Climate-driven stock accounts of inland Green-Blue Ecosystems: Coupling the carbon and water cycles via Reverse Engineering and Geodetectors

doi:10.21203/rs.3.rs-4681296/v1

Download PDF

Research Article

Climate-driven stock accounts of inland Green-Blue Ecosystems: Coupling the carbon and water cycles via Reverse Engineering and Geodetectors

https://doi.org/10.21203/rs.3.rs-4681296/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

The critical role of inland Green-Blue Ecosystems (GBE) in delivering goods and services underscores the need to understand their relationships. This research investigates the impact of climate variables on GBE stock accounts by estimating Tree Cover Density (TCD) and Water & Wetness Probability Index (WWPI). Using supervised machine learning and factor analysis, we measured GBE extent and condition. Key predictors for Blue Ecosystems were topographic variables, while vegetation indices were crucial for Green Ecosystems. In 2018, 33% of the inland area was covered by forests, freshwater, and wetlands. Key climate-driven factors for forests included precipitation (0.65), aridity index (0.54), and evapotranspiration (0.44). For freshwater and wetlands, precipitation (0.69), aridity index (0.55), and elevation (0.42) were significant. This research enhances our understanding of how climate impacts GBE, influencing biomass density and water availability. It bridges socio-environmental science with engineering by integrating advanced modelling techniques, promoting ecosystem resilience and sustainability.

Environmental Modelling

Machine Learning

Google Earth Engine

Data fusion

Copernicus Land Monitoring Services

Classification tasks measured GBE extent, regression estimated GBE condition
Blue Ecosystems classification top three feature importance: DSM, NDWI and SLOPE
Climate-driven stock accounts of GBE were analyzed through geodetectors
The most important drivers for Green Ecosystems were PREC, AI, and PET
The top three drivers for Blue Ecosystems were PREC, AI, and DEM

Understanding the terrestrial biosphere’s functioning requires an assessment of the magnitude and distribution of Green-Blue Ecosystems (GBE) (Crowther et al., 2015). Green Ecosystems (GE) support a considerable amount of global biodiversity, playing an important role in biogeochemical cycles, and providing several Ecosystem Services (ES) such as water quantity and quality regulation, food production, raw materials provisioning, genetic resources, and carbon storage and sequestration (Mulatu et al., 2017). Forest density influences soil and water retention rates, flood regulation, as well as competitive dynamics and habitat suitability for a variety of plant and animal species (Rijal et al., 2021). Blue Ecosystems (BE) provide the backbone of ecosystems, enabling agriculture, and influencing urban development decisions (Newbold et al., 2015). As forests are intricately connected to rainfall and water availability, changes in vegetation cover, as well as poorly planned and managed forests, have significant impacts on the carbon and water cycles, leading to pressures on the environment (Liquete et al., 2011). Extreme weather events and soil and water contamination are perceived as the top risk, directly impacting natural capital and biodiversity, and consequently decreasing human well-being (Lundqvist & Unver, 2018). Knowledge of the interactions between carbon and water cycles is vital for predicting ecosystem responses to climate change, as well as the influence on natural capital (Seddon et al., 2021). Despite significant advances, the coupling mechanisms of these cycles under socio-economic and climatic pressures remain poorly understood (Keith et al., 2021).

Ecosystem Accounting (EA) encompass a statistical framework for quantifying and valuing the interactions between nature, society, and the economy, providing physical and monetary measurements of ecosystem assets (Bateman & Mace, 2020). In March 2021, the United Nations Statistical Commission adopted the System of Environmental-Economic Accounting Ecosystem-Accounting (SEEA-EA) as a baseline to inform policy development by countries. It represents a huge milestone in the history of accounting as “it moves beyond GDP and takes better account of biodiversity and ecosystems in national economic planning” (Edens et al., 2022). Thus, it is susceptible to shaping the future of the national statistic apparatus toward the inclusion of natural capital (Edens et al., 2022; Lange et al., 2022). Among the challenges still to be overcome is the substantial wide variety and amount of data required to produce comprehensive accounts (Bordt, 2018; Hein et al., 2020). Besides, extending locally generated ES models to other locations or scales has constraints due to the necessity for parameterization, calibration, and validation, which is sometimes hampered by a lack of ground truth data (Cord et al., 2017; Gosal et al., 2022).

The opportunities associated with technological innovations include the potential for gathering, processing, analyzing and visualizing different types of data, and integrating socio-economic and environmental information (Fleming et al., 2022). Advancements in geospatial techniques, such as satellite imagery, geospatial mapping, drones, and sensors, are viewed as highly promising to improve decision-making on both individual and cross-sectoral levels facilitating the deployment of EA (Farrell et al., 2021). Besides, the increased computational power and the transition towards open-access data and open-source technologies are significantly contributing to enhancing natural capital accounting (del Río-Mena et al., 2023).

Satellite Earth Observation (SEO) data and technologies, coupled with Geographic Information Systems (GIS), provide advantages such as synoptic and repeated coverage, historical spatial data analysis, and cost-effectiveness (Ramirez-Reyes et al., 2019). Therefore, where data is unavailable, imputation data techniques through statistics and Machine Learning (ML) are employed to fill in missing or incomplete data (Fleming et al., 2022). Additionally, data fusion is commonly applied, integrating information from diverse sources such as monitoring, statistics, modelling, or interviews (Braun et al., 2018). These approaches enable more frequent, and consistent assessment of ecosystem dynamics, regardless spatial scale of analysis (Hossain & Hashim, 2019).

In developing models for quantifying natural stocks, accurate and timely data on the extent and status of ecosystems are critical and challenging to gather particularly due to the dynamic and intricate interactions between water/groundwater and vegetation (Ellison et al., 2017). Significant efforts have been made to unveil the potential of SEO in ES assessments (Willcock et al., 2018), such as the European Copernicus programme powered by the European Spatial Agency (2023). The platform yields free and open access to SEO imagery and datasets, providing insights into the status of natural resources not only for Europe but also on a global scale (Almeida & Cabral, 2023). Among the various products, it provides the Water & Wetness Probability Index (WWPI) and the Tree Cover Density (TCD) to assist in building policy-relevant ecological accounts. The WWPI is a biophysical indicator for freshwater and wetland ecosystems (Ludwig et al., 2019) utilized as a spatial proxy to support environmental mitigation, wetland protection, erosion control, flood monitoring, and streamflow regulation (Vargas et al., 2019). The dataset provides information on the status of BE reflecting their overall quality and conditions (Copernicus Programme, 2023). The TCD is an indicator describing forest ecosystem conditions through a continuous spectrum of crown cover information, population numbers, densities, and wood stocks (Crowther et al., 2015). It is acknowledged as an effective method for environmental analyses, forest management, supporting decision-making, and tracking changes in tree cover losses and gains (Chen et al., 2020). Therefore, these products have some limitations due to: a) the availability and quality of SEO data, b) the availability and quality of in-situ data, and c) the influence of climatic and topographic conditions (European Environment Agency, 2023). Satellite mission launch dates restrict SEO data availability, and the quality of imagery is influenced by the atmospheric effects on the light reflected off the Earth's surface and captured by the sensor. Another challenge is the lack of ground-truth data, which is expensive to acquire (Mairota et al., 2015).

Some studies have demonstrated improvements with the integration of climate and topographic variables when assessing GBE. Crowther et al. (2015) demonstrated the strong associations between climatic variables and forest density in their research. Ludwig et al. (2019) combined multi-temporal optical imagery, topographic data, and spectral indices to build an automated wetland mapping detection, based on the WWPI. Madrigal-González et al. (2023) used climatic and topographic characteristics to investigate the relationship between tree density and water availability. Han et al (2021) estimated carbon stock in forests by considering temperature, precipitation, elevation, and soil type as factors. Alqadhi et al (2022) demonstrated the impact of landscape and topographic characteristics on ES capacity provisioning, considering: elevation, slope, aspect, drainage density, and precipitation. However, none of these studies used Reverse Engineering (RE) to reproduce TCD and WWPI taking into consideration climatic and topographic factors to estimate inland GBE at the national level. RE refers to the process of extracting knowledge or design information from products and subsequently reproducing or recreating them based on the acquired information (Nieves-Chinchilla et al., 2018). Through RE, it is possible to learn about the products, search for inconsistencies, or overall vulnerabilities, and assess whether there is a more efficient way to improve them (Wood, 2009).

