We present an end-to-end deep learning framework for short-term forecasting of global sea surface dynamics based on sparse satellite altimetry data. Building on two state-of-the-art architectures: U-Net and 4DVarNet, originally developed for image segmentation and spatiotemporal interpolation, respectively, we adapt the models to forecast the sea level anomaly (SLA) and sea surface currents over a seven-day horizon using sequences of sparse nadir altimeters observations. The model is trained on data from the GLORYS12 operational ocean reanalysis, with synthetic nadir sampling patterns applied to simulate realistic observational coverage. The forecasting task is formulated as a sequence-to-sequence mapping, with the input comprising partial SLA snapshots and the target being the corresponding future full-field SLA maps. We evaluate model performance using (i) normalized root mean squared error (nRMSE), (ii) averaged effective resolution, (iii) percentage of correctly predicted velocities, magnitudes and angles, and benchmark results against the operational Mercator Ocean forecast product. Results show that end-to-end neural forecasts outperform the baseline across all lead times, with particularly notable improvements in high variability regions. Our framework is developed within the OceanBench benchmarking initiative, promoting reproducibility and standardized evaluation in ocean machine learning. These results demonstrate the feasibility and potential of end-to-end neural forecasting models for operational oceanography, even in data-sparse conditions.

Melt ponds are puddles of water on Arctic summer sea ice that play an important role in the Arctic climate system, mainly through their influence on the surface energy balance. The albedo of polar oceans is primarily determined by the area fractions of sea ice and darker open ocean. However, the presence of melt ponds also significantly reduces the reflective properties of sea ice. This allows more solar energy to be absorbed, accelerating regional ice melt in a phenomenon known as ice-albedo feedback. Despite their importance, there is a lack of observational datasets that explicitly parameterize melt ponds next to sea ice and open water, and the existing data are mainly based on optical imagery. In this study, we present a new helicopter-borne dataset based on (TIR) and employ a deep learning-based semantic segmentation approach to classify surface types (melt ponds, sea ice, and open water). We evaluate the performance of several model architectures: U-Net, UNet++, PSPNet, DeepLabV3+, SwinTransformer, SAM2-UNet, and AutoSAM. Among these models, UNet++ pre-trained on the AID dataset demonstrates the highest performance, achieving a melt pond (IoU) of 0.486, a sea ice IoU of 0.764, and an open water IoU of 0.660. Furthermore, we apply the most effective model to the datasets of five helicopter flights, resulting in (MPF) estimates with a bias of approximately 11%, which is higher than for melt pond classifications based on visible observations.

Species trait data are widely used in ecology and evolution to approximate ecological niche. Despite this, the data housed in species trait databases are often patchy or biased toward charismatic species and easy-to-study traits. Filling data gaps can be achieved by determining the most likely values via data imputation. We conducted a systematic literature review to identify data imputation methods for species trait data in ecological and evolutionary studies, as well as in data science studies, and compared the two domains to identify whether alternative methods might be suitable for adoption in ecology and evolution. We focus on novel and state-of-the-art deep learning (DL) imputation methods. We use evidence maps to identify missing data contexts where DL alternatives have been shown to be effective in data science. Differences in preferred methods were identified between the two domains. DL methods were rarely used in ecological and evolutionary studies, whereas they were frequently employed in data science. Several DL methods that have not been adopted in ecology and evolution demonstrated state-of-the-art performance when imputing data in contexts that are similar to those faced by ecologists wishing to impute species trait data. We identified a clear gap in the ecological and evolutionary literature in relation to the uptake and use of imputation methods based on DL. We then identify autoencoders, transformer-based architectures, Bayesian neural networks, and feed-forward neural networks as primary architectures worth further investigation for their capacity to impute missing species trait data and potential for multiple imputation.

