Description:DESCRIPTION:
Field of the invention:
[0001] The present disclosure generally relates to the technical field of an image processing technology, in specific, relates to a system and method for classifying and extracting spectral-spatial features of the hyper spectral image using a two-stage hybrid model.
Background of the invention:
[0002] Hyperspectral imagery (HSI), a high-resolution imaging technique that uses high-dimensional image cubes with numerous spectral bands, produces exceptionally accurate and detailed information about the surface materials and composition of the imaged scene. The correct prediction of the pixel values associated with different classes present in an image is the basis of HSI analysis, which is a broad field of study encompassing many applications such as image segmentation, object recognition, anomaly detection, and land cover classification. HSI is also used for greenery detection, environment analysis, crop analysis, and many others.

[0003] Hyperspectral images offer valuable spectral data and strong spatial correlations that benefit users in extracting distinctive features for land use and cover classification. Nevertheless, the excessive correlation among closely spaced spectral bands can introduce redundant information, leading to suboptimal classification results. To address this challenge in hyper-spectral image classification, numerous techniques have been devised, focusing on feature extraction and classification methods. A feature extraction technique transforms the initial feature space using nonlinear operations. Its primary goals include reducing the high-dimensional nature of hyper-spectral image pixels and extracting the most distinctive features or bands.

[0004] Hyperspectral imaging (HSI) is a technique that analyzes a wide spectrum of light instead of just assigning primary colors (red, green, and blue) to each pixel. The light striking each pixel is broken down into many different spectral bands in order to provide more information on what is being imaged.

[0005] Hyperspectral spectral Image classification is a task in the fields of remote sensing and computer vision. It involves the classification of pixels in hyperspectral images into different classes based on their spectral signature. Hyperspectral images contain information about the reflectance of objects in hundreds of narrow, contiguous wavelength bands, making them useful for a wide range of applications, including mineral mapping, vegetation analysis, and urban land-use mapping. The goal of this task is to accurately identify and classify different types of objects in the image, such as soil, vegetation, water, and buildings, based on their spectral properties.

[0006] Numerous studies have explored combinations of feature extraction (FE) methods with HSI classification algorithms. In existing approaches, several FE techniques have been developed, combining spectral and spatial data to capture the comprehensive characteristics of HSI data. The work employs the notion of spectral gradients for HSI classification. It begins by using the random forest algorithm to extract spatial features and then applies spectral gradients to obtain spectral features. Subsequently, a multi-scale fusion process combines these spatial-spectral features, enabling support vector machines (SVM) for classification. There are various existing approaches using spectral-spatial feature mining techniques for HSI classification that are available. Examples include simple linear iterative clustering (SLIC), the Gabor filter, extended morphological profiles (EMPs), and multiple kernel learning, which offer spectral and spatial feature-based classification frameworks.

[0007] Additionally, diverse deep learning-based methods are introduced to extract valuable spectral-spatial features for HSI classification. Following feature extraction, discriminative features are employed to train general-purpose classifiers. The existing approach contains numerous models, such as random forests, decision trees, and support vector machines (SVMs), for the classification phase. The existing approaches advocate the use of deep learning techniques, particularly convolutional neural networks (CNNs), for image data classification. The HSI classification involves various CNN models in addition to traditional feature-based approaches. Many of these models are rooted in 2D CNNs and 3D CNNs. Both independent models have shown good performance.

[0008] Fused-squeeze and excitation network (FuSENet) is a model designed for HSI classification. There are also attempts using Recurrent Neural Networks (RNN), Generative Adversarial Networks (GAN), Graph CNNs, and the Squeeze and Excitation Residual Network. RNNs, for instance, treat the spectral signature of the HSI as a sequence to extract discriminative features. While 3D-2D CNNs aim to capture spatial and spectral features from HSI data cubes, their performance across multiple datasets appears constrained. Additionally, 3D CNNs demand more computational resources compared to 2D CNNs. However, the existing techniques for HSI classification offer a number of limitations in providing better accuracy and efficiency in extracting spatial-spectral features. Therefore, there is a preference for methods that rely solely on 2D CNNs while retaining the ability to extract both spatial and spectral features. Over the last two decades, the wavelet transform has found extensive use across various domains, including image classification, computer vision, texture classification, and remote sensing. Notably, in the context of hyper-spectral image classification, researchers have introduced feature extraction techniques relying on the wavelet transform.

