The TRY initiative (https://www.try-db.org) was established in 2007 on request from IGBP and DIVERSITAS to develop a joint database of in-situ measured plant traits supporting vegetation modelling and biodiversity research. Based on the mandate from the two initiatives, TRY has received significant contributions of original and integrated datasets from the global research community and achieved unprecedented coverage. The TRY database is regularly updated, and since 2019, data have been publicly available under a CC BY license. 

Trait data from the TRY database are frequently used to map plant traits at global scales. We here provide an overview of the TRY database, briefly summarise trait mapping approaches highlighting caveats, and provide an outlook on upcoming developments in the context of the TRY database.

How to cite: Kattge, J., Boenisch, G., Dechant, B., Moreno-Martínez, Á., Kattenborn, T., Díaz, S., Lavorel, S., Prentice, I. C., Leadley, P., Wirth, C., and Consortium, T. T.: TRY - Plant Trait Database, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6174, https://doi.org/10.5194/egusphere-egu24-6174, 2024.

Natural ecosystem maps are a fundamental tool for describing natural habitats. They are used when analysing ecological networks, for studying ecological connectivity, for conservation planning or for the management of ecosystem services. While remote sensing information is commonly employed for mapping land cover, accurate ecosystem maps require going beyond the classification of the physical surface of the land and attempting to distinguish different communities of living beings. New research directions for ecosystem classification leverage species observations created by citizen scientists on social media or crowd-sourcing platforms, profiting from their extensive spatial and temporal coverage and their low acquisition cost.
In this work, we make complimentary use of crowd-sourced data with remote sensing imagery to produce ecosystem maps in alpine areas. Our study area covers the state of Valais, a territory of about 5, 000km2 in southwestern Switzerland. We retrieve nearly 3 million species observations from the Global Biodiversity Information Facility (GBIF) database that includes governmental, crowd-sourced and scientific observations. We combine the species data with high-resolution aerial remote sensing imagery provided by the Swiss Federal Office of Topography swisstopo. As reference ecosystem maps, we follow the European Nature Information System (EUNIS) at a 100m scale. To solve the task of classifying ecosystems, we propose a multi-modal data fusion approach based on a multi-modal transformer architecture. Such a model can handle redundant and complementary information coming from different data sources and provide an explicit and interpretable decision through the visualisation of its attention scores. Through our preliminary experiments, we observed that the unequal distribution of samples between the classes and also the sampling biases negatively impacted the performance of our approach. We are working towards a more informative inclusion of species observation and a more balanced learning of each ecosystem type.

How to cite: Zermatten, V., Castillo-Navarro, J., Marcos, D., and Tuia, D.: Ecosystem mapping with remote sensing images and ground observations, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8898, https://doi.org/10.5194/egusphere-egu24-8898, 2024.

With the advent of a large number of available satellite imagery for classification, several classification techniques have been developed to classify these images over the years. Each classification method has its own set of challenges that restrict its use. Hence, a novel research method is proposed using hybrid techniques (Machine Learning and Deep Learning) and geospatial techniques for enhancing land use and land cover mapping using satellite images. Furthermore, this study will provide insights into integrated agricultural management to achieve the UN’s Sustainable Development goals. This study uses freely available satellite images, i.e., Sentinel 1 and Sentinel 2, for classifying land use/ land cover over an agricultural area in Hissar, India. The major crops grown in this area include paddy, maize, cotton, and pulses during Kharif (summer) and wheat, sugarcane, mustard, gram, and peas during Rabi (winter) seasons. The datasets for the study area were pre-processed using SNAP and ArcGIS software. After pre-processing, a comprehensive feature set is identified consisting of Polarimetric features such as elements of covariance matrix, Entropy/scattering angle (alpha), and traditional geometric features such as shape, size, and area. After this phase, dimensionality reduction techniques (such as PCA) were applied to reduce the no. of features to the most important ones. Utilizing the feature set, a hybrid machine learning model is constructed using ground truth images by fine-tuning the model. For the best split, the test and train ratio is divided into 70:30. Optical (Sentinel 1) and Microwave (Sentinel 2) datasets are fused and the quality of the fused image was evaluated using several fusion metrics, including Erreur Relative Globale Adimensionnelle de Synthese (ERGAS), Spectral Angle Mapper (SAM), Relative Average Spectral Error (RASE), Universal Image Quality Index (UIQI), Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Correlation Coefficient (CC). After obtaining the resultant fused image hybrid technique is used for the final classification. The classification accuracy is represented using overall classification accuracy and kappa value. A comparison of classification results indicates a better performance by the hybrid technique with an overall accuracy of 92%.

