Novel Techniques in Hyperspectral Data Analysis for Endmember Extraction, Change Detection and Classification

Abstract

Hyperspectral image (HSI) analysis is a powerful technique in remote sensing that involves the acquisition and analysis of images captured across hundreds or even thousands of narrow and contiguous spectral bands. Unlike traditional remote sensing techniques that capture information in just a few broadbands, HSI provides detailed spectral information for each pixel in an image scene. This wealth of spectral data enables a more comprehensive understanding of the Earth’s surface and the objects it contains. By analyzing the unique spectral signatures of different materials, HSI enables the identification and discrimination of various land cover types, vegetation species, soil properties, and even specific minerals. However, the analysis of hyperspectral data presents several challenges that require specialized approaches. Endmember extraction (EE) is one such challenge, involving the identification of pure spectral signatures or reference spectra that represent specific target materials. Spectral matching is a vital component of HSI analysis that involves measuring the degree of closeness between the spectral signatures extracted from the image and those obtained through ground-based spectrometer measurements or known reference spectra. Spectral matching algorithms are useful not only for validating the accuracy of image-based signatures but also for feature extraction. This matching process helps in the identification, analysis, and interpretation of different materials or targets present in the HSI. Through the combination of endmember extraction and spectral matching in HSI, diverse applications in geology and related fields are empowered, enabling tasks such as geological mapping, mineral exploration, environmental monitoring, and land cover analysis with increased accuracy and efficiency. Although hyperspectral imaging was originally developed for mining and geology, mineral identification using hyperspectral data has not been addressed adequately yet and remains a challenging task. HSIs due to advancements in spatial-spectral resolutions and the availability of multi-temporal information are in demand for many applications. Change detection (CD), in particular, is an important and challenging problem in monitoring changes such as deforestation, urban development, and landslides using time series HSI data. Though several endmember extraction algorithms (EEAs) are developed, spectral matching algorithms (SMAs) have not been explored much in the extraction of spectrally distinct signatures. Therefore, in this work, similarity measures based EEAs (SM-EEAs) are proposed to explore the fundamental characteristic, i.e. spectrally distinctive in nature, of endmembers. Experimental results on proposed EEAs i.e., a similarity measures-based subtractive clustering algorithm (SM-SCA) and a similarity measures-based endmember initialization algorithm (SM-EIA) showed the applicability of SMAs in extracting spectrally distinct signatures as the endmembers and also hinted the importance of endmember initialization. The darkest pixel identified as a target pixel of interest (TPOI) in the further investigation on endmember initialization strategies emerges not only as a potential TPOI but also contributes to improving the performance of EEAs when combined with the brightest pixel as TPOIs. Experiments carried out on an improved SM-EIA to test its applicability in extracting pure endmembers present maximin-distance algorithm (MDA) do not able to identify vertices of the simplex with simple metrics like Euclidean distance (ED) and other simple SMAS but with higher dimensionality metrics like volume. The proposed corner-driven iterative clustering algorithm (CDIC) appears to perform better in EE by identifying the corner pixels and thereby providing training samples for HSI classification. Though few already developed spectral matching measures are available, the identification of diagnostic features of spectrally distinct signatures with the existing SMAs to discriminate them effectively is still a challenging task. Therefore, this work presents a gradientbased spectral similarity measure (GSSM) that captures the diagnostic (absorption) features to measure the degree of closeness between spectral signatures. The effectiveness of the proposed GSSM in distinguishing spectrally distinct signatures is studied with that of other spectral matching algorithms in (i) discriminating endmember signatures (ii) mixed pixel identification (iii) clustering spectral signatures of different classes and (iv) endmember extraction. The proposed GSSM was not only able to highlight diagnostic features of target signatures but also showed its effectiveness in discriminating spectrally distinct target signatures better than other SMAs. Further study on gradient correlation (GC) incorporated showed improved discrimination power with geometrical SMAs. Further, a meaningful way of measuring RSDPW is proposed. Reformulated RSDPW appears to be more meaningful in discriminating endmembers and obtaining the range of RSDPW values for different levels of discrimination than the former one. The high dimensionality of HSI data and limited availability of hyperspectral CD data sets with ground-truth change map make CD not so easy but a difficult task. Though there are many classical and deep learning (DL) based algorithms to detect changes, either their performance is not so better or the final performance depends on efficiency of pre-detection algorithms. In addition, there is not much comprehensive study on developing CD algorithms that not only simple to use but also as efficient as that of DL based algorithms is available. Therefore, an endmember related feature extraction is proposed for HSI-CD. Proposed ATGP based CD algorithms not only perform better than classical CD algorithms but also able to reach the performance of DL based CD algorithms. additionally, even a minimum number of features around three to five (3-5) also good enough to get high accuracy as that of DL models. Mineral identification remains a challenging task due to the subtle differences among the spectral signatures of minerals and insufficient ground truth. The classical spectral angle measure (SAM) classifier is a simple model and does not yield high accuracy and an expert system for hyperspectral data classification (ExHype) is a binary classifier and therefore complex to train binary classifier modules equal to the number of minerals to be classified and obtain thresholds to gain accuracy. Due to lack of training samples and sufficient data with ground truth to be tested, DL models have not been explored much in mineral classification so far. To overcome this, a virtual sample generation to be able to generate more training samples that provide a chance to explore DL models that need variations in training samples in mineral classification is proposed. Further, a one-dimensional convolutional neural network (1-D CNN) model, trained on training samples generated by virtual sample generation, designed to classify minerals performed well in classifying the tested mineral classes with high accuracy.