Detecting ships from satellite imagery is vital for maritime surveillance. Most current methods rely on deep learning (DL), which requires a large number of high-quality annotated images to train accurate models. Since satellite imagery comes from various sensors, DL-based ship detection algorithms need to perform well across different sensor types. However, privacy concerns, especially with commercial images, limit data, and annotation sharing. Federated learning (FL) offers a promising solution for collaborative learning while addressing these concerns. Despite its potential, research on FL for ship detection is still sparse. This study implements and evaluates three FL models for detecting ships using multi-source optical satellite images, spanning high to low resolution. Our experiments on two distinct datasets demonstrate that FL models significantly enhance detection performance without centralizing data. Source codes are publicly available athttps://github.com/ffyyytt/FLYOLO.

Object detection tasks still often require manual image analysis. Using Machine Learning (ML) instead creates accountability challenges, necessitating experts for model refinement, which is costly and takes time. We investigate integrating crowd knowledge as a cost-effective alternative. While human capabilities in recognizing complex patterns and perceiving variations can still outperform machines and improve an imperfect ML model, ML predictions can compensate for the crowd’s lack of expertise. Our investigation (N=28 non-expert) in oil spill detection shows that adopting an ML-assisted UI elevates precision and recall by over 11% and increases efficiency by 29% compared to a non-assisted UI. Considering agreement among non-expert crowd workers further improved precision by 8% and recall by almost 5%, which is also substantially beyond pure ML performance. Our work contributes an approach for combining crowd knowledge and ML to advance human-AI collaboration in oil spill detection.

This paper addresses the challenges of an early flood warning caused by complex convective systems (CSs), by using Low-Earth Orbit and Geostationary satellite data. We focus on a sequence of extreme events that took place in central Vietnam during October 2020, with a specific emphasis on the events leading up to the floods. In this critical phase, several hydrometeorological indicators could be identified thanks to Earth Observation satellites, which enable the characterization and monitoring of a CS, in terms of low-temperature clouds and heavy rainfall. Himawari-8 (H8) images, both individually and in time-series, allow identifying and tracking convective clouds. This is complemented by the observation of heavy/violent rainfall through GPM IMERG data, and the detection of strong winds using radiometers/scatterometers. Collectively, these datasets, along with the estimated intensity and duration of the event from each source, form a comprehensive dataset detailing the intricate behaviors of CSs. All of these factors are significant contributors to the magnitude of flooding and the short-term dynamics anticipated in the studied region.

Ship detection from satellite imagery using Deep Learning (DL) is an indispensable solution for maritime surveillance. However, applying DL models trained on one dataset to others having differences in spatial resolution and radiometric features requires many adjustments. To overcome this issue, this paper focused on the DL models trained on datasets that consist of different optical images and a combination of radar and optical data. When dealing with a limited number of training images, the performance of DL models via this approach was satisfactory. They could improve 5–20% of average precision, depending on the optical images tested. Likewise, DL models trained on the combined optical and radar dataset could be applied to both optical and radar images. Our experiments showed that the models trained on an optical dataset could be used for radar images, while those trained on a radar dataset offered very poor scores when applied to optical images.

This paper proposes the observations of oil drift and the changes in oil shape and size based on the collocation of Sentinel-1 descending and ascending images with a time lag of 12 hours offshore Nigeria. The oil slicks are first detected from the descending and ascending images using the hierarchical split-based approach to identify oil objects and non-linear filters (mean and standard deviation) to determine oil contours. Then, the detected oil is collocated to estimate the distance and direction of its movements. Finally, surface wind and current data are used for analyzing the relationship between met-ocean conditions and the evolution of oil slicks.

