Welcome to my personal website, where you can find my latest research, teaching tools and news.
Coming from Italy, I graduated in telecommunications with the University of Trento, where I received both my Bachelor's and Master's degree.
I decided to do my Erasmus stay with the University of Iceland, in the end staying for longer than three years and earning the title of PhD.
Since then, I've been a research assistant at JSC - Jülich Supercomputing Centre. My area of expertise is Remote Sensing (RS), High Performance Computing (HPC), and Machine Learning (ML).
You can contact me at
g.cavallaro@fz-juelich.de
Earth Observation has evolved dramatically in the last decades due to the technological advances incorporated into remote sensing instruments in the optical and microwave domains. Remote sensing data are now used in a wide-range of applications, aimed at monitoring and implementing new policies in the fields of agriculture, assessment of environmental resources, urban planning, defense, disaster management, etc.
Machine (deep) learning for big Earth Observation applications is going through a major revolution due to the unique convergence of easy access to large volumes of open remote sensing data and the large-scale computing capability.
Due to the insufficient memory size and number of cores available in commodity computers on the one hand and on the other hand the increasing number of applications that require data computing in near real time (i.e., supporting decision-makers), remote sensing data processing pipelines necessitate the use of algorithms that can run and scale on High-Performance Computing systems.
Here you can find my recent talks.
All talks
This webinar will introduce the new working group “High-performance and Disruptive Computing in Remote Sensing” (HDCRS) of the GRSS Earth Science Informatics Technical Committee (ESI TC).
High Performance Computing (HPC) has recently been attracting more attention in remote sensing applications due to the challenges posed by the increased amount of open data that are acquired daily by Earth Observation programs. The unique parallel computing environments and programming techniques that are integrated in HPC systems are able to solve large-scale problems such as the training of classification algorithms with large amounts of remote sensing data. This webinar will explain how to distribute the training of deep neural networks with parallel implementation techniques on HPC systems that include a large number of Graphics Processing Units. To show that distributed training can drastically reduce the training time and preserve the accuracy performance, the webinar will present recent experimental results performed on the HPC systems at the Jülich Supercomputing Centre
This work proposes a novel distributed deep learning model for Remote Sensing (RS) images super-resolution. High Performance Computing (HPC) systems with GPUs are used to accelerate the learning of the unknown low to high resolution mapping from large volumes of Sentinel-2 data. The proposed deep learning model is based on self-attention mechanism and residual learning. The results demonstrate that stateof- the-art performance can be achieved by keeping the size of the model relatively small. Synchronous data parallelism is applied to scale up the training process without severe performance loss. Distributed training is thus shown to speed up learning substantially while keeping performance intact.
Download my latest teaching materials
All courses
Invited lecture on ''Deep Learning driven by Big Data'' in the Cloud Computing and Big Data course at the University of Iceland. The lecture first introduces to the limitations of a linear perceptron model and how Artificial Neural Networks (ANNs) can solve non-linearly separable data sets. Then, it explains how Convolutional Neural Networks (CNNs) can solve image recognition tasks. Finally it is demonstrates how to tune the parameters of the networks with the backpropagation algorithm.
Recent advances in remote sensors with higher spectral, spatial, and temporal resolutions have significantly increased data volumes, which pose a challenge to process and analyse the resulting massive data in a timely fashion to support practical applications. Meanwhile, the development of computationally demanding Machine Learning (ML) and Deep Learning (DL) techniques (e.g., deep neural networks with massive amounts of tunable parameters) demand for parallel algorithms with high scalability performance. Therefore, data intensive computing approaches have become indispensable tools to deal with the challenges posed by applications from geoscience and remote sensing. In recent years, high-performance and distributed computing have been rapidly advanced in terms of hardware architectures and software. For instance, the popular graphics processing unit (GPU) has evolved into a highly parallel many-core processor with tremendous computing power and high memory bandwidth. Moreover, recent High Performance Computing (HPC) architectures and parallel programming have been influenced by the rapid advancement of DL and hardware accelerators as modern GPUs. This and more in my course on Scalable Machine Learning for Remote Sensing Big Data.
