Object detection is one of the most fundamental problems to tackle in the computer vision research area. Recent advances in multimodal data streams and deep learning architectures have prompted a fast growth in the field of multimodal learning, which brings several advantages over single-modality approaches for object detection, such as improved accuracy, robustness to noise and ambiguity, handling of complex scenarios and adaptability to diverse data. Some of the biggest challenges when implementing a multimodal learning approach are the selection of the fusion strategy, design of processing architecture, modality alignment/synchronization and interpretability of such high-dimensional representations. To address this challenge, we propose a feature-level fusion architecture for object detection based on extracting YOLO features from images, spectral and rhythm features from sound using Mel-frequency cepstral coefficients, and general descriptors from radar modalities that, after timestamp and homography transformation matrix alignment, are combined with an attention mechanism into a single classification network. Preliminary experiments indicate that the proposed architecture can constitute itself as a base pipeline for several different multimodal object detection tasks in real-world applications.

With the growing importance of the use of information from electronic patient records in the development of machine learning models, there is also a need for a holistic understanding of those records, in particular abridging the clinical notes so that important information is used in the training process without the repetition that is commonly found in such notes. This paper presents the pre-processing of clinical notes from the BRATECA Dataset, a Brazilian tertiary care data collection, aiming at removing repeated information resulting from the interaction between healthcare providers and patients, considering assigned values of semantic similarity between sentences in clinical notes.

Multi-agent systems have shown great promise in addressing complex problems that traditional single-agent approaches are not be able to handle. In this article, we propose a multi-agent system for the conception of a multimodal machine learning problem on edge devices. Our architecture leverages docker containers to encapsulate knowledge in the form of models and processes, enabling easy management of the system. Communication between agents is facilitated by Message Queuing Telemetry Transport, a lightweight messaging protocol ideal for Internet of Things and edge computing environments. Additionally, we highlight the significance of object detection in our proposed system, which is a crucial component of many multimodal machine learning tasks, by enabling the identification and localization of objects within diverse data modalities. In this manuscript an overall architecture description is performed, discussing the role of each agent and the communication protocol between them. The proposed system offers a general approach to multimodal machine learning problems on edge devices, demonstrating the advantages of multi-agent systems in handling complex and dynamic environments.

This paper deals with the problem of detecting sand dunes from remotely sensed images of the surface of Mars. We build on previous approaches that propose methods to extract informative features for the classification of the images. The intricate correlation structure exhibited by these features motivates us to propose the use of probabilistic classifiers based on R-vine distributions to address this problem. R-vines are probabilistic graphical models that combine a set of nested trees with copula functions and are able to model a wide range of pairwise dependencies. We investigate different strategies for building R-vine classifiers and compare them with several state-of-the-art classification algorithms for the identification of Martian dunes. Experimental results show the adequacy of the R-vine-based approach to solve classification problems where the interactions between the variables are of a different nature between classes and play an important role in that the classifier can distinguish the different classes.

Impact craters on Mars have been extensively modified by ancient geologic processes that may have included rainfall and surface runoff, snow and ice, denudation by lava flows, burial by eolian material, or others. Many of these processes can leave distinct signatures on the morphometry of the modified impact crater as well as the surrounding landscape. To look for signs of potential regional differences in crater modification processes, we conducted an analysis of different morphometric parameters related to modified impact craters located in the Margaritifer Sinus, Sinus Sabaeus, Iapygia, Mare Tyrrhenum, Aeolis, and Eridania quadrangles, including depth, crater wall slope, crater floor slope, the curvature between the interior wall and the crater floor slope, and the curvature between the interior wall and surrounding landscape. A Welch's t test analysis comparing these parameters shows that fresh impact craters (Type 4) have consistent morphologies regardless of their geographic location examined in this study, which is not unexpected. Modified impact craters both in the initial (Type 3) and terminal stages (Type 1) of modification also have statistically consistent morphologies. This would suggest that the processes that operated in the late Noachian were globally ubiquitous, and that modified craters eventually reached a stable crater morphology. However, craters preserved in advanced (but not terminal) stages of modification (Type 2) have morphologies that vary across the quadrangles. It is possible that these variations reflect spatial differences in the types and intensity of geologic processes that operated during the Noachian, implying that the ancient climate also varied across regions.

The main objective of this paper is to verify the accuracy of delineating and characterizing ice-wedge polygonal networks with features exclusively extracted from remotely sensed images of very high resolution. This kind of mapping plays a key role for quantifying ice-wedge degradation in warming permafrost. The evaluation of mapping a network is performed in this study with two sets of aerial images that are compared to ground reference data determined by fieldwork on the same network, located in Adventdalen, Svalbard (78°N). One aerial dataset is obtained from a photogrammetric survey with RGB+NIR imagery of 20cm/pixel, the other from an UAV (Unmanned Aerial Vehicle) survey that acquired RGB images of 6cm/pixel of spatial resolution. Besides evaluating the degree of matching between the delineations, the morphometric and topological features computed for the differently mapped versions of the network are also confronted, to have a more solid basis of comparison. The results obtained are similar enough to admit that remotely sensed images of very high resolution are an adequate support to provide extensive characterizations and classifications of this kind of patterned ground.

