The increasing scale and complexity of remote sensing observations demand distributed processing to effectively manage the vast volumes of data generated. However, distributed processing presents significant challenges, including bandwidth limitations, high latency, and privacy concerns, especially when transmitting high-resolution images. To address these issues, we propose a novel scheme leveraging the encoder of a masked autoencoder to generate associated embedding (CLS tokens) from masked images which enable training deep learning models under Federated Learning scenarios. This approach enables the transmission of compact image patches instead of full images to processing nodes, drastically reducing bandwidth usage. On the processing nodes, classifiers are trained with the CLS tokens, and model weights are aggregated using FedAvg and FedProx Federated Learning algorithms. Experimental results on benchmark datasets demonstrate that the proposed approach significantly reduces data transmission requirements while maintaining and even surpassing the accuracy of systems with access to full data.

The substantial volume of continuously gathered remote sensing data can serve as a valuable information source in mitigating the effects of natural disasters. This involves identifying changes in the time series of observations. Considering that the precise location of the changes may not be available in real-world scenarios, we propose an unsupervised method for detecting extreme events in multi-temporal satellite images. Specifically, we learn a basis matrix of each dimension of the feature space of the images using the tensor decomposition learning method. Then, each new image is represented in the feature space by expressing it as a multilinear combination of the learned tensor decomposition factors. The predicted changes can be obtained by comparing and thresholding the distance of the corresponding extracted features of the images before and after the event. Experimental results on real Sentinel-2 multi-temporal images demonstrate that the proposed method can efficiently detect the effects of fires and floods with low complexity.

Accurate assessment of areas burned in wildfires is vital for various monitoring, management, and spread modeling applications. Wildfires, especially in forested regions, pose immense challenges for precise mapping due to the inherent dynamics of fuel types and terrain complexities. While remote sensing, particularly satellite imagery, offers an approach to studying burned areas, reliance on such satellite sources introduces challenges in characterizing burned areas amidst dense vegetation and environmental variations. This paper presents a mapping of forested burned areas utilizing global navigation satellite system–reflectometry (GNSS-R) from Cyclone Global Navigation Satellite System (CYGNSS) with ancillary observations from Soil Moisture Active Passive (SMAP) mission and Shuttle Radar Topography Mission (SRTM) using machine learning approaches. We validate the results with existing burned area products and provide maps of representative California fires within CYGNSS coverage. Assimilation of GNSS-R data into the model provides near real-time and high temporal resolution, enabling rapid response and mitigation efforts to fire events.

Soil moisture is an essential climate variable that directly influences many hydrological, agricultural, and water-cycle processes. Many satellites have been launched and are still being launched to map soil moisture accurately at a global scale. Motivated by the coarse resolution of existing satellite products, many statistical, physics-based, and machine learning-based methods have been proposed to downscale soil moisture to much finer spatial scales. In this paper, we propose a novel deep learning approach that is constrained by the Tau-omega radiative transfer model to enhance the resolution of surface soil moisture. We demonstrate the proposed framework by downscaling Soil Moisture Active Passive (SMAP) L-band brightness temperature (TB) with C-band synthetic aperture radar (SAR) backscattering coefficient (σ0) imagery from Sentinel-1A/B and subsequently retrieving high spatial resolution (1km) soil moisture.

While microwave imaging satellites, such as the NASA Soil Moisture Active Passive (SMAP), can provide reliable estimates of surface soil moisture at km resolution, the temporal frequency of observations is on the order of days. To increase the temporal frequency of observations, a new class of approaches considers global navigation satellite system (GNSS)-reflectometry (GNSS-R) signals. In this work, we consider observations from the NASA Cyclone GNSS (CYGNSS) constellation, as well as auxiliary observations, and seek to provide instantaneous soil moisture estimates. To achieve accurate retrievals, a novel machine learning approach for probabilistic regression is considered, namely the NGBoost. In addition to achieving an accuracy comparable to previous approaches employing state-of-the-art machine learning methods, the considered framework also provides prediction intervals to quantify prediction uncertainty. Using observations from the Yanco SMAP core validation site in southeast Australia over a period of three years, we quantify the performance in terms of both retrieval accuracy and associated uncertainty. Furthermore, using noisy observations, we experimentally demonstrate the impact of input noise on the prediction uncertainty.