Deep Learning is emerging as the leading AI technique owing to the current convergence of scalable computing capability (i.e., HCP and Cloud computing), easy access to large volumes of data, and the emergence of new algorithms enabling robust training of large-scale deep neural networks. The tutorial aims at providing a complete overview for an audience that is not familiar with these topics.
Here you can find recent events and news.
All news and positions
This special session collects papers in the most advanced and trendy areas interested in exploiting new high-performance and distributed computing technologies and algorithms to expedite the processing and analysis of big remote sensing data.
Following the success of the 11th IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS 2020), this special issue is to provide an updated snapshot of recent advances in the use of pattern recognition techniques to address the challenges in new-generation remote sensing data.
Recent advances in satellite technology have led to a regular, frequent, and high-resolution monitoring of Earth at the global scale, providing an unprecedented amount of Earth observation (EO) data. This Special Issue aims at gathering a collection of papers in areas interested in learning from data with applications to remote sensing image analysis.
Read my latest publications
All conference papers
G. Cavallaro, D. Willsch, M. Willsch, K. Michielsen and M. Riedel
Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS)
Support Vector Machine (SVM) is a popular supervised Machine Learning (ML) method that is widely used for classification and regression problems. Recently, a method to train SVMs on a D-Wave 2000Q Quantum Annealer (QA) was proposed for binary classification of some biological data. First, ensembles of weak quantum SVMs are generated by training each classifier on a disjoint training subset that can be fit into the QA. Then, the computed weak solutions are fused for making predictions on unseen data. In this work, the classification of Remote Sensing (RS) multispectral images with SVMs trained on a QA is discussed. Furthermore, an open code repository is released to facilitate an early entry into the practical application of QA, a new disruptive compute technology.
R. Sedona, G. Cavallaro, J. Jitsev, A. Strube, M. Riedel and M. Book
Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS)
Similarly to other scientific domains, Deep Learning (DL) holds great promises to fulfil the challenging needs of Remote Sensing (RS) applications. However, the increase in volume, variety and complexity of acquisitions that are carried out on a daily basis by Earth Observation (EO) missions generates new processing and storage challenges within operational processing pipelines. The aim of this work is to show that High-Performance Computing (HPC) systems can speed up the training time of Convolutional Neural Networks (CNNs). Particular attention is put on the monitoring of the classification accuracy that usually degrades when using large batch sizes. The experimental results of this work show that the training of the model scales up to a batch size of 8,000, obtaining classification performances in terms of accuracy in line with those using smaller batch sizes.
R. Zhang, G. Cavallaro and J. Jitsev
Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS)
This work proposes a novel distributed deep learning model for RS images super-resolution. HPC systems with GPUs are used to accelerate the learning of the unknown low to high resolution mapping from large volumes of Sentinel-2 data. The proposed deep learning model is based on self-attention mechanism and residual learning. The results demonstrate that state-of-the-art performance can be achieved by keeping the size of the model relatively small. Synchronous data parallelism is applied to scale up the training process without severe performance loss. Distributed training is thus shown to speed up learning substantially while keeping performance intact.
