module;

and

wherein the rotational invariant HOG feature extraction module is operative to generate feature data by;

establishing a plurality of angular bins for an image in a gradient domain;

histogramming gradients from the image with respect to the plurality of angular bins;

selecting based on the histogramming the angular bin with the largest histogram value to define a primary direction;

setting a plurality of pooling windows with relative offsets from a pixel of interest;

defining a gradient orientation window encompassing the plurality of pooling windows;

configuring the gradient orientation window to align the configuration of the gradient orientation window relative to the primary direction;

shifting the positions of the pooling windows relative to the gradient orientation window based on the configuring; and

rotating the internal orientation of the pooling windows based on the configuring.

52. The system of claim 21, wherein at least one of the feature extraction modules comprises a clustering histogram feature extraction module for extraction of features based on clustered and classified spectra values;

and

wherein the clustering histogram feature extraction module is operative to generate feature data by;

assigning each pixel of an image into one of a plurality of a predefined classes;

identifying a kernel surrounding each pixel containing a plurality of others pixels from the image, wherein each of the plurality of other pixels have been assigned into one of the plurality of predefined classes;

building a histogram for each pixel based on the number of other pixels in each of the plurality of predefined classes relative to the pixel for which the histogram is built.

56. The system of claim 1, wherein the classification compute module is operable to utilize at least one machine learning algorithm to generate the classification model.

57. The system of claim 56, wherein the classification compute module is operable to utilize a plurality of machine learning algorithms to generate the classification model.

58. The system of claim 56, wherein a user is operable to specify at least one of a plurality of machine learning algorithms for use in generation of the classification model.

59. The system of claim 1, wherein the classification model comprises at least one model parameter corresponding to at least one of a geographic region or a temporal period.

60. The system of claim 59, wherein the classification computation module model is operable to determine the classification model for classification of image data based on the at least one model parameter and image data metadata.

61. The system of claim 1, wherein the one or more classes correspond to at least one of a land cover, a land use, or temperature.

62. The system of claim 1, further comprising:

a model evaluation module operable to provide quantifiable evaluation data regarding the classification module based on performance of the classification module with respect to a portion of training data received by a user.

63. The system of claim 62, wherein the model evaluation module is operable to generate a classifier accuracy value for the classification module.

64. The system of claim 63, wherein the model evaluation module is operable to generate at least one precision measurement for the classification module.

65. The system of claim 62, wherein the model evaluation module is operable to compare a first model to a second model to determine if the performance of the first model is statistically significant over the performance of the second model.

66. The system of claim 1, wherein the feature extraction module is operable to first produce feature information for pixels belonging to the training data prior to generation of the classification model by the classification compute model and is operable to second produce feature information for pixels other than the training data prior to application of the classification model to classify pixels other than the training data.

22. (canceled)

24-28. -28. (canceled)

30-38. -38. (canceled)

40-51. -51. (canceled)

53-55. -55. (canceled)

67. A method for land classification based on analysis of very high resolution (VHR) remotely-sensed multispectral Earth imagery, comprising:

storing VHR remotely-sensed multispectral Earth image data in an image store;

generating feature data based on the image data using at least one feature extraction module;

compiling a feature stack comprising the image data and corresponding feature data;

receiving training data regarding at least one or more pixels of the image data from a user regarding at least one class to which the one or more pixels belong; and

generating a classification model at least in part based on a portion of the feature stack corresponding to the one or more pixels and the training data, wherein the classification model related to classification of pixels of image data into one or more classes.

68-87. -87. (canceled)