Autonomous drone navigation in confined tubular environments remains a major challenge due to the constraining geometry of the conduits, the proximity of the walls, and the perceptual limitations inherent to such scenarios. We propose a reinforcement learning (RL) approach enabling a drone to navigate unknown three-dimensional tubes without any prior knowledge of their geometry, relying solely on local observations from a Light Detection and Ranging (LiDAR) sensor and a conditional visual detection of the tube center. In contrast, the Pure Pursuit algorithm, used as a deterministic baseline, benefits from explicit access to the centerline, creating an information asymmetry designed to assess the ability of RL to compensate for the absence of a geometric model. The agent is trained through a progressive curriculum learning strategy that gradually exposes it to increasingly curved geometries, where the tube center frequently disappears from the visual field. A turning-negotiation mechanism, based on the combination of direct visibility, directional memory, and LiDAR symmetry cues, proves essential for ensuring stable navigation under such partial observability conditions. Experiments show that the Proximal Policy Optimization (PPO) policy acquires robust and generalizable behavior, consistently outperforming the deterministic controller despite its limited access to geometric information. Validation in a high-fidelity three-dimensional environment further confirms the transferability of the learned behavior to continuous physical dynamics. In particular, this work introduces an explicit formulation of the turn negotiation problem in tubular navigation, coupled with a reward design and evaluation metrics that make turn-handling behavior measurable and analyzable. This explicit focus on turn negotiation distinguishes our approach from prior learning-based and heuristic methods. The proposed approach thus provides a complete framework for autonomous navigation in unknown tubular environments and opens perspectives for industrial, underground, or medical applications where progressing through narrow and weakly perceptive conduits represents a central challenge.

This study emanates from a simple observation: as specified by Vapnik [37] in his study, an artificial neural network cannot generate a universal approximator if the aggregation function chosen to design the artificial neuron does not include non-linearity. The usual option is to follow a linear aggregation by a non-linear function, or so-called activation function. We wonder if this approach could be replaced by one using a natively non-linear aggregation function.Among all of the available non-linear aggregation functions, here we are interested in aggregations based on weighted minimum and weighted maximum operations [8]. As these operators were originally developed within a possibility theory and fuzzy rule framework, such operators cannot be easily integrated into a neural network because the values that are usually considered belong to [0,1]. For gradient descent based learning, a neuron must be an aggregation function derivable with respect to its inputs and synaptic weights, whose variables (synaptic weights, inputs and outputs) must all be signed real values. We thus propose an extension of weighted maximum based aggregation to enable this learning process. We show that such an aggregation can be seen as a combination of four Sugeno integrals. Finally, we compare this type of approach with the classical one.

ABSTRACT We release photometric redshifts, reaching ∼0.7, for ∼14M galaxies at r ≤ 20 in the 11 500 deg2 of the SDSS north and south Galactic caps. These estimates were inferred from a convolution neural network (CNN) trained on ugriz stamp images of galaxies labelled with a spectroscopic redshift from the SDSS, GAMA, and BOSS surveys. Representative training sets of ∼370k galaxies were constructed from the much larger combined spectroscopic data to limit biases, particularly those arising from the over-representation of luminous red galaxies. The CNN outputs a redshift classification that offers all the benefits of a well-behaved PDF, with a width efficiently signalling unreliable estimates due to poor photometry or stellar sources. The dispersion, mean bias, and rate of catastrophic failures of the median point estimate are of order σMAD = 0.014, <Δznorm>=0.0015, $\eta (|\Delta z_{\rm norm}|\gt 0.05)=4{{\, \rm per\ cent}}$ on a representative test sample at r < 19.8, outperforming currently published estimates. The distributions in narrow intervals of magnitudes of the redshifts inferred for the photometric sample are in good agreement with the results of tomographic analyses. The inferred redshifts also match the photometric redshifts of the redMaPPer galaxy clusters for the probable cluster members.

Whether in times of conflict or during peacetime, the fight against underwater mines is an activity of paramount importance. For naval forces, it is a question of acquiring the means to eliminate the possible presence of underwater mines near the domestic coasts or in areas of amphibious and naval operations. For the civilian field, the off-shore wind in northern Europe is the example of a sector particularly affected by the threat of unexploded ordnances (UXO) dating from military conflicts of the past. The advances of recent decades in marine technologies have enabled the development of high-resolution acoustic imaging systems fitted on towed fish or autonomous underwater vehicle (AUV), capable of generating high-resolution imagery of the underwater environment. Based on deep learning approaches, the collected data is used to develop Automatic Target Recognition (ATR) algorithms to detect suspicious objects on the seafloor and classify each as an object of interest (e.g., a mine) or not. However, because obtaining labelled underwater images demands time and effort, applying deep learning based approaches in underwater environment remains a challenge due to the scarcity of training data. This paper presents a new approach for improving the detection of minelike objects. Our method is based on two steps: the first step consists in transforming our detection problem into a semantic segmentation task. The second step involves active learning. We demonstrate that our approach leads to a significant improvement in terms of area under the curve (AUC).