R. Sedona, L. Hoffmann, R. Spang, G. Cavallaro, S. Griessbach, M. Höpfner, M. Book, M. Riedel
Atmospheric Measurement Techniques
Polar stratospheric clouds (PSCs) play a key role in polar ozone depletion in the stratosphere. Improved observations and continuous monitoring of PSCs can help to validate and improve chemistry–climate models that are used to predict the evolution of the polar ozone hole. In this paper, we explore the potential of applying machine learning (ML) methods to classify PSC observations of infrared limb sounders. Two datasets were considered in this study. The first dataset is a collection of infrared spectra captured in Northern Hemisphere winter 2006/2007 and Southern Hemisphere winter 2009 by the Michelson Interferometer for Passive Atmospheric Sounding (MIPAS) instrument on board the European Space Agency's (ESA) Envisat satellite. The second dataset is the cloud scenario database (CSDB) of simulated MIPAS spectra. We first performed an initial analysis to assess the basic characteristics of the CSDB and to decide which features to extract from it. Here, we focused on an approach using brightness temperature differences (BTDs). From both the measured and the simulated infrared spectra, more than 10 000 BTD features were generated. Next, we assessed the use of ML methods for the reduction of the dimensionality of this large feature space using principal component analysis (PCA) and kernel principal component analysis (KPCA) followed by a classification with the support vector machine (SVM). The random forest (RF) technique, which embeds the feature selection step, has also been used as a classifier. All methods were found to be suitable to retrieve information on the composition of PSCs. Of these, RF seems to be the most promising method, being less prone to overfitting and producing results that agree well with established results based on conventional classification methods.
R. Sedona, G. Cavallaro, J. Jitsev, A. Strube, M. Riedel, J. A. Benediktsson
Remote Sensing
High-Performance Computing (HPC) has recently been attracting more attention in remote sensing applications due to the challenges posed by the increased amount of open data that are produced daily by Earth Observation (EO) programs. The unique parallel computing environments and programming techniques that are integrated in HPC systems are able to solve large-scale problems such as the training of classification algorithms with large amounts of Remote Sensing (RS) data. This paper shows that the training of state-of-the-art deep Convolutional Neural Networks (CNNs) can be efficiently performed in distributed fashion using parallel implementation techniques on HPC machines containing a large number of Graphics Processing Units (GPUs). The experimental results confirm that distributed training can drastically reduce the amount of time needed to perform full training, resulting in near linear scaling without loss of test accuracy.
J. M. Haut, J. A. Gallardo, M. E. Paoletti, G. Cavallaro, J. Plaza, A. Plaza and M. Riedel
IEEE Transactions on Geoscience and Remote Sensing
Advances in remote sensing hardware have led to a significantly increased capability for high-quality data acquisition, which allows the collection of remotely sensed images with very high spatial, spectral, and radiometric resolution. This trend calls for the development of new techniques to enhance the way that such unprecedented volumes of data are stored, processed, and analyzed. An important approach to deal with massive volumes of information is data compression, related to how data are compressed before their storage or transmission. For instance, hyperspectral images (HSIs) are characterized by hundreds of spectral bands. In this sense, high-performance computing (HPC) and high-throughput computing (HTC) offer interesting alternatives. Particularly, distributed solutions based on cloud computing can manage and store huge amounts of data in fault-tolerant environments, by interconnecting distributed computing nodes so that no specialized hardware is needed. This strategy greatly reduces the processing costs, making the processing of high volumes of remotely sensed data a natural and even cheap solution. In this paper, we present a new cloud-based technique for spectral analysis and compression of HSIs. Specifically, we develop a cloud implementation of a popular deep neural network for non-linear data compression, known as autoencoder (AE). Apache Spark serves as the backbone of our cloud computing environment by connecting the available processing nodes using a master-slave architecture. Our newly developed approach has been tested using two widely available HSI data sets. Experimental results indicate that cloud computing architectures offer an adequate solution for managing big remotely sensed data sets.
M. Goetz, G. Cavallaro, T. Geraud, M. Book and M. Riedel
IEEE Transactions on Parallel and Distributed Systems (TPDS)
Component trees are region-based representations that encode the inclusion relationship of the threshold sets of an image. These representations are one of the most promising strategies for the analysis and the interpretation of spatial information of complex scenes as they allow the simple and efficient implementation of connected filters. This work proposes a new efficient hybrid algorithm for the parallel computation of two particular component trees-the max- and min-tree-in shared and distributed memory environments. For the node-local computation a modified version of the flooding-based algorithm of Salembier is employed. A novel tuple-based merging scheme allows to merge the acquired partial images into a globally correct view. Using the proposed approach a speed-up of up to 44.88 using 128 processing cores on eight-bit gray-scale images could be achieved. This is more than a five-fold increase over the state-of-the-art shared-memory algorithm, while also requiring only one-thirty-second of the memory.