In this study, the WWPI and TCD will be modelled within Portugal's mainland in ArcGIS Pro (ESRI, 2023) using the Random Trees Regression algorithm, and spatial associations will be analyzed through geographical detectors utilizing the geodetectors package (version: 1.0–4) in R software (Wang et al., 2016). EA baselines (Kienast et al., 2009) will be followed to assess the climate-driven stock accounts of forests, water, and wetlands at the national level. The aim is to develop more integrated and updated versions of TCD and WWPI contributing to the broader scientific community by exploring the potential of SEO data and technologies combined with global climatic and topographic open-access data. Furthermore, this research will shed light on the interplay between aquatic and terrestrial ecosystems and foster innovative solutions for assessing the status and conditions of inland GBE. Advancing our understanding of the development of Copernicus products will deepen our knowledge of the replicability and applicability of such datasets in EA.

2.1. Study Area

Portugal is located on the Iberian Peninsula in southwestern Europe, comprising a continental portion and two autonomous regions - the Azores and Madeira archipelagos. Portugal's mainland spans an approximate area of 90000 square km located between 36°57' − 42°9'N latitude and 6°12' − 9°30'W longitude in a climatic Mediterranean region (Beck et al., 2018) (Fig. 1).

The environment changes importantly from north to south, with complex and diverse landscapes (Almeida & Cabral, 2021). Spatial and temporal variability is characterized by latitude, orography, and the effect of the Atlantic Ocean, which affects the water cycle primarily through oscillations in precipitation and temperature (Belo-Pereira et al., 2011). The territory is prone to climate variability, including droughts and desertification, like other southern European regions, notably in the southern sector, which receives the least precipitation and has the greatest temperatures (Oliveira et al., 2007).

The climatic conditions, influencing vegetation patterns, contribute to a diverse array of Atlantic, European, and Mediterranean plant species within the country (Santos et al., 2023). Forest ecosystems predominantly consist of broad-leaved and coniferous forests, including Cork oak, Holm oak, Eucalyptus, Maritime pine, and Stone pine (Costa et al., 2022). This ecological diversity is evident across the country, with the northern regions characterized by flourishing broadleaf and coniferous forests like Eucalyptus, Maritime pine, and Stone pine. In contrast, the southern landscape exhibits arid-adapted Mediterranean species such as cork oak, holm oak, and olive trees, representing distinct climatic zones in the country (Fonseca et al., 2019).

2.2. Data

Table 1 presents the data used in this study, describing their source, spatial resolution, and the role played by each dataset in the classification and regression tasks.

Table 1

Description of datasets utilized in the research. Legend: C – Classification; R – Regression; GEE – Google Earth Engine.
Data	Source	Pixel resolution	Task	Role
Satellite imagery 2018 - Single bands	European Space Agency (ESA) – GEE	10m	C	Feature
Satellite imagery 2018 - Spectral indices	European Space Agency (ESA) – GEE	10m	C	Feature
‍Global Digital Surface Model (DSM) (m)	Earth Observation Research Center Japan Aerospace Exploration Agency (JAXA EORC) - GEE	30m	C	Feature
Land Use and Land Cover Map (COS2018) 2018 (m)	Directorate General for Territory - DGT (https://snig.dgterritorio.gov.pt/)	10m	C	Target
Digital Elevation Model (DEM) (m) & Slope (SLOPE) (%)	NASA Shuttle Radar Topography Mission (https://www.earthdata.nasa.gov/)	30m	R	Predictor
Average Annual Precipitation 2018 (PREC) (mm)	WorldClim (https://worldclim.org/)	1km	R	Predictor
Global Potential Evapotranspiration 2018 (PET) (mm)	Consortium of Spatial Information - CSI (http://www.cgiar-csi.org)	1km	R	Predictor
Global Aridity Index 2018 (AI)	Consortium of Spatial Information -CSI (http://www.cgiar-csi.org)	1km	R	Predictor
Water & Wetneess Probability Index (WWPI %) 2018	Copernicus Land Monitoring Services - CLMS (https://land.copernicus.eu/en/products)	10m	R	Response variable
Tree Cover Density (TCD %) 2018	Copernicus Land Monitoring Services - CLMS (https://land.copernicus.eu/en/products)	10m	R	Response variable

In the GEE platform, cloud-free composites of Sentinel-2 (S2) imagery encompassing the entire mainland of Portugal for 2018 were selected. All available spectral bands from S2 imagery, satellite-derived indices, and topographic features, were employed as classification predictors. These included the Normalized Difference Vegetation Index (NDVI)(Rouse et al., 1973) and the Enhanced Vegetation Index (EVI)(Huete et al., 1997) used to highlight and distinguish vegetation of different types and stages throughout their seasonal cycles; the Soil Adjusted Vegetation Index (SAVI)(Huete, 1988) to mitigate the effects of soil reflectance and improve vegetation response, and the NDWI (Normalized Difference Water Index) used to differentiate between water and soil, based on the optical characteristics of water (Ji et al., 2015). The S2-derived indices were calculated using the band arithmetic functions: (1), (2), (3), and (4):

$$\:NDVI=\frac{B8-B4}{B8+B4}$$

$$\:EVI=2.5*\frac{B8-B4}{B8+6*B4-7.5*B2+1}$$

$$\:SAVI=\frac{B8-B4}{(B8+B4+L)*\left(1.0*L\right)}$$

$$\:NDWI=\:\frac{B3-B8}{B3+B8}$$

In the formulas, B refers to the spectral band, and L is the soil brightness correction that ranges between [0, 1]. In the SAVI calculation, L takes the value of 0.428, as suggested by Huete (1988). The spectral indices (1), (2) and (3) were used to classify forests, and (4) was employed in the freshwater and wetlands classification.

Topographic variables, such as the Digital Surface Model (DSM) and Slope, were provided by the Japan Aerospace Exploration Agency (JAXA, 2023). These datasets were added with the bands and spectral indices as an 18-band data stack to perform the classification tasks. The classification training datasets included reference vectorial data representing GBE (forest, freshwater, and wetlands) were extracted from the Portuguese land use and land cover map for the year 2018 (COS2018), a national product developed by the Directorate General for Territory (DGT, 2018).

Climatic variables, such as Precipitation (PREC), Potential Evapotranspiration (PET), Aridity Index (AI), and the Digital Elevation Model (DEM) and Slope (SLOPE) were further utilized as explanatory variables/factors in the regression tasks. Annual PREC was obtained from the WorldClim database (Fick & Hijmans, 2017); Global PET and the Global AI were obtained from the Consortium of Spatial Information, Global-AI and Global-PET Database (Zomer et al., 2022). Global PET and AI were derived from the WorldClim precipitation model and have the exact spatial resolution and temporal scales.

The response variables (WWPI and TCD) are from the High-Resolution Layers portfolio of the Copernicus. The WWPI is a raster that depicts the presence of water and wet areas as an index ranging from 0 to 100% (Copernicus Programme, 2023). The WWPI workflow includes satellite data fusion, supervised ML classification and an unsupervised thresholding technique for water/wetness detection. Data fusion is based on optical data from S2 and Landsat 5/7/8 and radar data from Sentinel-1 multi-temporal imagery from 2012–2018. The classification task was performed with the RF algorithm, and post-processing included quality assurance, control procedures, visual improvements and error correction. The TCD is a vertical projection of tree crowns onto a horizontal Earth’s surface providing pixel-level information on tree cover density (%). In the TCD development, Sentinel-2A + B time series (Level-2A data) were employed to build a supervised classification model using an RF classifier to create a binary Tree Cover Mask (TCM) (non-tree covered areas/tree cover). Following that, a multiple linear regression was carried out inside the bounds of the TCM to estimate TCD values.