The growing demand for reusable micrometeorological data in climate, environmental, and agricultural research, driven by challenges including weather-induced agricultural losses, water use for irrigation, emerging invasive species and disease vectors, deforestation, and increasing urbanization, highlights the need for datasets that comply with the FAIR (Findable, Accessible, Interoperable, and Reusable) principles. FAIR data and metadata should be easily discoverable by humans or machines, accessible under specific conditions or restrictions, conform to recognized formats and standards to be combined and exchanged, and licensed according to community norms, allowing users to know what kinds of reuse are permitted. To support researchers in creating FAIR-compliant datasets, the FAIR Micromet Portal was developed within the CA20108 COST Action. The portal provides a guided, step-by-step workflow for generating standardized metadata, linking datasets with persistent identifiers, and facilitating compliance with the FAIR principles. This paper presents the design and implementation of the FAIR Micromet Portal and describes how its metadata model and workflow satisfy the complete set of FAIR metrics. Using a micrometeorological case study, we demonstrate how the portal enables researchers to create FAIR-compliant metadata with minimal effort while supporting data discovery, interoperability, and long-term reusability.

Bias adjustment is standard preprocessing for climate model output used in crop-impact and damage-assessment work. Most validation studies still emphasize mean temperature, even though agricultural applications often depend on nonlinear heat metrics built from daily extremes. We evaluate how bias adjustment propagates into Extreme Degree Day (EDD) metrics, which crop-yield and damage functions use to summarize heat stress above specified temperature cutoffs. Using ERA5 reanalysis as reference, we compare EDD from raw and bias-adjusted CMIP6 models (GFDL-ESM4, MPI-ESM1-2-HR, and EC-Earth3) across European ADM1 regions for 1994–2014. We compute EDD at seven thresholds (0, 4, 8, 12, 28, 30, and 32 °C), aggregate it with harvested-area weights for seven crops under rainfed and irrigated management, and sum it over crop-specific phenological windows. Bias adjustment reduces mean absolute error at low and moderate thresholds (EDD $ \le $ 12 °C) by 37–86 °C$ \cdot $days at EDD-0 and 29–64 °C$ \cdot $days at EDD-8. At extreme thresholds (EDD $ \ge $ 28 °C), $ \Delta $Skill is small on average (1–4 °C$ \cdot $days at EDD-28), and bias adjustment increases error in about 40% of evaluated combinations. The split between helpful and unhelpful bias adjustment follows the EDD threshold more closely than model identity, which has implications for impact studies that treat bias-adjusted data as uniformly preferable.

With the ongoing effects of global warming, Antarctic supraglacial meltwater plays a critical role in ice-sheet instability. Spaceborne remote sensing provides a viable solution to detect surface melt in these vast and remote regions. Innovative deep learning image segmentation methods have the potential to improve supraglacial lake (SGL) mapping by adding spatial awareness to the mapping process, as currently established methods only consider the spectral variety of multispectral satellite data (Sentinel-1 and Landsat-8) on a per-pixel basis. In this study, we adapt a U-net image segmentation method to detect SGLs in Sentinel-2 data and quantify its performance in comparison to established mapping approaches in Antarctica. At a 90% probability threshold, the proposed workflow achieves a producer’s accuracy, user’s accuracy, and an F1 score of 0.915, 0.912, and 0.913, respectively. We compare the classification results to those of an established Random Forest classifier, and find that the U-net achieves better results in 13 of the 16 validation scenes. In a relative comparison with five established SGL products, the U-net approach maps, on average, the largest lake area, although there is disagreement between the datasets due to different data sources, cloud cover, and misclassification of some clouds as lakes. The main advantages of the developed approach are its capabilities to map lakes under thin clouds and ice, resulting in more continuous lake areas than conventional approaches, and a reduced number of input features required, which will be advantageous in future large-scale SGL mapping efforts.

Effective adaptation and mitigation strategies for climate change require high-resolution projections to inform strategic decision-making. Conventional global climate models, which typically operate at resolutions of 150 to 200 km, lack the capacity to represent essential regional processes. IPSL-AID is a global to regional downscaling tool based on a denoising diffusion probabilistic model designed to address this limitation. Trained on ERA5 reanalysis data, it generates 0.25° resolution fields for temperature, wind, and precipitation using coarse inputs and their spatiotemporal context. It also models probability distributions of fine-scale features to produce plausible scenarios for uncertainty quantification. The model accurately reconstructs statistical distributions, including extreme events, power spectra, and spatial structures. This work highlights the potential of generative diffusion models for efficient climate downscaling with uncertainty estimation. Climate downscaling generally comprises two components—super-resolution and bias correction. This work focuses on the former.