How to cite: Shakya, A.: Enhancement of Land Use and Land Cover Mapping using Satellite Imagery through Machine Learning Techniques , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10647, https://doi.org/10.5194/egusphere-egu24-10647, 2024.

This study conducts a comprehensive comparison of landscape feature data derived from different sources — Small Woody Features (SWF) product from Copernicus Land Monitoring system, LUCAS Landscape Feature (LF) Module, and LUCAS Transect Module — in EU agricultural landscapes. Additionally, we consider the European Monitoring of Biodiversity in Agricultural Landscapes (EMBAL) project's approach of in-situ data collection for land cover, landscape elements, and biodiversity. This approach offers a promising avenue for integrating detailed field survey data with broad-scale remote sensing observations. Furthermore the EMBAL project has the potential to enhance the previously mentioned datasets by incorporating additional information, such as the nature value of all surveyed land units, habitat types or pollination potential among others. This inclusion could contribute to having better insights  into the ecological significance and monitoring of agricultural landscapes. Our analysis further explores the potential of incorporating cutting-edge datasets for enhanced monitoring of landscape features. Specifically, we consider the high-resolution (3-meter) dataset from Liu et al. (2023), which presents a detailed canopy height map and quantifies tree cover and woody biomass across Europe. 

We critically assess the strengths and limitations of each source: SWF's remote sensing foundation provides broad coverage but focuses only on woody features, the LUCAS LF Module combines photo-interpretation with field survey for a more detailed typology, and the LUCAS Transect, though discontinued, offered valuable field data for linear features. Strategies for monitoring landscape features have been considered for a long time, but are continually updated with the latest available data and methods. A comprehensive comparison and evaluation of the new data sources has not yet been carried out. Our study aims to identify complementarities between the different datasets to  improve both quantitative and qualitative monitoring of landscape features informing sustainable agricultural practices and biodiversity conservation strategies.

How to cite: Musavi, T., Skøien, J., Czúcz, B., Hagyo, A., Sedano, F., Martinez-sanchez, L., Koeble, R., van der velde, M., Terres, J.-M., and D'andrimont, R. D.: Bridging Field Surveys and Remote Sensing for Enhanced Landscape Feature Analysis in EU Agriculture, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20365, https://doi.org/10.5194/egusphere-egu24-20365, 2024.

In recent years, multiple remote sensing technologies have been used to quantify aboveground biomass (AGB) at different scales, from regionally trained models to global maps. The former are capable of providing higher resolution and accurate estimations albeit for smaller regions. Tailoring model weights and inputs to a specific biome or region have proved to be effective to obtain better results. Global maps, on the other hand, provide an attempt to standardizing AGB mapping at slightly to moderately lower resolutions and tend to have differences between them. A similar standardization with the benefits of regional mapping, applied across biomes, has yet to be available. For that, the first step would be comprehensively comparing state-of-the-art regional and local studies. However, existing inconsistencies in the different models and inputs used, which are often region-specific, make it impracticable.