2.3. Methods

Forward Engineering goes from concept to product, while RE does the opposite, applying techniques to create a blueprint of an existing product (Vallero, 2014). RE is frequently used to replicate existing products, test, or improve functionalities, and introduce new data (Nieves-Chinchilla et al., 2018). It is defined as the process of examining an object to determine its components and their interrelationships for further redesign and recreation (Wood, 2009). There are a few steps to follow when applying RE, such as defining the purpose, reviewing product specifications, identifying limitations and vulnerabilities, disassembling, analysing, redesigning, and reporting (Chikofsky & Cross II, 2002).

The research development starts applying the RE approach to understand how TCD and WWPI were created, how they work, and how they could be improved, followed by employing the EA framework to quantify the physical measurements of inland GBE, through estimations of ecosystems extent (Fig. 2a) and condition (Fig. 2b). Ecosystem extent quantify the area per ecosystem type inside an accounting area. In contrast, ecosystem condition record the status of an ecosystem asset through indicators that reflect its condition (United Nations et al., 2021). The extent was estimated through classification tasks mapping GE and BE, while the condition of forest, freshwater and wetlands was predicted by regression tasks. Finally, there was the calculation of factor importance and interaction, and statistic tests between factors and within ecosystems, employing geodetectors.

2.3.1. Classification Task

Two distinct classification tasks were conducted: one dedicated to creating the Forest Cover Mask (FCM) and another to creating the Freshwater and Wetlands Mask (FWM). The training data were extracted from the COS18 dataset, and used as samples for a Random Forest (RF) classification (Breiman, 2001), in Google Earth Engine (GEE) (Google, 2023). The scripts are provided in Tables S1 and S2 in the Supplementary Materials.

Random point sampling was employed to provide training and validation data for the tasks. The RF ML algorithm evaluated the training features in terms of satellite-derived data, DSM, and slope. Additionally, to optimize the model and improve its accuracy, an analysis of feature importance and hyperparameter tuning was conducted for each classification. These analyses enabled the identification of the most relevant features and refining the RF algorithm, enhancing classification accuracy for each designated class.

The classification result was then subject to an accuracy assessment procedure using the validation points by calculating different metrics, such as overall accuracy (OA), confusion matrix, producers’ accuracy (PA), users’ accuracy (UA), kappa coefficient, and F1 score. Lastly, post-processing techniques were applied to eliminate isolated pixels and noise.

2.3.2. Regression Task

Predictive spatial ML regression was used to model the relationship between explanatory variables and response variables (WWPI and TCD) using ArcGIS Pro 2.9.0 software. The regression tasks were implemented using the tool Train Random Trees Regression Model.

The ML spatial regression models for TCD and WWPI were built considering five explanatory variables (PREC, PET, AI, DEM, and SLOPE). The Percent Samples for Testing option was set to 20%, which means that one-fifth of the training sample, referred to as test location points, was used to quantify the error for interpolation in space. Three types of errors were measured: errors on training points, errors on test points, and errors on test location points. The maximum number of trees, maximum tree depth, and maximum number of samples were respectively set to 50, 30, and 100000. The maximum depth of each tree refers to the number of rules that each tree can generate to decide.

It is critical to evaluate regression performance to understand how well the model is fitted and explained by independent variables (Hosmer & Lemeshow, 2000). The metrics utilized to detect bias and the amount of variance of the response variables were the coefficient of determination (R-squared) and regression error (Re). The tool outputs a table describing the importance of each predictor used in the model, as well as scatterplots of training data, test data, and test location data, and a regression definition file containing attribute information, statistics, and model performance.

2.3.3. Factor Analysis using Geodetectors

Factor analysis was carried out to estimate the drivers of GBE using geographical detectors. Geodetector is a tool for investigating spatial stratified heterogeneity (SSH) through (1) measuring SSH of a variable Y; (2) testing the power of a determinant X of a dependent variable Y based on the consistency of their spatial distributions; and (3) investigating the interaction between two explanatory variables X1 and X2 (Wang et al., 2016). The strata of Y are a partition of Y by a categorical explanatory variable X (h(X)) (J. Yang et al., 2022). In this study, the strata refers to each ecosystem, hence the q-statistic was calculated for GE and BE. The q-statistic value is between [0, 1], then q = 0 indicates no relationship between Y and X; q = 1 shows that Y is governed by X. The tool is implemented by the geographical detector q-statistic (5):

$$\:q=1-\left(\frac{1}{N{\sigma\:}^{2}}\right){\sum\:}_{h=1}^{a}{N}_{h}{\sigma\:}_{h}^{2}$$

where N and $\:{\sigma\:}^{2}$ represent the number of units and Y variance, respectively; the population Y is made up of a strata (h = 1, 2,..., a), and $\:{N}_{h}$ and $\:{\sigma\:}_{h}^{2}$ represents the number of units and Y variance in stratum h, respectively.

The factor analyses were calculated using the R geodetector package (Wang et al., 2016). The function returns the q statistic as well as the respective p-value. The ecological detector determines whether there are statistically significant differences between two risk factors X₁ ~ X₂. The F statistic is utilized in this function to test the differences at a significant threshold of 0.05. The interaction detector determines whether the factors interact with Y. The risk detector computes the average values for each stratum of the explanatory variable (X) and shows if there are differences between strata.

3.1. Forest, freshwater and wetlands mapping

A binary classification task mapping forest/non-forest areas was performed at the national scale. Also, a multilabel classification task delineating freshwater and wetlands areas was conducted. Both classification maps are required for the calculation of ecosystem extent and are employed as masks in the subsequent step, the regression tasks (Fig. 3).

The FCM and FWM classifications exhibit high accuracy, strong agreement, and reliability in their respective classifications. Table 2 presents the accuracy evaluation for forest classification (FCM) and freshwater and wetlands classification (FWM).

Table 2

Accuracy evaluation. OA: Overall Accuracy; PA: Producer’s Accuracy; UA: User’s Accuracy.
LULC	OA (%)	PA (%)	UA (%)	Kappa	F1 Score
GE	96.17	93.68	95.71	0.89	0.93
BE	94.66	95.83	96.67	0.94	0.94

The FWM classification achieved an overall accuracy (OA) of 94.66%, while the FCM classification achieved an even higher accuracy of 96.17%, indicating a high level of correctness in their respective classifications. For the producer’s accuracy (PA), the FCM classification demonstrated a result of 93.68%, while the FWM classification showed a slightly higher result (95.83%), indicating their ability to identify their respective categories accurately. As for user accuracy (UA), the FCM classification exhibited an accuracy of 95.71%, while the FWM classification demonstrated an accuracy of 96.67%, indicating the reliability of both models’ predictions.

The forest cover classification exhibited a robust agreement, boasting a Kappa value of 0.89. The classification of freshwater and wetlands demonstrated an even higher level of agreement, with a Kappa value of 0.94. These results underscore both models’ strong consistency and reliability in accurately categorizing vegetation, freshwater, and wetlands. The F1 scores further reinforce the models’ effectiveness in correctly identifying these land cover categories. The FCM classification achieved a commendable F1 score of 0.93, while the FWM classification excelled with an impressive F1 score of 0.94.

When looking at the mean percentage weight of feature importance for the water and wetlands classification at the national level (Fig. 4), the analysis provided insights into the features crucial for discriminating water bodies and wetland areas accurately. As such, among the 13 attributes assessed, the surface elevation (DSM) exhibited the highest significance (13.65%), followed by NDWI (9.70%) and Slope (9.30%). S2 Bands (8, 11, 12, 8A, 4, 6, 5 and 7) presented importance higher than 5%, while Bands (3 and 2) were less than 5%.

Regarding the forest cover classification, a total of 15 attributes were considered as predictors. The feature analysis demonstrates that the spectral indices, SAVI (9.14%), NDVI (8.19%) and EVI (6.89%), showcased notable significance, but also S2 Bands 11 (8.70%), 3 (8.09%) and 12 (7.91%). The Short Wave Infrared (SWIR) bands, B11 and B12, capture variations in vegetation water content and the spongy mesophyll structure within vegetation canopies. Meanwhile, B3 is sensitive to healthy vegetation. The S2 Band 8A was the only feature contributing with less than 5%. The S2 Bands (7, 8, 6, 5, 2 and 4), DSM and Slope presented an importance weight of more than 5% but less than 6.89%.

