Deep Ensembles for Hyperspectral Image Data Classification and Unmixing

Hyperspectral images capture very detailed information about scanned objects and, hence, can be used to uncover various characteristics of the materials present in the analyzed scene. However, such image data are difficult to transfer due to their large volume, and generating new ground-truth datasets that could be utilized to train supervised learners is costly, time-consuming, very user-dependent, and often infeasible in practice. The research efforts have been focusing on developing algorithms for hyperspectral data classification and unmixing, which are two main tasks in the analysis chain of such imagery. Although in both of them, the deep learning techniques have bloomed as an extremely effective tool, designing the deep models that generalize well over the unseen data is a serious practical challenge in emerging applications. In this paper, we introduce the deep ensembles benefiting from different architectural advances of convolutional base models and suggest a new approach towards aggregating the outputs of base learners using a supervised fuser. Furthermore, we propose a model augmentation technique that allows us to synthesize new deep networks based on the original one by injecting Gaussian noise into the model’s weights. The experiments, performed for both hyperspectral data classification and unmixing, show that our deep ensembles outperform base spectral and spectral-spatial deep models and classical ensembles employing voting and averaging as a fusing scheme in both hyperspectral image analysis tasks.

Recent Advances in Multi- and Hyperspectral Image Analysis

Current advancements in sensor technology bring new possibilities in multi- and hyperspectral imaging. Real-life use cases which can benefit from such imagery span across various domains, including precision agriculture, chemistry, biology, medicine, land cover applications, management of natural resources, detecting natural disasters, and more. To extract value from such highly dimensional data capturing up to hundreds of spectral bands in the electromagnetic spectrum, researchers have been developing a range of image processing and machine learning analysis pipelines to process these kind of data as efficiently as possible. To this end, multi- or hyperspectral analysis has bloomed and has become an exciting research area which can enable the faster adoption of this technology in practice, also when such algorithms are deployed in hardware-constrained and extreme execution environments; e.g., on-board imaging satellites.

A Study of Microplastic Particles in Danish Tap Water

Microplastics (MPs) are omnipresent in our surroundings and in the environment, with drinking water being a potential pathway for human exposure. This study investigated the presence of MPs in Danish drinking water from 17 different households and workplaces in Denmark. Samples of tap water were collected using a closed sampling system to decrease airborne contamination, and QA/QC measurements were performed to assess background contamination. Particles >100 µm were visually analysed by stereomicroscopy in combination with spectroscopy analysis (µ-FTIR) to evaluate morphology and chemical composition. An assessment of MP particles down to 10 µm was performed on water samples from three locations using hyperspectral image analysis. The results indicate a low level of MPs in Danish drinking water, with a total of seven MP particles across all samples, comprising PET, PP, PS, and ABS. Microfibers were the most common type of MP-like particles in both drinking water and blanks, but the concentration for all samples was below the limit of detection and could not be differentiated from background contamination. Most of the particles analysed by µ-FTIR were identified as cellulose fibres and a smaller subset as protein. Based on this work, we discuss the status of MP drinking water studies and address challenges and limitations regarding the analysis of MP in drinking water.

Determining the composition of an object based on its hyperspectral image

The article presents results of the development and research of a hyperspectral imaging spectrometer for analyzing borehole fluids in real operating conditions in the spectral range from 0.35 microns to 2.1 microns. A mathematical model and an algorithm for identifying the borehole fluid by composition and percentage content based on the results of hyperspectral image analysis are developed.

Hyperspectral image analysis using a simultaneous denoising and intrinsic order selection (DIOS) approach

Recent hyperspectral applications demand for higher accuracy and speed. This thesis develops a hyperspectral application analysis solution to address challenges in the different steps of denoising, order selection and unmixing of hyperspectral application data. Currently, all these steps process the data in cascade to achieve the optimum results. While in existing approaches the desired criterion is different in these steps, the proposed simultaneous Denoising and Intrinsic Order Selection (DIOS) method unifies these criteria. This property not only makes more sense for the desired optimization problem, but also leads to a faster processing algorithm. Consequently, DIOS avoids possible error propagation from the denoising stage to the dimension estimation stage, leading to more accurate results. The proposed method is based on minimizing the estimated Mean Square Error (MSE). The success rate of existing dimension estimation methods declines with the increase of image dimension and the decrease of Signal-to-Noise Ratio (SNR). The most competitive method fails to detect the correct dimension in 30% of cases around 2dB. However, in simulation results DIOS is shown to be successful with a failure rate of about 5%. The proposed unmixing method, based on a simple least square estimation, improves the speed performance least 10 times for an average-sized data cube of 2MB. Compared to some well known existing approaches, the unmixing method improves the estimated MSE up to 60% for SNR<10dB. A new whitening process for hyperspectral applications with coloured noise is also proposed. Since the proposed method avoids the inversion of large matrices, computational complexity is substantially decreased. In the presence of coloured noise, simulation results show that the proposed whitening method lowers the MSE of unmixing and outperforms the existing whitening methods particularly when the noise correction factors increase.

Hyperspectral image analysis using a simultaneous denoising and intrinsic order selection (DIOS) approach

Recent hyperspectral applications demand for higher accuracy and speed. This thesis develops a hyperspectral application analysis solution to address challenges in the different steps of denoising, order selection and unmixing of hyperspectral application data. Currently, all these steps process the data in cascade to achieve the optimum results. While in existing approaches the desired criterion is different in these steps, the proposed simultaneous Denoising and Intrinsic Order Selection (DIOS) method unifies these criteria. This property not only makes more sense for the desired optimization problem, but also leads to a faster processing algorithm. Consequently, DIOS avoids possible error propagation from the denoising stage to the dimension estimation stage, leading to more accurate results. The proposed method is based on minimizing the estimated Mean Square Error (MSE). The success rate of existing dimension estimation methods declines with the increase of image dimension and the decrease of Signal-to-Noise Ratio (SNR). The most competitive method fails to detect the correct dimension in 30% of cases around 2dB. However, in simulation results DIOS is shown to be successful with a failure rate of about 5%. The proposed unmixing method, based on a simple least square estimation, improves the speed performance least 10 times for an average-sized data cube of 2MB. Compared to some well known existing approaches, the unmixing method improves the estimated MSE up to 60% for SNR<10dB. A new whitening process for hyperspectral applications with coloured noise is also proposed. Since the proposed method avoids the inversion of large matrices, computational complexity is substantially decreased. In the presence of coloured noise, simulation results show that the proposed whitening method lowers the MSE of unmixing and outperforms the existing whitening methods particularly when the noise correction factors increase.