Global, continental and regional maps of concentrations, stocks and fluxes of natural resources provide baseline data to assess how ecosystems respond to human disturbance and global warming. They are also used as input to numerous modelling efforts. But these maps suffer from multiple error sources and hence it is good practice to report estimates of the associated map uncertainty, so that users can evaluate their fitness for use. In this presentation we address the question of how to obtain the uncertainty of spatial aggregates of the map predictions. This is needed when the mapped variable is reported as an average or total for subareas within the study area, such as for rectangular grid cells, administrative units or bioclimatic domains, or for the study area as a whole. We first explain why quantification of uncertainty of spatial aggregates is more complex than uncertainty quantification at point support, because it must account for spatial autocorrelation of the map errors. We describe how this can be done with block kriging and illustrate this method in a case study of mapping the topsoil organic carbon content at various administrative aggregation levels in mainland France. Next, we propose an approach that avoids the numerical complexity of block kriging and is feasible for large-scale studies where maps are typically made using machine learning. Our approach relies on Monte Carlo integration to derive the uncertainty of the spatial average or total from point support prediction errors. We account for spatial autocorrelation of the map error by geostatistical modelling of the standardized map error. The methodology is illustrated with mapping aboveground biomass and deriving the associated uncertainty for various block supports in a region in Western Africa. Both case studies clearly show the need to account for spatial autocorrelation in order to get realistic estimates of the uncertainty of spatial averages and totals.

How to cite: Heuvelink, G. and Wadoux, A.: Uncertainty of spatial averages and totals of natural resource maps, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2900, https://doi.org/10.5194/egusphere-egu23-2900, 2023.

Recent approaches for large-scale mapping of continuous environmental variables by combining ground observations, remote sensing and machine learning have proposed incorporating computer vision capabilities into the model, so that potentially-complex regression features may be learned automatically from covariate datasets, such as of terrain elevation and other satellite imagery (e.g. see Kirkwood et al 2022; 'Bayesian deep learning for spatial interpolation in the presence of auxiliary information').

Here we present new research using national-scale land-surface geochemical data to explore and compare how the incorporation of computer vision for automatic feature learning affects the predictive performance of geostastistical interpolators both within and beyond the spatial extents of the study areas in which ground observations are collected. We attempt to characterise empirically how well the predictive performance of different models is preserved with increasing distance from training observations in order to provide insights into the value of incorporating computer vision capabilities into geostatistical models, compared to more traditional approaches.

How to cite: Kirkwood, C., Economou, T., Odbert, H., and Pugeault, N.: Can learning regression features by computer vision improve the generalisation of geostastistical interpolators?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6656, https://doi.org/10.5194/egusphere-egu23-6656, 2023.

Our understanding of the Earth´s functional biodiversity and its imprint on ecosystem functioning, structure and resilience is still incomplete. Large-scale information on vegetation properties (‘plant traits’) is critical to assess functional diversity and its role in the Earth system. Such parameters are constantly changing due to variations in environmental conditions which makes extensive in-situ measurements in the field not logistically feasible. The advent of the upcoming space-borne hyperspectral missions will facilitate to map these properties. However, we are still lacking efficient and accurate methods to translate hyperspectral reflectance into large scale information on plant traits across biomes, land cover and sensor types. Yet, the absence of globally representative data sets on reflectance data and the corresponding in-situ measurements represents a bottleneck to develop empirical models for estimating plant traits from hyperspectral reflectance. Recent and ongoing initiatives (e.g. EcoSIS) provide a constantly growing source of hyperspectral data and plant trait observations from different vegetation types and sensors. In this study we integrated 29 data sets including four different ecosystem types spanning from Europe to north America. By combining these heterogeneous data sets, we propose multi-trait models based on Convolutional Neural Networks (CNNs) that simultaneously infer multiple plant traits from canopy spectra. We targeted a broad set of structural and chemical traits (n=20) related to light harvesting, growth, propagation and defense (e.g. leaf mass per area, leaf area index, pigments nitrogen, phosphorus). The performance of our multi-trait CNN models predicting these traits was compared to single-trait CNNs as well as single-trait partial least squares regression (PLSR) models. The results of the multi-trait models across a broad range of vegetation types (crops, forest, tundra, grassland, shrubland) and sensor types were promising and outcompeted state-of-the-art PLSR models. We found that the overall prediction performances significantly increased from single- to multi-trait CNN models and those of PLSR models. The key contribution of this study is to highlight the potential of weakly supervised approaches together with Deep Learning to overcome the scarcity of in-situ measurements and take a step forward in creating large-scale maps of Earth’s biophysical properties with the increase in availability of hyperspectral Earth observation data.