3.2. FCD and FWPI predictions

R-squared and residuals were used to assess the accuracy of FCD and FWPI predictions. The true values were compared to predicted values in terms of proportion of variance (R-squared), and residuals over targets. Figure 5 shows the scatterplot of predicted (FCD) over-reference values (TCD), and results for R-squared at training points (Tr), at test points (Te), and test location points (Tl), as well as the residual plot. The R-squared at Tr, Te and Tl, were respectively, 0.8746, 0.5456, and 0.3208. The R-squared for residuals is 0.4131 and the mean residual value is 0.04. These statistics were calculated using 100000 samples.

The scatterplot of predictions (FWPI) over observed values (WWPI) and the residual plot for the FWPI model are shown in Fig. 6. R-squared values at Tr, Te, and Tl were 0.8275, 0.6153, and 0.6555, respectively. The R-squared residual is 0.4765 and the mean residual value is 0.08. In the predictions of FWPI, the R-squared showcased superior results in Te and Tl. However, FCD performed better than FWPI, as R-squared at Tr is higher, and corresponds to 80% of samples.

Table 3 presents the equations for each sample set (Tr, Te, and Tl) used to assess the models’ performance.

Table 3

Performance evaluation by model (Tr: Training data at training locations, Te: Test data at training locations, and Tl: Test data at test locations).
Model	Residuals	Tr	Te	Tl
FCD	$\:y=0.32x-15.76$	$\:y=0.74x+13.68$	$\:y=0.54x+22.83$	$\:y=0.32x+24.62$
FWPI	$\:y=0.36x-\:19.82$	$\:y=0.71x+17.9$	$\:y=0.55x+24.71$	$\:y=0.6x+24.21$

The explanatory variables for both regression models were DEM, SLOPE, PREC, PET, and AI. Figure 7 describes the importance of each variable for FCD and FWPI. The predictors with the highest percentage values were more relevant to the model. All variables were shown to be important for both models, except the SLOPE. DEM and PREC were the factors that most contributed to the FWPI model, followed by PET and AI, and lastly, SLOPE. For the FCD model, PET and AI weighted the most, followed by PREC, DEM, and SLOPE.

3.3. Factor analysis using geodetectors

The factor detector q-statistic indicates if explanatory variables stratify Y (GE and BE). Table 4 depicts the factor detector results. In the prediction of BE, PREC and AI presented q-values higher than 0.5; DEM and PET higher than 0.25; and SLOPE close to 0, meaning there was no association between FWPI predictions and SLOPE. For GE, PREC and AI show the highest degree of association with values higher than 0.5, followed by PET (0.44), SLOPE (0.13) and DEM (0.11).

Table 4

Factor detector results for Blue Ecosystems (BE) and Green Ecosystems (GE) estimation.
Y	Statistics	SLOPE	DEM	PET	AI	PREC
BE	q-statistic	0.06	0.42	0.29	0.55	0.69
BE	p-value	1	9.04E-10	9.05E-10	9.35E-10	9.19E-10
GE	q-statistic	0.13	0.11	0.44	0.54	0.65
GE	p-value	1	9.50E-10	8.47E-10	9.66E-10	9.36E-10

The ecological detector determines if there are statistically significant differences between two covariates. This function utilizes the F-statistic to test for differences, employing a significance threshold of 0.05. In both GE and BE models, all variables were found to be significantly different.

The interaction detector determines whether the factors interact with the covariate. All interactions between factors had q-values greater than individual factors, demonstrating that combining carbon and water cycle phenomenon-related factors increases the explanatory power of GBE (Fig. 8). The strongest interactions (greater than 0.75) were observed between AI and PREC (0.83), DEM and PREC (0.82), AI and PET, PET and PREC (0.78), and between DEM and PET (0.75). It was not verified that there were any interactions between AI with SLOPE, AI with DEM, and SLOPE with PREC, but SLOPE interacted with DEM (0.59) and PET (0.53).

The risk detector computes the average values within ecosystems and determines if there are significant differences between the means of each stratum. Considering the country's extent, the mean value of GE corresponds to 51.47%. The mean value of BE referring to wetlands was 66.36% and 87.23% for freshwater ecosystems. All statistics were significantly different, considering a significant threshold of 0.05.

3.4. Inland GBE physical accounting: extent and condition

In 2018, forests, freshwater bodies, and wetlands covered around 62% of the territory of Continental Portugal. Forests occupied a total area of 5287284.95 hectares covering 60.32% of the total country area, freshwater bodies filled an area of 128391.01 ha, constituting roughly 1.7%, and wetlands 0.33% occupying an area of 21792.51 ha.

The condition of forests was assessed with the mean forest crown cover density (FCD), and the mean FWPI a proxy for the status of freshwater and wetlands. The GBE status assessment is calculated by pixel, meaning that, in an area of 100 square meters classified as forests, 51.47% of this area will account for aboveground biomass. In a pixel classified as freshwater and wetlands, respectively, 87.23% and 66.36% will effectively account for Blue Ecosystems. The area (hectares), proportional extent (%), status (%), and physical accounts (%) of forests, freshwater bodies and wetlands in 2018 at the national scale are presented in Table 5.

Table 5

Accounting for inland GBE extent and condition for Mainland Portugal in 2018.
2018	Forest	Freshwater	Wetlands
Area (ha)	5287285	128391.01	21792.51
Proportional extent (%)	60.36	1.7	0.33
Status (%)	51.47	87.23	66.36
Physical accounts (%)	31.07	1.48	0.22

We proposed a novel methodology that relies on RE techniques to disassemble TCD and WWPI datasets from the Copernicus products portfolio and redesign and reproduce them introducing new data and methods to enhance usability. Besides, our study delivered a hands-on approach to estimating climate-driven stock accounts of forest, freshwater, and wetland ecosystems. The findings on the drivers of biomass density and water availability shed new light on the potential implications of land cover conversions and climate conditions on inland GBE. This information is critical, given that these ecosystems are biodiversity hotspots capable of supporting a wide range of livelihoods, and able to provide a variety of goods and services.

4.1. Cloud-based automated classification of inland GBE

Open-access geospatial data streams and open-source software have made frequent, large-scale, high-resolution land cover mapping a reality (Arruda et al., 2021). These recent breakthroughs in acquiring and processing big earth data have made large-scale, high-resolution land cover classifications more viable for a diverse range of user groups, organizations, and scholars (DeLancey et al., 2019). GEE has evolved in recent years as an appealing high-performance computing platform for the cloud-based processing of petabytes of earth data, offering computational capability for planetary-scale data processing, as well as the ability to apply well-known ML algorithms (Gorelick et al., 2017).

The separation between GE and BE was implemented primarily to test the significant improvements in the accuracy and utility of such analyses (Koma et al., 2021). The fundamental motivation for segregating GE and BE into distinct categories arises from their inherently disparate spectral characteristics. Vegetation, characterized by its chlorophyll content and unique reflectance properties, displays pronounced spectral features in the green and near-infrared regions (Gitelson et al., 2003). Conversely, water bodies, including wetlands, exhibit spectral absorption patterns in the same regions (Khalid et al., 2021). These distinctive spectral properties lay the foundation for specialized classification approaches.

The classification analysis has underscored the improvement in classification accuracy achieved through this separation. The Kappa coefficient in both classifications showed compelling results. This enhancement in accuracy is not merely an academic achievement but has practical implications for informed decision-making (Jagannathan & Divya, 2021). Beyond accuracy, the separation of GE and BE enables tailored analysis for specific applications. For instance, GE classification supports studies related to vegetation health, land-use planning, and forest management, while BE classification aids in monitoring water bodies, wetland conservation, and hydrological assessments (Kulawardhana, 2011).

The F1 scores, indicative of the models' robustness in correctly identifying GE and BE categories, further emphasize the advantages of this separation (Alem & Kumar, 2022). GE achieved an F1 score of 0.93, while BE classification surpassed an F1 score of 0.94. These high F1 scores underscore the models' ability to effectively distinguish between these two categories.

A stand-alone ML classification model for mapping GE and BE has significant implications for environmental assessment, empowering researchers to conduct focused analyses on ecosystems, and supporting decision-makers in crafting sustainable land management and conservation strategies (Chignell et al., 2018; Z. Yang et al., 2021). Furthermore, the outputs of these classification tasks, the FCM from GE mapping and FWM from the BE were utilized to delineate ecosystem assets and quantify physical measurements of inland GBE (United Nations et al., 2021).