This study addresses the need for a comparison of a single methodology consistent across different biomes in order to understand the nuances that drive the estimation of AGB. We present a data fusion approach to mapping aboveground biomass at a 20m resolution using regionally trained, regression-enhanced Random Forests in 4 different biomes: semi-arid savannas in the Sahel region, dense tropical forests in Brazil, Mediterranean forests in coastal and pre-coastal areas of Catalonia, and temperate/boreal forests in Minnesota. GEDI L4A AGB data (25-m discrete footprints) is used to train a regression model in each study area. We derived predictors from Sentinel-1 SAR, Sentinel-2 multi-spectral and Digital Elevation Model (DEM) datasets, which are common for all locations. Additionally, auxiliary data such as proximity to coastlines, human-made structures or bio-climatic variables are used to enhance predictions in saturation-prone areas. Additional to the goodness of adjustment to the training data from GEDI, we carried out a thorough validation of the results using in-situ data from Forest Inventory plots gathered in all 4 study regions. This enables a comprehensive comparison of the capabilities of Machine Learning modelling to adapt to the particular characteristics of each ecosystem. An in-depth analysis is carried out to find the most important predictor variables in each biome, as well as to assess the accuracy that can be expected across a wide range of AGB values.

How to cite: Perpinyà-Vallès, M., Huertas, C., Escorihuela, M. J., Ameztegui, A., and Romero, L.: Regionally trained models for mapping aboveground biomass from Remote Sensing data fusion: a comparison of the capabilities of Machine Learning in 4 different biomes., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17559, https://doi.org/10.5194/egusphere-egu24-17559, 2024.

Estimating forest canopy height is a crucial aspect in quantifying forest biomass, carbon stocks, monitoring forest degradation, and succession or restoration initiatives. Operational forest structure monitoring on a large scale involves various satellite sensors and datasets as well as large geographical variations. Challenges remain in obtaining uniformly representative ground truth data across diverse areas, phenology phases and forest types. Recently deep learning models are more frequently used to map forest canopy height at large scales by e.g., using sample observations from space-borne LiDAR sensors as training data. Training deep learning models rely on large amounts of data but are often trained on limited source domain data that is confined to cover the above-mentioned aspects. However, during testing, models may encounter out-of-distribution samples, leading to unexpected model behaviour and predictions. This vulnerability reduces the reliability of deterministic Deep Learning architectures, and finally, reliance on predictions without confidence indicators can result in misleading scientific conclusions or potentially under-informed policy decisions. Due to the varying ways of data processing, differences in forest canopy height products emerge, and hence product inter-comparison is often difficult and debatable. Existing products often do not provide any information about the confidence of their predictions.
To address this lack of confidence, we employ evidential deep Learning, adapting non-Bayesian architectures to estimate forest height and associated evidence. This approach aims to capture both aleatoric and epistemic uncertainties in areas with different forest structures investigating study areas , by incorporating evidential priors into the Gaussian likelihood function. Our method involves training a myriad of deep learning architectures including a basic CNN and Residual neuronal nets for regression to infer the hyperparameters of the evidential distribution.

Alongside other current studies, Sentinel-2 and Sentinel-1 satellite imagery serve as predictors for forest canopy height, with reference data obtained from the Global Ecosystem Dynamics Investigation (GEDI) LiDAR instrument on the International Space Station. The research explores the effects of distributional data shifts on canopy height predictions and identifies footprint samples where prediction uncertainty increases, representing different forest structures.
The application of evidential deep learning could extend far beyond this study, potentially benefiting various tasks in estimating biophysical parameters from remote sensing.

How to cite: Kugler, L., Wessollek, C., and Forkel, M.: Examining the Reliability Gap: Insights into Forest Canopy Height Using Evidential Deep Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15542, https://doi.org/10.5194/egusphere-egu24-15542, 2024.

In Switzerland, detailed soil information is missing for many regions – although urgently needed. Soon, authorities will need to legally delineate areas for the high-quality arable land inventory based on high-resolution surveys that require sampling of large numbers of new locations within the next decade. Consequently, cost and time efficient surveying strategies are required which fulfil high quality demands for legally binding decisions. The soil functional evaluation used in these decisions requires soil attributes which cannot only be derived by objective laboratory measurement but are partly based on pedological field description by experts.

In our study area of about 1’000 hectares, we established first a feature space coverage sampling design to sample 1’500 locations based on elevation and land use data, geological information, and expert knowledge of soil scientists. 170 locations were determined by a stratified random sampling design and used for independent validation of mapping results.