Retrieval of Surface Soil Moisture (SSM) over large scales at high spatial resolution is crucial for numerous applications. Existing solutions rely on the analysis of remote sensing platforms or in situ measurements that are either too coarse in their resolution or too localized to address the aforementioned need. In this work, we propose a novel deep learning approach for reliably estimating SSM at a high spatial resolution of 1 km over broad regions. To achieve this objective, the proposed framework employs a Convolutional Neural Network that can capture both multi-modal and spatial correlations. Introducing a novel loss function, the proposed scheme can leverage limited in situ observations while also generating estimates consistent with physical models. This is achieved through the utilization of coarse-resolution data assimilation estimates. For training and assessing the performance of the proposed framework, a novel dataset is generated by combining information from remote sensing, in situ measurements, and data assimilation estimates. Experimental analysis demonstrates that the proposed approach can provide accurate retrieval of SSM, significantly outperforming existing products.

Numerous satellite sources measure global soil moisture, but the spatial resolution of these products is usually coarse making it challenging to use them for various science applications. In this work, we develop a conditional generative adversarial network-based model to downscale satellite measured brightness temperature. Augmenting a physics-based Tau-omega model to the framework, high resolution soil moisture is retrieved. For training and validation, upscaled in-situ soil moisture measurements are utilized. Experimental results suggests that the proposed framework captures the soil moisture variations from in-situ sensors and helps in substantially improving the resolution of passive microwave satellite soil moisture maps.

Multi-temporal remotely sensed observations acquired by multi-spectral sensors contain a wealth of information related to the Earth’s state. Deep learning methods have demonstrated a great potential in analyzing such observations. Traditional 2D and 3D approaches are unable to effectively extract valuable information encoded across all available dimensions. In this work, we propose the extension of current fully-convolutional models for multi-temporal remote sensing data classification to their high-dimensional analogs, which can naturally capture multi-dimensional dependencies and correlations. Experimental analysis on recently released observations from Landsat-8 reveals that our proposed high-dimensional fully-convolutional approach exhibits a clear classification performance improvement over state of the art low-order fully-convolutional models.

In this paper we develop a Convolutional Long Short-term memory (ConvLSTM) model, a time series deep learning neural network, to predict soil moisture, with an add-on module of passive microwave (radiometer) soil moisture retrieval using the Tau omega model. We incorporate antecedent observations, landscape properties, and forcing factors such as precipitation, landcover, clay fraction, and brightness temperature in the prediction scheme. A regularization Monte Carlo Dropout layer is added to the network to remove stochasticity and avoid overfitting during the training phase. This dropout layer also provides a Bayesian approximation to quantify uncertainty during forecasting. The model is validated at four Soil Moisture Active Passive (SMAP) Cal/Val locations using performance metrics such as Root Mean Square Error (RMSE) and Bias to evaluate effectiveness of the proposed method. This model is developed as a component of the Science Simulator within the Distributed Spacecraft with Heuristic Intelligence to Enable Logistical Decisions (D-SHIELD) project.

This work focuses on the problem of surface soil moisture estimation from multi-modal remote sensing observations. We focus on the scenario where both passive radiometer observations from NASA SMAP satellite, as well as active radar measurements from ESA Sentinel 1 are available. We formulate the problem as multi-source observation fusion and develop a deep learning model for SM estimation. To train and validate the performance of the proposed scheme, we consider observations from in-situ SM sensor networks over the continental USA. Experimental results demonstrate that the proposed model achieves high quality SM estimation, surpassing the performance of available products.

Nowadays, multidimensional data structures, known as tensors, are widely used in many applications like earth observation from remote sensing image sequences. However, the increasing spatial, spectral and temporal resolution of the acquired images, introduces considerable challenges in terms of data storage and transfer, making critical the necessity of an efficient compression system for high dimensional data. In this paper, we propose a tensor-based compression algorithm that retains the structure of the data and achieves a high compression ratio. Specifically, our method learns a dictionary of specially structured tensors using the Alternating Direction Method of Multipliers, as well as a symbol encoding dictionary. During run-time, a quantized and encoded sparse vector of coefficients is transmitted, instead of the whole multidimensional signal. Experimental results on real satellite image sequences demonstrate the efficacy of our method compared to a state-of-the-art compression method.