[0009] By addressing all the above mentioned problems, there is also a need for a system for classifying and extracting spectral-spatial features of the hyperspectral image using a two-stage hybrid model. There is also a need for a system that provides multi-resolution analysis by using wavelet transform (WT) and a two-dimensional convolutional neural network (2D-CNN) for capturing information at different spatial and spectral scales. There is also a need for a system that combines the discrimination capabilities of 2D-CNN with the signal analysis proficiency of wavelets in both time and frequency domains. There is also a need for a system that uses wavelet neural networks (WNN) for modeling complex non-linear relationships within data, potentially leading to improved classification performance compared to linear models. There is also a need for a system that offers better generalizability to unseen data due to the combined strengths of WT's multi-scale learning and WNN's non-linear modeling capability.
Objectives of the invention:
[0010] The primary objective of the present invention is to provide a system for classifying and extracting spectral-spatial features of the hyperspectral image using a two-stage hybrid model.

[0011] Another objective of the present invention is to provide a system that provides multi-resolution analysis by using wavelet transform (WT) and a two-dimensional convolutional neural network (2D-CNN) for capturing information at different spatial and spectral scales.

[0012] Yet another objective of the present invention is to provide a system that combines the discrimination capabilities of 2D-CNN with the signal analysis proficiency of wavelets in both time and frequency domains.

[0013] Another objective of the present invention is to provide a system that uses a wavelet neural network (WNN) for modeling complex non-linear relationships within HSI data, potentially leading to improved classification performance compared to linear models.

[0014] Further objective of the present invention is to provide a system that offers better generalizability to unseen data due to the combined strengths of WT's multi-scale learning and WNN's non-linear modeling capability.
Summary of the invention:
[0015] The present disclosure proposes a system for classifying and extracting spectral-spatial features of the hyperspectral image. The following presents a simplified summary in order to provide a basic understanding of some aspects of the claimed subject matter. This summary is not an extensive overview. It is not intended to identify key/critical elements or to delineate the scope of the claimed subject matter. Its sole purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is presented later.

[0016] In order to overcome the above deficiencies of the prior art, the present disclosure is to solve the technical problem to provide a system for classifying and extracting spectral-spatial features of the hyperspectral image using a two-stage hybrid model.

[0017] According to one aspect, the invention provides a system for classifying and extracting spectral-spatial features of the hyperspectral image. The system comprises a computing device having a processor and a memory for storing one or more instructions executable by the processor. The computing device is in communication with a server via a network. In one embodiment, the computing device includes a laptop, a computer, a smartphone, and other electronic devices.

[0018] In one embodiment, the processor is configured to control plurality of modules to execute image feature extraction and image classification functions. In one embodiment, the plurality of modules comprises an input module, an analyzing module, an extraction module, a classification module and a communication module.

[0019] In one embodiment, the input module is configured to extract and store one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI). In one embodiment herein, the at least one hyper spectral image (HSI) is pre-processed using a factor analysis method to reduce dimensionality.

[0020] In one embodiment, the analyzing module is configured to analyze the one or more HSI patches, thereby generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique.

[0021] In one embodiment, the extraction module is configured to extract one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using a two-dimensional convolutional neural network (CNN) model. The two-dimensional convolutional neural network (CNN) model is a discrete wavelet two-dimensional convolutional neural network (CNN). In one embodiment herein, the discrete wavelet two-dimensional convolutional neural network (CNN) uses a daubechies 4-tap (D4) orthogonal filter to extract one or more spectral-spatial features.

[0022] In one embodiment, the classification module is configured to receive the extracted one or more spectral-spatial features and classify the at least one HSI patch using a classification neural network model. The classification neural network model is a wavelet neural network (WNN) model.

[0023] In one embodiment, the communication module is configured to enable communication between the computing device and the server.

[0024] In one embodiment, the system uses plurality of datasets to encompass a range of HSI data, including agricultural, rural-urban, and urban scenes. The plurality of datasets includes Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset.

