Method and system for learning spectral features of hyperspectral data using DCNN
The embodiments herein provide a method and system that analyzes the pixel vectors by transforming the pixel vector into two-dimensional spectral shape space and then perform convolution over the image of graph thus formed. Method and system disclosed converts the pixel vector into image and provides a DCNN architecture that is built for processing 2D visual representation of the pixel vectors to learn spectral and classify the pixels. Thus, DCNN learn edges, arcs, arcs segments and the other shape features of the spectrum. Thus, the method disclosed enables converting a spectral signature to a shape, and then this shape is decomposed using hierarchical features learned at different convolution layers of the disclosed DCNN at different levels.
1. A processor implemented method for learning spectral features from hyperspectral data, the method comprising:
receiving, via one or more hardware processors, a hyperspectral imagery of a region of interest (ROI), wherein the hyperspectral data corresponding to the hyperspectral imagery comprises a plurality of pixel vectors corresponding to each pixel of the plurality of pixel vectors in the hyperspectral imagery;
generating, via the one or more hardware processors, a plurality of Two Dimensional (2D) representations for the plurality of pixel vectors, wherein for every pixel vector the 2D representation comprises values of each pixel in the pixel vector plotted against a corresponding spectrum in the hyperspectral imagery to generate a 2D graph, wherein each 2D graph captures spectral features of a corresponding pixel vector among the plurality of pixel vectors;
generating, via the one or more hardware processors, a plurality of 2D graph images from visual representations obtained from the plurality of 2D representations;
processing the plurality of 2D graph images by a Deep Convolutional Neural Network (DCNN) model executed by the one or more hardware processors, to learn the spectral features from a plurality of activation points in each of the plurality of 2D graphs, wherein the plurality of activation points comprise edges and arcs, that represent reflectance variations in the ROI, and
converting via the one or more hardware processors, a spectral signature to a shape, wherein the shape is decomposed using hierarchical features learned at a plurality of convolutional layers of the DCNN model at different levels.
2. The method as claimed in claim 1 , wherein the method further comprising:
utilizing the DCNN model to extract current spectral features of a current hyperspectral data of a second ROI by generating a second plurality of 2D graph images; and
utilizing the DCNN model by classifying the plurality of 2D graph images generated from the current hyperspectral data of the second ROI into at least one class among a plurality of predefined classes in accordance with the extracted current spectral features.
3. The method as claimed in claim 2 , wherein the DCNN model comprises:
a plurality of convolutions blocks
wherein each of the plurality of convolution blocks comprises the plurality of convolution layers with each of the plurality of convolution layers followed by a batch normalization layer,
wherein each of the plurality of convolution layers is kernel size and corresponding stride of a last convolution layer among the plurality of convolution layers is greater than the kernel size and the corresponding stride of the remaining convolution layers, wherein a value of the stride is designed to enable a kernel to sample 2D graph images without overlapping, and
wherein the last convolution layer among the plurality of convolution layers is followed by a dropout layer; and
a first fully connected layer and a second fully connected layer cascaded to a last convolution block, of the plurality of convolution blocks, wherein the second fully connected layer is a sparse layer.
4. The method as claimed in claim 2 , wherein the DCNN model is a capsule network comprising:
a convolution layer followed by a cascaded first capsule layer the cascaded first capsule layer to be followed by a second capsule layer,
wherein the convolution layer is followed by a RELU activation layer and detects low level features in each of the plurality of 2D graph images,
wherein the first capsule layer comprises a first set of capsules which processes the low level features detected by the convolution layer to produce combination of the low level features and generate a set of first feature vectors,
wherein the second capsule layer comprises a second set of capsules equal to a number of a plurality of predefined classes and processes the set of first features to vector to classify each of the plurality of 2D graph images into one of the plurality of predefined classes.
5. The method as claimed in claim 4 , wherein the capsule network utilizes a squash activation function and is followed by a decoder that manages reconstruction loss.
