Portail HAL ONERA
Not a member yet
12842 research outputs found
Sort by
Large-eddy simulation analysis of turbulent flame periodic flashbacks in a backward-facing step (BFS) combustor
No full textInternational audienceHighly turbulent premixed flame dynamics is studied in a backward-facing step combustor. Large-eddy simulations are performed for two operating conditions: (i) a stable case with the flame anchored in the vicinity of the sudden expansion and (ii) an unstable case with massive periodic flashbacks. This set of computational results is first assessed through comparisons with experimental data. Then, it is used to conduct a detailed analysis of the flashback process. Pressure and velocity fluctuations signals display (i) high frequency fluctuations associated to turbulence with similar amplitudes in both the stable and unstable cases, and (ii) a low frequency fluctuation, of acoustic origin, the amplitude of which is significantly higher in the unstable case. The occurrence of flame flashback is found to be closely related to the structure of the first longitudinal acoustic mode, which can be modified thanks to a throttling plug used to modulate the area of the combustor exit cross-section. In the unstable case, this corresponds to an axial-velocity anti-node at the backward-facing step location. Pressure variations are synchronized with the motions of the flame toward the combustor upper wall and the cross-correlation between pressure and heat-release rate oscillations is relevant to a thermoacoustic feedback
Relation between strength and local stresses in nanostructures as a function of their size and shape
No full textInternational audienceThe aim of this work is to understand how and why the strength of nanostructures depends on their size and shape. We performed extensive molecular dynamics simulations of the uniaxial compression of 2D FCC models with various sizes and shapes. It is found that the strength varies as a power law with size, with an exponent critically depending on shape. Plastic deformation is triggered by a unique mechanism, which is the nucleation of Shockley dislocation at local stress gradients in the vicinity of contact corners. This allows to unravel the effect of size and shape by examining the relation between local stresses and applied stress. Our analyses reveal that spatial variations of local stresses as a function of size and shape are complex and not simply related to the stress concentration as predicted by the elasticity theory. However, it is found that these local stresses linearly depend on the applied stress. Finally, we propose that the observed size effect on strength is caused by a screening interaction between the stress gradients occurring at all contact corners
Conception de commandes de vol haptiques intuitives suivant une approche basée sur l'opérateur humain
No full textThe dawn of Fly-by-Wire (FBW) with the introduction of Flight Control Computers (FCC) marked the end of mechanical linkage in flight controls. If this came with a bunch of advantages such as alleviated maintenance, it disrupted the kinesthesic feedback pilots used to have in their sidestick. With the recent maturation of active inceptors, following the path of robotic teleoperation to benefit from haptic sidesticks to fill in for the lack of kinesthesic feedback, seems a promising approach. Though, despite the wide range of designs, current works struggle to secure acceptability. Here, we investigate a new design strategy based on the perceptual cycle model of Situational Awareness (SA) (Smith & Hancock, 1995) and underpinning ecological approach (Gibson, 2014). To this end, we focus on a task of slalom derived from certification standard ADS-33F (Blanken et al., 2019) and propose two designs of haptic assistance to compare on a Pilot-in-the-Loop (PITL) flight simulator, for several slalom configurations and two control types. Both represent goal and obstacles as attractor and repellers but act on different quantities. The first design is potential field-based, negative and positive respectively. The second one is an implementation of Fajen andWarren's model of human steering dynamics and aims to embody a more intuitive control through more naturally correlated input