[0025] In one embodiment, the system explores plurality of mother wavelets, which include Haar wavelet, Daubechies wavelet, Mexican hat wavelet, and Morlet wavelet to improve the discriminative capabilities of the two-stage hybrid model for HSI classification.

[0026] According to another aspect, the invention provides a method for extracting and classifying a hyper spectral image using a two-stage hybrid model through a system. At one step, the input module extracts the one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI).

[0027] At other step, the analyzing module analyzes the one or more HSI patches for generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique.

[0028] At another step, the extraction module extracts one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using a discrete wavelet two-dimensional convolutional neural network (CNN).

[0029] Further at other step, the classification module receives the extracted one or more spectral-spatial features and classify the at least one HSI patch using a wavelet neural network (WNN).

[0030] Further, objects and advantages of the present invention will be apparent from a study of the following portion of the specification, the claims, and the attached drawings.
Detailed description of drawings:
[0031] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate an embodiment of the invention, and, together with the description, explain the principles of the invention.

[0032] FIG. 1 illustrates a block diagram of a system for extracting and classifying a hyper spectral image using a two-stage hybrid model, in accordance to an exemplary embodiment of the invention.

[0033] FIG. 2A illustrates a graph representing a Haar wavelet of the system for hyper spectral image (HSI) classification, in accordance to an exemplary embodiment of the invention.

[0034] FIG. 2B illustrates a graph representing a Morlet wavelet of the system for hyper spectral image (HSI) classification, in accordance to an exemplary embodiment of the invention.

[0035] FIG. 2C illustrates a graph representing a Daubechies 4-tap (D4) wavelet of the system for hyper spectral image (HSI) classification, in accordance to an exemplary embodiment of the invention.

[0036] FIG. 2D illustrates a graph representing a Mexican Hat wavelet of the system for hyper spectral image (HSI) classification, in accordance to an exemplary embodiment of the invention.

[0037] FIG. 3 illustrates a structural diagram of a two-stage hybrid model for spectral-spatial classification of hyper spectral image (HSI), in accordance to an exemplary embodiment of the invention.

[0038] FIG. 4 illustrates a network diagram of the two-stage hybrid model for spectral-spatial classification of hyper spectral image (HSI), in accordance to an exemplary embodiment of the invention.

[0039] FIGs. 5A-5C illustrate graphs between overall accuracies and the number of training samples for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset, in accordance to an exemplary embodiment of the invention.

[0040] FIG. 6 illustrates a graph representing kappa coefficients with 10% training data for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset, in accordance to an exemplary embodiment of the invention.

[0041] FIGs. 7A-7C illustrate graphs representing the comparison between the two-stage hybrid model and the state of the existing art methods by considering the kappa coefficients and the number of training samples for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset, in accordance to an exemplary embodiment of the invention.

[0042] FIGs. 8A-8F illustrate visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the IP dataset, in accordance to an exemplary embodiment of the invention.

[0043] FIGs. 9A-9F illustrate visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the PU dataset, in accordance to an exemplary embodiment of the invention.

[0044] FIGs. 10A-10F illustrate visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the SA dataset, in accordance to an exemplary embodiment of the invention.

[0045] FIGs. 11A-11F illustrate visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the IP dataset, in accordance to an exemplary embodiment of the invention.

[0046] FIGs. 12A-12F illustrate visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the PU dataset, in accordance to an exemplary embodiment of the invention.

[0047] FIGs. 13A-13F illustrate visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the SA dataset, in accordance to an exemplary embodiment of the invention.

[0048] FIG. 14 illustrates a flowchart of a method for extracting and classifying a hyper spectral image using a two-stage hybrid model through a system, in accordance to an exemplary embodiment of the invention.
Detailed invention disclosure:
[0049] Various embodiments of the present invention will be described in reference to the accompanying drawings. Wherever possible, same or similar reference numerals are used in the drawings and the description to refer to the same or like parts or steps.

[0050] The present disclosure has been made with a view towards solving the problem with the prior art described above, and it is an object of the present invention to provide a system and method for classifying and extracting spectral-spatial features of the hyper spectral image using a two-stage hybrid model.

