Search

Deep neural networks (NNs) trained on hyperspectral images are employed typically for the classification of new images collected from the same sensor, assuming similar characteristics to those of the training images. Creating, however, high-quality ground truth (GT) for training is rather complex, especially when attempting to classify multi-temporal images over seasonal changes. To overcome this difficulty, we propose a novel method that utilizes an additional, one-time collection of hyperspectral FENIX images in the Spring along with ground observations from the end of the Fall. The hyperspectral data are then used for simulation of GT for training. At the same time, the field campaign allows for fine-tuning of the NN to achieve enhanced, multi-seasonal hyperspectral image classification. Indeed, we demonstrate how the proposed method successfully classifies new VEN $\mu \mathrm{S}$ images obtained during different seasons.

Training a deep neural network for classification constitutes a major problem in remote sensing due to the lack of adequate field data. Acquiring high-resolution ground truth (GT) by human interpretation is both cost-ineffective and inconsistent. We propose, instead, to utilize high-resolution, hyperspectral images for solving this problem, by unmixing these images to obtain reliable GT for training a deep network. Specifically, we simulate GT from high-resolution, hyperspectral FENIX images, and use it for training a convolutional neural network (CNN) for pixel-based classification. We show how the model can be transferred successfully to classify new mid-resolution VENuS imagery.

In recent years, large datasets of high-resolution mammalian neural images have become available, which has prompted active research on the analysis of gene expression data. Traditional image processing methods are typically applied for learning functional representations of genes, based on their expressions in these brain images. In this paper, we describe a novel end-to-end deep learning-based method for generating compact representations of in situ hybridization (ISH) images, which are invariant-to-translation. In contrast to traditional image processing methods, our method relies, instead, on deep convolutional denoising autoencoders (CDAE) for processing raw pixel inputs, and generating the desired compact image representations. We provide an in-depth description of our deep learning-based approach, and present extensive experimental results, demonstrating that representations extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Our methods improve the previous state-of-the-art classification rate (Liscovitch, et al.) from an average AUC of 0.92 to 0.997, i.e., it achieves 96% reduction in error rate. Furthermore, the representation vectors generated due to our method are more compact in comparison to previous state-of-the-art methods, allowing for a more efficient high-level representation of images. These results are obtained with significantly downsampled images in comparison to the original high-resolution ones, further underscoring the robustness of our proposed method., Comment: arXiv admin note: substantial text overlap with arXiv:1711.09663

Several image registration methods, based on the scaled-invariant feature transform (SIFT) technique, have appeared recently in the remote sensing literature. All of these methods attempt to overcome problems encountered by SIFT in multimodal remotely sensed imagery, in terms of the quality of its feature correspondences. The deterministic method presented in this letter exploits the fact that each SIFT feature is associated with a scale, orientation, and position to perform mode seeking (in transformation space) to eliminate outlying corresponding key points (i.e, features) and improve the overall match obtained. We also present an exhaustive empirical study on a variety of test cases, which demonstrates that our method is highly accurate and rather fast. The algorithm is capable of automatically detecting whether it succeeded or failed.

The mean shift MS algorithm is based on a statistical approach to the clustering problem. Specifically, the method is a variant of density estimation. We revisit in this article the MS paradigm and its use for clustering of remotely sensed images. Specifically, we investigate further the classification accuracy of remotely sensed images as a function of various MS parameters, such as the variant used, kernel type, dimensionality, kernel bandwidth, etc. We provide empirical assessment of the algorithm based on experiments with multi-temporal and multi-spectral remotely sensed data sets, representing agricultural and land-cover data. Although the classification accuracy and reliability seem comparable to those obtained by other unsupervised methods e.g. ISODATA, the MS algorithm provides several important operational advantages. The adaptation of the procedure to a parallel computational environment is also discussed and demonstrated.

This paper presents a novel deep learning-based method for learning a functional representation of mammalian neural images. The method uses a deep convolutional denoising autoencoder (CDAE) for generating an invariant, compact representation of in situ hybridization (ISH) images. While most existing methods for bio-imaging analysis were not developed to handle images with highly complex anatomical structures, the results presented in this paper show that functional representation extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Using this CDAE representation, our method outperforms the previous state-of-the-art classification rate, by improving the average AUC from 0.92 to 0.98, i.e., achieving 75% reduction in error. The method operates on input images that were downsampled significantly with respect to the original ones to make it computationally feasible.

