Jordi Muñoz-Marí

High-resolution soil property maps have a wide range of applications, e.g. policymaking and precision agriculture. Previously published maps covering the EU-27 plus the UK range in resolution from 250 to 30 metres depending on the soil property. In this work, we produced 10 m resolution maps for soil texture (clay, silt, sand), coarse fragments, pH and bulk density. They are the products of a machine-learning workflow trained on LUCAS soil datasets and using AlphaEarth Foundations satellite embeddings as predictors. We performed hyperparameter tuning and training on Random Forests (RF) and Artificial Neural Networks (ANN). The best-performing models were ANN, with R ² values ranging from 0.21 for coarse to 0.69 for pH. The best-performing ANN hyperparameters for each soil property were used to produce soil property maps over Europe as well as uncertainty estimate maps obtained from Monte Carlo dropout. The generated maps aim to close the resolution gap for the EU-27 + UK and provide a promising baseline for precision agriculture, soil health assessment at field scale and spatial planning.

Optical remote sensing is fundamental for land monitoring, climate analysis, and agricultural applications; however, its effectiveness is often compromised by cloud contamination, which disrupts temporal continuity. Although cloud detection and masking techniques have advanced, additional gap-filling methods are needed to ensure consistent records of surface reflectance. This letter presents an enhanced version of the Harmonized and Improved Spatial and Temporal Adaptive Reflectance Fusion Model (HISTARFM) adapted for Sentinel-2 imagery. The method generates gap-filled reflectance at 10-m spatial resolution with a five-day temporal frequency, implemented on Google Earth Engine for large-scale processing. Validation across representative sites in North America, Europe, and East Asia demonstrates strong accuracy, with an average relative RMSE below 15%. The proposed approach offers a reliable framework for generating continuous, high-resolution optical datasets, thereby supporting diverse environmental, agricultural, and climate-related studies.

High-resolution maps of plant functional traits are crucial for understanding terrestrial ecosystem processes; however, their integration into ecosystem models has been hindered by uncertainties and a lack of spatially detailed data. Here we combine optical remote sensing, global crowd-sourced biodiversity records and plant trait databases to map community trait distributions worldwide at 1-km resolution, estimating community-weighted means (CWMs) and higher-order moments (standard deviation, skewness, and kurtosis) for specific leaf area (SLA), leaf nitrogen (LNC) and leaf phosphorus (LPC) concentrations. Benchmarking against sPlotOpen plot-level CWMs shows low explained variance (R² = 0.10–0.27 across traits), indicating limited plot-scale predictive skill under current limited open global benchmarks and scale mismatches. Agreement increases when using a canopy-weighted comparator (TWM; R² = 0.22–0.38; relative RMSE ≈ 12–18%), consistent with the top-of-canopy sensitivity of optical sensors. By providing spatially explicit trait distributions and their higher-order moments, our findings deliver improved detail for understanding biodiversity patterns and ecosystem functioning and provide landscape-scale insights into trait-mediated coexistence. This work enhances ecological modeling and offers a foundation for assessing the impacts of global environmental changes, advancing our understanding of plant functional diversity’s role in ecosystem resilience and sustainability.

Accurate monitoring of terrestrial CO₂ uptake is essential for Natural Climate Solutions and reducing carbon accounting uncertainty. Project-scale certification protocols can provide robust estimates but often depend on costly site-level measurements and are difficult to scale. Global carbon flux models provide continuous coverage, but their coarse resolution cannot represent heterogeneous land management. Bridging these approaches requires high-resolution, scalable carbon monitoring with transparent uncertainty estimates.Within the BenchFlux project, we present initial results from a High-Resolution Carbon Flux Monitoring framework based on the Global Estimation Model (CARBON-GEM). CARBON-GEM integrates (i) surface reflectances from the HISTARFM data-fusion approach (Moreno-Martinez et al., 2020), (ii) meteorological drivers, and (iii) eddy-covariance (EC) observations to estimate both gross primary production (GPP) and net ecosystem exchange (NEE) daily at 30 m resolution. The approach utilizes machine-learning methods, such as neural networks, to capture nonlinear responses, and is implemented in Google Earth Engine for scalable mapping. The workflow also delivers pixel-level uncertainty quantification, moving beyond categorical quality flags to support auditability and interpretation.In addition to standard out-of-sample cross-validation to assess robustness and generalization, we validate CARBON-GEM against independent, scale-aware FluxMapper ground truth (Metzger, S., 2018) provided by BenchFlux’s SpatialEddy component. FluxMapper couples next-generation EC processing with flux spatialization to enable explicit space-time matching and local-to-regional nesting. In this context, CARBON-GEM extends the FluxMapper-scale structure beyond individual stations, allowing continuity across diverse landscapes. Complementing this, FluxMapper provides a novel, independent benchmark for high-resolution carbon-flux estimates and serves as a robust reference point, alleviating standard EC spatial-resolution constraints and facilitating the decomposition of aggregate point measurements into fine-grained spatial patterns. CARBON-GEM and FluxMapper together establish a foundation for scalable, uncertainty-aware 30-meter monitoring of GPP and NEE. This approach captures essential spatial heterogeneity necessary for large-scale real-world auditing in NCS planning, reporting, and verification.Moreno-Martínez, Á., Izquierdo-Verdiguier, E., Maneta, M. P., Camps-Valls, G., Robinson, N., Muñoz-Marí, J., ... & Running, S. W. (2020). Multispectral high-resolution sensor fusion for smoothing and gap-filling in the cloud. Remote Sensing of Environment, 247, 111901.Metzger, S. (2018). Surface-atmosphere exchange in a box: Making the control volume a suitable representation for in-situ observations. Agricultural and Forest Meteorology, 255, 68-80.

