Abhinav Mathur

Denoising and Wavelet-Based Feature Extraction of MODIS Multi-Temporal Vegetation Signatures

Temporal vegetation signatures (i.e., vegetation indices as functions of time) generated using the MODIS imagery poses many challenges, primarily due to signal-to-noise-related issues. This article describes the use of MODIS time-series data for the detection of specific tropical invasive species vegetation types. Due to challenges with the MODIS quality assurance data, a significant level of noise was present in the temporal signatures. This study investigated methods for denoising the vegetation temporal signatures, followed by a comparative analysis of three denoising methods to generate signatures for vegetation target detection. The analytical approach focused on the use of wavelet-based versus Fourier-based feature extraction methods. Methods included the development of a novel wavelet-based feature extraction method that quantifies the fundamental shape of the temporal signatures.

Decision level fusion with best-bases for hyperspectral classification

In recent years, more intuitive understanding about the characteristics of higher dimensional space has influenced the development of subsequent data analysis and classification algorithms in the field of hyperspectral remote sensing. Earlier data analysis and classification algorithms rely on processing high dimensional space as a whole to extract a lower dimensional feature space. The major impediment on these techniques is the limited training data size, which does not confer with the large dimensionality of hyperspectral data. Previous work has shown that statistically reliable parameter estimation can be performed on lower dimensional subspaces that are formed by decomposing the entire dimension into a set of subspaces (bases), based on certain discrimination criterion. In this paper the authors present a classification technique that combines the feature level fusion capabilities of lower dimensional subspaces; with decision level fusion to improve the classification potential of hyperspectral data. In order to reduce the impact of conflicting decisions by individual bases, a voting scheme called Qualified Majority Voting (QMV) is used in combining the decisions. Each base is qualified to influence the final decision, based on its ability to predict the classes with respect to other bases. This information can be derived from training data, analyst inputs or feed back from prior applications. Unlike the traditional classification approaches, this technique not only utilizes the projected lower dimensional feature space, but also makes use of the reliability of the subspaces in classifying certain classes.

Discrimination of subtly different vegetative species via hyperspectral data

The authors of this paper investigate the use of hyperspectral reflectance curves for the discrimination of cogangrass (Imperata cylindrica) from other subtly different vegetation species. Receiver operating characteristics (ROC) curves are used to determine which spectral bands should be considered as candidate features. Multivariate statistical analysis is then applied to the candidate features to determine the optimum subset of spectral bands. Linear discriminant analysis (LDA) is used to compute the optimum linear combination of the selected subset to be used as a feature for classification. Similarly, ROC analysis, multivariate statistical analysis, and LDA are utilized to determine the most advantageous wavelet-based scalar feature for classification. Nearest-neighbor classification results show that cogongrass can be detected with an accuracy of /spl ap/90%.

Hyperspectral dimensionality reduction via localized discriminant bases

To overcome the dimensionality curse of hyperspectral data, the authors of the paper have investigated the use of grouping the spectral bands along with localized discriminant bases, followed by decision fusion to develop an ATR system for data reduction and enhanced classification of hyperspectral data. The proposed system is robust to the availability of limited training data. Initially, the entire span of spectral bands in the hyperspectral data is subdivided into subspaces or groups based on a performance metric. The groups are not allowed to grow beyond what is supported by the amount of available training data. Feature extraction is done using supervised methods as well as unsupervised methods. Further, decision level fusion is applied to the features extracted from each group. To reduce the effect of conflicting decisions by individual groups, a voting scheme called Qualified Majority Voting is adopted to combine decisions. The effectiveness of the proposed system is tested using a data set consisting of hyperspectral signatures of a target class, Cogongrass (Imperata Cylindrica), and a non-target class, Johnsongrass (Sorghum halepense). Cogongrass is an invasive species of plant whose monitoring has become important due to the extensive ecosystem damage that it causes. A comparison of target detection accuracies by the proposed system before and after decision fusion is done to illustrate the effect of the influence of each group of spectral bands on the final decision and to illustrate the benefit of using decision fusion with multiclassifiers.