In this paper we describe the problem of painter classification, and propose a novel approach based on deep convolutional autoencoder neural networks. While previous approaches relied on image processing and manual feature extraction from paintings, our approach operates on the raw pixel level, without any preprocessing or manual feature extraction. We first train a deep convolutional autoencoder on a dataset of paintings, and subsequently use it to initialize a supervised convolutional neural network for the classification phase. The proposed approach substantially outperforms previous methods, improving the previous state-of-the-art for the 3-painter classification problem from 90.44% accuracy (previous state-of-the-art) to 96.52% accuracy, i.e., a 63% reduction in error rate.

Performing standard unmixing of a hyperspectral image, while taking into account all of the potential endmembers (EMs) in a pixel, is known to be prone to error. Instead, determining first the set of EMs that actually reside in each pixel, leads to enhanced unmixing results. This important insight for achieving higher unmixing accuracy can be exploited efficiently by extracting relevant spatial information from a given image. In this work, we present a new method for spatially adaptive spectral unmixing, called the Gaussian based spatially adaptive unmixing (GBSAU) method. GBSAU takes advantage of the spatial arrangement of the image pixels and their spectral relations in order to determine an actual subset of EMs per pixel. It is based on spatial localization of the EMs by fitting, for each EM, the parameters of the series of spatial Gaussians whose sum represents the EM's fraction surface over the image.

The goal of the project described in this paper is to build a prototype of an operational system, which will provide registration within subpixel accuracy of multitemporal Landsat data, acquired by either Landsat-5 or Landsat-7 Thematic Mapper instruments. Integrated within an automated mass processing system for Landsat data, the input to our registration system consists of scenes that have been geometrically and radiometrically corrected, as well as preprocessed for detection of clouds and cloud shadows. Such preprocessed scenes are then georegistered relative to a database of Landsat chips. This paper describes the entire registration process, including the use of landmark chips, feature extraction performed by an overcomplete wavelet representation, and feature matching using statistically robust techniques. Knowing the approximate longitudes and latitudes or the UTM coordinates of the four corners of each incoming scene, a subset of the chips that represent landmarks included in the scene are selected to perform the registration. For each of these selected landmark chips, a corresponding window is extracted from the incoming scene, and each chip-window pair is registered using a robust wavelet feature-matching methodology. Based on the transformations from the chip-window pairs, a global transformation is then computed for the entire scene using a variant of a robust least median of squares estimator. Empirical results of this registration process, which provided subpixel accuracy for several multitemporal scenes from different study areas, are presented and discussed.

In k-means clustering, we are given a set of n data points in d-dimensional space R/sup d/ and an integer k and the problem is to determine a set of k points in Rd, called centers, so as to minimize the mean squared distance from each data point to its nearest center. A popular heuristic for k-means clustering is Lloyd's (1982) algorithm. We present a simple and efficient implementation of Lloyd's k-means clustering algorithm, which we call the filtering algorithm. This algorithm is easy to implement, requiring a kd-tree as the only major data structure. We establish the practical efficiency of the filtering algorithm in two ways. First, we present a data-sensitive analysis of the algorithm's running time, which shows that the algorithm runs faster as the separation between clusters increases. Second, we present a number of empirical studies both on synthetically generated data and on real data sets from applications in color quantization, data compression, and image segmentation.

In this paper we describe the problem of painter classification, and propose a novel hybrid approach incorporating genetic algorithms (GA) and deep restricted Boltzmann machines (RBM). Given a painting, we extract features using both generic image processing (IP) functions (e.g., fractal dimension, Fourier spectra coefficients, texture coefficients, etc.) and unsupervised deep learning (using deep RBMs). We subsequently compare several supervised learning techniques for classification using the extracted features as input. The results show that the weighted nearest neighbor (WNN) method, for which the weights are evolved using GA, outperforms both a support vector machine (SVM) classifier and a standard nearest neighbor classifier, achieving over 90% classification accuracy for the 3-painter problem (an improvement of over 10% relatively to previous results due to standard feature extraction only).

The nearest neighbor problem is that of preprocessing a set P of n data points in Rd so that, given any query point q, the closest point in P to q can be determined efficiently. In the chromatic nearest neighbor problem, each point of P is assigned a color, and the problem is to determine the color of the nearest point to the query point. More generally, given k⩾1, the problem is to determine the color occurring most frequently among the k nearest neighbors. The chromatic version of the nearest neighbor problem is used in many applications in pattern recognition and learning. In this paper we present a simple algorithm for solving the chromatic k nearest neighbor problem. We provide a query sensitive analysis, which shows that if the color classes form spatially well separated clusters (as often happens in practice), then queries can be answered quite efficiently. We also allow the user to specify an error bound ε⩾0, and consider the same problem in the context of approximate nearest neighbor searching. We present empirical evidence that for well clustered data sets, this approach leads to significant improvements in efficiency.

