Attribution of sea-level change to its different drivers is typically done using a sea-level budget approach. While the global mean sea-level budget is considered closed, closing the budget on a finer spatial scale is more complicated due to, for instance, limitations in our observational system and the spatial processes contributing to regional sea-level change. Consequently, the regional budget has been mainly analysed on a basin-wide scale. Here we investigate the sea-level budget at sub-basin scales, using two machine learning techniques to extract domains of coherent sea-level variability: a neural network approach (self-organizing map, SOM) and a network detection approach (δ-MAPS). The extracted domains provide more spatial detail within the ocean basins and indicate how sea-level variability is connected among different regions. Using these domains we can close, within 1σ uncertainty, the sub-basin regional sea-level budget from 1993–2016 in 100 % and 76 % of the SOM and δ-MAPS regions, respectively. Steric variations dominate the temporal sea-level variability and determine a significant part of the total regional change. Sea-level change due to mass exchange between ocean and land has a relatively homogeneous contribution to all regions. In highly dynamic regions (e.g. the Gulf Stream region) the dynamic mass redistribution is significant. Regions where the budget cannot be closed highlight processes that are affecting sea level but are not well captured by the observations, such as the influence of western boundary currents. The use of the budget approach in combination with machine learning techniques leads to new insights into regional sea-level variability and its drivers.
Over the last two decades, Vibrio vulnificus infections have emerged as an increasingly serious public health threat along the German Baltic coast. To manage related risks, near real-time (NRT) modelling of V. vulnificus quantities has often been proposed. Such models require spatially explicit input data, for example, from remote sensing or numerical model products. We tested if data from a hydrodynamic, a meteorological, and a biogeochemical model are suitable as input for an NRT model system by coupling it with field samples and assessing the models’ ability to capture known ecological parameters of V. vulnificus. We also identify the most important predictors for V. vulnificus in the Baltic Sea by leveraging the St. Nicolas House Analysis. Using a 27-year time series of sea surface temperature, we have investigated trends of V. vulnificus season length, which pinpoint hotspots mainly in the east of our study region. Our results underline the importance of water temperature and salinity on V. vulnificus abundance but also highlight the potential of air temperature, oxygen, and precipitation to serve as predictors in a statistical model, albeit their relationship with V. vulnificus may not be causal. The evaluated models cannot be used in an NRT model system due to data availability constraints, but promising alternatives are presented. The results provide a valuable basis for a future NRT model for V. vulnificus in the Baltic Sea.
Horizontal patchiness of water quality attributes in lakes substantially influences the ability to accurately determine an average condition of a lake from traditional in situ sampling. Monitoring programmes for lake water quality often rely on water samples from one or few locations but the assumption of representativeness is seldomly tested. Satellite observations can support environmental monitoring by detecting horizontal variability of water quality attributes over entire lakes. This article is a co-submission with Lehmann et al. (2021), who present a method to create a regional calibration of a satellite chlorophyll a algorithm and a spatial analysis of an image time series to detect recurring patchiness. Our method was developed on 13 lakes in the central North Island of New Zealand and this publication makes available the data used in our analysis and the spatial fields of results. These data are immediately valuable for practitioners operating within the region of interest providing a five year archive of synoptic water quality data and spatial fields to help optimize in situ monitoring efforts. In addition, there is value to the wider scientific community as the study lakes are a useful ‘natural lab’ for the development of aquatic remote sensing methods due to the range of trophic conditions and water colour in a single satellite image scene. Together with decades of in situ water quality records, our data is therefore useful for the development and validation of widely applicable methods of water quality retrieval from satellite data.
Attribution of sea-level change to its different drivers is typically done using a sea-level budget (SLB) approach. While the global mean SLB is considered closed, closing the SLB on a finer spatial scale is more complicated due to, for instance, limitations in our observational system and the spatial processes contributing to regional sea-level change. Consequently, the regional SLB has been mainly analysed on a basin-wide scale. Here we investigate the SLB at sub-basin scales, using two machine learning techniques to extract domains of coherent sea-level variability: a neural network approach (Self-Organising Maps) and a network detection approach (δ-MAPS). The extracted domains provide a higher level of spatial detail than entire ocean basins and besides indicating how sea-level variability is connected among different regions. Using these domains we can close the regional SLB world-wide on different spatial scales. Steric variations dominate the temporal sea-level variability and determine a significant part of the total regional change. Sea-level change due to mass transport between ocean and land has a relatively homogeneous contribution to all regions. In highly dynamic regions (e.g., Gulf Stream region) the dynamic mass redistribution is significant. Regions where the SLB cannot be closed highlight processes that are affecting sea level but are not well captured by the observations, such as the influence of western boundary currents. Hence, the use of the SLB approach in combination with machine learning techniques leads to new insights into regional sea-level variability and its drivers.
7 May 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
The increasing spatial resolution of satellite sensors– now reaching sub-meter scale– enables advanced applications in aquatic remote sensing, including highly detailed bathymetry or habitat mapping. However, this fine resolution also increases pixel-level susceptibility to wave glint and whitecap contami nation. These wave-induced reflectance anomalies (WIA) can severely degrade image quality and reduce the accuracy of derived products. Although strategic acquisition planning can help to minimize WIA, users often have limited control over satellite tasking. As a result, WIA contamination is difficult to prevent in practice. In this study, we present a novel WIA correction method for pan-sharpened Pl´ eiades-1 imagery that detects and reconstructs affected pixel values. WIA features are identified with an ensemble of similarity-weighted SVMs, which exploit spatial curvature and near-infrared reflectance. Detected anomalies are then corrected with a spatially adaptive gap filling technique. Our method operates on top-of-atmosphere reflectances, allowing users flexibility in further data preprocessing steps like atmospheric correction. Our approach demonstrates very high accuracy in WIA detection and shows comparable variance reductions in optically deep water compared to the popular Hedley glint correction and the Water Anomaly Filter. In optically shallow environments, our method outperforms these benchmarks by effectively removing WIA while preserving benthic spatial detail. Requiring minimal parameter tuning, the approach is well-suited for integration into automated preprocessing pipelines. It provides a practical tool for improving data quality and enhancing the reliability of downstream applications. Our approach is also easily adaptable to other very high-resolution sensors with similar spatial resolutions.