Wavelet transform for dimensionality reduction in hyperspectral linear unmixing

In Li et al. (2001), the authors investigated how dimensionality reduction using wavelet-based feature extraction can improve the classification of materials from hyperspectral reflectance. In this paper, a similar approach is suggested for the hyperspectral linear unmixing problem. The paper shows, both experimentally and theoretically, that the abundance estimation using the least squares estimation can be improved through appropriate feature extraction. The discrete wavelet transform is suggested for the feature extraction, and a wavelet-based unmixing system is designed and implemented. Two metrics, the root-mean-square error and the confidence of abundance estimation, are proposed to quantitatively evaluate the unmixing system performance.

Exploiting Hyperspectral Hypertemporal Imagery with Feature Clustering for Invasive Species Detection

This paper presents a feature extraction method for exploiting hyperspectral hypertemporal data and applies the new method to the problem of invasive species detection. By definition, hyperspectral hypertemporal imagery is very high dimensional data, and dimensionality reduction will play a critical role in utilizing such data. We present a feature extraction method that takes advantage of the high correlation among elements of the spectral and temporal feature space. This high correlation can be attributed to the premise that the changes in the reflectance of closely spaced wavelengths do not always change dramatically over short periods of time. The proposed feature clustering method is based on the assumption that adjacent elements in the spectro-temporal feature space are highly correlated and can be grouped together to form lower- dimensional feature spaces. The proposed hyperspectral hypertemporal feature clustering method is tested and validated within an invasive vegetation detection application. The hypothesis is that as time progresses, the spectral response of different plant species change differently. Thus, there should be hyperspectral hypertemporal features that can be used to discriminate between the vegetative species. Additionally, the results of the feature clustering method can be used to determine which regions of the spectrum and which collection dates are optimum for the given invasives detection problem.

Denoising and wavelet-based feature extraction of MODIS multi-temporal vegetation signatures

Temporal vegetation signatures (i.e., vegetation indices as functions of time) generated using the MODIS imagery poses many challenges, primarily due to signal to noise-related issues. This article describes the use of MODIS time-series data for the detection of specific tropical invasive species vegetation types. Due to challenges with the MODIS quality assurance data, a significant level of noise was present in the temporal signatures. This study investigated methods for denoising the vegetation temporal signatures, followed by a comparative analysis of three denois- ing methods to generate signatures for vegetation target detection. The analytical approach focused on the use of wavelet-based versus Fourier-based feature extrac- tion methods. Methods included the development of a novel wavelet-based feature extraction method that quantifies the shape of the fundamental in the temporal signa- tures.

Effects of watermarking on feature efficacy in remotely sensed data

Digital watermarking is gaining increasing popularity and attention in various research communities as a means of providing protection of ownership rights for multimedia data. Recently, researchers have begun to explore its feasibility for the same purposes for remotely sensed data. The suitability criterion for any watermarking algorithm is typically determined by the processing and analysis techniques to which the data is most likely to be subjected. In remotely sensed data, a primary application is target detection, which further leads to various secondary end products like land cover maps. In proposing a watermarking algorithm for remotely sensed data, it is necessary to minimize any adverse effect on classification, while simultaneously achieving reliable detection of the embedded watermark. This paper proposes two algorithms for this purpose

Multiresolutional texture analysis of multispectral imagery for automated ground cover classification

The authors provide an introduction to wavelet-based texture recognition. The authors compare conventional approaches, such as co-occurrence matrix methods and Fourier-based techniques, to wavelet-based approaches. Practical remote sensing applications are presented, including the detection of weeds in agricultural crops and the detection and analysis of noxious weeds for environmental applications. For these case studies, the authors present a theoretical discussion as well as experimental classification results. These applications demonstrate the benefits of multiresolutional methods to the texture classification task in multispectral imagery.

Automated texture recognition system based on 2D minimum variance spectral estimation

The primary feature of any image texture is the spatial frequency content. This paper proposed the use of a 2D minimum variance spectral estimation (MVSE) method for recognizing target multispectral image textures. The power spectral density of the target texture is estimated via MVSE. This estimate is then used as a feature to discriminate between target and nontarget textures. A remotely sensed multispectral image of a row crop agricultural field is analyzed and, the corresponding results are presented to illustrate the applicability of the proposed technique.