4.2. From Reverse Engineering to satellite-based climate-driven models

There is an emerging trend involving using SEO data to quantify ecosystem structure and functional traits which serve as more accurate indicators of ES than simplistic cover-based proxies (Schirpke et al., 2023). Their combination with regression models contributes to a comprehensive understanding of the social-ecological processes influencing the provision of ES and yields reliable data and results (C. Ramirez-Reyes et al., 2019) and thereby, provide cost-effective data to supply ecological accounts (Vargas et al., 2019).

The WWPI and TCD have the potential to provide data to build ecological-economic accounts for GBE. Redesigning and reproducing these datasets focused on forest, freshwater and wetlands evidenced the interdependency of some ecosystem functions, such as biomass density, water yield and soil moisture water content, and climatic and topographic drivers, as shown by Mengist et al. (2021). Topography and climate are key elements in explaining changes in vegetation abundance at the local (i.e. pixel, plot, or parcel scale) and regional levels (Mpakairi et al., 2022).

FCD predictions were better than FWPI when comparing the results of R-squared, regression errors and residuals. The independent variables considered in both models were topographic characteristics (SLOPE and DEM) and climatic factors (PREC, PET and AI). Using the same factors to estimate SSH within carbon and water cycles demonstrated the intrinsic relationships between aquatic and terrestrial ecosystems (Cantonati et al., 2020; Dwire et al., 2018; Hose et al., 2021; Huang et al., 2020; Kløve et al., 2014; Koit et al., 2021). The aboveground biomass density represented by FCD estimates varies spatially and temporally according to ecosystem types, altitude, latitude, climatic conditions, and water presence (Tabacchi et al., 2011). GE supports biodiversity and many livelihoods around lakes, perennial streams, and wetlands, which benefit from BE’s water and nutrient availability (Harris, 1984). Mpakairi et al. (2022) have shown that the density of healthy trees decreased due to limited water and nutrient availability, with water presence driving about 40% of biological processes in forests.

Regarding the contribution of each variable to the models, all of them showed to be relevant, a part of SLOPE presenting very low importance to both predictions. DEM and PREC were the factors that most contributed to the FWPI model, followed by PET, AI, and SLOPE. For the FCD model, PET and AI contributed the most, followed by PREC, DEM, and SLOPE. The results are aligned with those presented in the studies of Crowther et al. (2015) and Madrigal-González et al. (2023), analysing the influence of climatic and topographic conditions in assessing GBE. These findings expanded our understanding of nature stock climate drivers, landscape behaviours, and increasing accuracy in ecosystem accounts. Therefore, unleashing the capacity to analyze the status and conditions of GBE helps to execute integrated natural resource management.

The geographical detector analysis used to verify SSH enriched the knowledge of the drivers of GE and BE. The factor detectors indicate that both response variables are stratified by the exploratory variables, with more evidence of stratification shown by PREC, PET and AI. DEM showed a higher association with the BE than the GE model. Lastly, the physical accounts of inland GBE were assessed using the EA methodological framework. The national GBE extent account represented their magnitude in terms of area in hectares, while the GBE condition described their integrity and status (Czúcz et al., 2021). The aboveground biomass in the forest was measured by calculating the area classified as forest, and the mean estimated value of FCD was applied to assess GE condition. BE were evaluated by determining the extent of freshwater bodies and wetlands and the mean estimated values of FWPI for each ecosystem. The physical accounts evidenced that in 2018, 33% of the total inland area of Portugal was represented by forest, freshwater and wetlands ecosystems. The physical accounts offer means to understand how sustainable the production is through measurements of natural stocks and yields, providing baselines for the assessment of ecosystems’ capacity to deliver goods and services at multiscale (Keith et al., 2020).

4.3. Implications for future research

There are challenges with reverse engineering, especially when it comes to digital products (Chikofsky & Cross II, 2002). Other issues outside understanding the dataset itself include determining how it was created when we have little or no access to inputs, intermediate outputs, or even the code, algorithms, and tools used in each task (Wood, 2009). Without clear documentation of the model architecture, hyperparameters, and training process, it may be difficult for others to reproduce methods or understand the model's behaviour, and how the model arrives at its predictions. The transparency through reporting the entire ML pipeline, including data preprocessing, model architecture, hyperparameters, and evaluation metrics enables others to reproduce and validate the research findings and improve the robustness and reliability of ML models, leading to better accuracy and more trustworthy results. Additionally, continuous monitoring and updating of models as new data becomes available can help maintain their performance over time.

Nowadays, a wide range of technologies exists to aid in the forward/reverse engineering process, especially in the production of digital products. A good improvement would be to provide more information about these products' development and make version control available to the public. This practice helps in keeping track of dataset modifications, comparing several dataset versions, and revert to whichever previous version is required. Future research would further take these products apart to continue improving functionality, introducing new data and testing algorithms, as well as replicating to other areas in the globe. Additionally, it would be interesting to explore other products developed by Copernicus, which comprises a wide variety of products, such as Soil Water Index, Imperviousness Density and Impervious Built-up, Dominant Leaf Type and Forest Type, Coastal Zones and Protected Areas. These datasets are valuable resources for a wide range of geospatial services and environmental analysis.

This research demonstrated the potential of the TCD and WWPI products provided by Copernicus as fundamental data for ecological-economic studies and presented an approach to overcome limitations and enhance their usability. Besides, it underscores the significance of climatic (PREC, PET, AT) and topographic (DEM and SLOPE) variables influencing the status and conditions of GBE, by analyzing spatial relationships in the TCD and WWPI predictions. RE proved to be a valuable tool for innovation and problem-solving while examining the product's functionality and behaviour to determine how it works and identifying any vulnerabilities, weaknesses, or areas for improvement. The outcomes demonstrate the synergy of combining global open-access data with open-source technologies in creating predictive models that offer actionable insights, thereby facilitating evidence-based decision-making across various environmental and societal challenges.

Funding

Open Access funding was provided by NOVA Information Management School NOVA University Lisbon. This study was funded by the research project MaSOT – Mapping Ecosystem Services from Earth Observations, funded by the Portuguese Science Foundation – FCT [EXPL/CTA-AMB/0165/2021], and by national funds through FCT (Fundação para a Ciência e a Tecnologia), under the project - UIDB/04152/2020 - Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS.

Author Contribution

B.A.: Conceptualization, Methodology, Investigation, Formal analysis, Writing - original draft, preparation, Writing - review and editing. L.M.: Conceptualization, Methodology, Investigation, Writing - review and editing. P.S.: Conceptualization, Methodology, Investigation, Writing - review and editing. P.C.: Conceptualization, Methodology, Investigation, Writing - review and editing.

Acknowledgement

This study was supported by the research project MaSOT – Mapping Ecosystem Services from Earth Observations, funded by the Portuguese Science Foundation – FCT [EXPL/CTA-AMB/0165/2021], and by national funds through FCT (Fundação para a Ciência e a Tecnologia), under the project - UIDB/04152/2020 - Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS. We are grateful to the Institut français du Portugal for fostering Franco-Portuguese collaborations through scientific periods in France (BIC-2023), and the IUEM for hosting.