P. Ghamisi, G. Cavallaro, Dan, Wu, J. A. Benediktsson and A. Plaza
Computer Vision and Pattern Recognition
In this paper, an approach is proposed to fuse LiDAR and hyperspectral data, which considers both spectral and spatial information in a single framework. Here, an extended self-dual attribute profile (ESDAP) is investigated to extract spatial information from a hyperspectral data set. To extract spectral information, a few well-known classifiers have been used such as support vector machines (SVMs), random forests (RFs), and artificial neural networks (ANNs). The proposed method accurately classify the relatively volumetric data set in a few CPU processing time in a real ill-posed situation where there is no balance between the number of training samples and the number of features. The classification part of the proposed approach is fully-automatic.
G. Cavallaro, N. Falco, M. D. Mura and J. A. Benediktsson
IEEE Transactions on Image Processing (TIP)
Morphological attribute profiles are multilevel decompositions of images obtained with a sequence of transformations performed by connected operators. They have been extensively employed in performing multiscale and region-based analysis in a large number of applications. One main, still unresolved, issue is the selection of filter parameters able to provide representative and non-redundant threshold decomposition of the image. This paper presents a framework for the automatic selection of filter thresholds based on Granulometric Characteristic Functions (GCFs). GCFs describe the way that non-linear morphological filters simplify a scene according to a given measure. Since attribute filters rely on a hierarchical representation of an image (e.g., the Tree of Shapes) for their implementation, GCFs can be efficiently computed by taking advantage of the tree representation. Eventually, the study of the GCFs allows the identification of a meaningful set of thresholds. Therefore, a trial and error approach is not necessary for the threshold selection, automating the process and in turn decreasing the computational time. It is shown that the redundant information is reduced within the resulting profiles (a problem of high occurrence, as regards manual selection). The proposed approach is tested on two real remote sensing data sets, and the classification results are compared with strategies present in the literature.
All journals
J. A. Benediktsson, G. Cavallaro, N. Falco, I. Hedhli, V. A. Krylov, G. Moser, S. B. Serpico and J. Zerubia
Mathematical Models for Remote Sensing Image Processing
Current and forthcoming sensor technologies and space missions are providing remote sensing scientists and practitioners with an increasing wealth and variety of data modalities. They encompass multisensor, multiresolution, multiscale, multitemporal, multipolarization, and multifrequency imagery. While they represent remarkable opportunities for the applications, they pose important challenges to the development of mathematical methods aimed at fusing the information conveyed by the input multisource data. In this framework, the present chapter continues the discussion of remote sensing data fusion, which was opened in the previous chapter. Here, the focus is on data fusion for image classification purposes. Both methodological issues of feature extraction and supervised classification are addressed. On both respects, the focus is on hierarchical image models rooted in graph theory. First, multilevel feature extraction is addressed through the latest advances in Mathematical Morphology and attribute profile theory with respect to component trees and trees of shapes. Then, joint supervised classification of multisensor, multiscale, multiresolution, and multitemporal imagery is formulated through hierarchical Markov random fields on quad-trees. Examples of experimental results with data from current VHR optical and SAR missions are shown and analysed.