[0051] According to one exemplary embodiment of the invention, FIG. 1 refers to a block diagram of a system 100 for extracting and classifying a hyper spectral image using a two-stage hybrid model. In one embodiment herein, the system 100 comprises a computing device 102 having a processor 104 and a memory 106 for storing one or more instructions executable by the processor 104. The computing device 102 is in communication with a server 124 via a network 122.

[0052] In one embodiment herein, the computing device 102 is used generally herein to refer to any computing device configured to perform operations of the various embodiments, including one or all of personal computers, cellular telephones, smart phones, personal data assistants (PDA's) laptop computers, tablet computers, smart books, palm-top computers, wireless electronic mail receivers, multimedia Internet enabled cellular telephones, and similar personal electronic devices. In one embodiment herein, the server 124 is configured to store data related to hyper spectral images.

[0053] In another embodiment herein, the network 122 could be, but is not limited to, Wi-Fi, Bluetooth, a wireless local area network (WLAN) /Internet connection, and radio communication. In some embodiments, the computing device 102 could be a touchscreen and/or a non-touchscreen and adopted to run on any type of OS, such as iOS, Windows, Android, Unix, Linux and/or others. In an embodiment herein, the server 124 is at least one of a general or special purpose computer or a server. The server 124 could be operated as a single computer, which can be a hardware and/or software server, a workstation, a desktop, a laptop, a tablet, a mobile phone, a mainframe, a supercomputer, a server farm, and so forth.

[0054] In one embodiment herein, the processor 104 is configured control plurality of modules 108 to execute image feature extraction and image classification functions. In one embodiment herein, the plurality of modules 108 comprises an input module 110, an analyzing module 112, an extraction module 114, a classification module 116 and a communication module 118.

[0055] In one embodiment herein, the input module 110 is configured to extract and store one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI). In one embodiment herein, the at least one hyper spectral image (HSI) is pre-processed using a factor analysis method to reduce dimensionality. The system 100 comprises a database 120 to store data related to hyper spectral images (HSI).

[0056] In one embodiment herein, the analyzing module 112 is configured to analyze the one or more HSI patches, thereby generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique.

[0057] In one embodiment herein, the extraction module 114 is configured to extract one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using discrete wavelet two-dimensional convolutional neural network (CNN). In one embodiment herein, the discrete wavelet two-dimensional convolutional neural network (CNN) uses a daubechies 4-tap (D4) orthogonal filter to extract one or more spectral-spatial features.

[0058] In one embodiment herein, the classification module 116 is configured to receive the extracted one or more spectral-spatial features and classify the at least one HSI patch using a wavelet neural network (WNN).

[0059] In one embodiment herein, the communication module 118 is configured to enable communication between the computing device 102 and the server 124.

[0060] In one embodiment herein, the system 100 uses plurality of datasets to encompass a range of HSI data, including agricultural, rural-urban, and urban scenes. The plurality of datasets includes Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset.

[0061] In one embodiment herein, the system 100 explores plurality of mother wavelets, which include Haar wavelet, Daubechies wavelet, Mexican hat wavelet, and Morlet wavelet to improve the discriminative capabilities of the two-stage hybrid model for HSI classification.

[0062] According to another exemplary embodiment of the invention, FIG. 2A refers to a graph 200 representing a Haar wavelet of the system for hyper spectral image (HSI) classification. The Haar wavelet is characterized by a piecewise constant waveform with abrupt transitions. Because of its simplicity and convenience, the Haar wavelet offers a straightforward and computationally efficient method for examining the local characteristics of a signal. As a part of the discrete wavelet transform (DWT), Haar wavelets are particularly suitable for analysing signals with sudden and sharp transitions.

[0063] According to another exemplary embodiment of the invention, FIG. 2B refers to a graph 202 representing a Morlet wavelet of the system for hyper spectral image (HSI) classification. In one embodiment herein, the system 100 focuses on exploring plurality of mother wavelets, which include Haar wavelet, Daubechies wavelet, Mexican hat wavelet, and Morlet wavelet to improve the discriminative capabilities of the two-stage hybrid model for HSI classification. The mother wavelets are used to improve the performance of two-dimensional convolutional neural network (CNN).

[0064] The Morlet wavelet is a popular choice for examining time-frequency features in signals with oscillations. In image processing, it finds application in tasks like edge detection, texture analysis, and feature extraction. Various test signals, including noise, phase shift, bump, and a minor spike, are examined to evaluate the Morlet wavelet's performance under different parameter settings. The Morlet wavelet demonstrates a commendable equilibrium between time and frequency localization.

