Search

Introduced in this paper is the utilization of support vector machines (SVMs) to semiautomatically perform endmember extraction from hyperspectral data. The strengths of SVM are exploited to provide a fast and accurate calculated representation of high-dimensional data sets that may consist of multiple distributions. Once this representation is computed, the number of distributions can be determined without prior knowledge. For each distribution, an optimal transform can be determined that preserves informational content while reducing the data dimensionality and, hence, the computational cost. Finally, endmember extraction for the whole data set is accomplished. Results indicate that this SVM-based endmember extraction algorithm has the capability of semiautonomously determining endmembers from multiple clusters with computational speed and accuracy while maintaining a robust tolerance to noise. Index Terms--Endmember extraction, hyperspectral imaging, remote sensing, support vector machines (SVMs).

Continuum removal (CR) is often used for geologic mapping; however, more research is needed to better establish the utility of CR for vegetation classification, particularly when used with artificial neural networks (ANNs). In this paper, fuzzy learning vector quantization (FLVQ) was applied to hyperspectral Airborne Visible/Infrared Imaging Spectrometer imagery for coastal vegetation classification. FLVQ performance was compared with that of a multilayer perceptron (MLP), a self-organizing map (SOM), and an endmember-based algorithm [spectral feature fitting (SFF)]. The objective was to assess the effect of CR as an input vector-preprocessing step for ANN model development on classification accuracy. Compared with a related study, continuum intact (CI) reflectance data generally yielded higher classification accuracies than those based on CR. Thus, CR may not be a preferred preprocessing method for coastal vegetation mapping over broad wavelength ranges. MLP slightly outperformed FLVQ when applied to CI data, but FLVQ yielded higher accuracy than MLP with CR. However, there was no significant difference between them for both data treatments at the 95% confidence level. All ANNs tested yielded significantly higher classification accuracies than SFF. For model development, the 588-neuron FLVQ required only 8.2% of MLP training time, 27.8% of the 400-neuron SOM time, and 8.8% of the 729-neuron 3-D SOM time. Index Terms--Continuum removal (CR), fuzzy neural networks, hyperspectral, learning vector quantization (LVQ), remote sensing, wetlands.

To purchase or authenticate to the full-text of this article, please visit this link: http://dx.doi.org/10.1111/j.1467-9272.2007.00610.x Byline: Anthony M. Filippi (1) Keywords: artificial neural network; bathymetry; derivative; hyperspectral; remote sensing Abstract: Bathymetry is an important variable in scientific and operational applications. The research objectives in this study were to estimate bathymetry based on derivative reflectance spectra used as input to a multilayer perceptron artificial neural network (ANN) and to evaluate the efficacy of field and simulated training/testing data. ANNs were used to invert reflectance field data acquired in optically shallow coastal waters. Results indicate that for the simulation-based models, nonderivative spectra yielded more accurate bathymetry retrievals than the derivative spectra used as ANN input. However, for the empirical field-based models, derivative spectra were superior to nonderivative spectra as ANN input. This study identifies circumstances under which derivative spectra are useful in bathymetry estimation, and thus increases the likelihood of obtaining accurate inversions. Author Affiliation: (1)Texas A&M University Article History: Initial submission, July 2005; revised submission, January, May, and July 2006; final acceptance, August 2006.

Accurate characterization of the direction of land change is a neglected aspect of land dynamics. Knowledge on direction of historical land change can be useful information when understanding relative influence of different land-change drivers is of interest. In this study, we present a novel perspective on land-change analysis by focusing on directionality of change. To this end, we employed Maximum Cross-Correlation (MCC) approach to estimate the directional change in land cover in a dynamic river floodplain environment using Landsat 5 Thematic Mapper (TM) images. This approach has previously been used for detecting and measuring fluid and ice motions but not to study directional changes in land cover. We applied the MCC approach on land-cover class membership layers derived from fuzzy remote-sensing image classification. We tested the sensitivity of the resulting displacement vectors to three user-defined parameters--template size, search window size, and a threshold parameter to determine valid (non-noisy) displacement vectors--that directly affect the generation of change, or displacement, vectors; this has not previously been thoroughly investigated in any application domain. The results demonstrate that it is possible to quantitatively measure the rate of directional change in land cover in this floodplain environment using this particular approach. Sensitivity analyses indicate that template size and MCC threshold parameter are more influential on the displacement vectors than search window size. The results vary by land-cover class, suggesting that spatial configuration of land-cover classes should be taken into consideration in the implementation of the method. [ABSTRACT FROM AUTHOR]

New computationally efficient inversion techniques are needed to better estimate depth values and inherent optical properties (IOPs) of the water column from apparent optical properties (AOPs). Improved methods are needed [...]