Previous references indicated that surface wind gusts associated with deep convection can be observed and estimated from Sentinel-1 images. They also presented the relationship between surface wind patterns and deep convective clouds observed on Meteosat geostationary (GEO) images. To strengthen this relationship, this paper presents the comparison between surface wind speed retrieved from Sentinel-1 data, wind magnitude measured by the weather stations, and deep convective clouds observed on GOES-16 GEO images over the Gulf of Mexico. The results show that a mesoscale surface wind pattern (a squall line) observed on Sentinel-1 images corresponds to deep convective cloud locations. In particular, the peaks of wind intensity measured by the weather stations match the Sentinel-1 wind gusts and the deep convective clouds.

Based on the combination of the matching images from Meteosat geostationary, Aeolus Lidar, and Sentinel-1 polar-orbiting satellites, this study proposes a new way to observe convective systems (CS) in three-dimensions. This 3D observation follows the schematic CS dynamics: an intense downdraft associated with a deep convection cloud may induce wind gust from moderate to high intensity (10–25 m/s) when it hits the sea surface. In addition, the combination of Sentinel-1 and SMAP enables observation of the CS surface wind pattern displacement. The motion direction and speed of the wind patterns seem to be similar to those of the deep convection clouds.

Based on the combination of low altitude and geostationary satellite sensor large database, the use of machine learning algorithm now let to produces some kind of virtual ocean surface wind fields under clouds - observed by geostationary satellites -while there is no actual observation of these winds. As a result, it is then possible to extrapolate ocean surface wind database from a smart combination of low altitude satellites and the high frequency acquired data from geostationary satellites. This database is then available to investigate climatologic relevant location of wind powerful areas. The same process may be available to monitor the wind energy plant in real time by using geostationary satellite every few minutes refresh.

Geostationary meteorological satellites (METEOSAT, GOES, Himawari) allow the detection of Convective Systems (CS) to become more accurate. In the meantime, thanks to the Synthetic Aperture Radar (SAR) onboard the polar-orbit satellites, one may propose the detection of high-resolution surface wind patterns associated with the CS. The detected wind patterns in this study have the shape of a squall line and wind intensity varying from 10-25 m/s. In particular, the wind patterns have close locations and comparable shapes to those of the deep convective clouds observed by the geostationary satellites (METEOSAT).

The use of a single source of data - very high-resolution SAR data from polar earth orbiters - allows the computation of several major environmental parameters over ocean surface: high-resolution winds (with a brand-new method), ship detection, oil spill detection, sea-ice detection. These computations are based both on deterministic and machine learning methods The monitoring of all these parameters is gathered in a single operational Web platform that allows and triggers computation on demand.

This paper puts forward a practical application of emergent cloud computing technologies that addresses is-sues encountered in the earth observation and meteorological domain. Ever increasing satellite capture technologies result in larger and larger data that is being processed by more and more complex algorithms for the generation of weather reports, forecasts, etc. Companies and researchers that work on this kind of processing chains struggle with storage and computing issues that cloud computing technologies can overcome. This kind of architecture becomes all the more relevant as cloud providers propose accessible pay-per-use services that make such infrastructure economically viable. This article relates the complexity of satellite data processing with the solutions that cloud computing can offer. It details use cases that highlight the need for a centralized access to data between processing resources with localization constraints, for scalable storage and distributed computing capabilities, and for hardware optimization to minimize both costs and processing time.