High-dimensional data structures, known as tensors, are fundamental in many applications, including multispectral imaging and color video processing. Compression of such huge amount of multidimensional data collected over time is of paramount importance, necessitating the process of quantization of measurements into discrete values. Furthermore, noise and issues related to the acquisition and transmission of signals frequently lead to unobserved, lost or corrupted measurements. In this paper, we introduce a tensor robust principal component analysis algorithm in order to recover a tensor with real-valued entries from a partly observed set of quantized and sparsely corrupted entries. We formulate the problem as a constrained maximum likelihood estimation of the sum of a low-rank tensor and a sparse tensor, through matricizations in each mode, in combination with a quantization and statistical measurement model. Experimental results on satellite derived land surface time-series demonstrate that directly operating with the quantized measurements, rather than treating them as real values, results in a low recovery error, while the proposed method is also capable of detecting temperature anomalies (e.g., forest fires).

In this work, we propose a novel Convolutional Neural Network architecture for increasing the low spatial resolution SMAP radiometer based soil moisture estimations from 36 km to 3 km resolution by using time-series of observations from both SMAP's radiometer and Sentinel-1 radar. By simultaneously extracting features from both current low-resolution input and residuals between high and low resolution at previous time instances, the proposed network is capable of accurately estimating soil moisture using coarse resolution observations. Experimental results on three different locations demonstrate that the proposed scheme is able to estimate soil moisture with accuracy in the range of the requirements set by the SMAP science team.

In this paper, we address the problem of spectroscopic redshift estimation in Astronomy. Due to the expansion of the Universe, galaxies recede from each other on average. This movement causes the emitted electromagnetic waves to shift from the blue part of the spectrum to the red part, due to the Doppler effect. Redshift is one of the most important observables in Astronomy, allowing the measurement of galaxy distances. Several sources of noise render the estimation process far from trivial, especially in the low signal-to-noise regime of many astrophysical observations. In recent years, new approaches for a reliable and automated estimation methodology have been sought out, in order to minimize our reliance on currently popular techniques that heavily involve human intervention. The fulfilment of this task has evolved into a grave necessity, in conjunction with the insatiable generation of immense amounts of astronomical data. In our work, we introduce a novel approach based on Deep Convolutional Neural Networks. The proposed methodology is extensively evaluated on a spectroscopic dataset of full spectral energy galaxy distributions, modelled after the upcoming Euclid satellite galaxy survey. Experimental analysis on observations of idealistic and realistic conditions demonstrate the potent capabilities of the proposed scheme.

Bracketed High Dynamic Range (HDR) imaging architectures acquire a sequence of Low Dynamic Range (LDR) images in order to either produce a HDR image or an “optimally” exposed LDR image, achieving impressive results under static camera and scene conditions. However, in real world conditions, ghost-like artifacts and noise effects limit the quality of HDR reconstruction. We address these limitations by introducing a post-acquisition snapshot HDR enhancement scheme that generates a bracketed sequence from a small set of LDR images, and in the extreme case, directly from a single exposure. We achieve this goal via a sparse-based approach where transformations between differently exposed images are encoded through a dictionary learning process, while we learn appropriate features by employing a stacked sparse autoencoder (SSAE) based framework. Via experiments with real images, we demonstrate the improved performance of our method over the state-of-the-art, while our single-shot based HDR formulation provides a novel paradigm for the enhancement of LDR imaging and video sequences.

This paper presents a method for reducing the Diffusion Weighted Magnetic Resonance Imaging (DW-MRI) examination time based on the mathematical framework of sparse representations. The aim is to undersample the b-values used for DW-MRI image acquisition which reflect the strength and timing of the gradients used to generate the DW-MRI images since their number defines the examination time. To test our method we investigate whether the undersampled DW-MRI data preserve the same accuracy in terms of extracted imaging biomarkers. The main procedure is based on the use of the k-Singular Value Decomposition (k-SVD) and the Orthogonal Matching Pursuit (OMP) algorithms, which are appropriate for the sparse representations computation. The presented results confirm the hypothesis of our study as the imaging biomarkers extracted from the sparsely reconstructed data have statistically close values to those extracted from the original data. Moreover, our method achieves a low reconstruction error and an image quality close to the original.

High resolution astronomical imagery plays a critical role in multiple remote sensing applications. In this work, we introduce a novel post-acquisition computational technique aiming to recover the high-quality versions of blurry and degraded astronomical observations. Additionally, the proposed scheme is able to retrieve significant information regarding the characteristic properties of the blurring kernel, i.e point spread function (PSF). In order to accomplish this goal, we exploit the mathematical frameworks of Sparse Reprentations, and the Alternating Direction Method of Multipliers (ADMM). Experimental results demonstrate the ability of the proposed approach to synthesize high-quality astronomical imagery.