Optical remote sensing imagery offers a cost-effective alternative to echo sounding and bathymetric light detection and ranging surveys for deriving high density bottom depth estimates for coastal and inland water bodies. The common practice of previous studies has been to calibrate a single global bathymetric inversion model for an entire image scene. The performance of conventional global models is limited when the bottom type and water quality vary spatially within the scene. To address the inadequacy of the conventional global models, this paper presents a geographically adaptive inversion model to better estimate bottom depth. Although the general mathematical form of the geographically adaptive model is the same, model parameters are optimally determined within a geographical region or a local area, in contrast to the entire scene in the global inversion model. By using high-resolution IKONOS and moderate-resolution Landsat satellite images, we demonstrated that regionally and locally calibrated inversion models can effectively address the problems introduced by spatial heterogeneity in water quality and bottom type, and provide significantly improved bathymetric estimates for more complex coastal waters. [ABSTRACT FROM AUTHOR]

Remote sensing (RS) is an important technology for Earth observation, where detailed physical and chemical properties of the Earth can be remotely characterized. Hyperspectral images (HSIs), which are collected from a sensor where the spectral width of each band is narrow compared with other (broadband) optical sensors, yield continuous spectra that can be used to distinguish those ground objects with similar spectral features. In order to process and analyse HSIs, classification, where each pixel in a HSI is assigned to a predefined label based on a specific machine-learning model, for example, is a widely-used technology. Over the past decade, deep learning, a subset of machine-learning methods, has been increasingly employed due to its accurate performance in various research subfields, including computer vision, medical image analysis, speech recognition, and nature-language processing. For the RS and HSI communities, deep learning has been successfully utilized since 2014. Since then, varied deep learning-based approaches have been proposed, including deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Among those deep-learning models, RNNs have been increasingly been applied to the problem of RS image processing, especially multi-temporal image analysis, due to their great capacity for handling sequential data. However, for the particular case of RNN-based classification of single RS images, further research is required. The key question in this scenario is how to extract the sequential feature for each pixel. To address this issue, various researchers have explored and proposed different solutions which can be summarized into two main categories: Given a target pixel, the first strategy is to utilize the spectral feature as its sequential feature. The second strategy is to extract "similar" pixels from the image and exploit those "similar" pixels to construct a sequential feature of a target pixel. Based on our previous work [1], we proposed a similarity-measurement-based RNN HSI classification model where the sequential feature of a target pixel is encoded by first utilizing several "similar" pixels extracted from the whole image. Although the classification performance of this proposed framework is much better than other standard algorithms, its computational time cost is still a critical issue. Therefore, how to develop a more computationally-efficient sequential feature-extraction method within the RNN-based classification domain still needs more investigation for the singleimage classification problem. Based on our previous work [1], we propose an improved sequential feature-extraction framework for single HSI-based classification using RNN. The main contribution of the present research effort is to modify the searching range of the similarity measurement such that it is shrunk from the whole-image pixel set to one or two similar segments created by an object-based segmentation algorithm. The effect is that similar pixels of a target pixel will not be extracted from all pixels of the image, but rather from similar segments generated from a segmentation map derived from the fractal net evolution approach (FNEA) model, which is a well-known multiresolution segmentation algorithm. Specifically, for our proposed method, we design a two-phase similarity measurement: 1) segment-similarity is first calculated in order to select similar segments; and 2) pixel-similarity is then obtained, utilizing those pixels collected from selected "similar" segments. After pixel-similarity calculation, similar pixels can be extracted to construct sequential features in the same manner as what we proposed in our previous work [1]. During the calculation of segment-based similarity, both spectral and geometry features of those segments are considered in order to better characterize properties of the segments. For the selection of similar segments, we develop three different strategies, including: 1) a local segmentbased approach where only the segment that contains the current target pixel is taken into consideration; 2) a non-local segment-based approach, where the second most similar segment is selected instead of the current segment; and 3) a mixed segment-based approach, where the two aforementioned segments are utilized simultaneously. Moreover, pixelmatching (PM) and block-match (BM) proposed in our previous work [1] are employed as well when measuring similarity between two individual pixels. To assess classification performance, three benchmark HSIs are utilized including the Pavia University HSI, the Salinas HSI, and the Indian Pines HSI. Meanwhile, the models proposed in [1] are assessed as well, as they are considered to be a baseline for the evolution of computational time costs. In addition, other benchmark algorithms, including support vector machine (SVM), the 1-dimensional convolutional neural network (1DCNN), and the 1-dimensional RNN (1DRNN), are utilized as well. The results demonstrate that our proposed models obtain better classification accuracies compared with other evaluated algorithms. For the Pavia University and Salinas HSIs, using the mixed segment-based approach yields the highest accuracy, with 96.89% and 96.64% overall accuracy (OA), respectively. Regarding the characterization of the classification maps, apparent improvements in classification performance can be visually observed. For the Indian Pines HSI, the best results are obtained from the local segment-based model instead of the mixed segment-based approach, with 94.04% OA. We also evaluate the computational costs for both the proposed models and the ones in our previous work [1]. We find that such time costs are reduced significantly, by more than one hundred times. For example, for the Pavia University HSI, the time cost of our previous model is 4155.98 seconds. However, such a cost for the proposed model utilizing the local segment-based approach is 40.63 seconds. Such results demonstrate that those proposed approaches are capable of achieving satisfactory classification accuracy, while still obtaining markedly lower computational time cost. [ABSTRACT FROM AUTHOR]