[0065] According to another exemplary embodiment of the invention, FIG. 2C refers to a graph 204 representing a Daubechies 4-tap (D4) wavelet of the system for hyper spectral image (HSI) classification. In one embodiment herein, the daubechies wavelets have most notable advantage in their use of a finite impulse response conjugate mirror filter (FIR-CMF). In FIR-CMF, the filter bank design technique revolves around creating a set of filters that divide the signal into multiple frequency sub-bands, enabling effective processing across various frequency ranges. This results in perfect reconstruction, minimal delay, and efficient multi-resolution analysis. As orthogonal wavelets, the daubechies wavelets exhibit strong time-frequency localization properties. They possess superior regularity and finite support, meaning they are non-zero only within a limited interval. This attribute enables the efficient and localized representation of signals and images.

[0066] According to another exemplary embodiment of the invention, FIG. 2D refers to a graph 206 representing a Mexican Hat wavelet of the system for hyper spectral image (HSI) classification. The Mexican Hat wavelet is frequently used in image processing applications, including tasks such as edge detection, feature extraction, and the identification of specific-sized features in an image (such as blobs or circular structures). This is because the Mexican Hat wavelet possesses the capacity to pinpoint information in both the time and frequency domains. It is also used in signal processing and wavelet analysis to highlight edges and features in the signal.

[0067] According to another exemplary embodiment of the invention, FIG. 3 refers to a structural diagram 300 of a two-stage hybrid model for spectral-spatial classification of hyper spectral image (HSI). In the two-stage hybrid model, the initial stage, which is the feature extraction stage, includes an input layer, a lambda layer (used for wavelet transform), five convolution layers, one pooling layer, and a flatten layer. At step 302, the discrete wavelet transform base has been constructed by adopting the lifting scheme based model. The lifting scheme offers faster computation and consumes less memory. It preserves wavelet properties while addressing the initial constraints. Next at step 304, the operation of a convolution layer in a CNN for extracting spatial features can be expressed mathematically as an equation Y = X*W. In this equation, X represents an input image tensor, Y represents the output labels and W represents a set of learnable filters/kernels. In this setup, each filter is equipped with its unique set of weights designed to capture distinct features from the input. In the proposed invention a 3x3 size filters/kernels are considered, and the five convolution layers employ a varying number of filters: 64, 128, 256, 256, and 128, respectively.

[0068] Additionally, a Batch Normalization layer is employed after each convolutional layer and before the ReLU activation function. Its primary role is to normalize the inputs for each layer, which helps stabilize and expedite the training process. At step 306, the two dimensional CNN hierarchically decomposes the input vector with different kernels, a concatenation layer is used to concatenate the input feature maps from different branches of the network along the channel axis. This concatenated tensor used as an input for subsequent layers. At step 308, the average pooling technique is applied in the pooling layer to filter the data and decrease the number of outputs. At step 310, the flatten layer is employed to generate a set of 128 spectral-spatial features. Finally, at step 312 the spectral-spatial features are classified using the wavelet neural network (WNN).

[0069] According to another exemplary embodiment of the invention, FIG. 4 refers to a network diagram 400 of the two-stage hybrid model for spectral-spatial classification of hyper spectral image (HSI). In one embodiment herein, FIG. 4 refers depicts the WNN module with Xj (j = 1,2,…,128) forming the input layer, ?k (k= 1,2, ….,L) comprising the hidden layer with wavelet basis functions, and yo,(o= 1,2,…,N) serving as the output layer. The weight signifies the connection strength between the jth neuron in the input layer and the kth neuron in the hidden layer, while denotes the connection between the kth neuron in the hidden layer and the oth neuron in the output layer. The output layer employs the softmax function for the final classification. Alongside these layers, we've implemented Dropout and L2 regularization techniques to prevent over fitting.

[0070] Generally, classifying the HSI images involves extracting spectral-spatial features from each HSI image, constructing a deep learning-based classifier and training the classifier, and assessing the classifier's classification efficiency. In the HSI feature extraction phase the DWT is employed to capture spectral features and incorporated the CNN module to extract spatial features. The resulting spectral-spatial feature vector is fed into a WNN for the classification task.