Photometric redshift estimation plays a crucial role in modern cosmological surveys for studying the universe's large-scale structures and the evolution of galaxies. Deep learning has emerged as a powerful method to produce accurate photometric redshift estimates from multiband images of galaxies. Here, we introduce a multimodal approach consisting of the parallel processing of several subsets of prior image bands, the outputs of which are then merged for further processing through a convolutional neural network (CNN). We evaluate the performance of our method using three surveys: the Sloan Digital Sky Survey (SDSS), the Canada-France-Hawaii Telescope Legacy Survey (CFHTLS), and the Hyper Suprime-Cam (HSC). By improving the model's ability to capture information embedded in the correlation between different bands, our technique surpasses state-of-the-art photometric redshift precision. We find that the positive gain does not depend on the specific architecture of the CNN and that it increases with the number of photometric filters available.

Automated pollen analysis is not yet efficient on environmental samples containing many pollen taxa and debris, which are typical in most pollen-based studies. Contrary to classification, detection remains overlooked although it is the first step from which errors can propagate. Here, we investigated a simple but efficient method to automate pollen detection for environmental samples, optimizing workload and performance. We applied the YOLOv5 algorithm on samples containing debris and c. 40 Mediterranean plant taxa, designed and tested several strategies for annotation, and analyzed variation in detection errors. About 5% of pollen grains were left undetected, while 5% of debris were falsely detected as pollen. Undetected pollen was mainly in poor-quality images, or of rare and irregular morphology. Pollen detection remained effective when applied to samples never seen by the algorithm, and was not improved by spending time to provide taxonomic details. Pollen detection of a single model taxon reduced annotation workload, but was only efficient for morphologically differentiated taxa. We offer guidelines to plant scientists to analyze automatically any pollen sample, providing sound criteria to apply for detection while using common and user-friendly tools. Our method contributes to enhance the efficiency and replicability of pollen-based studies.

Collembola are very abundant organisms in soils (several thousand individuals per square meter) and are considered to be good indicators of soil quality. These indicators are mainly based on the number of individuals observed (abundance per square meter of soil), but also the singularity and number of species present (species richness). A limitation that comes with the usage of collembola as an indicator is the complexity of the identification of the species under a microscope, how time-consuming it is, and the morphological similarity between some species. Deep learning approaches have been very successful in the resolution of image-based problems. Still, no work yet exists that uses deep learning in the recognition of collembola on a microscope slide. This could be a valuable tool for experts seeking to use Collembola as a metric on a larger scale. In this work, we explore and evaluate the performance of state-of-the-art deep learning techniques over the identification of Collembola on a new manually annotated dataset.

Aims. The treatment of astronomical image time series has won increasing attention in recent years. Indeed, numerous surveys following up on transient objects are in progress or under construction, such as the Vera Rubin Observatory Legacy Survey for Space and Time (LSST), which is poised to produce huge amounts of these time series. The associated scientific topics are extensive, ranging from the study of objects in our galaxy to the observation of the most distant supernovae for measuring the expansion of the universe. With such a large amount of data available, the need for robust automatic tools to detect and classify celestial objects is growing steadily.Methods. This study is based on the assumption that astronomical images contain more information than light curves. In this paper, we propose a novel approach based on deep learning for classifying different types of space objects directly using images. We named our approach ConvEntion, which stands for CONVolutional attENTION. It is based on convolutions and transformers, which are new approaches for the treatment of astronomical image time series. Our solution integrates spatio-temporal features and can be applied to various types of image datasets with any number of bands.Results. In this work, we solved various problems the datasets tend to suffer from and we present new results for classifications using astronomical image time series with an increase in accuracy of 13%, compared to state-of-the-art approaches that use image time series, and a 12% increase, compared to approaches that use light curves.