Convolutional Neural Networks (CNNs) represent the cutting edge in signal analysis tasks like classification and regression. Realization of such architectures in hardware capable of performing high throughput computations, with minimal energy consumption, is a key enabling factor towards the proliferation of analysis immediately after acquisition. Our driving problem is a satellite-based remote sensing platform in which onboard signal processing and classification tasks must take place, given strict bandwidth and energy limitations. In this work, we exploit the implementation of a CNN on Field Programmable Gate Array (FPGA) platforms and explore different ways to minimize the impact of different hardware restrictions to performance. We compare our results against competing technologies such as Graphics Processing Units (GPU) in terms of throughput, latency and energy consumption. In actual experimental runs we demonstrate competitive latency and throughput of the FPGA platform vs. GPU technology at an order-of-magnitude energy savings, which is especially important for space-borne computing.

Land cover classification is a flourishing research topic in the field of remote sensing. Conventional methodologies mainly focus either on the simplified single-label case or on the pixel-based approaches that cannot efficiently handle high-resolution images. On the other hand, the problem of multilabel land cover scene categorization remains, to this day, fairly unexplored. While deep learning and convolutional neural networks have demonstrated an astounding capacity at handling challenging machine learning tasks, such as image classification, they exhibit an underwhelming performance when trained with a limited amount of annotated examples. To overcome this issue, this paper proposes a data augmentation technique that can drastically increase the size of a smaller data set to copious amounts. Our experiments on a multilabel variation of the UC Merced Land Use data set demonstrate the potential of the proposed methodology, which outperforms the current state of the art by more than 6% in terms of the F-score metric.

High-spectral resolution imaging systems play a critical role in the identification and characterization of objects in a scene of interest. Unfortunately, multiple factors impair spectral resolution, as in the case of modern snapshot spectral imagers that associate each hyperpixel with a specific spectral band. In this paper, we introduce a novel postacquisition computational technique aiming to enhance the spectral dimensionality of imaging systems by exploiting the mathematical frameworks of sparse representations and dictionary learning. We propose a coupled dictionary learning model which considers joint feature spaces, composed of low- and high-spectral resolution hypercubes, in order to achieve spectral superresolution performance. We formulate our spectral coupled dictionary learning optimization problem within the context of the alternating direction method of multipliers, and we manage to update the involved quantities via closed-form expressions. In addition, we consider a realistic spectral subsampling scenario, taking into account the spectral response functions of different satellites. Moreover, we apply our spectral superresolution algorithm on real satellite data acquired by Landsat-8 and Sentinel-2 sensors. Finally, we have investigated the problem of hyperspectral image unmixing using the recovered high-spectral resolution data cube, and we are able to demonstrate that the proposed scheme provides significant value in hyperspectral image understanding techniques. Experimental results demonstrate the ability of the proposed approach to synthesize high-spectral-resolution 3-D hypercubes, achieving better performance compared to state-of-the-art resolution enhancement methods.

Bleed-through is a pervasive degradation in ancient documents, caused by the ink of the opposite side of the sheet that has seeped through the paper fiber, and appears as an extra, interfering text. Bleed-through severely impairs document readability and makes it difficult to decipher the contents. Digital image restoration techniques have been successfully employed to remove or significantly reduce this distortion. The main theme is to identify the bleedthrough pixels and estimate an appropriate replacement for them, in accordance to their surrounding. This paper proposes a two-step image restoration method, exploiting information from the recto and verso images. First, based on a non-stationary linear model of the two texts overlapped in the recto-verso pair, the bleed through pixels are identified. In the second step, a sparse representation based image inpainting technique, with a non-negative sparsity constraint, is used to find an appropriate replacement for the bleedthough pixels. Thanks to the power of dictionary learning and sparse image reconstruction methods, the natural texture of the background is well reproduced in the bleed-through areas, and even a their possible overestimation is effectively corrected, so that the original appearance of the document is preserved. The experiments are conducted on the images of a popular database of ancient documents, and the results validate the performance of the proposed method compared to the state of the art.