6. A system for learning spectral features from hyperspectral data, the system comprising:
a memory storing instructions;
one or more Input/Output (I/O) interfaces; and
one or more hardware processors coupled to the memory via the one or more I/O interfaces, wherein the one or more hardware processors are configured by the instructions to:
receive a hyperspectral imagery of a region of interest (ROI), wherein the hyperspectral data corresponding to the hyperspectral imagery comprises a plurality of pixel vectors corresponding to each pixel of the plurality of pixel vectors in the hyperspectral imagery;
generate a plurality of Two Dimensional (2D) representations for the plurality of pixel vectors, wherein for every pixel vector the 2D representation comprises values of each pixel in the pixel vector plotted against a corresponding spectrum in the hyperspectral imagery to generate a 2D graph, wherein each 2D graph captures spectral features of a corresponding pixel vector among the plurality of pixel vectors;
generate a plurality of 2D graph images from visual representations obtained from the plurality of 2D representations;
process the plurality of 2D graph images to via a DCNN model to learn the spectral features from a plurality of activation points in each of the plurality of 2D graphs, wherein the plurality of activation points comprise edges and arcs that represent reflectance variations in the ROI, and
converting via the one or more hardware processors, a spectral signature to a shape, wherein the shape is decomposed using hierarchical features learned at a plurality of convolutional layers of the DCNN model at different levels.
7. The system as claimed in claim 6 , wherein the one or more hardware processors are configured to:
utilize the DCNN model, to extract spectral current spectral features of a current hyperspectral data of a second ROI by generating a second plurality of 2D graph images; and
utilize the DCNN model by classifying the plurality of 2D graph images generated from the current hyperspectral data of the second ROI into at least one class among a plurality of predefined classes in accordance with the extracted current spectral features.
8. The system as claimed in claim 7 , wherein the DCNN model comprises:
a plurality of convolutions blocks,
wherein each of the plurality of convolution blocks comprises the plurality of convolution layers with each of the plurality of convolution layers followed by a batch normalization layer,
wherein each of the plurality of convolutional layers is kernel size and corresponding stride of a last convolution layer among the plurality of convolution layers is greater than the kernel size and the corresponding stride of the remaining convolution layers, wherein a value of the stride is designed to enable a kernel to sample 2D graph images without overlapping, and
wherein the last convolution layer among the plurality of convolution layers is followed by a dropout layer; and
a first fully connected layer and a second fully connected layer cascaded to a last convolution block, of the plurality of convolution blocks, wherein the second fully connected layer is a sparse layer.
9. The system as claimed in claim 7 , wherein the DCNN model is a capsule network comprising:
a convolution layer followed by a cascaded first capsule layer and the cascaded first capsule layer followed by a second capsule layer,
wherein the convolution layer is followed by a RELU activation layer and detects low level features in each of the plurality of 2D graph images,
wherein the first capsule layer comprises a first set of capsules which processes the low level features detected by the convolution layer to produce combination of the low level features and generate a set of first feature vectors,
and wherein the second capsule layer comprises a second set of capsules equal to a number of a plurality of predefined classes and processes the set of first vectors features to classify each of the plurality of 2D graph images into one of the plurality of predefined classes.
10. The system as claimed in claim 9 , wherein the capsule network utilizes a squash activation function and is followed by a decoder that manages reconstruction loss.
11. One or more non-transitory machine readable information storage mediums comprising one or more instructions, which when executed by one or more hardware processors causes:
receiving a hyperspectral imagery of a region of interest (ROI), wherein the hyperspectral data corresponding to the hyperspectral imagery comprises a plurality of pixel vectors corresponding to each pixel of the plurality of pixel vectors in the hyperspectral imagery;
generating, a plurality of Two Dimensional (2D) representations for the plurality of pixel vectors, wherein for every pixel vector the 2D representation comprises values of each pixel in the pixel vector plotted against a corresponding spectrum in the hyperspectral imagery to generate a 2D graph, wherein each 2D graph captures spectral features of a corresponding pixel vector among the plurality of pixel vectors;
generating, a plurality of 2D graph images from visual representations obtained from the plurality of 2D representations;
processing the plurality of 2D graph images by a Deep Convolutional Neural Network (DCNN) model executed, to learn the spectral features from a plurality of activation points in each of the plurality of 2D graphs, wherein the plurality of activation points comprise edges and arcs that represent reflectance variations in the ROI; and
converting via the one or more hardware processors, a spectral signature to a shape, wherein the shape is decomposed using hierarchical features learned at a plurality of convolutional layers of the DCNN model at different levels.