How to cite: Cherif, E., Feilhauer, H., Berger, K., Ewald, M., Hank, T. B., Kovach, K. R., Townsend, P. A., Wang, Z., and Kattenborn, T.: From spectra to functional plant traits: Transferable multi-trait models from heterogeneous and sparse data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10901, https://doi.org/10.5194/egusphere-egu23-10901, 2023.

Machine learning algorithms have become widely used for geospatial applications, including spatial mapping and upscaling ecological variables and traits. Multivariate splines, random forests, and neural networks have been widely used to upscale a few sparse measurements to larger areas. Machine learning models, however, cannot offer reliable predictions in out-of-the-sample areas, which is often the case in such applications [1,2]. In [3], an area of applicability is proposed as an extrapolation index based on the minimum distance to the training data in the multidimensional predictor space with predictors being weighted by their respective importance in the model. We propose Gaussian Processes (GPs) to derive such extrapolation indicator [4].  A GP is a popular method in machine learning and multivariate statistics for regression problems. It provides a probabilistic description of the predictive function, so one can derive both predictive mean and variance for the predictions on new data. We here suggest using the predictive variance as an indicator for extrapolation and show the relation with a customized dissimilarity index computed that follows the Area of Applicability methodology proposed in [3]. We show the relation and in some cases the generalization in a set of controlled synthetic experiments and for vegetation traits global mapping using remote sensing, meteorological variables and the (huge yet sparse and biased) TRY database. This relation opens the door to a more sound way of identifying and characterizing extrapolation regimes through GPs in geospatial and upscaling applications.

References

[1] Deep learning for the Earth Sciences: A comprehensive approach to remote sensing, climate science and geosciences. Gustau Camps-Valls, Devis Tuia, Xiao Xiang Zhu, Markus Reichstein (Editors). Wiley & Sons 2021

[2] Perspective on Deep Learning for Earth Sciences. Camps-Valls, Gustau. Generalization with Deep Learning: for Improvement on Sensing Capability, World Scientific Pub Co Inc 2021

[3] Meyer, H., & Pebesma, E. (2021). Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods in Ecology and Evolution, 12, 1620–1633. https://doi.org/10.1111/2041-210X.13650

[4] A Survey on Gaussian Processes for Earth Observation Data Analysis: A Comprehensive Investigation. Camps-Valls, G. and Verrelst, J. and Muñoz-Marí, J. and Laparra, V. and Mateo-Jiménez, F. and Gómez-Dans, J. IEEE Geoscience and Remote Sensing Magazine 2016

How to cite: Martínez-Ferrer, L., Moreno-Martínez, Á., Muñoz-Marí, J., Meyer, H., Ludwig, M., and Camps-Valls, G.: Gaussian Processes for vegetation traits global mapping, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5331, https://doi.org/10.5194/egusphere-egu23-5331, 2023.

Ecological research on the responses of biotic systems to climate change is largely based on coarse-gridded climate data, derived from standardized meteorological weather stations. These stations measure temperatures in open areas at 1.2 to 2 m above ground covered by short grass, thus measuring macroclimatic conditions. Those macroclimate conditions are, however, not representative for the microclimates most organisms experience. In closed canopy forests, for instance, hot and cold temperature extremes are buffered by several degrees and forests thus provide microclimatic shelters for many forest-dwelling species under climate change. Yet, with increasing forest disturbances opening up forest canopies globally, the future buffering capacity of forests remains uncertain. We aim at closing this knowledge gap by modelling the temperature buffering capacity of forests from remote sensing data. Models were based on boosted regression trees and a set of forest structural and topographic predictors derived from an airborne LiDAR acquisition. Buffering capacity was estimated from in situ microclimatic loggers across 150 plots and a network of meteorological weather stations in the Berchtesgaden National Park – a 20,000 ha landscape located in southern Germany. Spatial models of temperature buffering yielded high predictive accuracies, ranging from R²=0.62 to R²=0.74 depending on the month of observation. Forest structure was consistently more important than topography in explaining temperature buffering. Spatial predictions of temperature buffering revealed a clear elevational gradient, with less buffering in higher-elevation forests (open canopy structures) compared to low-elevation forests (mostly closed canopies). We also found strong variation in temperature buffering over the successional trajectory, with no or even inverted buffering in disturbed sites and a recovery of the buffering capacity within 30 years after disturbance on most sites. Our results will help better understanding the impacts of climate change on forest dwelling species by improving species distribution models and other models of key life history traits. Our approach further provides the ability to be expanded to other regions.

