US 12,749,304 B1Granted

Method and system for aerial land recognition based on remote sensing images

AssigneePrince Mohammad Bin Fahd University

InventorsMuhammad Attique Khan, Faisal Yousif Al Anezi

Life of this patent

Estimating the term…

Loading patent…

Abstract

A method and system for land use classification includes capturing remote sensing (RS) images by sensors on various remote sensing devices. The system includes a satellite constellation and an image database system. The database system stores classified remote sensing images in the image database system by downstream application. The system is provided with an input interface and built-in database functions for performing analysis on remote-sensing images. The method includes receiving the RS images by server communication circuitry and storing the RS images in memory. The method includes iteratively training a deep learning network that integrates an input layer, a residual block, inception blocks, bottleneck blocks and inverted bottleneck blocks, dense blocks, having integrated squeeze-excitation blocks. The method includes inputting the RS images, preventing a training gradient from vanishing, extracting features at multiple scales, balancing computational cost, and promoting reuse of the extracted features.

Claims20

Dependent claims

A system for analyzing land use and land cover (LULC), comprising:

network communication circuitry configured to receive a plurality of remote sensing (RS) images from multiple sources, wherein each RS image includes a geographic region, a GPS location and a timestamp,
a memory configured to store the plurality of RS images,
processing circuitry configured to iteratively train a squeeze-excitation multiscale fusion network, including:
input, to an input layer, the plurality of RS images from the memory,
maintain, by at least one residual block, a training gradient by integrating a shortcut connection that skips one or more layers in the at least one residual block,
extract, by a plurality of inception blocks, where each inception block includes parallel layers with different kernel sizes, features at multiple scales,
enhance the extracted features, by a dense block incorporated after a bottleneck residual block, wherein the dense block includes a plurality of convolutional layers, a concatenation layer that merges feature of a current layer with subsequent layers, and a squeeze and excitation block, and
determine, by a classification layer, at least one class label and a respective confidence score for the at least one class label based on the enhanced extracted features for a RS image of the plurality of RS images
store, in an image database, the plurality of remote sensing images with respective class label;
a system interface configured to
input a selection of whether general or specific LULC images are of interest;
when general is selected,
input a date and time,
input whether count of number of each class is selected,
input whether area of each class is selected,
input whether there is a change in class for a location is selected,
when specific LULC is selected,
input class name,
input a region of interest,
input a location,
input whether change detection is selected,
a database compute component configured to
determine a count of images in each respective class for a date/time,
determine a total area of the images in each respective class for a data/time,
determine a change in an image for a region,
determine a change in an image for a location.

23 more elements

A non-transitory computer-readable storage medium including computer executable instructions, wherein the instructions, when executed by a computer, cause the computer to perform a method for analyzing land use and land cover (LULC), the method comprising:

receiving, by network communication circuitry, a plurality of remote sensing (RS) images from multiple sources, wherein each RS image includes a geographic region, a GPS location and a timestamp;
storing, by a memory, the plurality of RS images;
iteratively training, by processing circuitry, a squeeze-excitation multiscale fusion network including:
inputting, to an input layer, the plurality of RS images from the memory,
maintaining, by at least one residual block, a training gradient by integrating a shortcut connection that skips one or more layers in the at least one residual block,
extracting, by a plurality of inception blocks, where each inception block includes parallel layers with different kernel sizes, features at multiple scales,
enhancing the extracted features, by a dense block incorporated after a bottleneck residual block, wherein the dense block includes a plurality of convolutional layers, a concatenation layer that merges feature of a current layer with subsequent layers, and a squeeze and excitation block,
determining, by a classification layer, at least one class label and a respective confidence score for the at least one class label based on the enhanced extracted features for a RS image of the plurality of RS images;
storing, in an image database, the plurality of remote sensing images with respective class label;
inputting, by a system interface, a selection of whether general or specific LULC images are of interest;
when general is selected,
inputting, a date and time;
inputting whether count of number of each class is selected;
inputting whether area of each class is selected;
inputting whether there is a change in class for a location is selected;
when specific LULC is selected,
inputting class name;
inputting a region of interest;
inputting a location;
inputting whether change detection is selected;
determining, by a database compute component, a count of images in each respective class for a date/time;
determining, by the database compute component, a total area of the images in each respective class for a data/time;
determining, by the database compute component, a change in an image for a region;
and
determining, by the database compute component, a change in an image for a location.

22 more elements

Dependencies are inferred from the claim text.