While deep learning (DL) methods have expanded the approach to solving state estimation problems in data assimilation (DA), a robust approach requires extending them for uncertainty quantification. In this work, we propose a method to represent location-based uncertainty through random, spatiotemporally coherent displacements in physical space, which are used to sample states in the vicinity of a given reference state. We integrate this strategy into an end-to-end variational DA framework, termed 4DVarNet-LU, and compare its performance with a UNet baseline, demonstrating improved reconstruction skill through the explicit use of prior state information. Trained deterministically in output and stochastically in input, we demonstrate that at inference, we see improvement in skill when using information from initial conditions. Using observation system simulation experiments based on a single-layer quasi-geostrophic model with sparse pseudo-nadir altimetry observations, we evaluate reconstruction mean and ensemble skill. Results show that 4DVarNet-LU consistently outperforms the UNet baseline in terms of root mean square error (RMSE) and continuous ranked probability score (CRPS), with increased robustness to initial condition bias and improved temporal stability. While the method does not propose to account for all different uncertainty sources in DA, it provides a structured way for modeling and training DL models for dealing with location-based uncertainty. Future work may extend this approach to account for additional sources of uncertainty and end-to-end ensemble generation.

Coral reef islands are known to exhibit significant sub- and inter-annual coastal variability. Maldivian coastlines are not systematically or frequently monitored, so modes of variability and morphological responses to climate forcings are not well understood and characterized. Here, a data processing pipeline was developed to reveal spatiotemporal patterns of coastal change by leveraging medium-resolution satellite imagery. The pipeline was applied to the islands of Huvadhoo Atoll, Maldives, but is also applicable to other Small Island Developing States. Coastline positions were extracted monthly from Sentinel-2, Landsat-8 and 9 images, using an image segmentation algorithm. Trend and seasonal signals were disentangled from time series of coastline position, to reveal short- and long-term modes of variability. Results show island- and atoll-scale spatial patterns of coastal change. 28% of the atoll’s studied islands show detectable oscillating coastline position changes, suggesting sub-annual alongshore sediment transport, coinciding with the phases of the Indian Monsoon and regional changes in met-ocean conditions. Islands located in the lagoon experience greater variability than those on the atoll rim. Seasonal erosion and accretion periods were found to be asymmetrical and aligned with the duration of the Indian Monsoon phases. The seasonal signal is modulated by inter-annual variability in met-ocean conditions (e.g., Indian Ocean Dipole events). Therefore, in order to forecast the impact of climate change on coral reef island equilibrium, drivers likely to compound sea level rise need to be identified and their effects quantified. These results can be used to support evidence-based environmental policies, coastal planning, and marine infrastructure development.

Understanding how anthropogenic environmental signals reorganize over time is central to Environmental Data Science, particularly because long-run emission trajectories reflect interacting technological, institutional, behavioral, and macroeconomic processes whose effects need not appear as smooth changes in level. This study develops a reproducible analytical framework that combines rolling spectral entropy with Bayesian change-point inference to examine structural transitions in annual transport-related CO2 emissions across the G7 economies from 1970 to 2023. The analysis adopts a structure-aware interpretation of the series while specifying what spectral entropy can and cannot establish in short annual windows. Entropy is interpreted as the concentration or dispersion of detrended within-window variability across resolvable Fourier components, rather than as a direct physical measure of resilience. The baseline window is $ W=10 $ years, with $ W=8 $ and $ W=12 $ used for sensitivity assessment. Country-specific Bayesian single-mean-shift models are estimated independently, and posterior means are reported together with 90% credible intervals and the shape of the complete posterior distribution. The results reveal recurrent entropy troughs and collapse-rate spikes in the early 1980s, around 1998–2000, and in the late 2000s. The late-1990s episode constitutes a conspicuous descriptive cross-country alignment, but the Bayesian evidence is heterogeneous: several countries exhibit broad or multimodal posteriors, so the analysis does not treat a common G7 change year as established. The recent-collapse calculation is interpreted in light of the exceptional COVID-19 shock. Overall, the evidence supports the value of spectral descriptors as exploratory indicators of changing temporal organization, provided that short-window frequency resolution, autocorrelation, transient shocks, and posterior uncertainty are made explicit.