We predicted a large range of soil attributes (clay, silt, humus, pH, moisture regime, rootable soil depth) in multiple depth (0-20 cm, 0-30 cm, 20-30 cm, 30-50 cm, 50-100 cm), using the first 1’200 samples, random forest, and a wide range of environmental covariates. The predicted spatial information showed low to medium accuracy and maps exhibited further deficiencies, particularly poor performance for values at the tail of the soil attribute distributions. By visual inspection of prediction interval maps we found high model uncertainties in some specific areas like geological transition zones and anthropogenic altered zones through drainage and covering layers. To improve the quality of the maps we increased the total number of sampling locations up to 2’200 by two in-fill sampling design strategies in two zones:

• To complement the feature space coverage sampling design potentially based on incomplete environmental factors, experienced surveyors directly added additional sampling locations based on their expert knowledge. Those covered landscape features not contained in the primary sampling design such as local extrema or transition zones.
• In the second zone a two-level infill sampling design was created. A first general level complemented the feature space further as spanned in the initial sampling design. The second level consisted of additional 250 sampling points within zones of high model uncertainties.

Subsequently generated maps showed increased accuracy with increasing sampling density for most attributes, e.g., an increase of 0.1 for the clay content in the topsoil at a sampling density of 1.7 observations per ha compared to a sampling density of 1.1 Our results further displayed increasing accuracy of 0.05 with higher-weighted data collected by experts and simultaneous lowering the sampling density to 1.5 per ha by ignoring data with the lowest quality, collected before the internal calibration and synchronisation of the field survey.
To upscale the soil mapping in such high resolution and with expert-based parameters it is crucial to have synchronized data in high and stable quality across different soil formation regions.

How to cite: Tanner, S., Nussbaum, M., Oechslin, S., and Burgos, S.: Multi-stage soil surveying complementing statistical sampling designs to provide high-resolution soil maps for policy, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17268, https://doi.org/10.5194/egusphere-egu24-17268, 2024.

There is an increasing need for dynamic soil information, especially focused on monitoring soil health indicators such as soil organic carbon, soil chemical properties, soil pollution / soil degradation and status of soil micro-, meso- macro-fauna. To detect change over time, AI4SoilHealth project (https://cordis.europa.eu/project/id/101086179) is developing a spatiotemporal Machine Learning framework based on large EO data cubes (Witjes et al., 2023) to model soil health indicators e.g. to map them at high spatial resolution (30-m) with annual increments (2000–2022+) and for standard depth intervals (e.g. 0–30 cm, 30–60 cm, 60–100 cm). Produced time-series of predictions are then analyzed for trends and slope and similar coefficients are derived (per pixel) showing positive and negative changes in soil health indicators over longer periods of time (25+ years) (for the time-series method see: Hackländer et al. 2024). Areas where the trends are especially negative (e.g. significant decrease in soil carbon, significant salinization, significant loss of land cover / FAPAR etc) are flagged as requiring further soil sampling and detection of drivers of soil degradation, which should be ideally done jointly with national soil monitoring system in Europe.

The difference between spatiotemporal vs purely 2D / 3D mapping is in the three main aspects: (1) points and covariate layers are matched in spacetime (usually month-year period or at least year), (2) covariate layers are based on time-series data and include also accumulative indices (e.g. cumulative rainfall, cumulative snow cover, cumulative cropping fraction and similar) and derivatives, (3) during model training and validation, points are subset in both spacetime to avoid overfitting and bias in predictions. The rationale for using spatiotemporal machine learning is fitness of data for reliable time-series analysis: the predictions for anywhere in the spacetime cube need to be unbiased, with objectively quantified prediction errors (uncertainty), so that hence changes can be derived without a risk for serious over-/under-estimation. This framework has been tested on local and regional data sets (e.g. LUCAS soil samples covering 2009, 2012, 2015, 2018 for Europe) and can be now potentially applied using global compilations of soil points (https://opengeohub.github.io/SoilSamples/). Spatiotemporal machine learning could also potentially be used for predicting future states of soil, e.g. by extrapolating models to future climate scenarios and future land use systems (Bonannella et al., 2023). We are currently building a Soil Health Data Cube for Europe that will include some 15–20 biophysical indices (annual tillage index, bare surface cover, bimonthly FAPAR, NDWI, SAVI), climatic, terrain and parent material covariates and including the time-series of predictions of the key soil properties. This data will be made available under open data license through https://EcoDataCube.eu.