sidestick activity and output trajectory. Three types of objective criteria of intuitiveness are identified: straight-forward, input-domain on sidestick activity (Mottet & Bootsma, 1999), and output-domain on the trajectory (Fajen & Warren, 2003). Along those objective criteria, perceived task load is assessed through the NASA Task Load indeX (NASA-TLX) (Hart & Staveland, 1988). Results show that with similar levels of performance, the human-based design calls for lower mental and physical demands, and smoother sidestick activity and trajectory, which is interpreted as better intuitiveness hence acceptability.Avec l'introduction des ordinateurs de bord, l'avènement des commandes de vol électriques a marqué la fin du lien mécanique qui traversait autrefois l'ensemble de la chaîne de contrôle. Si ce choix présente de multiples atouts dont l'allégement notoire des procédures de maintenance, ce changement a entraîné une rupture du retour kinesthésique que les pilotes pouvaient ressentir dans le manche. Avec la maturation récente des organes de commande actifs, il semble naturel de suivre la voie de la téléopération robotique pour reconstruire ce lien au travers des manches haptiques. Toutefois, malgré un large éventail de designs, les travaux actuels peinent à être adoptés en raison de l'opacité des designs étudiés. Nous proposons donc dans le cadre de cette thèse une nouvelle stratégie de conception basée sur le modèle de Situational Awareness (SA) par cycle perceptuel (Smith & Hancock, 1995), laquelle étaye l'approche écologique (Gibson, 2014). Pour cela, nous nous concentrons sur une adaptation de la tâche de slalom définie dans le standard de certification ADS-33F (Blanken et al., 2019) et proposons deux designs d'assistance haptique, lesquels sont testés sur un simulateur de vol avec humain dans la boucle, dans diverses configurations de slalom et avec deux modes de contrôle de l'aéronef. Un modèle attraction-répulsion des effets des objectif et obstacles sous-tend les deux designs qui divergent sur la grandeur d'intérêt d'application. Le premier design s'ancre dans une représentation par potentiels respectivement négatif et positifs, tandis que le second implémente un modèle de marche humaine par poursuite d'objectifs (Fajen & Warren, 2003). Trois catégories de critères d'intuitivité sont identifiées afin de quantifier l'activité de pilotage : directs portant sur la performance globale, sur l'activité au manche dans le domaine d'entrée (Mottet & Bootsma, 1999), et sur la trajectoire dans le domaine de sortie (Fajen & Warren, 2003). En complément de ces critères objectifs, la charge de travail perçue est estimée au travers du questionnaire NASA Task Load indeX (NASA-TLX) (Hart & Staveland, 1988). Les résultats montrent que pour des niveaux similaires de performance, le design basé sur le modèle de marche humaine requiert des taux plus faibles d'exigences mentale et physique, ainsi qu'une activité au manche et donc une trajectoire plus souple, ce qui peut s'interpréter comme un contrôle plus intuitif et qui sera donc mieux accepté
Using temporal interpolation on optical-derived labels improves snow detection on SAR images using deep learning method
No full textInternational audienceSnow detection is important in many domains as it is a key variable for climate monitoring [Aguirre2018]. It also allows us to assess available water resources for human consumption or hydroelectricity generation [Rouhier2018]. Using optical data, snow can be detected because of its high reflectance in the visible spectrum and the low reflectance in the shortwave infrared spectrum. The Normalized Differential Snow Index (NDSI) which exploits these spectral characteristics [Hall2002] is commonly used to create binary or fractional snow cover maps, but is highly sensible to cloud cover resulting in unevenly spaced time series. Synthetic Aperture Radar (SAR) data can be acquired at night and through clouds. However, snow detection with SAR is challenging, as dry snow is almost transparent to SAR and most of the observed signal comes from the ground. A method to retrieve dry snow depth using ratios between VV and VH backscatter compared to a reference created using means of snow-free acquisitions was proposed by [Lievens2019], but