High-resolution soil property maps (SPMs) are of high relevance on multiple spatial scales. At field-scale, highly resolved knowledge of soil properties can inform soil management zone delineation for precision agriculture applications like irrigation, fertilisation or compaction risk management. On regional to national scales, high-resolution SPMs can aid in formulating soil policy with accurate baselines and realistic improvement goals. They may assist in soil unit definition, aiding in the balancing act between scientific precision and geographic cohesion, i.e. administrative effectiveness. On a continental, European scale, these SPMs can inform on the adequacy of Soil Monitoring Law implementation and can increase the representativity of continental soil sampling campaigns.Contributing to this effort, we produced 10 m-resolution maps of soil pH, bulk density, coarse fragments, and texture (sand, silt, clay). This represents a large increase in spatial resolution compared to previously published maps, whose cells cover at least 250 x 250 m. We generated our maps by training shallow property-specific artificial neural networks (ANNs) on soil sample data across the EU-27 + the United Kingdom. We used soil data from the Land Use/Coverage Area frame statistical Survey (LUCAS): texture and coarse fragments data from the 2009 campaign, and pH and bulk density data from 2018. Instead of collecting a large number of spatial covariate data layers, we used the AlphaEarth Foundations satellite embeddings, produced by Google DeepMind. The satellite embeddings are the result of a representation learning model trained to efficiently compress Earth observation data. In this way, data from various sources is represented as a 64-dimensional vector per 10 m pixel. We extracted the 2018 embeddings as predictors for all soil properties. After training the models, we produced the full maps on Google Earth Engine (GEE) using a hand-implemented ANN with five neurons in the hidden layer, followed by a dropout layer. The dropout layer gave us the opportunity to additionally provide a prediction uncertainty map per soil property. A single output model was used to produce pH, bulk density and coarse fragments, while a multioutput NN generated the three texture components. Model performances varied, with R² ranging from 0.21 (coarse fragments) to 0.69 (pH). We validated our maps using a variety of field- and regional-scale soil datasets, furthermore by comparison of the value distributions with LUCAS data on European and national scales, and by visually contrasting with published soil property maps. We look forward to testing the temporal stability of our models once LUCAS 2022 data are available. Our soil maps will be necessary and useful for the scientific community across scales, from the field to the continent.

Maximum light use efficiency (LUEmax) is a key parameter in state-of-the-art global carbon models (GCMs), representing the maximum conversion rate of absorbed photosynthetically active radiation into vegetation biomass under non-stress conditions. Despite its significance, LUEmax is often oversimplified in most GCMs, where its variation is constrained by a limited number of plant functional types (PFTs). This coarse classification overlooks well-documented variability within PFTs and fails to account for adaptation and acclimation processes, introducing substantial uncertainty in carbon cycle estimates.Recent studies suggest that replacing PFT-based parameterization with spatially explicit LUEmax maps could significantly enhance ecosystem productivity modeling. In this study, we explore the potential of symbolic regression, an emerging machine learning technique based on genetic algorithms for deriving explicit mathematical relationships, alongside Kolmogorov-Arnold Networks (KANs) based on parameterized neural networks, which facilitate interpretable functional discovery, to estimate LUEmax from climatic data and key ecosystem traits.Using novel plant trait datasets and multiannual flux tower eddy covariance observations combined with MODIS data, we assess the ability of symbolic regression techniques and KANs to derive equations linking LUEmax to ecosystem traits. Our findings demonstrate that these approaches improve the generalization of LUEmax estimation and enhance interpretability, offering significant implications for global-scale environmental modeling and remote sensing applications.