Cited references:

• Bonannella, C., et al. (2023). Biomes of the world under climate change scenarios: increasing aridity and higher temperatures lead to significant shifts in natural vegetation. PeerJ, 11, e15593. https://doi.org/10.7717/peerj.15593 
• Hackländer, J., et al. (2023). Land potential assessment and trend-analysis using 2000–2021 FAPAR monthly time-series at 250 m spatial resolution. PeerJ, in review. https://doi.org/10.21203/rs.3.rs-3415685/v1
• Witjes, M., et al. (2023). Ecodatacube. eu: Analysis-ready open environmental data cube for Europe. PeerJ, 11, e15478. https://doi.org/10.7717/peerj.15478 

How to cite: Hengl, T., Minarik, R., Parente, L., and Tian, X.: Mapping dynamic soil properties at high spatial resolution using spatio-temporal Machine Learning: towards a consistent framework for monitoring soil health across borders, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17704, https://doi.org/10.5194/egusphere-egu24-17704, 2024.

Soil moisture is a key factor that influences the productivity and energy balance of ecosystems and biomes. Global soil moisture measurements have coarse native resolutions of 36km and infrequent revisits of around three days. However, these limitations are not present for many variables connected to soil moisture such as land surface temperature and evapotranspiration. For this reason many previous studies have aimed to discern the relationships between these higher resolution variables and soil moisture to produce downscaled soil moisture products.

In this study, we test four ensembles of simple machine learning models for this downscaling task. These ensembles use a dataset of over 1,000 sites across the US to predict soil moisture at sub-km scales. We find that all ensembles, particularly one with a very simple structure, can outperform SMAP on  a cross-fold analysis of the 1,000+ sites. This ensemble has an average ubRMSE of 0.058 vs SMAPs 0.065 and an average R of 0.639 vs SMAPs 0.562. However, not all ensembles are beneficial, with some architectures performing better with different training weights than with ensemble averaging. Additionally, we find that although general improvements over SMAP are observed, there appears to be difficulty in consistently doing so in cropland regions with high clay and low sand content.

Key Points:

• Ensembles of simple ML architectures can downscale SWC predictions to sub 1km resolutions
• Simpler architectures can outperform these ensembles and may be further enhanced with an improved weighting scheme during training
• Training the models on temporally padded data provides more benefits than drawbacks in terms of overall performance.
• The top performing ensemble is unreliable on croplands with higher than average clay and lower than average sand content.

How to cite: Poehls, J., Alonso, L., Koirala, S., Carvalhais, N., and Reichstein, M.: Downscaling Soil Moisture with Simple Machine Learning Ensembles, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10504, https://doi.org/10.5194/egusphere-egu24-10504, 2024.

High water table depths (WTD) and water-saturated soil are important elements for peatlands to remain healthy and are crucial targets in peatland restoration projects. Recent studies have suggested that Earth Observation data might be applicable for this task, but proposed models often lack either the spatial extension or the temporal dimension. This study has been developing a spatio-temporal model of peatland water table depth by combining time series of satellite data, namely Sentinel-1, Sentinel-2; aerial data, namely Getmapping Digital Surface Model (DSM); and field collected water table depth measurements to provide the ability to evaluate WTD changes both spatially and over time. 

Water table depth measurements were collected from 59 loggers between February and September of 2018, with loggers covering peatlands in various condition clustered around four research sites in the North of Scotland. NDVI, NDWI and OPTRAM indices were derived from reconstructed cloud-free imagery of Sentinel-2 at 5-day intervals. Similarly, VV, VH, and incidence angle values were obtained from Sentinel-1 imagery at the same time interval. Finally, a Topographic Wetness Index (TWI) was calculated using GetMapping DSM data. A Generalised Additive Model (GAM) was then fitted using all above mentioned inputs with 70% training and 30% testing split method. The model showed a good overall fit (R²=0.59 for training data; R²=0.49 for testing data), with optical covariates outperforming the radar covariates. The model was then applied spatially using the R terra package, providing raster imagery of predicted WTD for 24 unique dates with clear distinction in wetness both over different seasons, and spatially in the landscape.