need prior information about snow presence. When snow melts, its liquid water content increases and most of the signal is scattered in the specular direction, strongly attenuating the backscattered signal. In [Nagler2016], the authors use this attenuation by combining ratios of both polarization backscatter with their reference to detect wet snow with a thresholding method. As it is a pixel-wise decision, it can be noisy. Deep learning methods can be used to detect wet snow with optical-derived labels in a semantic segmentation task using ratios between backscatter and reference as input as it presents the advantage of being independent on incident observation angle [Lê2023]. In [Montginoux2023], [Nagler2016] and [Lievens2019] ratios were concatenated to detect wet and dry snow and topological information was added in [Gallet2024]. These previous methods show promising results, however optical-derived label maps used for training the network are always patchy due to high cloud cover. In this study, we investigate whether increasing the amount of labels by temporal interpolation of the NDSI improves wet and dry snow detection results even if these labels are uncertain. We compare two temporal interpolation methods on NDSI: Closest Neighbours Interpolation (CNI) using a three days window and a Kalman smoother. CNI only fills small gaps taking little risk while Kalman smoother estimates a NDSI value for each date regardless of the gap size. The NDSI maps are then thresholded to get binary snow cover maps and projected on SAR geometry using LabSAR algorithm [Weissgerber2022]. To conduct this study, we first assess which set of input channels gives the best results training a network with non-interpolated labels. We consider four sets of input channels: the one used in [Montginoux2022] (A), the one used in [Lê2023] (B), a concatenation of VV and VH backscatter (C), and channel set C concatenated with their references (D). After identifying the best set, we use it to assess label interpolation effect. As machine learning methods performances are very dependent on training data, we test the robustness of our method under spatial and temporal domain shift. We use couples of SAR acquisitions and optical label maps from the Guil basin located in the Queyras massif in the French Alps from September 2018 to June 2019 as our main domain and split it temporally in a training set, a validation set to avoid overfitting and tune model hyperparameters and a test set to evaluate it. To evaluate temporal shift robustness, we use acquisitions from the same basin between September 2019 and June 2020, and for spatial shift the Gyronde basin in the Ecrins massif between September 2018 and June 2019. The Sentinel-1 data is acquired in interferometric wide swath mode, with a range-azimuth ground resolution of 5x20 m and a temporal resolution of 6 days. Three orbits go over each basin, so we have 6 acquisitions every 12 days combining both ascending and descending orbits. Reference images are computed for each year, basin and orbit using snow-free acquisitions between the month of June and August of the respective year. The optical data is from the MOD10A1 dataset [Hall2021] from the National Snow and Ice Data Center (NSIDC), which provides daily NDSI and cloud cover maps for both basins at 500m ground resolution. When investigating channel sets, each lends accuracies over 0.85 without domain transfer with D performing best with 0.899 accuracy. With temporal transfer, performances do not change much as we have more mono class dates which are easy to segment and D remains the best channel set. Spatial transfer is a harder task but D still is the best channel set with an accuracy of 0.866. With qualitative evaluation on predicted maps, we see that models trained with channel set C can miss snow as reference information about snow-free ground is needed. Model trained with channel set D always predicts better maps than those trained with channel set A and B, which are less precise and noisier. Keeping the reference as an independent channel using channel set D allows better segmentation, as using a ratio between backscatter and reference removes incident angle variability thus topography information. For the rest of