Stream water temperature strongly influences aquatic ecosystem health, affecting dissolved oxygen, species distributions, and thermal stress for sensitive taxa. We present a data-driven framework to predict water temperatures across rivers in the Canton of Vaud, Switzerland, explicitly quantifying the influence of riparian vegetation. Meteorological and topographic features were extracted at each water temperature station. Vegetation features were derived from satellite imagery data, using Vegetation Height Models (VHM) and Specific Leaf Area Vegetation Index (SLAVI), and aggregated along upstream river segments to capture the influence of riparian zones on water temperature. These features were incorporated into a machine-learning model using grouped inputs with Softmax-based weighting for meteorology and a Gaussian attention layer for vegetation. Spatial cross-validation ensured realistic generalization. The model achieves a median mean absolute error of 1.3 °C for daily 90th percentile water temperatures. Model analyses show that dense riparian vegetation can reduce predicted stream temperatures by nearly 3 °C on warm days. Moreover, the spatial influence of vegetation varies with river slope: in flatter reaches, vegetation close to the water temperature station exerts the strongest cooling effect, whereas in steeper reaches, vegetation located further upstream becomes more influential, consistent with faster flow transporting thermal signals downstream. These results underscore the importance of incorporating spatially explicit riparian vegetation metrics in predictive models. The framework offers actionable insights for river management, showing that targeted riparian preservation or restoration, even in narrow corridors, can effectively mitigate thermal stress and support climate adaptation strategies.

Accurate and high-resolution estimation of land surface temperature (LST) is crucial in estimating evapotranspiration, a measure of plant water use and a central quantity in agricultural applications. In this work, we develop a novel statistical method for downscaling LST data obtained from NASA’s ECOSTRESS mission, using high-resolution data from the Landsat 8 mission as a proxy for modeling agricultural field structure. Using the Landsat data, we identify the boundaries of agricultural fields through edge detection techniques, allowing us to capture the inherent block structure present in the spatial domain. We propose a block-diagonal Gaussian process (BDGP) model that captures the spatial structure of the agricultural fields, leverages independence of LST across fields for computational tractability, and accounts for the change of support present in ECOSTRESS observations. We use the resulting BDGP model to perform Gaussian process regression and obtain high-resolution estimates of LST from ECOSTRESS data, along with uncertainty quantification. Our results demonstrate the practicality of the proposed method in producing reliable high-resolution LST estimates, with potential applications in agriculture, urban planning, and climate studies.

Machine learning models can represent climate processes that are nonlocal in horizontal space, height, and time, often by combining information across these dimensions in highly nonlinear ways. While this can improve predictive skill, it makes learned relationships difficult to interpret and prone to overfitting as the extent of nonlocal information grows. We address this challenge by introducing data-driven integration kernels, a framework that adds structure to nonlocal operator learning by explicitly separating nonlocal information aggregation from local nonlinear prediction. Each spatiotemporal predictor field is first integrated using learnable kernels (defined as continuous weighting functions over horizontal space, height, and/or time), after which a local nonlinear mapping is applied only to the resulting kernel-integrated features and optional local inputs. This design confines nonlinear interactions to a small set of integrated features and makes each kernel directly interpretable as a weighting pattern that reveals which horizontal locations, vertical levels, and/or past timesteps contribute most to the prediction. We demonstrate the framework for South Asian monsoon precipitation using a hierarchy of neural network models with increasing structure, including baseline, nonparametric kernel, and parametric kernel models. Across this hierarchy, kernel models achieve near-baseline performance with far fewer trainable parameters, indicating that much of the relevant nonlocal information can be captured through a small set of interpretable integrations when appropriate structural constraints are imposed.