Several image registration methods, based on the scaled-invariant feature transform (SIFT) technique, have appeared recently in the remote sensing literature. All of these methods attempt to overcome problems encountered by SIFT in multi-modal remotely sensed imagery, in terms of the quality of its feature correspondences. The method presented in this paper performs mode seeking (in transformation space) to eliminate outlying corresponding key-points (i.e., features) and improve the overall match obtained. Preliminary experimental results seem to indicate that our method achieves high accuracy and is rather fast in a variety of test cases.

This paper describes the problem of painter classification. We propose solving the problem by using genetic algorithms, which yields very promising results. The proposed methodology combines dimensionality reduction (via image preprocessing) and evolutionary computation techniques, by representing preprocessed data as a chromosome for a genetic algorithm (GA). The preprocessing of our scheme incorporates a diverse set of complex features (e.g., fractal dimension, Fourier spectra coefficients, and texture). The training phase of the GA employs a weighted nearest neighbor (NN) algorithm. We provide initial promising results for the 2- and 3-class cases, which offer significant improvement in comparison to a standard nearest neighbor classifier.

We consider various algorithmic solutions to image registration based on the alignment of a set of feature points. We present a number of enhancements to a branch-and-bound algorithm introduced by Mount, Netanyahu, and Le Moigne (Pattern Recognition, Vol. 32, 1999, pp. 17–38), which presented a registration algorithm based on the partial Hausdorff distance. Our enhancements include a new distance measure, the discrete Gaussian mismatch, and a number of improvements and extensions to the above search algorithm. Both distance measures are robust to the presence of outliers, that is, data points from either set that do not match any point of the other set. We present experimental studies, which show that the new distance measure considered can provide significant improvements over the partial Hausdorff distance in instances where the number of outliers is not known in advance. These experiments also show that our other algorithmic improvements can offer tangible improvements. We demonstrate the algorithm’s efficacy by considering images involving different sensors and different spectral bands, both in a traditional framework and in a multiresolution framework.

This paper presents an integrated approach, combining a state-of-the-art commercial object detection system and genetic algorithms (GA)-based learning for automatic object classification. Specifically, the approach is based on applying weighted nearest neighbor classification to feature vectors extracted from the detected objects, where the weights are evolved due to GA-based learning. Our results demonstrate that this GA-based approach is considerably superior to other standard classification methods.

We present in this paper a number of new enhancements to a branch-and-bound algorithm given by Mount, Netanyahu, and Le Moigne [8] for feature matching. We introduce a new distance measure, which is more robust to the presence of outliers than the previously used measure, as well as a new variant of the search algorithm and a new search strategy. We also present experimental results, which show that these enhancements offer significant tangible improvements in performance.

With the recent explosion in the amount of remotely sensed imagery and the corresponding interest in temporal change detection and modeling, image registration has become increasingly important as a necessary first step in the integration of multi-temporal and multi-sensor data for applications such as the analysis of seasonal and annual global climate changes, as well as land use/cover changes. The task of image registration can be divided into two major components: (1) the extraction of control points or features from images; and (2) the search among the extracted features for the matching pairs that represent the same feature in the images to be matched. Manual control feature extraction can be subjective and extremely time consuming, and often results in few usable points. Automated feature extraction is a solution to this problem, where desired target features are invariant, and represent evenly distributed landmarks such as edges, corners and line intersections. In this paper, we develop a novel automated registration approach based on the following steps. First, a mathematical morphology (MM)-based method is used to obtain a scale-orientation morphological profile at each image pixel. Next, a spectral dissimilarity metric such as the spectral information divergence is applied for automated extraction of landmark chips, followed by an initial approximate matching. This initial condition is then refined using a hierarchical robust feature matching (RFM) procedure. Experimental results reveal that the proposed registration technique offers a robust solution in the presence of seasonal changes and other interfering factors. Keywords-Automated image registration, multi-temporal imagery, mathematical morphology, robust feature matching.

The goal of our project is to develop and evaluate image analysis methodologies for use on the ground or on-board spacecraft particularly spacecraft constellations. Our focus is on developing methods to perform automatic registration and fusion of multisensor data representing multiple spatial, spectral and temporal resolutions, as well as dimension reduction of hyperspectral data. Feature extraction methods such as wavelet decomposition, edge detection and mutual information are combined with feature matching methods such as cross-correlation, optimization, and statistically robust techniques to perform image registration. The approach to image fusion is application-based and involves wavelet decomposition, dimension reduction, and classification methods. Dimension reduction is approached through novel methods based on principal component analysis and wavelet decomposition, and implemented on Beowulf-type parallel architectures. Registration algorithms are tested and compared on several multi-sensor datasets, including one of the EOS Core Sites, the Konza Prairie in Kansas, utilizing four different sensors: IKONOS, Landsat-7/ETM+, MODIS, and SeaWIFS. Fusion methods are tested using Landsat, MODIS and SAR or JERS data. Dimension reduction is demonstrated on A VIRIS hyperspectral data.

