AU2021106697A4

AU2021106697A4 - A Superpixel Based Deep Neural Network System For Satellite Imagery

Info

Publication number: AU2021106697A4
Application number: AU2021106697A
Authority: AU
Inventors: K. V. Kale; Parminder KAUR
Original assignee: Individual
Current assignee: Individual
Priority date: 2021-08-24
Filing date: 2021-08-24
Publication date: 2022-01-06
Anticipated expiration: 2029-08-24

Abstract

Deep neural networks are compute intensive due to large number of layers, so in the proposed work, a deep convolutional neural network (DCNN) system with minimum number of layers is proposed. In this system, training images are annotated using ground truth data for a sample dataset and using this, an innovative algorithm is developed to generate the training dataset for training DCNN. Superpixel based algorithm is further extended for use by the DCNN to classify the remote sensed images. The algorithm reduces errors introduced in creation of dataset due to human intervention and higher accuracies are obtained for low spatial resolution images and hyper spectral images which is not reported in earlier solutions. 12 c :: I'd 4- C-

Description

c ::

I'd

4- C-

A Superpixel Based Deep Neural Network System For Satellite Imagery FIELD OF THE INVENTION

The present invention relates to remote sensed image classification, and in particularly relates to a superpixel based deep neural network system for satellite imagery.

BACKGROUND OF THE INVENTION

It has been observed that higher classification accuracies are obtained for large scale remote sensed images if the training input image is completely hand-labelled i.e. per-pixel. The hand-crafted dataset creation process is time consuming and error prone due to manual intervention. The existing solutions however do not focus on automating this process to train the deep neural network. In addition, the existing systems do not involve completely annotated training images for deep neural network that are recommended for higher classification accuracies. The existing systems lack in accuracy. In view thereof, there exist a need for a more efficient and accurate systems that works on a smaller set of training dataset to achieve higher accuracies.

SUMMARY OF THE INVENTION

The proposed system focuses on automating this process to train the deep neural network. Completely annotated training images for deep neural network are recommended for higher classification accuracies. The proposed system works on a smaller set of training dataset to achieve higher accuracies. Majority of the work carried out discusses that training DCNN is very compute intensive due to large number of layers present in the network. Deep neural networks are compute intensive due to large number of layers, so in the proposed work, a deep convolutional neural network (DCNN) system with minimum number of layers is proposed.

In the present invention, training images are annotated using ground truth data for a sample dataset and using this, and an innovative process algorithm is developed to generate the training dataset for training DCNN. Superpixel based algorithm is further extended for use by the DCNN to classify the remote sensed images. The algorithm reduces errors introduced in creation of dataset due to human intervention and higher accuracies are obtained for low spatial resolution images and hyper spectral images which is not reported in earlier work.

To further clarify advantages and features of the present disclosure, a more particular description of the invention will be rendered by reference to specific embodiments thereof, which is illustrated in the appended drawings. It is appreciated that these drawings depict only typical embodiments of the invention and are therefore not to be considered limiting of its scope. The invention will be described and explained with additional specificity and detail with the accompanying drawings.

BRIEF DESCRIPTION OF FIGURES

These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

Figure 1 illustrates a block diagram of a Superpixel based deep learning system in accordance with an embodiment of the present invention; Figure 2 illustrates a Superpixel based deep learning system in accordance with an embodiment of the present invention; and Figure 3 illustrates results are presented for the test image in accordance with an embodiment of the present invention.

Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help to improve understanding of aspects of the present disclosure. Furthermore, in terms of the construction of the device, one or more components of the device may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having benefit of the description herein.

DETAILED DESCRIPTION:

Figure 1 illustrates a block diagram of a Superpixel based deep learning system. The system 100 comprises an input unit 102 configured to receive a training image as an input; a first processing unit 104 which is used by the superpixel based algorithm to create image patch-based training dataset; a deep convolutional neural network 106, wherein the training dataset is further used by the optimized deep convolutional neural network 106, wherein the deep network has few convolutional layers to attain the desired classification accuracy; a second processing unit 108 wherein learning algorithm terminates and the network is saved with updated weight values and other hyper parameters of the deep network; a classification unit 110 based on trained deep network to classify the complete remote sensed image.

In an embodiment, the deep convolutional neural network makes use of the stochastic gradient descent learning method by using this training dataset for extracting the spatial and spectral features from the image patches.

In an embodiment, the image patch size can also be. increased and reduced as per the spatial resolution of the required application.

In an embodiment, the classified image is used for accuracy assessment using ground truth dataset.

The Superpixel based deep learning system developed is as shown in Figure 2. The remote sensed dataset used in this system is from Landsat-8, medium spatial resolution multispectral sensor and AVIRIS, low spatial resolution hyperspectral sensor. The superpixel based system uses dataset from hyperspectral sensor is the innovative concept introduced here. The algorithm works for all types of remote sensed data i.e. panchromatic, multispectral and hyperspectral sensors.