Following this successful application, work is in progress to test the model on additional sites across Scotland and on European level to further test the applicability of the model to a wider range of northern peatlands.

How to cite: Toca, L., Gimona, A., Donaldson-Selby, G., Smart, C., Sideris, K., Ball, J., and Artz, R.: The bigger peat picture: Spatio-temporal modelling of peatland water table depth using Sentinel-1, Sentinel-2, DSM, and field collected data., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19789, https://doi.org/10.5194/egusphere-egu24-19789, 2024.

Peat makes up roughly 28% of Scotland’s soil and is critical in many areas, including biodiversity and habitat support, water management, and carbon sequestration. The latter is only possible in healthy, undisturbed peatland habitats where the water table is sufficiently high, otherwise this potential carbon sink becomes a carbon source, that if left untreated will disappear forever. Drainage and erosion features are crucial indicators of peatland condition and are key for estimating national greenhouse gas emissions.
    
Previous work on mapping peat depth and condition in Scotland has provided maps with reasonable accuracy at 100 metre resolution, allowing land managers and policymakers to both plan and manage these soils and to work towards identifying priority peat sites for restoration. However, the spatial variability of the surface condition is much finer than this scale, limiting the ability to inventory greenhouse gas emissions or develop site-specific restoration and management plans.
    
This work involves an updated set of mapping using high-resolution (25 cm) aerial imagery which provides the ability to identify and segment individual drainage channels and erosion features. Combining this imagery with a classical deep learning-based segmentation model, enables high spatial resolution, national scale mapping to be carried out allowing for a deeper understanding of Scotland’s peatland resource and which will enable various future analyses using this data.

How to cite: Macfarlane, F., Robb, C., McKeen, M., Aitkenhead, M., and Matthews, K.: Using Deep Learning and High-Resolution Imagery to Map the Condition of Scotland’s Peatland Resource., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19664, https://doi.org/10.5194/egusphere-egu24-19664, 2024.

In this study, we generate a global, long-term, hourly terrestrial water-energy-carbon fluxes, using model simulations, in-situ measurements, and physics-informed machine learning. The soil-plant model, STEMMUS-SCOPE was deployed to simulate land surface fluxes over 170 Fluxnet sites (STEMMUS – Simultaneous Transfer of Energy, Mass, and Momentum in Unsaturated Soil; SCOPE - Soil Canopy Observation, Photochemistry and Energy fluxes radiative transfer model). The model input and output data were then used as training data-pairs to develop the STEMMUS-SCOPE emulator using multivariate random forests regression algorithm. Here, physics-informed machine learning refers to the fact that the emulator was trained and constrained by the physical consistency represented by the soil-plant model. We compared the physics-informed emulator (RF_S-S) with the one trained using only Fluxnet in-situ measurements (RF_in-situ), and found that the land surface fluxes predicted by RF_S-S are less scattered than that by RF_in-situ.

We estimate six variables simultaneously: net radiation, latent heat flux, sensible heat flux, gross primary productivity, solar-induced fluorescence in 685 nm and 740 nm. Results show that RF_S-S can estimate fluxes with Pearson Correlation Coefficient score (r-score) higher than 0.97 for these six variables. The testing result using independent stations (not included for developing emulators) show a r-score higher than 0.94. The feature importance shows that incoming shortwave radiation, surface soil moisture, and leaf area index are top predictor variables that determine the prediction performance. We further explore the performance of RF_S-S in predicting soil heat flux, root zone soil moisture, and leaf water potential, which assist the understanding of ecosystems’ drought responses to climate change.

How to cite: Han, Q., Zeng, Y., Wang, Y., Alidoost, F. (., Nattino, F., Liu, Y., and Su, B.: Global long-term hourly 9 km terrestrial water-energy-carbon fluxes with physics-informed machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5488, https://doi.org/10.5194/egusphere-egu24-5488, 2024.

