Abstract: Systems and methods for large-scale geospatial mosaic generation with image processing in a cloud computing environment. The approaches described herein specifically leverage scalable cloud computing features to facilitate highly parallel, granular image processing. Front-end image processing techniques allow for generation of a user interface that may provide automated material selection with human operator refinement. The user interface may be a web-based design that provides browse version images at lower resolutions to improve performance of the user interface and generating mosaic recipe. In turn, the mosaic recipe may be provided to back-end image processing that coordinates strip-level jobs and tile-level jobs and highly parallel fashion scalable cloud computing nodes. In turn, very large-scale mosaic images may be generated from geospatial images in computationally and cost-effective manner.

Abstract: Systems and methods for detecting vehicle behavior patterns in real-time. One embodiment is a method of continuous vehicle behavior detection. The method includes receiving a vehicle behavior profile including one or more travel patterns that define a vehicle behavior, receiving track data of one or more vehicles, hashing the track data as it is received to generate hash values that uniquely identify cells that approximate locations of the one or more vehicles, and storing the hash values in a hash library. The method also includes analyzing the hash library as the hash values are stored to compare the cells with the one or more travel patterns in the vehicle behavior profile, and in response to determining a group of the cells match the one or more travel patterns, generating a message indicating the vehicle behavior is detected.

Abstract: A system for vector extraction comprising a vector extraction engine stored and operating on a network-connected computing device that loads raster images from a database stored and operating on a network-connected computing device, identifies features in the raster images, and computes a vector based on the features, and methods for feature and vector extraction.

Abstract: Onboard geolocation of target pixels in images captured by a spacecraft is disclosed. By performing the geolocation onboard the spacecraft considerable time is saved. The geolocation may be based on timestamped ephemeris data and timestamped attitude data of the spacecraft. In one aspect, initial geolocation data is refined based on a digital construct. In one aspect, the accuracy of the ephemeris data and/or the attitude data are improved based on refinements made to the initial geolocation data. For example, corrections may be made to attitude and/or orbit filter parameters based on the refinements made to the initial geolocation data. In one aspect, the accuracy of the attitude data and/or the ephemeris data may be improved by the use of state information. The state information may be determined at a ground computing station based on information not available to the spacecraft such as Ground Control Points and/or Star Control Points.

Abstract: A method is provided for estimating velocity of a moving object. The method includes generating a raw velocity for the moving object from a position change between a first image generated by a first sensor and a second image generated by a second sensor of a sensor assembly and selecting a plurality of stationary background features that are captured with the moving object in one or more images. A background misregistration value is generated from apparent movement of the plurality of stationary background features and is used to correct the raw velocity.

Abstract: Estimating absolute geospatial accuracy in input images without the use of surveyed control points is disclosed. For example, the absolute geospatial accuracy of a satellite images may be estimated without the use of control points (GCPs). The absolute geospatial accuracy of the input images may be estimated based on a statistical measure of relative accuracies between pairs of overlapping images. The estimation of the absolute geospatial accuracy may include determining a root mean square error of the relative accuracies between pairs of overlapping images. For example, the absolute geospatial accuracy of the input images may be estimated by determining a root mean square error of the shears of respective pairs of overlapping images. The estimated absolute geospatial accuracy may be used to curate GCPs, evaluate a digital elevation map, generate a heatmap, or determine whether the adjust the images until a target absolute geospatial accuracy is met.

Abstract: Estimating absolute geospatial accuracy in input images without the use of surveyed control points is disclosed. For example, the absolute geospatial accuracy of a satellite images may be estimated without the use of control points (GCPs). The absolute geospatial accuracy of the input images may be estimated based on a statistical measure of relative accuracies between pairs of overlapping images. The estimation of the absolute geospatial accuracy may include determining a root mean square error of the relative accuracies between pairs of overlapping images. For example, the absolute geospatial accuracy of the input images may be estimated by determining a root mean square error of the shears of respective pairs of overlapping images. The estimated absolute geospatial accuracy may be used to curate GCPs, evaluate a digital elevation map, generate a heatmap, or determine whether the adjust the images until a target absolute geospatial accuracy is met.

Abstract: A set of input images from satellites (or other remote sensors) can be orthorectified and stitched together to create a mosaic. If the resulting mosaic is not of suitable quality, the input images can be adjusted and the processes of orthorectifying and creating the mosaic can be repeated. However, orthorectifying and creating the mosaic uses a large amount of computational resources and takes a lot of time. Therefore, performing numerous iterations is expensive and sometimes not practical. To overcome these issues, it is proposed to generate an indication of accuracy of the resulting mosaic prior to orthorectifying and creating the mosaic by accessing a set of points in the plurality of input images, projecting the points to a model, determining residuals for the projected points, and generating the indication of accuracy of the orthorectified mosaic based on the determined residuals.

Abstract: Techniques for automatically determining, on a pixel by pixel basis, whether imagery includes ground images or is obscured by cloud cover. The techniques include training a Neural Network, making an initial determination of cloud or ground by using the Neural Network, and performing a max-flow, min-cut operation on the image to determine whether each pixel is a cloud or ground imagery.