The input to the system is the training image, which is used by the superpixel based algorithm to create image patch-based training dataset. This training dataset is further used by the optimized deep convolutional neural network which is also another innovative attempt where the deep network has few convolutional layers to attain the desired classification accuracy. The deep convolutional neural network makes use of the stochastic gradient descent learning method by using this training dataset for extracting the spatial and spectral features from the image patches. As the network converges for desired parameter values, the learning algorithm terminates and the network is saved with updated weight values and other .hyper parameters of the deep network. This trained deep network is further used to classify the complete remote sensed image. The classified image is used for accuracy assessment using ground truth dataset. As deep convolutional neural networks are compute-intensive due to many operations performed in convolutional layers, the model proposed is having minimumnumber of layers compared to existing deep networks. The proposed system works for the dataset from hyperspectral, multispectral and panchromatic sensors. This can be compute-intensive for the hyperspectral dataset as there are a large number of spectral bands. This issue can be addressed by training the deep network on a computer system with GPU processors and increasing the RAM. The Superpixel based method is an innovation where the image patch size can also be. increased and reduced as per the spatial resolution of the required application. This system is useful for agriculture land area monitoring, where agencies can use the generated output to make decisions. This system can also be useful for insurance companies related to crop yield and crop production studies. The results are presented in figure 23for the test image. The results show higher User's Accuracy and Producer's Accuracy for Sugarcane crop which we are identifying for two stages i.e. at harvest stage (Crop-Hv) 95.2 %, 98 % respectively and at crop grown (Crop-Gr) 98.6%, 99% stage. The water body also shows higher accuracies with values of UA and PA as 94.7%, 99.7% followed by land cover, settlement with values of 75.4%, 93.2%. The crop land area under sugarcane crop is identified with very high accuracy value. This work can be used by government or private agencies who have to make decision regarding agricultural land monitoring as this system produces classified image as output. The classified image can be used to make estimates of land area under crop cultivation, man-made structures, water bodies, vegetation etc. This invention can also be used by insurance agencies for making estimates regarding insurance calculations based on crop land area and crop yield.

In an embodiment, the Superpixel based deep learning system comprises a remote sensed dataset. The Landsat-8, medium spatial resolution multispectral sensor and AVIRIS, low spatial resolution hyperspectral sensor. The superpixel based system uses dataset from hyperspectral sensor. The input to the system is the training image, which is used by the superpixel based algorithm to create image patch-based training dataset.

In an embodiment, the training dataset is further used by the optimized deep convolutional neural network which is also another innovative attempt where the deep network has few convolutional layers to attain the desired classification accuracy.

In an embodiment, the network converges for desired parameter values, the learning algorithm terminates and the network is saved with updated weight values and other hyper parameters of the deep network. In an embodiment, the trained deep network is further used to classify the complete remote sensed image.

In an embodiment, the image patch size can also be increased and reduced as per the spatial resolution of the required application which is useful for agriculture land area monitoring where agencies can use the output generated to make decision regarding this also useful for insurance companies related to crop yield and crop production studies.

For the purpose of promoting an understanding of the principles of the invention, reference will now be made to the embodiment illustrated in the drawings and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention as illustrated therein being contemplated as would normally occur to one skilled in the art to which the invention relates.

It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be restrictive thereof.

Reference throughout this specification to "an aspect", "another aspect" or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Thus, appearances of the phrase "in an embodiment", "in another embodiment" and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or elements or structures or components proceeded by "comprises...a" does not, without more constraints, preclude the existence of other devices or other sub systems or other elements or other structures or other components or additional devices or additional sub-systems or additional elements or additional structures or additional components.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The system, methods, and examples provided herein are illustrative only and not intended to be limiting. Embodiments of the present disclosure will be described below in detail with reference to the accompanying drawings.

The scope of embodiments is by no means limited by these specific examples. Numerous variations, whether explicitly given in the specification or not, such as differences in structure, dimension, and use of material, are possible. The scope of embodiments is at least as broad as given by the following claims. Benefits, other advantages, and solutions to problems have been described above with regard to specific embodiments. However, the benefits, advantages, solutions to problems, and any component(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential feature or component of any or all the claims.

Claims

WE CLAIM: 1. A Superpixel based deep learning system, said system comprising: an input unit configured to receive a training image as an input; a processing unit which is used by the superpixel based algorithm to create image patch-based training dataset;

a deep convolutional neural network, wherein the training dataset is further used by the optimized deep convolutional neural network, wherein the deep network has few convolutional layers to attain the desired classification accuracy;

a processing unit wherein learning algorithm terminates and the network is saved with updated weight values and other hyper parameters of the deep network;

a classification unit based on trained deep network to classify the complete remote sensed image.
2. The system as claimed in claim 1, wherein the deep convolutional neural network makes use of the stochastic gradient descent learning method by using this training dataset for extracting the spatial and spectral features from the image patches.
3. The system as claimed in claim 1, wherein the image patch size can also be. increased and reduced as per the spatial resolution of the required application.
4. The system as claimed in claim 1, wherein deep convolutional neural network makes use of the stochastic gradient descent learning method by using this training dataset for extracting the spatial and spectral features from the image patches.
5. The system as claimed in claim 1, wherein the classified image is used for accuracy assessment using ground truth dataset.

input unit 102 first processing unit 104

deep convolutional neural second processing unit 108 network 106

classification unit 110

Figure 1

Figure 2