Plant hydraulic traits are critical in regulating plant–water interactions and essential for understanding vegetation responses to environmental stress. Building on our earlier methodology for mapping global plant functional traits, we now incorporate a newly-available dataset of hydraulic traits for 55,779 tree species. This integrated framework leverages remotely sensed imagery, crowdsourced biodiversity data, and trait databases to estimate and map key hydraulic parameters, including maximum stomatal conductance (gsMAX), xylem pressure at 50% and 88% conductance loss (P50, P88), and photosynthetic water use efficiency (WUE).The tree trait data underlying our study accounts for the large phylogenetic signals inherent in these hydraulic traits by leveraging phylogenetically-informed machine learning models and novel trait imputation methods. These enhanced predictions of hydraulic traits are subsequently integrated into our trait-mapping workflow, which has previously demonstrated high accuracy (r > 0.5; rME < 6%; rRMSE < 11%) for leaf-level traits at a 1 km spatial resolution. While the hydraulic trait maps have not yet been validated due to a lack of independent validation data, the observed patterns are consistent with a meta-analysis based on recent literature.We also capture the full distribution of hydraulic traits (standard deviation, skewness, and kurtosis) at the grid-cell level to reflect the non-Gaussian variability of community-level traits. This added detail helps elucidate the ecological strategies of species assemblages and refines our understanding of ecosystem vulnerability to climate extremes. Overall, this work offers a new avenue for improving global ecosystem models and Earth system simulations by providing spatially explicit community-level hydraulic trait estimates at large scales. Our results highlight the importance of merging global remote sensing data with state-of-the-art trait imputation and phylogenetic information to advance research on plant functioning and ecosystem dynamics.

Plant functional traits play a crucial role in determining how terrestrial ecosystems function. However, most Earth system models (ESMs) oversimplify this information, representing it with a limited number of static, empirically fixed values assigned to a selection of plant functional types (PFTs). This results in a reduction of the diversity of plant communities into a relatively small number of categories and the loss of key variability within individual PFTs. As a result, local processes occurring within ESM grid cells are not well represented, leading to uncertainties in predicting ecosystem functions.The TRY global plant traits database is home to the most extensive collection of in-situ trait observations for a broad spectrum of species across the globe. Nonetheless, despite the numerous species and samples included in TRY, it still falls short compared to the overall richness and diversity of species and ecosystem functions worldwide. As a result, various initiatives have emerged to create global maps of plant traits. In this study, we created maps of essential plant traits, such as specific leaf area (SLA), leaf nitrogen content (LNC), and leaf phosphorus content (LPC), at a spatial resolution of 1 km. We took an innovative approach by leveraging the use of biodiversity, trait databases, and remote sensing data as primary sources of information. Additionally, we provide ancillary data layers that indicate regions where data gaps currently exist and where more samples are needed to improve trait representation in TRY.We compared our results to plot-level estimates for thousands of sites globally. The comparison demonstrated strong correlations (r > 0.5) and low relative errors (rME < 6% and rRMSE < 11%) for all considered traits despite the challenges in scaling up from local to global scales. Our results reveal the non-Gaussian nature of trait distributions at a global scale when computing community representative mean trait values and further statistical descriptors, including standard deviation, skewness, and kurtosis estimations. These higher-order moments provide a more detailed and nuanced view of plant functional diversity and distribution. Using these new data to parameterize global ecological models could lead to more accurate predictions and a better understanding of the main drivers of different ecosystem processes.

Medicanes, tropical-like cyclones in the Mediterranean Sea, pose unexpected challenges to unprepared areas due to their projected increases in intensity. To address these challenges, we proposed: 1) the development of an automatic tracking method in the absence of a comprehensive tracking database for Medicanes; 2) the implementation of a forecasting model for extreme cyclones utilizing artificial intelligence techniques. This is especially beneficial when traditional numerical models struggle to account for nonlinear interactions. We use a K-means algorithm and mean sea level pressure reanalysis data to track storm centers, determining maximum wind speed and position throughout each case’s lifetime. This information categorizes our dataset into storm-like and extreme Medicanes, and facilitates the extraction of spatiotemporal data from infrared satellite images. These features enable us to predict the final classification of Medicanes (whether they are storm-like or extreme) 6 to 36 h before peak wind speed, using an optimized combination of Convolutional Neural Network and Random Forest binary classification methods. By training and testing on Mediterranean data from 1984 to 2020, we successfully diagnosed between 72% and 87% of extreme Medicanes in the studied cases, depending on the lead-time. Our study is the first to employ artificial intelligences for both tracking and forecasting Medicanes, offering a foundational approach to enhance Medicanes preparedness and awareness.

Medicanes are rare cyclones in the Mediterranean Sea, with intensifying trends partly attributed to climate change. Despite progress, challenges persist in understanding and predicting these storms due to limited historical tracking data and their infrequent occurrence, which make monitoring and forecasting difficult. In response to this issue, we present an AI-based system for tracking and forecasting Medicanes, employing machine learning techniques to identify cyclone positions and key evolving spatio-temporal structural features of the cloud system that are associated with their intensification and potential extreme development. While the forecasting model currently operates with limited training data, it can predict extreme Medicane events up to two days in advance, with precision rates ranging from 65% to 80%. These innovative data-driven methods for tracking and forecasting provide a foundation for refining AI models and enhancing our ability to respond effectively to such events.