[0071] Initially, one or more factor analysis techniques is used to reduce the dimensionality of the HSI patch/cube, transforming the HSI patch/cube into a 2D format suitable for further processing. Following this, we extract (M × M) sized patches from the 2D image, which serve as the data source during the training of the two-stage hybrid model.

[0072] At stage-1, when a HSI patch X, with an MxM window size, undergoes the discrete wavelet transformation, it breaks down the input into sub-bands. The sub bands are then sent through a convolution layer within a multi-resolution CNN. At each layer, the multi-resolution CNN hierarchically decomposes the wavelet transformed input vector with different kernels. And the part of the sub-band is once again subjected to decomposition by the wavelet transformation in the subsequent layer before being sent through the convolution layer. Spatial feature vector x of length 128, and forwards to a WNN classifier in stage-2.

[0073] According to another exemplary embodiment of the invention, FIGs. 5A-5C refer to graphs (500, 502, 504) between overall accuracies and the number of training samples for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset. For analytical purposes, the experiments are conducted with various mother wavelets (Haar, D4, Mexican hat, and Morlet) as activation functions in the hidden layer of the WNN within the proposed two-stage hybrid model. All the experiments are conducted using the same dataset. The D4 wavelet among the plurality of mother wavelets achieves the highest performance.

[0074] According to another exemplary embodiment of the invention, FIG. 6 refers to a graph 600 representing kappa coefficients with 10% training data for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset. By analysing the three datasets (IP, SA, and PU), it becomes evident that wavelets from the DWT category, specifically D4 and Haar, deliver superior performance compared to Morlet and Mexican Hat wavelets, which fall under the CWT category. This difference in performance can be attributed to the fact that practical applications like HSI feature extraction and classification typically involve discrete, sampled data, where only a finite number of values are available. The D4 wavelets are renowned for their superior capacity to effectively capture both smooth and oscillatory features in images across various scales, surpassing the capabilities of Haar wavelets. This advantage contributes to better preservation of fine image details.

[0075] According to another exemplary embodiment of the invention, FIGs. 7A-7C refer to graphs (700, 702, 704) representing the comparison between the two-stage hybrid model and the state of the existing art methods by considering the kappa coefficients and the number of training samples for Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset. By considering overall accuracy and kappa coefficient, the proposed invention with the D4 wavelet activation consistently outperformed all others. The two-stage hybrid model with D4 wavelet activation consistently outperforms across the three distinct HSI datasets, achieving successful classification outcomes. The results clearly demonstrate that the two-stage hybrid model with the D4 wavelet activation outperforms state-of-the-art methods in terms of Kappa values for all the HSI datasets.

[0076] According to another exemplary embodiment of the invention, FIGs. 8A-8F refer to visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the IP dataset.

[0077] According to another exemplary embodiment of the invention, FIGs. 9A-9F refer to visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the PU dataset.

[0078] According to another exemplary embodiment of the invention, FIGs. 10A-10F refer to visualization maps for the two-stage hybrid model with different activation functions, 100 fixed epochs, and 10% training data on the SA dataset.

[0079] According to another exemplary embodiment of the invention, FIGs. 11A-11F refer to visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the IP dataset.

[0080] According to another exemplary embodiment of the invention, FIGs. 12A-12F refer to visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the PU dataset.

[0081] According to another exemplary embodiment of the invention, FIGs. 13A-13F refer to visualization maps for the two-stage hybrid model with D4 activation function, alongside state-of-the-art methods with 100 fixed epochs, and 10% training Data on the SA dataset.

[0082] According to another exemplary embodiment of the invention, FIG. 14 refers to a flowchart 1400 of a method for extracting and classifying a hyper spectral image using a two-stage hybrid model through a system 100. At step 1402, the input module 110 extracts the one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI).

[0083] At step 1404, the analyzing module 112 analyzes the one or more HSI patches for generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique.

[0084] At step 1406, the extraction module 114 extracts one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using discrete wavelet two-dimensional convolutional neural network (CNN).

[0085] Further at step 1408, the classification module 116 receives the extracted one or more spectral-spatial features and classify the at least one HSI patch using a wavelet neural network (WNN).