How to cite: Senf, C. and Vandewiele, M.: Mapping forest temperature buffering from LiDAR, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2065, https://doi.org/10.5194/egusphere-egu23-2065, 2023.

Drought periods will become more frequent, more severe, and longer in the coming decades in many regions of the world as a consequence of climate change. In Central Europe, grassland vegetation substantially deteriorated immediately in response to extreme droughts in recent years with major impacts on livestock farming. A deeper understanding of the grassland response to drought under different environmental and land management characteristics is required for adapting to future droughts. Fractional cover time series of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), and soil from remote sensing provide essential information to characterize grassland dynamics and impacts on grassland vitality during drought periods based on continuous, physically meaningful variables across larger areas.

Based on the methodological developments in our previous studies (Kowalski et al. 2022, Kowalski et al. 2023), we here present a regression modeling framework that enabled the retrieval of a consistent, multidecadal time series of PV, NPV, and soil fractional cover from a data cube comprising all available Landsat and Sentinel-2 imagery for Germany. Fractional cover time series retrieval relied on spatially and temporally generalized regression models. The generalization step relevant for applying models across space (i.e. entire Germany), time (i.e. 1984 – 2021), and sensors (i.e. Landsat 5, 7 & 8, Sentinel-2A & 2B) was based on synthetic training data generated from a global spectral library which integrated laboratory and image-based spectral measurements. Investigation of the multidecadal feature spaces of the Landsat and Sentinel-2 sensor families confirmed the compatibility and global applicability of the spectral library as a training source in our regression modeling framework. The application of the generalized regression models to the data cube produced consistent time series of PV, NPV, and soil fractional cover independent from the underlying sensor. This was confirmed by comparing pairs of estimated cover fractions from different sensors for similar dates (± 2 days) relative to each other and to ground reference fractions. We further demonstrate the value of the multidecadal fractional time series as a means for drought monitoring in temperate grasslands. Periods of anomalous vegetation browning could be consistently linked to meteorological and soil moisture drought in the past four decades. Our study demonstrates the value of integrating spectral measurements from various sources, including image-based data and existing in-situ networks, as a means for consistent grassland fractional cover time series retrieval based on generalized regression models and multisensor data cubes.

References

• Kowalski, K., Okujeni, A., Brell, M., Hostert, P., 2022. Quantifying drought effects in Central European grasslands through regression-based unmixing of intra-annual Sentinel-2 time series. Remote Sensing of Environment 268, 112781. https://doi.org/10.1016/j.rse.2021.112781
• Kowalski, K., Okujeni, A., Hostert, P., 2023. A generalized framework for drought monitoring across Central European grassland gradients with Sentinel-2 time series. Remote Sensing of Environment 286, 113449. https://doi.org/10.1016/j.rse.2022.113449

How to cite: Okujeni, A., Kowalski, K., and Hostert, P.: National-scale monitoring of grassland vitality – combining spectral databases, machine learning regression modeling, and multidecadal Landsat/Sentinel-2 time series, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9837, https://doi.org/10.5194/egusphere-egu23-9837, 2023.

The consistent monitoring of trees both inside and outside of forests is key to mitigating climate change. However existing large-scale tree cover maps primarily quantify forest cover and do not include isolated trees, as these are not discernible in lower resolution satellite images. In many dryland countries these non-forest trees constitute the main form of tree cover, and play a vital role in ecological stability, local economies, livelihoods, and food security.