the study, we use channel set D as input, and trained models using CNI and Kalman smoother with different regularization parameter values. All the interpolation methods improve from using non-interpolated labels, and CNI performs best with accuracies over 0.9 for all domains. Qualitatively, we see a clear improvement on the predicted snow maps which are smoother due to better spatial regularization learned during training, where the network sees less patchy label maps. Using any label interpolation is better than none, but the Kalman smoother performs worse than CNI. While we get more labels than using CNI, it increases the risk of introducing label noise by misclassifying more frequently a pixel by filling all the gaps in its timeseries. To improve our method, we can use the estimation variance the smoother outputs to model the confidence in a label and use this information during training. References: [Aguirre2018] F. Aguirre et al., « Snow Cover Change as a Climate Indicator in Brunswick Peninsula, Patagonia », Front. Earth Sci., vol. 6, sept. 2018, doi: 10.3389/feart.2018.00130. [Gallet2024] M. Gallet, A. Atto, F. Karbou, et E. Trouvé, « Wet Snow Detection From Satellite SAR Images by Machine Learning With Physical Snowpack Model Labeling », IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing, vol. 17, p. 2901 2917, 2024, doi: 10.1109/JSTARS.2023.3342990. [Hall2021] D. K. Hall et G. A. Riggs, « MODIS/Terra Snow Cover Daily L3 Global 500m SIN Grid, Version 61 ». NASA National Snow and Ice Data Center Distributed Active Archive Center, 2021. doi: 10.5067/MODIS/MOD10A1.061. [Hall1995] D. K. Hall, G. A. Riggs, et V. V. Salomonson, « Development of methods for mapping global snow cover using moderate resolution imaging spectroradiometer data », Remote Sensing of Environment, vol. 54, no 2, p. 127 140, nov. 1995, doi: 10.1016/0034-4257(95)00137-P. [Lê2023] T. T. Lê, A. Atto, E. Trouvé, et F. Karbou, « Deep Semantic Fusion of Sentinel-1 and Sentinel-2 Snow Products for Snow Monitoring in Mountainous Regions », in IGARSS 2023 - 2023 IEEE International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA: IEEE, juill. 2023, p. 6286 6289. doi: 10.1109/IGARSS52108.2023.10282065. [Lievens2019] H. Lievens et al., « Snow depth variability in the Northern Hemisphere mountains observed from space », Nat Commun, vol. 10, no 1, p. 4629, oct. 2019, doi: 10.1038/s41467-019-12566-y. [Montginoux2023] M. Montginoux, F. Weissgerber, S. Lobry, et J. Idier, « Évaluation du couvert neigeux à partir d’images SAR par apprentissage profond basé sur des images optiques de référence », in 29e colloque GRETSI, Grenoble (38000), France, août 2023. Consulté le: 18 septembre 2024. [En ligne]. Disponible sur: https://hal.science/hal-04256105 [Nagler2016] T. Nagler, H. Rott, E. Ripper, G. Bippus, et M. Hetzenecker, « Advancements for Snowmelt Monitoring by Means of Sentinel-1 SAR », Remote Sensing, vol. 8, no 4, Art. no 4, avr. 2016, doi: 10.3390/rs8040348. [Rouhier2018] L. Rouhier, « Régionalisation d’un modèle hydrologique distribué pour la modélisation de bassins non jaugés. Application aux vallées de la Loire et de la Durance », phdthesis, Sorbonne Université, 2018. Consulté le: 28 novembre 2024. [En ligne]. Disponible sur: https://theses.hal.science/tel-02409965 [Weissgerber 2022] F. Weissgerber, L. Charrier, C. Thomas, J.-M. Nicolas, et E. Trouvé, « LabSAR, a one-GCP coregistration tool for SAR–InSAR local analysis in high-mountain regions », Front. Remote Sens., vol. 3, p. 935137, sept. 2022, doi: 10.3389/frsen.2022.935137
3D Radiative Transfer Modeling for Maize Traits Retrieval Across Different Growth Stages: Exploring the Complementarity of Sentinel-2 and CHIME
