1. Introduction
Nature-based solutions (NBS) are increasingly recognized as an essential tool in combating global warming as they remove greenhouse gases (GHG) from the atmosphere [
1], reduce GHG emissions through sustainable land use practices, and store carbon in the soil and biomass [
2]. While NBS are becoming quite advanced and continuously going through significant developments, the methods for accurately quantifying their resulting benefits are still lagging [
3]. A vast majority of the existing projects use biometric soil or tree survey data to estimate carbon stock changes [
4]. These methods provide solid estimates of the sequestered carbon but come with multiple important limitations. They typically describe either soil layer or partial canopy layer (e.g., above ground), with time resolution limited to multiple months and years and spatial resolution limited to a particular plot where such measurements have been performed. As a result of such limitations, these methods may not be best suited for large-scale ongoing NBS quantification [
2]. Moreover, the measurements required by these methods for a meaningful time-resolved, large-scale assessment are expensive and unaffordable for most users.
A pure modeling approach employing either Earth system models or terrestrial biosphere models has been used as an alternative to direct measurements on a large scale [
5,
6], but the spatial resolution of these models is too coarse, while the uncertainty remains high due to their inability to respond timely to actual changes in the ecosystem. Likewise, satellites that measure carbon dioxide (CO
2) concentrations and are used in inverse modeling to map GHG emission and sequestration [
7,
8] have coarse spatial resolution and hence can hardly be used for NBS assessment at the farm or forest stand level. Optical, radar, and lidar satellites provide a scalable solution for mapping soil carbon [
9] and carbon fluxes at a much finer spatial resolution [
10,
11] and do have the potential to help large-scale space- and time-resolved NBS assessments if anchored in accurate ground measurements. The current remote-sensing-based studies typically use allometric/biometric data for machine learning (ML) model training [
2]. The amount of ground data required for capturing the landscape heterogeneity and temporal variation is challenging to sustain and scale. The models thus inherit the disadvantages of the biometric measurements.
There is, however, another potential solution that combines many advantages of the above methods while overcoming some of their limitations and can result in a high spatial resolution, scalability to the regional level and practical affordability. This can be achieved by blending the direct highly time-resolved small-scale observations from the numerous eddy covariance flux towers [
2] and the less direct but large-scale remote sensing methods. Considerable efforts have been made in this area lately to improve the quantification of the spatiotemporal CO
2 patterns, employing multiple different approaches vs. existing estimates to help resolve the majority of issues with current carbon accounting methods in the land sector. For example, [
12] evaluate how seven light-use-efficiency (LUE) models captured the spatiotemporal gross primary productivity (GPP) variations from the LaThuile FLUXNET dataset. This study used data from 157 eddy covariance (EC) towers located in the six major terrestrial biomes (evergreen broadleaf forest (EBF), deciduous broadleaf forest (DBF), mixed forest (MF), evergreen needle leaf forest (ENF), shrubland (SHR), and grassland (GRA)). None of the seven models matched well with estimated GPP based on overall EC measurements. The models explained 41–57% GPP variations in the study sites. Reference [
13] evaluated four LUE models for GPP estimation and compared them with data obtained by EC in 51 sites with diverse vegetation types. The average daily
for all models and all sites was about 66%. Reference [
14] used three different models to estimate GPP in a local deciduous forest in the US in combination with flux measurements and remote sensing data. Comparisons were made yearly, but on average, the correlation coefficient was much better, at 0.91. Reference [
15] used two enhanced vegetation index (EVI)-based models to estimate the GPP in 15 sites across the USA. The model’s results vary according to the vegetation; in some cases, coefficients of determination can vary from 0.27 up to 0.94 for specific sites. Using remote sensing data, disturbance information, and a support vector regression model, [
16] upscaled the CO
2 fluxes from a network of 21 EC towers to estimate the regional-scale Alaskan CO
2 budget from 2000 to 2011 and compared the results with an inverse model. The predicted values were consistent with the field observations (
= 0.79 for NEE) and the magnitude of the inverse model. However, the interannual variations of the results achieved by the remote-sensing-based model and those of the inverse model were inconsistent and therefore require additional validation. Reference [
17] used three remote sensing data-driven models to evaluate and understand GPP estimations in 119 EC sites in the northern hemisphere. Considering all the study sites, the mean
was 0.63. However, the values can vary significantly depending on the model used and the vegetation, and in some cases, they can reach 0.78. Usually, the highest values were found for deciduous broadleaf and mixed forest. Reference [
18] mapped the soil organic carbon (SOC) in Italy using ground observations and ancillary remote sensing data. Compared with all observations, the geographically weighted regression using MODIS normalized difference vegetation index (NDVI) obtained an
of 0.45 for SOC. Reference [
19] conducted comprehensive work generalizing and mapping NEE, GPP, and respiration data for all types of classification as a collaborative FLUXCOM initiative. The overall coefficients of determination for GPP and respiration were more than 0.71 and 0.64, respectively, while about 0.48 for NEE.