Here we make use of the PlanetScope nanosatellite constellation, which delivers global very high-resolution daily imagery, to map both forest and non-forest tree cover for continental Africa using images from a single year. We composite 232,053 4-band scenes from 2019 into 1x1° mosaics and apply a Convolutional Neural Network to segment canopy cover of all trees and shrubs. To train the network we use a combination of manually annotated 1 m labels and source images upsampled from 3 m to 1 m, resulting in a final model that maps tree cover at 1 m across the continent, segmenting closed canopies in forest areas and individual scattered trees in savannah areas.

Our prototype map demonstrates that a precise assessment of all tree-based ecosystems is possible at continental scale, and reveals that 29% of tree cover is found outside areas previously classified as tree cover in state-of-the-art maps, such as in croplands and grassland. This analysis lays the groundwork towards global scale studies of tree cover at individual tree level and annual temporal scale, which is crucial for improved managing of woody resources, monitoring of TOF in relation to agroforestry, tree planting and restoration efforts, and assessing land use impacts in non-forest landscapes.

How to cite: Reiner, F., Brandt, M., Tong, X., Kariryaa, A., and Fensholt, R.: An African-wide map of tree cover at individual tree level, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15488, https://doi.org/10.5194/egusphere-egu23-15488, 2023.

Passive microwave observations at different frequencies suffer extinction effects of the different vegetation components (branches, leaves, trunk) across the canopy of the soil’s microwave emission. These effects are often represented as a frequency-dependent variable called the Vegetation Optical Depth (VOD), which has been used (recently) to estimate Above-Ground Biomass (AGB). Low frequency observations, more particularly at L-band (1.4 GHz), have been shown to be sensitive to the woody components of plants (and thus to AGB), hence the growing interest in their use to monitor carbon stocks evolution.

In this study, and thanks to the multi-angle capabilities of the SMOS mission, a new approach to estimate AGB maps directly from multi-angular passive L-band Brightness temperatures (TBs) is proposed, thus surpassing the dependence on intermediate variables like the VOD. Biomass estimates are produced from Artificial Neural Networks (ANN), using as reference the three AGB maps of the Climate Change Initiative (CCI) for the years 2010, 2017 and 2018; the SMOS multi-angle TBs for the same years were selected as inputs. The best set of predictors for ANNs and the optimal learning data-set configuration to estimate AGB are proposed based on a sensitivity analysis; the use of TBs in both Vertical and Horizontal polarization, plus a polarization ratio provided the closest biomass estimates to the reference AGB maps.

ANNs trained from a purely data-driven approach explained 76% of AGB variability globally (incidence angles >35º showed high synergies with AGB); a hybrid approach (coupling ANN with variables derived from physically based models) slightly increased this value (+3%). However, when the trained models are applied to datasets from years different than those used during the training stage, a decrease in retrieval’s quality was observed; a new training scheme based on multi-year training sets is presented, results showed more stability from this kind of training schemes for temporal analyses.

Finally, ANN- and VOD-based estimates were compared with respect to different AGB reference maps, the former outperformed the latter in all evaluation metrics. VOD-based inversions tend to underestimate AGB due to their quick saturation (around 200 Mg/ha) on densely forested regions. Additionally, a strong simplification of the spatial variations of AGB was observed; maps produced from this methodology present abrupt transitions between densely and sparsely vegetated areas, a characteristic that was not observed in the reference maps. When using VOD-derived maps these limitations should be considered, especially when employing them to study the temporal evolution of carbon stocks. The ANN methodology here proposed proves to be a promising technique for the estimation of global AGB maps, with robust results both in the spatial representation and in the temporal reproduction of AGB maps.

How to cite: Salazar-Neira, J.-C., Rodríguez-Fernández, N., Mialon, A., Richaume, P., Mermoz, S., Kerr, Y., Bouvet, A., and Le Thoan, T.: Above-Ground Biomass estimation: a machine learning approach based on multi-angular L-Band passive microwaves brightness temperatures, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13701, https://doi.org/10.5194/egusphere-egu23-13701, 2023.