We are proposing a novel method that makes it possible to analyze high-dimensional data with arbitrary shaped projected clusters and high noise levels. At the core of our method lies the idea of subspace validity. We map the data in a way that allows us to test the quality of subspaces using statistical tests. Experimental results, both on synthetic and real data sets, demonstrate the potential of our method.

The conventional ordinary least squares (OLS) method of fitting a line to a set of data points is very unreliable when the amount of random noise in the input (such as an image) is significant compared with the amount of data that is correlated with the lane itself. In this paper we present an analytic method of separating the data of interest from the outliers. We assume that the overall data (i.e., the line data plus the noise) can be modeled as a mixture of two statistical distributions. Applying a variant of the method of moments (MoM) to the assumed model yields an analytic estimate of the desired line.

Automated image clustering/classification is a task of considerable importance. To apply this task to remotely sensed imagery, the authors have pursued an unsupervised clustering scheme based on principles of robust (statistical) estimation. A description of the module employed and results obtained are provided.

For many Earth and space science applications, automatic geo-registration at sub-pixel accuracy has become a necessity. In this work, we are focusing on building an operational system, which will provide a sub-pixel accuracy registration of Landsat-5 and Landsat-7 data. The input to our registration method consists of scenes that have been geometrically and radiometrically corrected. Such preprocessed scenes are then geo-registered relative to a database of Landsat chips. The method assumes a transformation composed of a rotation and a translation, and utilizes rotation- and translation-invariant wavelets to extract image features that are matched using statistically robust feature matching and a partial Hausdorff distance metric. The registration process is described and results on four Landsat input scenes of the Washington, D.C., area are presented.

The goal of our project is to build an operational system, which will provide a sub-pixel accuracy registration of Landsat-5 and Landsat-7 data. Integrated within an automated mass processing/analysis system for Landsat data (REALM), the input to our registration method consists of scenes that have been geometrically and radiometrically corrected, as well as pre-processed for the detection of clouds and cloud shadows. Such pre-processed scenes are then geo-registered relative to a database of Landsat chips. This paper describes our registration process, including the use of a database of landmark chips, a feature extraction performed by an overcomplete wavelet representation, and feature matching using statistically robust techniques. Knowing the approximate longitudes and latitudes of the four corners of the scene, a subset of chips which represent landmarks included in the scene are extracted from the database. For each of these selected landmark chips, a corresponding window is extracted from the scene, and each chip-window pair is registered using our robust wavelet feature matching. First results and future directions are presented in the paper.

As the amount of multidimensional remotely sensed data is growing tremendously, Earth scientists need more efficient ways to search and analyze such data. In particular, extracting image content is emerging as one of the most powerful tools to perform data mining. One of the most promising methods to extract image content is image classification, which provides a labeling of each pixel in the image. In this paper, we concentrate on neural classifiers and show how information obtained through wavelet transform can be integrated in such a classifier. After a systematic dimensionality reduction by a principal component analysis technique, we apply a local spatial frequency analysis. This local analysis with a composite edge/texture wavelet transform provides statistical texture information of the landsat imagery testset. The network is trained with both radiometric landsat/thematic mapper bands and with the additional texture bands provided by the wavelet analysis. The paper describes the type of wavelets chosen for this application, and several sets of results are presented.© (1997) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

As the amount of multidimensional remotely sensed data is growing tremendously, Earth scientists need more efficient ways to search and analyze such data. In particular, extracting image content is emerging as one of the most powerful tools to perform data mining. One of the most promising methods to extract image content is image classification, which provides a labeling of each pixel in the image. In this paper, we concentrate on neural network classifiers and we show how information obtained with a wavelet transform can be integrated to such a classifier. We apply a local spatial frequency analysis, a wavelet transform, to account for statistical texture information in Landsat/TM imagery. Statistical texture is extracted with a continuous edge-texture composite wavelet transform. We show how this approach relates to texture information computed from a co-occurrence matrix. The network is then trained with both the texture information and the additional pixel labels provided by the ground truth data. Theory and preliminary results are described in this paper. In future work, this pre-processing will automatically generate a texture index which will be fed into the artificial neural network, thus providing better discriminants of Landsat/TM imagery.

In this paper, we investigate how to further exploit the various characteristics of mean shift, in an attempt to achieve a robust and efficient clustering module for remotely sensed data. A mean shift algorithm has shown o be promising in various image-processing applications, specifically in cluster analysis.