No full textInternational audience# Context Maize (Zea mays) is one of the world's most important food crops, facing growing uncertainties, such as climatic events and the consequent rise in vulnerability to pests and diseases. The cereal industry constantly faces uncertainties regarding harvest dates, grain quantity, and quality. Consequently, it has become essential to monitor crops more accurately, both spatially and temporally. Satellite Remote Sensing (RS) offers significant potential for near real-time crop monitoring at a global scale through the tracking of biophysical and biochemical crop traits. Existing multispectral instruments such as Sentinel-2 (S2) with high spatial resolution (10-20 m) and revisit rate (approx. 5 days) are already used to monitor the temporal dynamic of key variables such as the Leaf Area Index (LAI) or Leaf Chlorophyll Content (LCC). The spectral information of these instruments is however limiting in key spectral regions (typically the SWIR domain) to monitor Leaf dry Matter Content (LMA), Leaf Water Content (LWC) and Leaf Nitrogen Content (LNC). Upcoming earth observation mission CHIME [1] will provide richer information on crop development thanks to more than 200 contiguous narrow spectral bands at the cost of a lower 11-day revisit rate and 30 m spatial resolution. Hyper- and multi-spectral instruments thus appear highly complementary to monitor crop development, support informed decision-making and ensure global food security. A common crop traits retrieval approach consists in inverting RS data with hybrid methods combining Radiative Transfer Models (RTMs) and Machine Learning (ML) techniques. RTMs help reducing costly and time-consuming field data collection and are used in the inversion approach to link spectral reflectance to crop biophysical and biochemical properties on the [0.4−2.5] μm domain. RTM simulations usually represent vegetation as a horizontally homogeneous environment, while crop fields like maize typically grow on varied terrain slopes with different row orientations and spacing. Advanced three-dimensional (3D) RTMs such as DART [2] are able to simulate RS observations of complex 3D scenes with topography and atmosphere. The use of 3D RTMs provide an opportunity for a growth stage-dependent understanding of the complex interactions between crop field parameters and RS observations. It is however essential for operational crop traits mapping with 3D RTMs to develop semi-automatic scene generation methods able to capture the natural variability of plant architecture at different growth stages. This study has two objectives: - Couple a dynamic 3D plant growth model with the DART RTM to accurately simulate maize field RS observations of S2 and CHIME sensors, for different growth stages and field geometries. - Validate the complementarity of hyper- and multi-spectral instruments for maize field development monitoring from multi-temporal S2 and synthetic CHIME data. # Data The study site includes irrigated maize fields at different phenological stages, located north of Grosseto (42°49′47.02′′N, 11°04′10.27′′E), Italy. The RS dataset consists of two airborne HyPlant-DUAL hyperspectral images (acquired on July 7 and July 30, 2018) collected during the ESA FLEXSENSE campaign [3] and a series of bare soil ASD FieldSpec 4 spectroradiometer measurements (on July 7, 2018). Field measurements were conducted for both flights to measure leaf (LCC, LMA, LNC, LWC) and canopy (LAI) traits as well as architectural features such as plant height and density. Both HyPlant images are used to generate synthetic S2 and CHIME data. Additional level 2A S2 data are selected for the entire maize season from May to August 2018. # Method DART model is combined with the DLAmaize [4] dynamic plant growth model to simulate S2- and CHIME-like maize field reflectances across multiple growth stages, leaf biochemical quantities and field geometries. A Look-Up Tables (LUT) composed of 5000 samples and covering the growth of maize (LAI: 0.0 - 7.0 m²/m²) is simulated using HyPlant-DUAL bands (wavelengths and FWHM), then resampled to S2 and CHIME bands. A global sensitivity analysis (GSA) on the simulated LUTs is used to select a series of optimal vegetation indices (VIs) from the S2 multispectral broad bands (MBB) and the CHIME hyperspectral narrow bands (HNB). The hybrid inversion approach is performed on the selected VIs using the Kernel Ridge Regression (KRR) ML algorithm. The training data for the KRR is first optimized by selecting a subset of ≈1000 samples from the original LUT using an active learning (AL) strategy. The models trained on S2 and CHIME LUTs are then validated against ground measurements for each of LAI, LCC, LNC, LMA and LWC, and used to generate biophysical and biochemical traits maps from synthetic S2 and