Sensor-based activity recognition is encountered in innumerable applications of the arena of pervasive healthcare and plays a crucial role in biomedical research. Nonetheless, the frequent situation of unobserved measurements impairs the ability of machine learning algorithms to efficiently extract context from raw streams of data. In this paper, we study the problem of accurate estimation of missing multimodal inertial data and we propose a classification framework that considers the reconstruction of subsampled data during the test phase. We introduce the concept of forming the available data streams into low-rank two-dimensional (2-D) and 3-D Hankel structures, and we exploit data redundancies using sophisticated imputation techniques, namely matrix and tensor completion. Moreover, we examine the impact of reconstruction on the classification performance by experimenting with several state-of-the-art classifiers. The system is evaluated with respect to different data structuring scenarios, the volume of data available for reconstruction, and various levels of missing values per device. Finally, the tradeoff between subsampling accuracy and energy conservation in wearable platforms is examined. Our analysis relies on two public datasets containing inertial data, which extend to numerous activities, multiple sensing parameters, and body locations. The results highlight that robust classification accuracy can be achieved through recovery, even for extremely subsampled data streams.

Obtaining an up-to-date high-resolution description of land cover is a challenging task due to the high cost and labor-intensive process of human annotation through field studies. This work introduces a radically novel approach for achieving this goal by exploiting the proliferation of remote sensing satellite imagery, allowing for the up-to-date generation of global-scale land cover maps. We propose the application of multilabel classification, a powerful framework in machine learning, for inferring the complex relationships between the acquired satellite images and the spectral profiles of different types of surface materials. Introducing a drastically different approach compared to unsupervised spectral unmixing, we employ contemporary ground-collected data from the European Environment Agency to generate the label set and multispectral images from the MODIS sensor to generate the spectral features, under a supervised classification framework. To validate the merits of our approach, we present results using several state-of-the-art multilabel learning classifiers and evaluate their predictive performance with respect to the number of annotated training examples, as well as their capability to exploit examples from neighboring regions or different time instances. We also demonstrate the application of our method on hyperspectral data from the Hyperion sensor for the urban land cover estimation of New York City. Experimental results suggest that the proposed framework can achieve excellent prediction accuracy, even from a limited number of diverse training examples, surpassing state-of-the-art spectral unmixing methods.

Fingerprint-based location sensing technologies play an increasingly important role in pervasive computing applications due to their accuracy and minimal hardware requirements. However, typical fingerprint-based schemes implicitly assume that communication occurs over the same channel (frequency) during the training and the runtime phases. When this assumption is violated, the mismatches between training and runtime fingerprints can significantly deteriorate the localization performance. Additionally, the exhaustive calibration procedure required during training limits the scalability of this class of methods. In this work, we propose a novel, scalable, multi-channel fingerprint-based indoor localization system that employs modern mathematical concepts based on the Sparse Representations and Matrix Completion theories. The contribution of our work is threefold. First, we investigate the impact of channel changes on the fingerprint characteristics and the effects of channel mismatch on state-of-the-art localization schemes. Second, we propose a novel fingerprint collection technique that significantly reduces the calibration time, by formulating the map construction as an instance of the Matrix Completion problem. Third, we propose the use of sparse Bayesian learning to achieve accurate location estimation. Experimental evaluation on real data highlights the superior performance of the proposed framework in terms of reconstruction error and localization accuracy.

This paper addresses the problem of efficient sampling and reconstruction of sparse spike signals, which have been convolved with low-pass filters. A modified compressed sensing (CS) framework is proposed, termed dictionary-based deconvolution CS (DDCS) to achieve this goal. DDCS builds on the assumption that a low-pass filter can be represented sparsely in a dictionary of blurring atoms. Identification of both the sparse spike signal and the sparsely parameterized blurring function is performed by an alternating scheme that minimizes each variable independently, while keeping the other constant. Simulation results reveal that the proposed DDSS scheme achieves an improved reconstruction performance when compared to traditional CS recovery.

Dynamic recognition of gestures from video sequences is a challenging task due to the high variability in the characteristics of each gesture with respect to different individuals. In this work, we propose a novel representation of gestures as linear combinations of the elements of an overcomplete dictionary, based on the emerging theory of sparse representations. We evaluate our approach on a publicly available gesture dataset of Palm Grafti Digits and compare it with other state-of-the-art methods, such as Hidden Markov Models, Dynamic Time Warping and the recently proposed distance metric termed Move-Split-Merge. Our experimental results suggest that the proposed recognition scheme offers high recognition accuracy in isolated gesture recognition and a satisfying robustness to noisy data, thus indicating that sparse representations can be successfully applied in the field of gesture recognition.