Alem, A., & Kumar, S. (2022). Applied Artificial Intelligence End-to-End Convolutional Neural Network Feature Extraction for Remote Sensed Images Classification. https://doi.org/10.1080/08839514.2022.2137650
Almeida, B., & Cabral, P. (2021). Water yield modelling, sensitivity analysis and validation: A study for Portugal. ISPRS International Journal of Geo-Information, 10(8). https://doi.org/10.3390/ijgi10080494
Almeida, B., & Cabral, P. (2023). Data-Driven Modelling of Freshwater Ecosystems: A Multiscale Framework Based on Global Geospatial Data. Proceedings of the 9th International Conference on Geographical Information Systems Theory, Applications and Management, 2023-April, 104–111. https://doi.org/10.5220/0012037800003473
Arruda, V. L. S., Piontekowski, V. J., Alencar, A., Pereira, R. S., & Matricardi, E. A. T. (2021). An alternative approach for mapping burn scars using Landsat imagery, Google Earth Engine, and Deep Learning in the Brazilian Savanna. Remote Sensing Applications: Society and Environment, 22, 100472. https://doi.org/10.1016/j.rsase.2021.100472
Bateman, I. J., & Mace, G. M. (2020). The natural capital framework for sustainably efficient and equitable decision making. Nature Sustainability, 3(10), 776–783. https://doi.org/10.1038/s41893-020-0552-3
Beck, H. E., Zimmermann, N. E., McVicar, T. R., Vergopolan, N., Berg, A., & Wood, E. F. (2018). Present and future Köppen-Geiger climate classification maps at 1-km resolution. Scientific Data, 5(1), 180214. https://doi.org/10.1038/sdata.2018.214
Belo-Pereira, M., Dutra, E., & Viterbo, P. (2011). Evaluation of global precipitation data sets over the Iberian Peninsula. Journal of Geophysical Research: Atmospheres, 116(D20). https://doi.org/https://doi.org/10.1029/2010JD015481
Bordt, M. (2018). Discourses in Ecosystem Accounting: A Survey of the Expert Community. Ecological Economics, 144, 82–99. https://doi.org/10.1016/J.ECOLECON.2017.06.032
Braun, D., Damm, A., Hein, L., Petchey, O. L., & Schaepman, M. E. (2018). Spatio-temporal trends and trade-offs in ecosystem services: An Earth observation based assessment for Switzerland between 2004 and 2014. Ecological Indicators, 89, 828–839. https://doi.org/10.1016/j.ecolind.2017.10.016
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324
Cantonati, M., Stevens, L. E., Segadelli, S., Springer, A. E., Goldscheider, N., Celico, F., Filippini, M., Ogata, K., & Gargini, A. (2020). Ecohydrogeology: The interdisciplinary convergence needed to improve the study and stewardship of springs and other groundwater-dependent habitats, biota, and ecosystems. Ecological Indicators, 110. https://doi.org/10.1016/j.ecolind.2019.105803
Chen, G., Singh, K. K., Lopez, J., & Zhou, Y. (2020). Tree canopy cover and carbon density are different proxy indicators for assessing the relationship between forest structure and urban socio-ecological conditions. Ecological Indicators, 113, 106279. https://doi.org/10.1016/J.ECOLIND.2020.106279
Chignell, S. M., Luizza, M. W., Skach, S., Young, N. E., & Evangelista, P. H. (2018). An integrative modeling approach to mapping wetlands and riparian areas in a heterogeneous Rocky Mountain watershed. Remote Sensing in Ecology and Conservation, 4(2), 150–165. https://doi.org/https://doi.org/10.1002/rse2.63
Chikofsky, E., & Cross II, J. H. (2002). Reverse Engineering. In Encyclopedia of Software Engineering. https://doi.org/https://doi.org/10.1002/0471028959.sof285
Copernicus Programme. (2023). CLC 2018 — Copernicus Land Monitoring Service. https://land.copernicus.eu/pan-european/corine-land-cover/clc2018
Cord, A. F., Brauman, K. A., Chaplin-Kramer, R., Huth, A., Ziv, G., & Seppelt, R. (2017). Priorities to Advance Monitoring of Ecosystem Services Using Earth Observation. Trends in Ecology & Evolution, 32(6), 416–428. https://doi.org/10.1016/j.tree.2017.03.003
Costa, H., Benevides, P., Moreira, F. D., Moraes, D., & Caetano, M. (2022). Spatially Stratified and Multi-Stage Approach for National Land Cover Mapping Based on Sentinel-2 Data and Expert Knowledge. Remote Sensing 2022, Vol. 14, Page 1865, 14(8), 1865. https://doi.org/10.3390/RS14081865
Crowther, T. W., Glick, H. B., Covey, K. R., Bettigole, C., Maynard, D. S., Thomas, S. M., Smith, J. R., Hintler, G., Duguid, M. C., Amatulli, G., Tuanmu, M. N., Jetz, W., Salas, C., Stam, C., Piotto, D., Tavani, R., Green, S., Bruce, G., Williams, S. J., … Bradford, M. A. (2015). Mapping tree density at a global scale. Nature, 525(7568), 201–205. https://doi.org/10.1038/nature14967
Czúcz, B., Keith, H., Driver, A., Jackson, B., Nicholson, E., & Maes, J. (2021). A common typology for ecosystem characteristics and ecosystem condition variables. One Ecosystem, 6, 1–16. https://doi.org/10.3897/oneeco.6.e58218
del Río-Mena, T., Willemen, L., Vrieling, A., & Nelson, A. (2023). How remote sensing choices influence ecosystem services monitoring and evaluation results of ecological restoration interventions. Ecosystem Services, 64. https://doi.org/10.1016/J.ECOSER.2023.101565
DeLancey, E. R., Kariyeva, J., Bried, J. T., & Hird, J. N. (2019). Large-scale probabilistic identification of boreal peatlands using Google Earth Engine, open-access satellite data, and machine learning. PLoS ONE, 14(6), 1–23. https://doi.org/10.1371/journal.pone.0218165
DGT. (2018). Technical specifications of the land use and land cover map for Portugal Mainland .
Dwire, K. A., Mellmann-Brown, S., & Gurrieri, J. T. (2018). Potential effects of climate change on riparian areas, wetlands, and groundwater-dependent ecosystems in the Blue Mountains, Oregon, USA. In Climate Services (Vol. 10, pp. 44–52). Elsevier B.V. https://doi.org/10.1016/j.cliser.2017.10.002
Edens, B., Maes, J., Hein, L., Obst, C., Siikamaki, J., Schenau, S., Javorsek, M., Chow, J., Chan, J. Y., Steurer, A., & Alfieri, A. (2022). Establishing the SEEA Ecosystem Accounting as a global standard. Ecosystem Services, 54, 101413. https://doi.org/10.1016/J.ECOSER.2022.101413
Ellison, D., Morris, C. E., Locatelli, B., Sheil, D., Cohen, J., Murdiyarso, D., Gutierrez, V., Noordwijk, M. van, Creed, I. F., Pokorny, J., Gaveau, D., Spracklen, D. V., Tobella, A. B., Ilstedt, U., Teuling, A. J., Gebrehiwot, S. G., Sands, D. C., Muys, B., Verbist, B., … Sullivan, C. A. (2017). Trees, forests and water: Cool insights for a hot world. Global Environmental Change, 43, 51–61. https://doi.org/10.1016/J.GLOENVCHA.2017.01.002
ESRI. (2023). ArcGIS Pro - ESRI (Environmental Systems Research Institute). https://www.arcgis.com/index.html
European Environment Agency. (2023). Copernicus Land Monitoring Service.
European Spatial Agency. (2023). ESA - Space Science. https://www.esa.int/Science_Exploration/Space_Science
Farrell, C., Coleman, L., Kelly-Quinn, M., Obst, C., Eigenraam, M., Norton, D., ODonoghue, C., Kinsella, S., Delargy, O., & Stout, J. (2021). Applying the System of Environmental Economic Accounting-Ecosystem Accounting (SEEA-EA) framework at catchment scale to develop ecosystem extent and condition accounts. One Ecosystem, 6. https://doi.org/10.3897/oneeco.6.e65582
Fick, S. E., & Hijmans, R. J. (2017). WorldClim 2: new 1-km spatial resolution climate surfaces for global land areas. International Journal of Climatology, 37(12), 4302–4315. https://doi.org/https://doi.org/10.1002/joc.5086
Fleming, A., O’Grady, A. P., Stitzlein, C., Ogilvy, S., Mendham, D., & Harrison, M. T. (2022). Improving acceptance of natural capital accounting in land use decision making: Barriers and opportunities. Ecological Economics, 200, 107510. https://doi.org/10.1016/J.ECOLECON.2022.107510
Fonseca, F., de Figueiredo, T., Vilela, Â., Santos, R., de Carvalho, A. L., Almeida, E., & Nunes, L. (2019). Impact of tree species replacement on carbon stocks in a Mediterranean mountain area, NE Portugal. Forest Ecology and Management, 439, 181–188. https://doi.org/https://doi.org/10.1016/j.foreco.2019.03.002
Gitelson, A. A., Gritz, Y., & Merzlyak, M. N. (2003). Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves. Journal of Plant Physiology, 160(3), 271–282. https://doi.org/10.1078/0176-1617-00887
Google. (2023). Google Earth Engine. https://earthengine.google.com/
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27. https://doi.org/https://doi.org/10.1016/j.rse.2017.06.031
Gosal, A. S., Evans, P. M., Bullock, J. M., Redhead, J., Charlton, M. B., Cord, A. F., Johnson, A., & Ziv, G. (2022). Understanding the accuracy of modelled changes in freshwater provision over time. Science of The Total Environment, 833, 155042. https://doi.org/10.1016/j.scitotenv.2022.155042
Harris, L. D. (1984). Island Biogeography Theory and the Preservation of Biotic Diversity. University of Chicago Press. https://doi.org/doi:10.7208/9780226219950
Hein, L., Bagstad, K. J., Obst, C., Edens, B., Schenau, S., Castillo, G., Soulard, F., Brown, C., Driver, A., Bordt, M., Steurer, A., Harris, R., & Caparrós, A. (2020). Progress in natural capital accounting for ecosystems. Science, 367(6477), 514–515. https://doi.org/10.1126/science.aaz8901
Hose, G. C., Dabovic, J., Nelson, T., Pollitt, L., & Korbel, K. L. (2021). Groundwater Dependent Aquatic and Terrestrial Ecosystems. Reference Module in Earth Systems and Environmental Sciences. https://doi.org/10.1016/B978-0-12-819166-8.00118-3
Hosmer, D. W., & Lemeshow, S. (2000). Interpretation of the Fitted Logistic Regression Model. In Applied Logistic Regression (pp. 47–90). Wiley. https://doi.org/10.1002/0471722146.ch3
Hossain, M. S., & Hashim, M. (2019). Potential of Earth Observation (EO) technologies for seagrass ecosystem service assessments. International Journal of Applied Earth Observation and Geoinformation, 77, 15–29. https://doi.org/10.1016/j.jag.2018.12.009
Huang, F., Chunyu, X., Zhang, D., Chen, X., & Ochoa, C. G. (2020). A framework to assess the impact of ecological water conveyance on groundwater-dependent terrestrial ecosystems in arid inland river basins. Science of the Total Environment, 709. https://doi.org/10.1016/j.scitotenv.2019.136155
Huete, A. R. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25(3), 295–309. https://doi.org/10.1016/0034-4257(88)90106-X