The North Atlantic Oscillation (NAO) is the dominant mode of atmospheric variability over the North Atlantic, influencing temperature and precipitation across Europe. While the NAO’s impact on North Atlantic sea surface temperatures (SSTs) is well understood, the NAO can also be driven by SST anomalies, although this response is believed to be weak and nonlinear. Former studies highlight that during early winter (November–December), El Niño-Southern Oscillation (ENSO) events modulate the NAO, with El Niño (La Niña) events being linked to positive (negative) NAO phases, and an opposite effect in late winter (January–February). Indian Ocean SSTs and the North Atlantic Horseshoe SST anomaly have also been suggested as contributors to early winter NAO variability. However, climate models often struggle to capture these SST-NAO teleconnections, particularly in early winter. To address this, a statistical framework based on convolutional neural networks (CNNs) is developed to predict the early winter NAO using observed SST fields one, two, and three months before. A linear model serves as a benchmark, and both are trained on ERA5 reanalysis data from 1940 to 2023. A sensitivity analysis interprets the CNN’s decision-making process, revealing that it focuses on regions such as the tropical Pacific and North Atlantic, confirming earlier findings. The CNN outperforms the linear model, highlighting the value of capturing nonlinear SST-NAO relationships. Prediction skill appears to be linked to ENSO, with strong events associated with greater skill than neutral events. These findings underscore the potential of deep learning to build NAO prediction.

Ambitious decarbonization targets are rapidly increasing the commissioning of new offshore wind farms. For these newly commissioned plants to run, accurate power forecasts are needed from the onset. These allow grid stability, good reserve management, and efficient energy trading. Despite machine learning models having strong performance, they tend to require large volumes of site-specific data that new farms do not yet have. To overcome this data scarcity, we propose a novel transfer learning framework that clusters power output according to covariate meteorological features. Rather than training a single, general-purpose model, we thus forecast with an ensemble of expert models, each trained on a cluster. As these pretrained models each specialize in a distinct weather pattern, they adapt efficiently to new sites and capture transferable, climate-dependent dynamics. Our contributions are two-fold—we propose this novel framework and comprehensively evaluate it on eight offshore wind farms, achieving accurate cross-domain forecasting with under 5 months of site-specific data. Our experiments achieve an MAE of 3.52%, providing empirical verification that reliable forecasts do not require a full annual cycle. Beyond power forecasting, this climate-aware transfer learning method opens new opportunities for offshore wind applications such as early-stage wind resource assessment, where reducing data requirements can significantly accelerate project development whilst effectively mitigating its inherent risks.

This article contributes a machine learning approach to detecting environmental crimes in Brazil. Detecting these crimes is important for the health of the environment and public safety. The measurement of environmental crimes is challenging, in part due to efforts of those involved to remain hidden. However, administrative data on embargoes—legal liens on parcels of land—identify locations where the government documented some of these crimes. Using these data, the analysis develops a method for detecting yet-unidentified locations of these crimes. Embargo data are merged with a wide range of predictors that capture features of the physical environment, as well as demographic, socioeconomic, political, and public safety attributes. After training a series of models on these data, the random forest (RF) model performs best (AUC = 0.92). Overall, a combination of changes in vegetation, homicide rates, the prevalence of young men, human development, employment, the presence of a poverty reduction program, and political factors that capture local electoral competition and participation are important predictors of illegal deforestation. These findings—some unsurprising and some surprising—show the value of multicausal, cross-disciplinary approaches to understanding and detecting environmental crime. The practical gains from integrating remote sensing, a wide range of administrative data, and machine learning help build more effective crime detection, enhancing the government’s ability to conduct forest enforcement more safely and efficiently. Future research is promising as more and better data become available, and the approach can be extended to detect other illicit behavior, in Brazil and elsewhere.

Artificial intelligence (AI) can rapidly process large ecological datasets, uncover patterns, and inform conservation decisions, but responsible adoption depends as much on human-centric skills as on technical methods. Ecologists face steep learning curves, an overwhelming and fast-evolving model landscape, uneven access to data and computing, and a growing transparency deficit. These challenges require human-centric skills like time management, critical thinking, collaboration, communication, creativity, and project management to select, implement, interpret, and responsibly translate AI outputs into ecological insight and action. We led a workshop, EcoViz+AI: Visualization and AI for Ecology, that brought together 35 experts to synthesize practical guidance for navigating these challenges across the AI pipeline. Using workshop discussions and experiences as a foundation, this position paper proposes practical solutions and complementary human-centric skill development to address these challenges: (1) educational resources that support opportunity-cost reasoning (time management) and methodological judgment (critical thinking), (2) communities of practice that build inclusive shared expertise (collaboration and mentorship), (3) effective visualizations that improve interpretability and strengthen transparency of model behavior and uncertainty (creativity and communication), and (4) computational resources that reduce implementation burden through shared data, extensible code, and accessible infrastructure (project management and problem-solving). Our workshop compiled resources, including science communication videos for five AI use cases and repositories for ecology-related AI models and communities of practice. Emphasizing human-centric skills and working in tandem with efforts to promote open science and computational literacy can make AI in ecology more rigorous, equitable, and ecologically relevant, advancing research and conservation.