We discuss the development of high-resolution (1 km) estimates of terrestrial carbon and water fluxes over the Australian continent by empirical upscaling of a regional network of flux tower measurements (“AusEFlux” v1.1 https://zenodo.org/records/7947265). We detail our ensemble learning approach for estimating the per-pixel epistemic uncertainty in flux predictions.  Our investigations demonstrate that regional or continental upscaling has several advantages over global upscaling, including: the ability to use regionally derived covariable datasets tailored to the regional environmental context; reduced computational constraints allowing for higher-resolution predictions, thus reducing the impacts of sub-cell landscape heterogeneity; ameliorating spatial biases present in global datasets that often have a strong northern hemisphere bias; simpler interpretation of results due the reduced requirement to generalise across vastly different climates, ecosystem types, and plant functional traits; and increased relevance to local stakeholders.  We compare AusEFlux with estimates from nine other products that cover the three broad categories that define current methods for estimating the terrestrial carbon cycle. We argue that consiliences between datasets derived using different methodologies offer alternative value for assessing the quality of an upscaling product than any given cross-validation technique, especially where training datasets have spatial or temporal biases that are difficult to mitigate. Lastly, we discuss the benefits of regularly updating our upscaling product to arrive at a systematic monitoring of terrestrial carbon and water fluxes.

How to cite: Burton, C., Renzullo, L., Rifai, S., and Van Dijk, A.: Empirical upscaling of OzFlux eddy covariance for accurate, high-resolution monitoring of terrestrial carbon and water fluxes in Australia., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4823, https://doi.org/10.5194/egusphere-egu24-4823, 2024.

Peat soil degradation in rural areas contributes to 3% of the annual GHG emissions in the Netherlands. In the 2019 climate agreement, the Dutch government set a goal to cut these emissions by 25% by 2030 and initiated a research consortium to achieve this, the Dutch National Research Programme on Greenhouse Gases in Peatlands (NOBV). The NOBV established a GHG monitoring network to map emissions based on the diversity of peat, edaphic conditions, grassland management, and water table management.

Eddy-Covariance plays an essential role in this monitoring network. More than 20 sites are part of it, including permanent and mobile EC towers, while an intensive airborne measurement campaign co-occurs above the studied areas. The first offers a high-temporal resolution monitoring of flux, covering a limited spatial landscape diversity; the latter provides a comparative map embracing the whole heterogeneity of the landscape during short timeslots.

A data-driven bottom-up model will be implemented. This model will focus on considering site-specific factors and characterizing the heterogeneities of the surroundings through footprint analysis. This approach is a progression from previous efforts that compared annual carbon budgets across various locations based on external data sources, which combined remote sensing and hydrological models and implemented a machine-learning framework to gap-fill time series.

Validation of the model includes a comparison with the airborne datasets. Beyond evaluating the quality of the model, the objective is to assess the strengths and limitations of this bottom-up approach when compared to real, independent datasets describing accurately spatial fluctuation of fluxes in heterogeneous landscapes.

The insights are essential for future developments, including models that leverage ground network and airborne measurements in the training process for more robust and accurate CO₂ Fluxes modelling.

How to cite: Bataille, L., Hutjes, R., Kruijt, B., van der Poel, L., Franssen, W., Biermann, J., Jans, W., Ingle, R., Rietman, A., Buzacott, A., van Giersbergen, Q., and Nouta, R.: Data-Driven Modelling and Comparative Analysis of CO2 Exchanges in Dutch Peatlands via Eddy-Covariance: Ground-calibrated bottom-up model vs airborne flux measurements, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6067, https://doi.org/10.5194/egusphere-egu24-6067, 2024.