Deep-learning models have been increasingly exploited in astrophysical studies, but these data-driven algorithms are prone to producing biased outputs that are detrimental for subsequent analyses. In this work, we investigate two main forms of biases: class-dependent residuals, and mode collapse. We do this in a case study, in which we estimate photometric redshift as a classification problem using convolutional neural networks (CNNs) trained with galaxy images and associated spectroscopic redshifts. We focus on point estimates and propose a set of consecutive steps for resolving the two biases based on CNN models, involving representation learning with multichannel outputs, balancing the training data, and leveraging soft labels. The residuals can be viewed as a function of spectroscopic redshift or photometric redshift, and the biases with respect to these two definitions are incompatible and should be treated individually. We suggest that a prerequisite for resolving biases in photometric space is resolving biases in spectroscopic space. Experiments show that our methods can better control biases than benchmark methods, and they are robust in various implementing and training conditions with high-quality data. Our methods hold promises for future cosmological surveys that require a good constraint of biases, and they may be applied to regression problems and other studies that make use of data-driven models. Nonetheless, the bias-variance tradeoff and the requirement of sufficient statistics suggest that we need better methods and optimized data usage strategies.

Many fields are interested in mapping and monitoring the sea floor natural structure and biology, for environmental surveys, including identifying macro waste or detection of submerged artifacts such as cars, tires, wrecks, and even military applications, e.g., mine warfare whose detection depends heavily on the seabed structure. In this paper, we propose a new active learning method to improve seabed segmentation by deep learning. We perform segmentation of the sea floor using two data sources, sonar, and bathymetry. We train a network to fuse these two modalities and segment each sea floor pixel into nine fine ecological classes, then into three gross class sets, alive/not alive, and, in two different ways, whether mines can be hunted for or not. Once this training is done, a second stage involving a new active learning method based on network uncertainties greatly improves the performance.

The advances of recent decades in marine technologies have enabled the development of high-resolution acoustic imaging systems capable of generating high-resolution imagery of the underwater environment. Based on deep learning approaches, the collected data is used to develop Automatic Target Recognition (ATR) algorithms to detect suspicious objects on the seafloor and classify each as an object of interest (e.g., a mine) or not. However, because obtaining labelled underwater images demands time and effort, applying deep learning based approaches in underwater environment remains a challenge due to the scarcity of training data. This paper presents a work on the improvement of the realism of synthetic sonar data generated by a simulator capable of massively producing labelled images. The simulator is based on the numerical modeling of the propagation, reverberation and reflection of acoustics waves in an artificial underwater environment. The reduction of the gap of realism with real sonar images is conducted by applying Cycle Generative Adversarial Networks. Then, we study the successive addition of a growing quantity of these refined synthetic data in a set of real training sonar image of limited size and the impact of this data augmentation on the performance of a CNN-based object detector.

Image-to-image translation with Deep Learning neural networks, particularly with Generative Adversarial Networks (GANs), is one of the most powerful methods for simulating astronomical images. However, current work is limited to utilizing paired images with supervised translation, and there has been rare discussion on reconstructing noise background that encodes instrumental and observational effects. These limitations might be harmful for subsequent scientific applications in astrophysics. Therefore, we aim to develop methods for using unpaired images and preserving noise characteristics in image translation. In this work, we propose a two-way image translation model using GANs that exploits both paired and unpaired images in a semi-supervised manner, and introduce a noise emulating module that is able to learn and reconstruct noise characterized by high-frequency features. By experimenting on multi-band galaxy images from the Sloan Digital Sky Survey (SDSS) and the Canada France Hawaii Telescope Legacy Survey (CFHT), we show that our method recovers global and local properties effectively and outperforms benchmark image translation models. To our best knowledge, this work is the first attempt to apply semi-supervised methods and noise reconstruction techniques in astrophysical studies.

Improving distance measurements in large imaging surveys is a major challenge to better reveal the distribution of galaxies on a large scale and to link galaxy properties with their environments. Photometric redshifts can be efficiently combined with the cosmic web (CW) extracted from overlapping spectroscopic surveys to improve their accuracy. We apply a similar method using a new generation of photometric redshifts based on a convolution neural network (CNN). The CNN is trained on the SDSS images with the main galaxy sample (SDSS-MGS, $r \leq 17.8$) and the GAMA spectroscopic redshifts up tor $\sim 19.8$. The mapping of the CW is obtained with 680,000 spectroscopic redshifts from the MGS and BOSS surveys. The redshift probability distribution functions (PDF), which are well calibrated (unbiased and narrow, $\leq 120$ Mpc), intercept a few CW structure along the line of sight. Combining these PDFs with the density field distribution provides new photometric redshifts, $z_{web}$, whose accuracy is improved by a factor of two (i.e.,${\sigma} \sim 0.004(1+z)$) for galaxies with $r \leq 17.8$. For half of them, the distance accuracy is better than 10 cMpc. The narrower the original PDF, the larger the boost in accuracy. No gain is observed for original PDFs wider than 0.03. The final $z_{web}$ PDFs also appear well calibrated. The method performs slightly better for passive galaxies than star-forming ones, and for galaxies in massive groups since these populations better trace the underlying large-scale structure. Reducing the spectroscopic sampling by a factor of 8 still improves the photometric redshift accuracy by 25%. Extending the method to galaxies fainter than the MGS limit still improves the redshift estimates for 70% of the galaxies, with a gain in accuracy of 20% at low $z$ where the resolution of the CW is the highest.