Biomass burning aerosols (BBAs) from Southern Africa seasonally overlie the semi-permanent South-East Atlantic (SEA) stratocumulus deck, impacting the region’s energy budget through complex aerosol-cloud-radiation-meteorology interactions. Climate model intercomparison initiatives, like the Aerosol Comparisons between Observations and Models (AeroCom), have highlighted the large inter-model variability for BBA radiative effects, especially over the SEA, due to parameterization of emission modeling and smoke properties. Observational constraints are needed to reduce these uncertainties, but correlative observational studies are typically affected by confounding meteorological influences. We propose a physically informed statistical approach, based on causal graphs applied to satellite observations, to disentangle BBA influences on shortwave radiation over the SEA and identify the main sources of statistical biases plaguing observational studies. We find that, during the fire season, BBAs cause a regional shortwave cooling of −2.5 W m−2, which can be decomposed into equal contributions from three physical pathways: aerosol-radiation interactions (ARI), adjustments to ARI, and aerosol-cloud interactions (ACI). We also perform ablation experiments with graph variants to investigate the main sources of confounding—like large-scale winds, humidity-biased retrievals, or spatial aggregation of data—and show that they result in biased radiative effect estimates (between −50% and +15%). Once free of such biases, our derived causal estimates of smoke radiative effects can be used as observational constraints to improve climate models.

Machine learning-based weather forecasting models now surpass state-of-the-art numerical weather prediction systems, but training and operating these models at high spatial resolution remains computationally expensive. We present a modular framework that decouples forecasting from spatial resolution by applying learned generative super-resolution as a post-processing step to coarse-resolution forecast trajectories. We formulate super-resolution as a stochastic inverse problem, using a residual formulation to preserve large-scale structure while reconstructing unresolved variability. The model is trained with flow matching exclusively on reanalysis data and is applied to global medium-range forecasts. We evaluate (i) design consistency by re-coarsening super-resolved forecasts and comparing them to the original coarse trajectories and (ii) high-resolution forecast quality using standard ensemble verification metrics and spectral diagnostics. Results show that super-resolution preserves large-scale structure and variance after re-coarsening, introduces physically consistent small-scale variability, and achieves competitive probabilistic forecast skill at 0.25° resolution relative to an operational ensemble baseline while requiring only a modest additional training cost compared with end-to-end high-resolution forecasting.

Wildfires are one of the most devastating natural disasters with wide impacts across economic sectors and society in Europe. It is, however, unclear how wildfire risk will evolve in the future under different climate scenarios, as it depends on future climate change scenarios as well as on land use and population changes. The aim of the study is to study end-of-century wildfire risk in Southern Europe, under different Shared Socioeconomic Pathways. A data-driven model is trained using historical data and a combination of other data sources representing hydro-meteorological variables, land cover, human activities, and topography. The best-performing model is the XGBoost model with an AUC of 0.93 and F1 score of 0.85. Using SHapley Additive exPlanations values, we identify the most influential predictors. The trained model is run daily using future climate projections from the Canadian Regional Climate Model version 5 under the SSP1–2.6 and SSP3–7.0 scenarios, including explicit changes in population and land cover. Future projections confirm that the wildfire risk will increase by the end of the century (2081–2100) in the SSP3–7.0, with limited changes in the SSP1–2.6 scenario. The results further show that including future land cover and population changes modulates the projected wildfire risk in scenario-dependent ways, highlighting the need to jointly account for climatic and socio-environmental drivers when assessing future wildfire risk.