The Eddy Covariance (EC) method is a state-of-the-art ecosystem flux measurement technique. EC data is used to calibrate and validate biogeochemical models for downstream Earth Observation (EO) products. Despite being a point measurement, the EC method samples an area around the measurement tower, called a flux footprint (FFP). The extent and direction of an FFP depend on the wind speed and direction and the state of the atmosphere, thus it changes continuously. Although most of the EC towers are placed in homogeneous areas, where the footprint direction should not play a role, at the modern 10-m spatial resolution of EO satellites such as Sentinel-2, the area under FFP may be seen as heterogeneous because of phenological differences or temporal changes in vegetation cover. The land cover heterogeneity under the FFP may cause challenges in interpreting the EC data and discrepancies in the calibration and validation of EO downstream products.

In this study, we investigated the FFP heterogeneity of 72 European EC sites from the ICOS Warm Winter 2020 dataset (https://doi.org/10.18160/2G60-ZHAK). Firstly, the footprint size was statistically estimated according to the dominant plant functional type. Secondly, the heterogeneity was assessed for a circular buffer and twelve sub-sectors of the buffer on Sentinel-2-derived normalised difference vegetation index (NDVI) with several statistical criteria (e.g. Tukey’s test, Z-score). This approach was demonstrated to be an alternative to precise FFP models, as the latter might require parameters not available to the end users, such as standard deviation of the v-wind component or atmospheric boundary layer height. The tool was developed as a stand-alone application and can be used for spatial heterogeneity assessment beyond the tested EC footprints. 

How to cite: Prikaziuk, E., Van der Tol, C., and Tomelleri, E.: A rapid approach for integrating high-resolution remote sensing and eddy flux observations, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19181, https://doi.org/10.5194/egusphere-egu24-19181, 2024.

Estimation of regional-scale carbon budget still faces considerable uncertainties, whereby reconciling bottom-up estimates and top-down inversions are essential steps towards reducing such uncertainties. Here,  we make full use of available flux tower observations and novel satellite land surface observations (e.g., soil moisture and solar-induced chlorophyll fluorescence) to constrain the spatiotemporal regimes of net ecosystem carbon exchange (NEE) for North America based on the advanced Long Short-term Memory (LSTM) networks. The LSTM-based model is capable of incorporating the memory effects of climate and environments on NEE variations, thus more accurately simulating the interannual variations (IAVs) and long-term trends of NEE. We produced gridded NEE at a spatial resolution of 0.1° × 0.1°and monthly time-step over 2001–2021 for North America.

The annual total NEE during 2001–2021 is -1.74 ± 0.10 Pg C yr^-1, which is much closer to the top-down inversions (-0.73 and -0.63 Pg C yr^-1 for CarbonTracker2022 over the same period and the Orbiting Carbon Observatory-2 (OCO-2) v10 MIP ensemble mean over 2015–2020 respectively) than the global flux upscaling products (-3.04 and -2.75 Pg C yr^-1 by FLUXCOM2020 RS+CRUJRA1.1 and RS+ERA5 respectively and -3.30 Pg C yr^-1 by NIES2020). Moreover, the spatial patterns matched with the OCO-2 v10 MIP ensemble mean reasonably well, which indicated the largest carbon uptake in the Midwest Corn-Belt area during peak growing seasons and in the Southeast on an annual basis. The estimated IAV of NEE is highly correlated with that by CarbonTracker2022. The LSTM estimate captured the NEE anomalies caused by the droughts in 2011, 2012, 2017, 2020–2021, and the 2019 Midwest floods. The performances are clearly better than existing global flux upscaling products. In addition, the estimated annual NEE exhibited a significant decline at the rate of -0.008 Pg C yr^-2 (p < 0.05), indicating an overall enhanced carbon sink during the recent two decades, which is in line with contemporary estimates.

These results suggest that the LSTM-based NEE upscaling provides an improved estimation of North American NEE for both spatial and temporal characteristics, narrowing down the gap between bottom-up estimates and top-down inversions, which is an important step towards robust regional carbon budget estimations.

How to cite: He, W., Huang, C., Liu, J., Nguyen, N. T., Yang, H., and Zhao, M.: Improved estimation of net ecosystem CO2 exchange for North America in eddy flux upscaling with memory-based deep learning , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13485, https://doi.org/10.5194/egusphere-egu24-13485, 2024.