As seen in the examples above, despite the significant advances in the use of remote sensing for CO
2 flux estimations, the results point to mixed success and lack of consistency, so there is room for major improvements. The key areas of such improvements include increasing spatial resolution, capturing temporal variability, improving the data quality (primarily by reducing uncertainties and increasing consistency) and developing better algorithms to enable the simultaneous use of multiple relevant parameters and different spatial and temporal scales. In order to achieve such improvements, we propose to rely on the significant advances which have been made in recent years in the fields of machine learning, artificial intelligence, and neural networks to extract patterns and insights from remote sensing data, processing large volumes of remotely sensed data with minor human interference [
20]. Here, we present the successful use of one such approach for globally scaled remote-sensing-based carbon accounting models anchored by the direct carbon flux measurements using eddy covariance flux stations. We use input data on the net ecosystem exchange (NEE, [
21,
22,
23,
24]) as targets for the training process and then combine the remote sensing and meteorological variables [
19] as predictors to reproduce variations of NEE on a daily and monthly scale to demonstrate the increased predictive power of the proposed approach over more traditional methods.
An overarching goal of this work is to bridge the domains of science and technology by integrating multi-scale remote sensing and eddy covariance measurements to support projects aimed at avoiding GHG emissions or to remove CO2 from the atmosphere. For this purpose, we target the following methodological improvements: (1) increasing spatial resolution, (2) capturing temporal variability, (3) improving the data quality (primarily by reducing uncertainties and increasing consistency), and (4) developing better algorithms to enable the simultaneous use of multiple relevant parameters and different spatial and temporal scales.
3. Results
As a result of
N-fold cross-validation for each land cover (
Section 2.2.3), we received an ensemble of
N different models trained on different independent datasets with the same inputs. Thus, we can use the mean value over the ensemble as a prediction.
First, selected metrics were applied to estimate the model’s accuracy by reproducing the original daily NEE values and then used to estimate the accuracy of the predicted NEE behavior. The results of the training and cross-validation experiments based on independent stations are presented in
Table 4. The trained models can adequately reproduce the behavior of NEE at stations with different land cover (overall mean across land-cover NEE,
> 0.6 with a range of 0.42 to 0.76). The lowest predicted values are for EBFs because such forests are mostly located in the tropics, where there are few Fluxnet stations, and there could be problems with cloud conditions during some seasons. The largest variation in error (standard deviation) is observed for CRO, possibly due to the difference in crop types and significant local anthropogenic influence on specific fields. An increase or decrease in NEE can be explained by the state of the biomass and the influence of agricultural management. Comparisons of observed values and predicted values are shown in
Figure 4 and
Figure 5.
To compare monthly NEE, additional daily footprint calculations similar to
Section 2.2.1 were carried out. In this case, the calculation period of one week before and one week after the available satellite images were selected. Usually, periods of 8 days are used [
19,
34], but in our case, we decided to use a similar averaging backward and forwards. Furthermore, remote sensing data from the original image was used for the additional dates, but with meteorological variables and weighting functions corresponding to the actual date. Monthly model NEE data were then calculated for those months where it was possible to predict more or equal to 2/3 days of the month. The resulting monthly NEE was compared with monthly NEE values from stations for which the quality flag was greater than or equal to 0.8. The results of the comparison are presented in
Table 5.