MORPHOLOGY ON MARS. Robert A. Craddock, Lourenço Bandeira, and Alan D. Howard, Center for Earth and Planetary Studies, National Air and Space Museum, Smithsonian Institution, Washington, DC 20560 craddockb@si.edu, Centre for Natural Resources and the Environment, Instituto Superior Técnico. University of Lisbon, Av. Rovisco Pais 1049-001, Lisbon, Portugal, lpcbandeira@ist.utl.pt, Department of Environmental Sciences, PO Box 400123 Clark 205, University of Virginia, Charlottesville, VA 22904, http://erode.evsc.virginia.edu

The ice-free areas of Maritime Antarctica show complex mosaics of surface covers, with wide patches of diverse bare soils and rock, together with various vegetation communities dominated by lichens and mosses. The microscale variability is difficult to characterize and quantify, but is essential for ground-truthing and for defining classifiers for large areas using, for example high resolution satellite imagery, or even ultra-high resolution unmanned aerial vehicle (UAV) imagery. The main objective of this paper is to verify the ability and robustness of an automated approach to discriminate the variety of surface types in digital photographs acquired at ground level in ice-free regions of Maritime Antarctica. The proposed method is based on an object-based classification procedure built in two main steps: first, on the automated delineation of homogeneous regions (the objects) of the images through the watershed transform with adequate filtering to avoid an over-segmentation, and second, on labelling each identified object with a supervised decision classifier trained with samples of representative objects of ice-free surface types (bare rock, bare soil, moss and lichen formations). The method is evaluated with images acquired in summer campaigns in Fildes and Barton peninsulas (King George Island, South Shetlands). The best performances for the datasets of the two peninsulas are achieved with a SVM classifier with overall accuracies of about 92% and kappa values around 0.89. The excellent performances allow validating the adequacy of the approach for obtaining accurate surface reference data at the complete pixel scale (sub-metric) of current very high resolution (VHR) satellite images, instead of a common single point sampling.

Introduction: In a little more than one decade, crater detection algorithms (CDA) have greatly evolved in their conception and in the methodological ingredients used, from the more classic image analysis and pattern recognition operators [1, 2, 3, 4] to the more up-to-date and adaptive tools [5, 6, 7, 8, 9], providing more robust processing sequences able to successfully deal with the large variety of cratered landscapes all over the Solar System. These have naturally been mainly developed and tested for those surfaces where imagery is more abundant, Mars and the Moon. More recently, CDA are also being applied on Mercury [10, 11], Phobos [12] and Vesta [10]. In this way, the robustness of the CDA has been proved undoubtedly in a wider type of surfaces, also contributing to update crater catalogues on Mars [13], Moon [14] and Phobos [12]. But even in these studies the figures involved, when optical images are concerned, are around some few thousands of craters. Our main objective in this work is to demonstrate that the detection of a huge amount of craters (hundreds of thousands) with an automated approach in relatively large regions covered by the assemblage of several adjacent images (mosaics), captured in distinct time periods, can be trusted. Dataset and Methodology: In this abstract we focus our study on the Moon, in particular in Sinus Iridum region (44.1° N, 31.5° W), a mare filled crater of about 236 km in diameter, through the analysis of a Kaguya (SELENE) [15] Terrain Camera (TC) Evening illumination tile set with spatial resolution of 7.4 m/pixel, released by the SELENE team [16] and rereleased by Astrogeology/USGS [17] to detect craters with a dimensional range of diameters between 100 and 1500 meters. Instead of processing the entire region at once, we analyzed smaller areas at each time, with dimension (2048 x 2048 pixels with a 200 pixels overlap between adjacent tiles) that permitted an efficient computational performance and the detection of the entire impact structure within the same tile. A total of 480 tiles in the mare region were generated this way, being 12 of them selected to train and test our approach: 6 of the tiles were concentrated around the same area, located in the southern part of the mare, while the other 6 tiles were selected from dispersed locations of Sinus Iridum to contain the diversity of the mare surface (Figure 1), where we have manually cataloged almost 190,000 craters. Figure 1 Sinus Iridum mare on a mosaic built from images

This paper presents the automated detection of impact craters on large regions of Mercury. The processing sequence is composed by three main phases: the first consists on creating the image mosaics of the large areas of interest, the second by finding crater candidates on these mosaics, and finally by extracting a set of features that are used in the classification by SVM-Support Vector Machine in the third phase. The detections are performed on images acquired by the MDIS-NAC camera of MESSENGER probe covering three large basins on Mercury (Rachmaninoff, Mozart and Raditladi).

In our current objective of making large scale crater detections on Mercury, we present preliminary results achieved with a method of ours in MDIS images of MESSENGER in Rachmaninoff basin.

In a large majority of lunar and planetary surface images, impact craters are the most abundant geological features. Therefore, it is not surprising that crater detection algorithms (CDAs) are one of the most studied subjects of image processing and analysis in lunar and planetary science. In this work we are proposing an Integrated CDA, consisting of: (1) utilization of DEM (digital elevation map)-based CDA; (2) utilization of an optical-based CDA; (3) re-projection of used datasets and crater coordinates from normal to rotated view and back; (4) correction of the brightness and contrast of a used optical image; and (5) tile generation for the optical-based CDA and an assembling of results with an elimination of multiple detections, in combination with a pyramid approach down to the resolution of the available DEM image; and (6) a final integration of the results of DEM-based and optical-based CDAs, including a removal of duplicates. The proposed CDA is applied to one specific asteroid-like body, the small Martian moon Phobos. The experimental evaluation of the proposed CDA is done by a manual verification of crater-candidates and a search for uncatalogued craters. The evaluation has shown that the proposed CDA was used successfully for cataloging Phobos craters. The major result of this paper is the PH9224GT – currently the most complete global catalogue of the 9224 Phobos craters. The possible applications of the new catalogue are: (1) age estimations for any selected location; and (2) comparison/evaluation of the different chronology and production functions for Phobos. This confirms the practical applicability of the new Integrated CDA – an additional result of this paper, which can be used in order to considerably extend the current crater catalogues.