Huete, A. R., Liu, H. Q., & van Leeuwen, W. J. D. (1997). Use of vegetation indices in forested regions: Issues of linearity and saturation. International Geoscience and Remote Sensing Symposium (IGARSS), 4, 1966–1968. https://doi.org/10.1109/IGARSS.1997.609169
Jagannathan, J., & Divya, C. (2021). Deep learning for the prediction and classification of land use and land cover changes using deep convolutional neural network. Ecological Informatics, 65, 101412. https://doi.org/10.1016/J.ECOINF.2021.101412
JAXA. (2023). ALOS World 3D - 30m. https://www.eorc.jaxa.jp/ALOS/en/dataset/aw3d30/aw3d30_e.htm
Ji, L., Geng, X., Sun, K., Zhao, Y., & Gong, P. (2015). Target Detection Method for Water Mapping Using Landsat 8 OLI/TIRS Imagery. Water, 7(2), 794–817. https://doi.org/10.3390/w7020794
Keith, H., Czúcz, B., Jackson, B., Driver, A., Nicholson, E., & Maes, J. (2020). A conceptual framework and practical structure for implementing ecosystem condition accounts. One Ecosystem, 5, 1–54. https://doi.org/10.3897/oneeco.5.e58216
Keith, H., Vardon, M., Obst, C., Young, V., Houghton, R., & Mackey, B. (2021). Evaluating nature-based solutions for climate mitigation and conservation requires comprehensive carbon accounting. Science of The Total Environment, 769, 144341. https://doi.org/10.1016/j.scitotenv.2020.144341
Khalid, H. W., Khalil, R. M. Z., & Qureshi, M. A. (2021). Evaluating spectral indices for water bodies extraction in western Tibetan Plateau. The Egyptian Journal of Remote Sensing and Space Science, 24(3), 619–634. https://doi.org/10.1016/J.EJRS.2021.09.003
Kienast, F., Bolliger, J., Potschin, M., de Groot, R. S., Verburg, P. H., Heller, I., Wascher, D., & Haines-Young, R. (2009). Assessing Landscape Functions with Broad-Scale Environmental Data: Insights Gained from a Prototype Development for Europe. Environmental Management, 44(6), 1099–1120. https://doi.org/10.1007/s00267-009-9384-7
Kløve, B., Ala-Aho, P., Bertrand, G., Gurdak, J. J., Kupfersberger, H., Kværner, J., Muotka, T., Mykrä, H., Preda, E., Rossi, P., Uvo, C. B., Velasco, E., & Pulido-Velazquez, M. (2014). Climate change impacts on groundwater and dependent ecosystems. Journal of Hydrology, 518(PB), 250–266. https://doi.org/10.1016/J.JHYDROL.2013.06.037
Koit, O., Tarros, S., Pärn, J., Küttim, M., Abreldaal, P., Sisask, K., Vainu, M., Terasmaa, J., Retike, I., & Polikarpus, M. (2021). Contribution of local factors to the status of a groundwater dependent terrestrial ecosystem in the transboundary Gauja-Koiva River basin, North-Eastern Europe. Journal of Hydrology, 600. https://doi.org/10.1016/j.jhydrol.2021.126656
Koma, Z., Seijmonsbergen, A. C., & Kissling, W. D. (2021). Classifying wetland-related land cover types and habitats using fine-scale lidar metrics derived from country-wide Airborne Laser Scanning. Remote Sensing in Ecology and Conservation, 7(1), 80–96. https://doi.org/https://doi.org/10.1002/rse2.170
Kulawardhana, R. W. (2011). Remote sensing of vegetation: principles, techniques and applications. By Hamlyn G. Jones and Robin A Vaughan. Journal of Vegetation Science, 22(6), 1151–1153. https://doi.org/https://doi.org/10.1111/j.1654-1103.2011.01319.x
Lange, S., Campagne, C. S., Comte, A., Bank, E., Santos-Martín, F., Maes, J., & Burkhard, B. (2022). Progress on ecosystem accounting in Europe. Ecosystem Services, 57, 101473. https://doi.org/10.1016/J.ECOSER.2022.101473
Liquete, C., Maes, J., Notte, A., & Bidoglio, G. (2011). Securing water as a resource for society: an ecosystem services perspective. Ecohydrology & Hydrobiology, 11(3–4), 247–259. https://doi.org/10.2478/V10104-011-0044-1
Ludwig, C., Walli, A., Schleicher, C., Weichselbaum, J., & Riffler, M. (2019). A highly automated algorithm for wetland detection using multi-temporal optical satellite data. Remote Sensing of Environment, 224, 333–351. https://doi.org/10.1016/J.RSE.2019.01.017
Lundqvist, J., & Unver, O. (2018). Alternative pathways to food security and nutrition – water predicaments and human behavior. Water Policy, 20(5), 871–884. https://doi.org/10.2166/WP.2018.171
Madrigal-González, J., Calatayud, J., Ballesteros-Cánovas, J. A., Escudero, A., Cayuela, L., Marqués, L., Rueda, M., Ruiz-Benito, P., Herrero, A., Aponte, C., Sagardia, R., Plumptre, A. J., Dupire, S., Espinosa, C. I., Tutubalina, O. V., Myint, M., Pataro, L., López-Sáez, J., Macía, M. J., … Stoffel, M. (2023). Global patterns of tree density are contingent upon local determinants in the world’s natural forests. Communications Biology 2023 6:1, 6(1), 1–6. https://doi.org/10.1038/s42003-023-04419-8
Mairota, P., Cafarelli, B., Didham, R. K., Lovergine, F. P., Lucas, R. M., Nagendra, H., Rocchini, D., & Tarantino, C. (2015). Challenges and opportunities in harnessing satellite remote-sensing for biodiversity monitoring. Ecological Informatics, 30, 207–214. https://doi.org/10.1016/j.ecoinf.2015.08.006
Mengist, W., Soromessa, T., & Feyisa, G. L. (2021). Monitoring Afromontane forest cover loss and the associated socio-ecological drivers in Kaffa biosphere reserve, Ethiopia. Trees, Forests and People, 6. https://doi.org/10.1016/j.tfp.2021.100161
Mpakairi, K. S., Dube, T., Dondofema, F., & Dalu, T. (2022). Spatio–temporal variation of vegetation heterogeneity in groundwater dependent ecosystems within arid environments. Ecological Informatics, 69, 101667. https://doi.org/10.1016/j.ecoinf.2022.101667
Mulatu, K. A., Mora, B., Kooistra, L., & Herold, M. (2017). Biodiversity monitoring in changing tropical forests: A review of approaches and new opportunities. In Remote Sensing (Vol. 9, Issue 10). MDPI AG. https://doi.org/10.3390/rs9101059
Newbold, T., Hudson, L. N., Hill, S. L. L., Contu, S., Lysenko, I., Senior, R. A., Börger, L., Bennett, D. J., Choimes, A., Collen, B., Day, J., De Palma, A., Díaz, S., Echeverria-Londoño, S., Edgar, M. J., Feldman, A., Garon, M., Harrison, M. L. K., Alhusseini, T., … Purvis, A. (2015). Global effects of land use on local terrestrial biodiversity. Nature, 520(7545), 45–50. https://doi.org/10.1038/nature14324
Nieves-Chinchilla, J., Martínez, R., Farjas, M., Tubio-Pardavila, R., Cruz, D., & Gallego, M. (2018). Reverse engineering techniques to optimize facility location of satellite ground stations on building roofs. Automation in Construction, 90, 156–165. https://doi.org/10.1016/J.AUTCON.2018.02.019
Oliveira, M., Novo, M. emilia, & Lobo-Ferreira, J.-P. (2007). Models to predict the impact of the climate changes on aquifer recharge. Water in Celtic Countries: Quantity, Quality and Climate Variability (Proceedings of the Fourth InterCeltic Colloquium on Hydrology and Management of Water Resources, 103–110. https://iahs.info/uploads/dms/13794.15-103-110-05-IC310-23-Oliveira.pdf
Ramirez-Reyes, C., Brauman, K. A., Chaplin-Kramer, R., Galford, G. L., Adamo, S. B., Anderson, C. B., Anderson, C., Allington, G. R. H., Bagstad, K. J., Coe, M. T., Cord, A. F., Dee, L. E., Gould, R. K., Jain, M., Kowal, V. A., Muller-Karger, F. E., Norriss, J., Potapov, P., Qiu, J., … Wright, T. M. (2019). Reimagining the potential of Earth observations for ecosystem service assessments. Science of The Total Environment, 665, 1053–1063. https://doi.org/10.1016/j.scitotenv.2019.02.150
Rijal, S., Rimal, B., Acharya, R. P., & Stork, N. E. (2021). Land use/land cover change and ecosystem services in the Bagmati River Basin, Nepal. Environmental Monitoring and Assessment, 193(10), 651. https://doi.org/10.1007/s10661-021-09441-z
Rouse, J., Haas, R., Schell J., & Deering D. (1973). Monitoring Vegetation Systems in the Great Plains with ERTS. Third ERTS Symposium, NASA.
Santos, F. L. M., Couto, F. T., Dias, S. S., Ribeiro, N. A., & Salgado, R. (2023). Vegetation fuel characterization using machine learning approach over southern Portugal. Remote Sensing Applications: Society and Environment, 101017. https://doi.org/10.1016/J.RSASE.2023.101017
Seddon, N., Smith, A., Smith, P., Key, I., Chausson, A., Girardin, C., House, J., Srivastava, S., & Turner, B. (2021). Getting the message right on nature-based solutions to climate change. Global Change Biology, 27(8), 1518–1546. https://doi.org/10.1111/GCB.15513
Tabacchi, G., Di Cosmo, L., & Gasparini, P. (2011). Aboveground tree volume and phytomass prediction equations for forest species in Italy. European Journal of Forest Research, 130(6), 911–934. https://doi.org/10.1007/S10342-011-0481-9/TABLES/8
United Nations et al. (2021). System of Environmental-Economic Accounting-Ecosystem Accounting (SEEA-EA). https://seea.un.org/ecosystem-accounting.
Vallero, D. (2014). Grand Challenges. Fundamentals of Air Pollution, 953–961. https://doi.org/10.1016/B978-0-12-401733-7.00033-5
Vargas, L., Willemen, L., & Hein, L. (2019). Assessing the Capacity of Ecosystems to Supply Ecosystem Services Using Remote Sensing and An Ecosystem Accounting Approach. Environmental Management, 63(1), 1–15. https://doi.org/10.1007/s00267-018-1110-x
Wang, J.-F., Zhang, T.-L., & Fu, B.-J. (2016). A measure of spatial stratified heterogeneity. Ecological Indicators, 67, 250–256. https://doi.org/10.1016/j.ecolind.2016.02.052
Willcock, S., Martínez-López, J., Hooftman, D. A. P., Bagstad, K. J., Balbi, S., Marzo, A., Prato, C., Sciandrello, S., Signorello, G., Voigt, B., Villa, F., Bullock, J. M., & Athanasiadis, I. N. (2018). Machine learning for ecosystem services. Ecosystem Services, 33, 165–174. https://doi.org/10.1016/J.ECOSER.2018.04.004
Wood, W. H. (2009). Computational Representations of Function in Engineering Design. Philosophy of Technology and Engineering Sciences, 543–564. https://doi.org/10.1016/B978-0-444-51667-1.50024-0
Yang, J., Wang, J., Liao, X., Tao, H., & Li, Y. (2022). Chain modeling for the biogeochemical nexus of cadmium in soil-rice-human health system. Environment International, 167, 107424. https://doi.org/10.1016/j.envint.2022.107424
Yang, Z., Bai, J., & Zhang, W. (2021). Mapping and assessment of wetland conditions by using remote sensing images and POI data. Ecological Indicators, 127, 107485. https://doi.org/10.1016/J.ECOLIND.2021.107485
Zomer, R. J., Xu, J., & Trabucco, A. (2022). Version 3 of the Global Aridity Index and Potential Evapotranspiration Database. Scientific Data, 9(1), 409. https://doi.org/10.1038/s41597-022-01493-1