4. Conclusions
There has been a drastic rise in NBS development in recent years [
3]. Although the methods for evaluating NBS are advancing, there are still certain gaps that need to be addressed [
2]. Previous studies found that most of the current methods for quantifying NBS benefits use biometric soil or tree survey data to estimate carbon stock changes. These methods do not enable seamless large-scale assessments, while associated regular measurements make the methods financially unsustainable for most players in the land sector. Modeling approaches based on Earth system models or terrestrial biosphere models, as well as inverse modeling based on CO
2 concentrations data, have spatial resolution too coarse to be used locally. The current remote-sensing-based approaches mainly rely on allometric data as ground truth and therefore involve challenges with capturing landscape heterogeneity and temporal variation that negatively impact scalability. As the result of the analysis from the recent studies, it is fair to say that the approaches based on observations from eddy covariance towers and remote sensing are the best fit to enable affordable, high spatial resolution, globally scalable CO
2 estimates. We analyzed the studies representing a vast amount of scientific efforts targeted on these approaches and identified the significant challenges. Here, we presented an improved approach to quantifying carbon fluxes by blending eddy covariance data with moderate resolution satellite data and weather data. Our method (1) is scalable due to the validation design based on a separate set of eddy covariance stations, (2) features a high spatial and temporal resolution of the Landsat data and set of meteorological variables, and (3) delivers robust and accurate predictions due to improved data quality control, advanced machine learning techniques, and rigorous validation.
Fluxnet towers are expected to estimate NEE values for homogeneous land covers that are stable over time. Therefore, in order to apply our method the regions should be thoroughly controlled for these parameters. In addition, models can only be applied to the areas with known land cover. Application to areas with different land cover may skew the results significantly. At the same time, the eddy covariance system captures the net ecosystem exchange, which accounts for total CO2 fluxes, including those related to the heterotrophic respiration process. So if the AI remote sensing model was trained with EC NEE, it consequently accounts for heterotrophic respiration.
In terms of global coverage, it is true that the sites we use may not be representative for all possible climate zones. However, since we use several validation folds that are blindly compiled (we do not know which station from which region is in which fold), we can assume that the regression we have constructed (ensemble of models) may be able to describe the underlying behavior processes for NEE. Such estimates then can be applied globally. A exact conclusion can be drawn using new EC data from independent sources and other climate zones.
Further in the Discussion, we suggest the key reasons for these improvements and also list our opinion on the current limitations of the proposed approach. Our methodology employed eddy covariance data from the Fluxnet network. Since the network has a standardized methodology for data acquisition and processing, we collected a robust and consistent training dataset, typically unavailable for more traditional carbon accounting methods based on biomass and soil sampling. The availability of a large amount of consistent and uniform data allowed for setting a fairly conservative threshold of 80% for the NEE data coverage and resulted in substantially reduced noise. On the remote sensing side of the approach, the Landsat archive provided a consistent and robust set of data with great temporal depth and spatial resolution. The FLEXPART model was instrumental in mapping the area that contributed to the NEE values captured by flux stations, helping to match these with remote sensing products better to account for landscape heterogeneity. Utilization of the KRR and its parameterization tailored to the goals of this study, and strict feature selection, enabled high prediction accuracy.
Still, the metrics of our regression estimators were not uniform and varied across the land-cover classes. The for the DBF was the highest, likely due to clear seasonality patterns and relatively low heterogeneity within the class. The same metric for the EBF class was the lowest, likely due to other carbon exchange process drivers and lower data availability due to high cloud cover in tropical regions.
Current results are presented for each land-cover class globally, providing visibility on the unified, scalable calculation approaches and the global accuracy baselines for various land covers (
Figure 6). Overall, the achieved accuracy was in line with similar studies but outperformed most of those targeting global geographical coverage (overall land cover cross-validation
= 0.73,
RMSE = 1.53 gC/m
2/day). Daily data accuracy has the
range between 0.42 for EBF and 0.76 for DBF (monthly data accuracy is between 0.6 for GRA and 0.84 for DBF). The models have better accuracy in those parts of the world where the number of stations is higher and, conversely, slightly worse accuracy in areas of lower station density.
To resolve or minimize these deficiencies and improve accuracy for specific cases or regions, the current system allows additional training (not a local re-training) in the presence of the additional local flux observations, as suggested by [
2]. Another way to improve the confidence of model data is to add other multispectral, hyperspectral, and radar satellite products, e.g., Sentinel-1, 2, or EnMap. It is expected that these new data streams can improve the accuracy of the daily data slightly but can contribute significantly to the monthly and annual data improvement.
Finally, since in this study we have trained N independent models for each land cover (where N = 3, 5, 7). In the future, it may be better to use ensemble predictions to allow additional characteristics such as mean, median, and standard deviation. These and other activities are currently in development but not advanced enough to be included in this manuscript.
Given the urgent need to decarbonize the atmosphere and the critical role of NBS in doing so, a fusion approach of eddy covariance and remote sensing data proved to be an efficient and cost-effective solution for quantifying these efforts. We hope our methods could contribute to mitigating global climate change in a cost-effective and transparent way.