G. Cavallaro, M. Dalla Mura and J. A. Benediktsson
Handbook of Pattern Recognition and Computer Vision, 5th edition
Very high resolution (VHR) images provide a detailed representation of the surveyed scene with a geometrical resolution that at present can be up to 30 cm (WorldView-3). One of the most promising strategy for the analysis and the interpretation of a scene relies on hierarchical representations of the spatial content of an image. The hierarchical structure of a tree is useful for gathering the heterogeneous characteristics of the objects among different spatial scales (i.e., from the pixel level up to the entire image). In the remote sensing literature, there are several contributions addressing the use of hierarchical representations in many tasks such as filtering, segmentation, classification, object detection, change detection, and considering different types of images (e.g., panchromatic, multispectral, hyperspectral). The structure of each representation can vary significantly and can be efficiently adopted for a specific remote sensing application. After providing an overview of the different hierarchical representations, this chapter focuses on the tree structures on which image filtering (here we consider morphological attribute filters) can be efficiently implanted. Particular attention is paid to the tree of shapes, which is an important morphological structure that represents images in a self-dual way. Moreover, the identification of structures with heterogeneous characteristics (i.e., scale and shape) can be effectively done when computing attribute filters in multi-scale architectures, for instance Self-Dual Attribute Profiles (SDAPs). In this chapter we focus on the use of multilevel filtering based on hierarchical representations of the image for land cover classification. In particular, the experimental results reported in the literature for the classification of SDAPs computed on remotely sensed images are presented. Since the effectiveness of SDAPs is strictly correlated to the set of the filter parameters selected for the filtering we report on a technique for the automatic selection on the filter's parameters in order to obtain a profile that is representative and a non-redundant.
J. A. Benediktsson, G. Cavallaro, N. Falco, I. Hedhli, V. A. Krylov, G. Moser, S. B. Serpico and J. Zerubia
Mathematical Models for Remote Sensing Image Processing
Current and forthcoming sensor technologies and space missions are providing remote sensing scientists and practitioners with an increasing wealth and variety of data modalities. They encompass multisensor, multiresolution, multiscale, multitemporal, multipolarization, and multifrequency imagery. While they represent remarkable opportunities for the applications, they pose important challenges to the development of mathematical methods aimed at fusing the information conveyed by the input multisource data. In this framework, the present chapter continues the discussion of remote sensing data fusion, which was opened in the previous chapter. Here, the focus is on data fusion for image classification purposes. Both methodological issues of feature extraction and supervised classification are addressed. On both respects, the focus is on hierarchical image models rooted in graph theory. First, multilevel feature extraction is addressed through the latest advances in Mathematical Morphology and attribute profile theory with respect to component trees and trees of shapes. Then, joint supervised classification of multisensor, multiscale, multiresolution, and multitemporal imagery is formulated through hierarchical Markov random fields on quad-trees. Examples of experimental results with data from current VHR optical and SAR missions are shown and analysed.
G. Cavallaro, M. Dalla Mura and J. A. Benediktsson
Handbook of Pattern Recognition and Computer Vision, 5th edition
Very high resolution (VHR) images provide a detailed representation of the surveyed scene with a geometrical resolution that at present can be up to 30 cm (WorldView-3). One of the most promising strategy for the analysis and the interpretation of a scene relies on hierarchical representations of the spatial content of an image. The hierarchical structure of a tree is useful for gathering the heterogeneous characteristics of the objects among different spatial scales (i.e., from the pixel level up to the entire image). In the remote sensing literature, there are several contributions addressing the use of hierarchical representations in many tasks such as filtering, segmentation, classification, object detection, change detection, and considering different types of images (e.g., panchromatic, multispectral, hyperspectral). The structure of each representation can vary significantly and can be efficiently adopted for a specific remote sensing application. After providing an overview of the different hierarchical representations, this chapter focuses on the tree structures on which image filtering (here we consider morphological attribute filters) can be efficiently implanted. Particular attention is paid to the tree of shapes, which is an important morphological structure that represents images in a self-dual way. Moreover, the identification of structures with heterogeneous characteristics (i.e., scale and shape) can be effectively done when computing attribute filters in multi-scale architectures, for instance Self-Dual Attribute Profiles (SDAPs). In this chapter we focus on the use of multilevel filtering based on hierarchical representations of the image for land cover classification. In particular, the experimental results reported in the literature for the classification of SDAPs computed on remotely sensed images are presented. Since the effectiveness of SDAPs is strictly correlated to the set of the filter parameters selected for the filtering we report on a technique for the automatic selection on the filter's parameters in order to obtain a profile that is representative and a non-redundant.