The forest age can be defined as the time since the last stand-replacement event. Within the local context, it determines the forest successional stage, a fundamental variable for diagnosing the net carbon fluxes in terrestrial ecosystems. To accurately quantify the carbon sink-source strength in these ecosystems and inform adaptation-mitigation strategies, it is essential to be able to infer forest age at high resolution.

This study presents an updated version of the MPI-BGC forest age product, featuring global distributions of forest age for 2010, 2017, 2018, and 2020 at 100m spatial resolution. We employed two machine learning approaches, XGBoost and a multi-layer perceptron model, to create data-driven estimates of forest age based on over 40,000 forest inventory plots, biomass, remote-sensing, and climate data. One key innovation of our approach is the incorporation of Landsat-based disturbance history metrics as input variables. Our updated estimates show better precision in identifying old-growth forests and reduce overestimation biases in young forests and underestimation biases in old forests, but not completely. Additionally, we found substantial regional variations related to changes in covariate strength and improvement in the model. Also, we discussed the uncertainty layers, created using model ensembles, that materialize the quantification of methodological uncertainty in the forest age estimates.

An analysis of the global distribution of forest age reveals significant variations across the years studied. We also quantified the changes in forest age in regions with high deforestation or forest degradation rates, where younger stands are becoming more prevalent. We discuss the challenges and limitations of using regression-based mapping approaches, including the choice of machine learning algorithm, spatial cross-validation techniques, and the caveats of extrapolation, given data limitations. Our research highlights the complementary biomass-based approaches for determining forest age and underscores the importance of detailed global estimates at high spatial resolutions. Overall, this study advances our understanding of forest age, a key variable for understanding the carbon cycle in terrestrial ecosystems.

How to cite: Besnard, S., Santoro, M., Herold, M., Cartus, O., Gütter, J., Herault, B., Kassi, J., Koirala, S., N'Guessan, A., Neigh, C., Nelson, J., Poulter, B., Weber, U., Zhang, T., and Carvalhais, N.: Mapping Forest Age at High-Resolution Using Inventory Data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6473, https://doi.org/10.5194/egusphere-egu23-6473, 2023.

Soil heterotrophic respiration (RH) is one of the largest and most uncertain components of terrestrial carbon cycle, which reflects the carbon loss from soils to the atmosphere due to microbial decomposition of soil organic carbon and litter debris. However, RH estimates vary greatly using different approaches, requiring RH estimates from different angles. Therefore, in current study, we first modelled RH from 1980 to 2016 using a Random Forest (RF) algorithm with the linkage of field observations from the Global Soil Respiration Database and global environmental drivers at 0.5 degree; second, we analyzed the spati-temporal patterns of RH; and third, we compared RF-driven RH with estimates from Global Dynamic Vegetation Model (GDVM) and Data-Model Integration approaches. Results showed that RF could satisfactorily capture the spati-temporal patterns of RH with a model efficiency of 50% and root mean square error of 143 g C m^-2 a^-1. RF-driven RH showed a large spatial variability and decreased with the increasing latitude. Total RF-driven RH was 57 Pg C a^-1 (1 Pg = 10¹⁵ g) with an average increasing trend of 0.036 Pg C a^-2 from 1980 to 2016 (p < 0.001). However, the temporal trend of RH varied with climatic zones that RH increased in boreal and temperate areas, while no temporal trend in tropical regions. RH from seven GDVMs changed from 34.8 Pg C a^-1 for ISAM model to 59.9 Pg C a^-1 with an average of 47.6 Pg C a^-1, underestimating RH by 9.6 Pg C a^-1 (16%) in comparison to RF-driven RH. Furthermore, RH estimates from data-driven approach of Hashimoto et al. (2015) and Yao et al. (2021) were 51 and 47 Pg C a^-1, which were lower than our estimate. Such difference was mainly attributed to different modelling algorithms and observational datasets. Therefore, given the potential uncertainties remaining in RH products, new approaches, e.g. deep learning or better representation of soil carbon cycling processes in GDVMs, are encouraged.  

*This study was supported by the National Natural Science Foundation of China (32271856). 

How to cite: Tang, X., Yang, Z., and Zhou, T.: Estimates of soil heterotrophic respiration in global terrestrial ecosystems, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7784, https://doi.org/10.5194/egusphere-egu23-7784, 2023.