Abstract: A system and method for automatically (without human intervention) identifying a material in an image that comprises a building material for buildings in the image. Building side polygons which may be used to identify building sides in off-nadir imagery are generated. Off-nadir, multispectral images, building footprint data and elevation data for a geographic area are taken as input. Building heights for buildings in the geographic area are determined by clipping the elevation data using the building footprint data and then calculating building heights. A candidate set of polygons representing visible side faces of each building in the images is created from the known building heights, and based on the viewpoint direction, using vector analysis. After culling occluded polygons and polygons too small for analysis, the polygons are associated with a building footprint. Building materials for each building having visible polygons can then be identified.

Abstract: A set of input images from satellites (or other remote sensors) can be orthorectified and stitched together to create a mosaic. If the resulting mosaic is not of suitable quality, the input images can be adjusted and the processes of orthorectifying and creating the mosaic can be repeated. However, orthorectifying and creating the mosaic uses a large amount of computational resources and takes a lot of time. Therefore, performing numerous iterations is expensive and sometimes not practical. To overcome these issues, it is proposed to generate an indication of accuracy of the resulting mosaic prior to orthorectifying and creating the mosaic by accessing a set of points in the plurality of input images, projecting the points to a model, determining residuals for the projected points, and generating the indication of accuracy of the orthorectified mosaic based on the determined residuals.

Abstract: Technology is provided for identifying concrete and/or asphalt (or other materials) in a multispectral satellite image that has multiple bands including a first set of bands from a first sensor and a second set of bands from a second sensor. The second sensor is at a different position on a focal plane as compared to the first sensor so that a single location depicted in the multispectral image will have been sensed at different times by the first sensor and the second sensor. The system identifies moving vehicles in the multispectral image and subsequently identifies sample pixels in the multispectral image that are near the moving vehicles. These pixels are high confidence samples of roads made of concrete and/or asphalt. Additional pixels are identified in the multispectral image having spectral characteristics that are within a threshold of spectral characteristics of the sample pixels. These additional pixels also depict concrete and/or asphalt.

Abstract: An automated system is provided for classifying materials in remotely-sensed imagery based on automated construction of a dynamic classifier—namely, a classifier that is automatically trained on the same image to which it is then subsequently applied. A first automated process identifies high confidence exemplars of each class using tailored classification techniques. This data is then used to train a supervised classification model (e.g., discriminant analysis), and the resultant classifier is applied to other pixels in the image that are unclassified or uncertain. Dynamic classification is automatically customized to the current image and can yield a more accurate and efficient material classification versus a static (image-independent) or manually trained classifier.

Abstract: A system and method for automatically (without human intervention) identifying a material in an image that comprises a building material for buildings in the image. Building side polygons which may be used to identify building sides in off-nadir imagery are generated. Off-nadir, multispectral images, building footprint data and elevation data for a geographic area are taken as input. Building heights for buildings in the geographic area are determined by clipping the elevation data using the building footprint data and then calculating building heights. A candidate set of polygons representing visible side faces of each building in the images is created from the known building heights, and based on the viewpoint direction, using vector analysis. After culling occluded polygons and polygons too small for analysis, the polygons are associated with a building footprint. Building materials for each building having visible polygons can then be identified.

Abstract: An automated system is provided for classifying materials in remotely-sensed imagery based on automated construction of a dynamic classifier—namely, a classifier that is automatically trained on the same image to which it is then subsequently applied. A first automated process identifies high confidence exemplars of each class using tailored classification techniques. This data is then used to train a supervised classification model (e.g., discriminant analysis), and the resultant classifier is applied to other pixels in the image that are unclassified or uncertain. Dynamic classification is automatically customized to the current image and can yield a more accurate and efficient material classification versus a static (image-independent) or manually trained classifier.

Abstract: Described herein are methods and systems for detecting clouds in satellite imagery captured using first and second sensor arrays that are carried by a satellite and physically offset from one another on the satellite. Movement mask data is produced based first image data and the second image data, obtained, respectively, using the first and second sensor arrays carried by the satellite. Cloud mask data is produced based on spectral information included in one of the first and second image data. Cloud detection data is produced based on the movement mask data and the cloud mask data, the cloud detection data indicating where it is likely, based on both the movement mask data and the cloud mask data, that one or more clouds are represented within one of the first and second image data. The cloud detection data can be used in various ways to account for the clouds included within the satellite imagery.

Abstract: Technology is provided for identifying concrete and/or asphalt (or other materials) in a multispectral satellite image that has multiple bands including a first set of bands from a first sensor and a second set of bands from a second sensor. The second sensor is at a different position on a focal plane as compared to the first sensor so that a single location depicted in the multispectral image will have been sensed at different times by the first sensor and the second sensor. The system identifies moving vehicles in the multispectral image and subsequently identifies sample pixels in the multispectral image that are near the moving vehicles. These pixels are high confidence samples of roads made of concrete and/or asphalt. Additional pixels are identified in the multispectral image having spectral characteristics that are within a threshold of spectral characteristics of the sample pixels. These additional pixels also depict concrete and/or asphalt.

Abstract: Described herein are methods and systems for detecting clouds in satellite imagery captured using first and second sensor arrays that are carried by a satellite and physically offset from one another on the satellite. Movement mask data is produced based first image data and the second image data, obtained, respectively, using the first and second sensor arrays carried by the satellite. Cloud mask data is produced based on spectral information included one of the first and second image data. Cloud detection data is produced based on the movement mask data and the cloud mask data, the cloud detection data indicating where it is likely, based on both the movement mask data and the cloud mask data, that one or more clouds are represented within one of the first and second image data. The cloud detection data can be used in various ways to account for the clouds included within the satellite imagery.