CHIME data on July 7 and July 30. Finally, the KRR models trained on S2 LUT only are used to monitor the temporal dynamic of LAI and LCC from actual S2 data acquisitions from May to August 2018. # Results The approach was already tested for high spatial-spectral resolution HyPlant-DUAL airborne images at 10 m resolution using the full reflectance spectrum instead of VIs. Preliminary results using the proposed method achieved high retrieval accuracy for LAI (R²=0.91, RMSE=0.42 m²/m²), LCC (R²=0.61, RMSE=3.89 µg/cm²), LNC (R²=0.86, RMSE=1.13×10⁻² mg/cm²), LMA (R²=0.84, RMSE=0.15 mg/cm²), and LWC (R²=0.78, RMSE=0.88 mg/cm²) and a good traits maps spatiotemporal consistency for moderate to high LAI (> 1.5 m²/m²). The findings demonstrated that integrating 3D RTMs with dynamic growth models allow high trait retrieval accuracy from hyperspectral data on heterogeneous row crops at various growth stages. S2 is expected to provide accurate LAI and LCC temporal dynamics from maize emergence to maturation. CHIME at a lower temporal frequency is expected to improve LAI and LCC retrieval accuracy for spatially homogeneous pixels and help to broaden the panel of accessible variables with the estimation of LMA, LNC and LWC challenging to estimate with S2 limited spectral information. # References [1] Celesti, M., Rast, M., Adams, J., Boccia, V., Gascon, F., Isola, C., & Nieke, J. (2022). The Copernicus Hyperspectral Imaging Mission for the Environment (Chime) : Status and Planning. IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, 5011‑5014. https://doi.org/10.1109/IGARSS46834.2022.9883592 [2] Gastellu-Etchegorry, J. (1996). Modeling radiative transfer in heterogeneous 3-D vegetation canopies. Remote Sensing of Environment, 58(2), https://doi.org/10.1016/0034-4257(95)00253-7 [3] Candiani, G., Tagliabue, G., Panigada, C., Verrelst, J., Picchi, V., Rivera Caicedo, J. P., & Boschetti, M. (2022). Evaluation of Hybrid Models to Estimate Chlorophyll and Nitrogen Content of Maize Crops in the Framework of the Future CHIME Mission. Remote Sensing, 14(8). https://doi.org/10.3390/rs14081792 [4] Zhen, Z., Chen, S., Yin, T., Han, C., Chavanon, E., Lauret, N., Guilleux, J., & Gastellu, J.-P. (2024). A Dynamic L-System-Based Architectural Maize Model for 3-D Radiative Transfer Simulation. IEEE Transactions on Geoscience and Remote Sensing, 62, 1‑20. https://doi.org/10.1109/TGRS.2023.334851
Experimental demonstration of coupled nano-Fabry–Perot groove resonators
No full textInternational audienceFabry–Perot (FP) resonances are ubiquitous in plasmonic resonators; however, their quality factor is mostly driven by the losses of the material and has a typical value of 10 for noble metals. The coupling of two FP nanocavities (cFP) was theoretically shown by Opt. Lett. 42, 5062 (2017) to enable the control of the quality factor, thanks to a geometrical parameter. Here, we experimentally demonstrate in the infrared range the cFP resonance with a quality factor of 30 in a structure made of metallic grooves with a trapezoidal-shaped profile that is simultaneously easier to fabricate and more complex to model. An analytical one-mode model is derived that undoubtedly attributes the resonance to the coupling of interferences between the two nanocavities. Finally, we also experimentally demonstrate the combination of two cFP either in orthogonal polarizations or for distinct wavelengths
Impact of Distributed Roughness on Tollmien-Schlichting Instabilities in Airfoil Boundary Layers
No full textInternational audienc
Caractérisation thermique d'un matériau composite thermoplastique
No full textInternational audienceLes travaux présentés concernent la caractérisation des propriétés thermiques d'un matériau composite thermoplastique. Un modèle de dégradation thermique basé sur trois réactions est proposé. Les paramètres des lois d'Arrhenius associées à ces réactions sont identifiés par des essais ATG/DSC. Les propriétés thermo-optiques sont mesurées. Enfin les propriétés thermophysiques sont identifiées en fonction de la température, pour deux stratifications : unidirectionnelle et quasi-isotrope. Ces mesures n'ont pu être réalisées que dans l'état vierge du matériau et devraient être complétées par la caractérisation de l'état dégradé