No competing interests reported.

SupplementaryMaterials.docx

Download PDF

Reviews received at journal
23 Aug, 2024
Reviewers agreed at journal
19 Aug, 2024
Reviewers invited by journal
16 Aug, 2024
Editor assigned by journal
06 Aug, 2024
Submission checks completed at journal
06 Aug, 2024
First submitted to journal
03 Jul, 2024

You are reading this latest preprint version

Model	Residuals	Tr	Te	Tl
FCD	\(\:y=0.32x-15.76\)	\(\:y=0.74x+13.68\)	\(\:y=0.54x+22.83\)	\(\:y=0.32x+24.62\)
FWPI	\(\:y=0.36x-\:19.82\)	\(\:y=0.71x+17.9\)	\(\:y=0.55x+24.71\)	\(\:y=0.6x+24.21\)

Climate-driven stock accounts of inland Green-Blue Ecosystems: Coupling the carbon and water cycles via Reverse Engineering and Geodetectors

Status:

Version 1

Abstract

Figures

Highlights

1. Introduction

2. Materials and Methods

2.1. Study Area

2.2. Data

2.3. Methods

2.3.1. Classification Task

2.3.2. Regression Task

2.3.3. Factor Analysis using Geodetectors

3. Results

3.1. Forest, freshwater and wetlands mapping

3.2. FCD and FWPI predictions

3.3. Factor analysis using geodetectors

3.4. Inland GBE physical accounting: extent and condition

4. Discussion

4.1. Cloud-based automated classification of inland GBE

4.2. From Reverse Engineering to satellite-based climate-driven models

4.3. Implications for future research

5. Conclusion

Declarations

Funding

Author Contribution

Acknowledgement

References

Additional Declarations

Supplementary Files

Status:

Version 1