In this paper, we address the problem of video classification from a set of compressed features. In particular, the properties of linear random projections in the framework of compressive sensing are exploited to reduce the task of classifying a given video sequence into a problem of sparse reconstruction, based on feature vectors consisting of measurements lying in a low-dimensional compressed domain. This can be of great importance in decision systems with limited power, processing, and bandwidth resources, since the classification is performed without handling the original high-resolution video data, but working directly with the set of compressed measurements. The experimental evaluation verifies the efficiency of the proposed scheme and illustrates that the compressed measurements in conjunction with an appropriate decision rule result in an effective video classification scheme, which meets the constraints of systems with limited resources.

Research on facial expression recognition has steadily been moving from analysis of deliberative frontal expressions to analysis of unconstrained spontaneous expressions. This shift has spawned complex 3D models and computationally expensive geometric methods that prevent usage on resource constrained platforms such as smart phones. This paper presents manifold learning techniques for accurate multi-view facial expression on low resolution 2D images. Our results indicate that mixed class local pose and expression manifold methods perform better than global expression techniques and work just as well as fusing together results from multiple manifolds.

Tracking an object without any prior information regarding its appearance is a challenging problem. Modern tracking algorithms treat tracking as a binary classification problem between the object class and the background class. The binary classifier can be learned offline, if a specific object model is available, or online, if there is no prior information about the object's appearance. In this paper, we propose the use of online distance metric learning in combination with nearest neighbor classification for object tracking. We assume that the previous appearances of the object and the background are clustered so that a nearest neighbor classifier can be used to distinguish between the new appearance of the object and the appearance of the background. In order to support the classification, we employ a distance metric learning (DML) algorithm that learns to separate the object from the background. We utilize the first few frames to build an initial model of the object and the background and subsequently update the model at every frame during the course of tracking, so that changes in the appearance of the object and the background are incorporated into the model. Furthermore, instead of using only the previous frame as the object's model, we utilize a collection of previous appearances encoded in a template library to estimate the similarity under variations in appearance. In addition to the utilization of the online DML algorithm for learning the object/background model, we propose a novel feature representation of image patches. This representation is based on the extraction of scale invariant features over a regular grid coupled with dimensionality reduction using random projections. This type of representation is both robust, capitalizing on the reproducibility of the scale invariant features, and fast, performing the tracking on a reduced dimensional space. The proposed tracking algorithm was tested under challenging conditions and achieved state-of-the art performance.

In this paper, we propose the combination of manifold learning and distance metric learning for the generation of a representation that is both discriminative and informative, and we demonstrate that this approach is effective for face recognition. Initial dimensionality reduction is achieved using random projections, a computationally efficient and data independent linear transformation. Distance metric learning is then applied to increase the separation between classes and improve the accuracy of nearest neighbor classification. Finally, a manifold learning method is used to generate a mapping between the randomly projected data and a low dimensional manifold. Face recognition results suggest that the combination of distance metric learning and manifold learning can increase performance. Furthermore, random projections can be applied as an initial step without significantly affecting the classification accuracy.

Face detection is a key component in numerous computer vision applications. Most face detection algorithms achieve real-time performance by some form of dimensionality reduction of the input data, such as Principal Component Analysis. In this paper, we are exploring the emerging method of Random Projections (RP), a data independent linear projection method, for dimensionality reduction in the context of face detection. The benefits of using random projections include computational efficiency that can be obtained by implementing matrix multiplications with a small number of integer additions or subtractions. The computational savings are of great significance in resource constrained environments, such as wireless video sensor networks. Experimental results suggest that RP can achieve performance that is comparable to that obtained with traditional dimensionality reduction techniques for face detection using support vector machines.

The multiple description coding framework is well-suited for the transmission of images in the presence of transmission errors. However, it has been mainly explored for idealized channels, where packets are either received correctly or not at all. In more realistic scenarios, bitstreams are affected by bit-level noise and error correcting codes must be employed. In the proposed scheme, the input image is decomposed into wavelet coefficients that are encoded using lattice vector quantization for the generation of multiple descriptions. Low-density parity-check (LDPC) codes are used for the protection of each stream, while joint decoding exploits the correlation between the multiple descriptions. The proposed joint source-channel decoding scheme can provide robust transmission of images by introducing a controlled amount of redundancy that can be used for error recovery. Results are presented for image transmission over binary symmetric and Gaussian channels.