This chapter introduces several state of the art techniques that could help to make deep underwater archaeological photogrammetric surveys easier, faster, more accurate, and to provide more visually appealing representations in 2D and 3D for both experts and public. We detail how the 3D captured data is analysed and then represented using ontologies, and how this facilitates interdisciplinary interpretation and cooperation. Towards more automation, we present a new method that adopts a deep learning approach for the detection and the recognition of objects of interest, amphorae for example. In order to provide more readable, direct and clearer illustrations, we describe several techniques that generate different styles of sketches out of orthophotos developed using neural networks. In the same direction, we present the Non-Photorealistic Rendering (NPR) technique, which converts a 3D model into a more readable 2D representation that is more useful to communicate and simplifies the identification of objects of interest. Regarding public dissemination, we demonstrate how recent advances in virtual reality to provide an accurate, high resolution, amusing and appropriate visualization tool that offers the public the possibility to ‘visit’ an unreachable archaeological site. Finally, we conclude by introducing the plenoptic approach, a new promising technology that can change the future of the photogrammetry by making it easier and less time consuming and that allows a user to create a 3D model using only one camera shot. Here, we introduce the concepts, the developing process, and some results, which we obtained with underwater imaging.

We developed a deeP architecturE for the LIght Curve ANalysis (PELICAN) for the characterization and the classification of supernovae light curves. It takes light curves as input, without any additional features. PELICAN can deal with the sparsity and the irregular sampling of light curves. It is designed to remove the problem of non-representativeness between the training and test databases coming from the limitations of the spectroscopic follow-up. We applied our methodology on different supernovae light curve databases. First, we tested PELICAN on the Supernova Photometric Classification Challenge for which we obtained the best performance ever achieved with a non-representative training database, by reaching an accuracy of 0.811. Then we tested PELICAN on simulated light curves of the LSST Deep Fields for which PELICAN is able to detect 87.4% of supernovae Ia with a precision higher than 98%, by considering a non-representative training database of 2k light curves. PELICAN can be trained on light curves of LSST Deep Fields to classify light curves of the LSST main survey, which have a lower sampling rate and are more noisy. In this scenario, it reaches an accuracy of 96.5% with a training database of 2k light curves of the Deep Fields. This constitutes a pivotal result as type Ia supernovae candidates from the main survey might then be used to increase the statistics without additional spectroscopic follow-up. Finally we tested PELICAN on real data from the Sloan Digital Sky Survey. PELICAN reaches an accuracy of 86.8% with a training database composed of simulated data and a fraction of 10% of real data. The ability of PELICAN to deal with the different causes of non-representativeness between the training and test databases, and its robustness against survey properties and observational conditions, put it in the forefront of light curve classification tools for the LSST era.

Cosmologists are facing the problem of the analysis of a huge quantity of data when observing the sky. The methods used in cosmology are, for the most of them, relying on astrophysical models, and thus, for the classification, they usually use a machine learning approach in two-steps, which consists in, first, extracting features, and second, using a classifier. In this paper, we are specifically studying the supernovae phenomenon and especially the binary classification I.a supernovae versus not-I.a supernovae. We present two Convolutional Neural Networks (CNNs) defeating the current state-of-the-art. The first one is adapted to time series and thus to the treatment of supernovae light-curves. The second one is based on a Siamese CNN and is suited to the nature of data, i.e. their sparsity and their weak quantity (small learning database).

Cultural heritage (CH) resources are very heterogeneous since the information was collected from vast diversity of cultural sites and digitally recorded in different formats. With the progress of 3D technologies, photogrammetry techniques become the adopted solution for representing CH artifacts by turning photos from small finds, to entire landscapes, into accurate 3D models. To meet knowledge representation with cultural heritage photogrammetry, this paper proposes an ontology-profiling method for modeling a real case of archaeological amphorae. The ontological profile consists of all needed information to represent a CH resource including typology attributes, geo-spatial information and photogrammetry process. An example illustrating the applicability of this profiling method to the problem of CH resources conceptualization is presented. We also outline our perspectives for using ontologies in data-driven science, in particular on modeling a complete pipeline that manages both the photogrammetric process and the archaeological knowledge.