[0086] Numerous advantages of the present disclosure may be apparent from the discussion above. In accordance with the present disclosure system 100 for extracting and classifying a hyper spectral image, is disclosed. The proposed invention provides a system 100 for classifying and extracting spectral-spatial features of the hyperspectral image using a two-stage hybrid model. The system 100 provides multi-resolution analysis by using wavelet transform (WT) and a two-dimensional convolutional neural network (2D-CNN) for capturing information at different spatial and spectral scales.

[0087] The system 100 combines the discrimination capabilities of 2D-CNN with the signal analysis proficiency of wavelets in both time and frequency domains. The system 100 uses a wavelet neural network (WNN) for modeling complex non-linear relationships within HSI data, potentially leading to improved classification performance compared to linear models. The system 100 offers better generalizability to unseen data due to the combined strengths of WT's multi-scale learning and WNN's non-linear modeling capability.

[0088] It will readily be apparent that numerous modifications and alterations can be made to the processes described in the foregoing examples without departing from the principles underlying the invention, and all such modifications and alterations are intended to be embraced by this application.
, Claims:CLAIMS:
I/We Claim:
1. A system (100) for extracting and classifying hyper spectral images using a two-stage hybrid model, comprising:
a computing device (102) having a processor (104) and a memory (106) for storing one or more instructions executable by the processor (104), wherein the computing device (102) is in communication with a server (124) via a network (122),
wherein the processor (104) is configured to control plurality of modules (108) to execute image feature extraction and image classification functions, wherein the plurality of modules (108) comprises:
an input module (110) configured to extract and store one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI);
an analyzing module (112) configured to analyze the one or more HSI patches, thereby generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique;
an extraction module (114) configured to extract one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using a two-dimensional convolutional neural network (CNN) model; and
a classification module (116) configured to receive the extracted one or more spectral-spatial features and classify the at least one HSI patch using a classification neural network model.
2. The system (100) as claimed in claim 1, wherein the two-dimensional convolutional neural network (CNN) model is a discrete wavelet two-dimensional convolutional neural network (CNN), wherein the classification neural network model is a wavelet neural network (WNN) model.
3. The system (100) as claimed in claim 1, wherein the at least one hyper spectral image (HSI) is pre-processed using a factor analysis method to reduce dimensionality.
4. The system (100) as claimed in claim 1, wherein the two-dimensional convolutional neural network (CNN) model uses a daubechies 4-tap (D4) orthogonal filter to extract one or more spectral-spatial features.
5. The system (100) as claimed in claim 1, wherein the system (100) uses plurality of datasets to encompass a range of HSI data, including agricultural, rural-urban, and urban scenes.
6. The system (100) as claimed in claim 4, wherein the plurality of datasets includes Indian pines (IP) dataset, Salinas (SA) dataset and University of Pavia (PU) dataset.
7. The system (100) as claimed in claim 1, wherein the system (100) explores plurality of mother wavelets, which include Haar wavelet, Daubechies wavelet, Mexican hat wavelet, and Morlet wavelet to improve the discriminative capabilities of the two-stage hybrid model for the HSI classification.
8. The system (100) as claimed in claim 1, wherein the system (100) comprises a communication module (118), which is configured to enable communication between the computing device (102) and the server (124).
9. The system (100) as claimed in claim 1, wherein the computing device (102) includes a laptop, a computer, a smartphone, a tablet, a personal data assistant (PDA), and electronic devices.
10. A method for extracting and classifying a hyper spectral image using a two-stage hybrid model through a system (100), comprising:
extracting, by an input module (110), one or more hyper spectral image (HSI) patches from at least one pre-processed hyper spectral image (HSI);
analyzing, by an analyzing module (112), the one or more HSI patches for generating spectral signatures by breaking the at least one HSI patch into plurality of sub bands using a discrete wavelet decomposition technique;
extracting, by an extraction module (114), one or more spectral-spatial features from the plurality of sub bands of the at least one HSI patch using a discrete wavelet two-dimensional convolutional neural network (CNN); and
receiving, by a classification module (116), the extracted one or more spectral-spatial features and classify the at least one HSI patch using a wavelet neural network (WNN).