Surface wind speed estimation from synthetic aperture radar (SAR) data is principally based on empirical (EP) approaches, e.g., CMOD functions. However, it is necessary and significant to compare radar backscattering modeling based on EP and electromagnetic (EM) approaches for enhancing the understanding of the physical processes between radar signal and sea surface, which is important for the design of radar sensors (e.g., cyclone global navigation satellite system). Indeed, through comparisons, it is worth noticing that the scattering of wave breaking is not taken into account in the physical modeling of radar backscattering. Surface wind speed is selected here as a reference parameter for investigating the difference between EP and EM models, due to its important role in radar backscattering modeling. In addition, wind speed estimates can be easily compared to in situ measurements. For EP approach, CMOD5.N and Komarov's model are selected for wind speed estimation from Sentinel-1 images. The CMOD5.N can offer wind speed estimates up to 25-35 m/s, while wind speed estimation based on Komarov's model does not require wind direction input. For EM approach, the asymptotic models, i.e., composite two-scale model, small-slope approximation (SSA), and resonant curvature approximation (RCA), are investigated for wind speed retrieval. They are studied with two models of surface roughness spectrum: semi-EP spectrum and EP model. In general, normalized radar cross section (NRCS) calculated by CMOD5.N and SSA/RCA is quite similar for incidence angles below 40° in vertical polarized and below 30° in horizontal polarized. For larger ones, significant NRCS deviations between two approaches are demonstrated, due to the lack of wave breaking scattering in EM models. As a result, wind speed estimates by CMOD5.N and SSA/RCA are very close for low and moderate incidence angles, while SSA-/RCA-based wind speeds are overestimated for larger ones.

Despite based on different approaches and objectives, it is reasonable to compare near-surface wind speed estimated by the empirical (EP) and electromagnetic (EP) geophysical model function (GMF), since both of them describe the relation between radar scattering and wind vector (directly for EP GMF and via surface roughness spectrum for EM GMF). In general, the EP and EM models give quite similar normalized radar cross section (NRCS) for radar incidence angle below 40°. Consequently, wind speed estimated by the EP and EM GMF is very close. However, for incidence angles above 40°, the EM models show poor performance of wind speed estimation.

Based on an empirical model without wind direction input for the retrieval of C-band HH-polarization wind speed, we propose a modified model for wind speed estimation in VV-polarization. The obtained wind speed is then applied for the CMOD5.N to estimate wind directions. The comparisons with the scatterometry-based approach demonstrate that the estimated wind speed by the proposed model is closer to in situ measurements than that obtained with the CMOD5.N. Likewise, the extracted wind directions from the CMOD5.N are more accurate that those obtained with the local gradient method.

Retrieval of sea surface wind speed from Synthetic Aperture Radar (SAR) data is one of the most widely used methods, since it can give higher resolutions than the other available surface wind sources. For this approach, two principal methods can be found: one is based on electromagnetic (EM) models and the other is based on empirical (EP) ones. In both indicated ways, the Geophysical Model Functions (GMFs) are used to describe the dependency of radar scattering on wind speed and the geometry of observations. By knowing radar scattering and geometric parameters from SAR data, it is possible to invert the GMFs to retrieve sea wind speed. Wind speed estimated by two studied models is compared together and evaluated with measured data. Based on the comparisons, the advantages and limits of the studied models are analyzed and discussed.

Wind direction is a crucial parameter in many inversion models to estimate wind speed from Synthetic Aperture Radar (SAR) data. Compared to the other available wind sources, i.e. measured data, numeric weather data, etc., the retrieval of wind directions from SAR data is more widely used, since it can give wind directions at different scales. Nevertheless, there are not a lot of studies which report about the sensitivity of wind direction retrieval, particularly with regard to the spatial resolution (or acquisition mode) of SAR images, speckle noise, and wind regimes. In order to investigate this issue, the Local Gradient method is selected to retrieve wind directions from the Sentinel-1 images at different scales, since it can give high wind resolution cells. Then, the impact of speckle noise and wind regimes on retrieved wind directions is assessed.

Because radar systems and TV reception are affected by wind turbines, the important development of hybrid-powered BTS (fuel, solar and wind energy) raises questions about possible radio impact between cellular radio signals and the small wind turbines (SWTs) that contribute to BTS supply. SWTs come in a wider variety of designs than bigger wind turbines. In the context of the OPERA-Net2 European project, a new generic parameterized model is developed to support electromagnetic interaction studies. It is notably scalable depending on the SWT nominal power to be considered. Monostatic Radar Cross Section (RCS) calculation in the wind turbine principal planes is done to select among different possible shapes and dimensions. Then an original study of the near field coupling between SWT and BTS panel antenna is carried out.