Predicting Abiotic Soil Characteristics Using Sentinel-2 at Nature-Management-Relevant Spatial Scales and Extents

Abstract

Knowledge of local plant community characteristics is imperative for practical nature planning and management, and for understanding plant diversity and distribution drivers. Today, retrieving such data is only possible by fieldwork and is hence costly both in time and money. Here, we used nine bands from multispectral high-to-medium resolution (10–60 m) satellite data (Sentinel-2) and machine learning to predict local vegetation plot characteristics over a broad area (approx. 30,000 km²) in terms of plants’ preferences for soil moisture, soil fertility, and pH, mirroring the levels of the corresponding actual soil factors. These factors are believed to be among the most important for local plant community composition. Our results showed that there are clear links between the Sentinel-2 data and plants’ abiotic soil preferences, and using solely satellite data we achieved predictive powers between 26 and 59%, improving to around 70% when habitat information was included as a predictor. This shows that plants’ abiotic soil preferences can be detected quite well from space, but also that retrieving soil characteristics using satellites is complicated and that perfect detection of soil conditions using remote sensing—if at all possible—needs further methodological and data development.

1. Introduction

Knowledge of local plant community characteristics is imperative for practical nature planning and management, and for understanding local-scale plant diversity and distribution drivers. Today, such data of a high enough quality for these purposes can only be derived through targeted field campaigns, which are costly both in time and money. Here, we use multispectral high-to-medium resolution (10–60 m) satellite data [1] to predict the local composition of plant species over a large area (approx. 30,000 km²) in terms of their preference for soil moisture, soil fertility, and pH, which are among the most important factors for local plant community composition [2,3].

In recent years, massive vegetation plot databases providing large amounts of data have enabled studies of local plant diversity at national and continental scales [4,5,6]. Combining these with indicator values such as Ellenberg’s opens avenues for novel insights into how local plant communities are distributed across broad spatial scales [7,8]. In Denmark, the monitoring of protected natural and semi-natural habitats (protected under the EU Habitats Directive Annex I, i.e., the nature deemed the most important to the EU) generates a significant amount of vegetation plot data each year [9,10,11].

At the same time, remote sensing techniques are slowly maturing and are approaching a stage where they can be used to predict local plant diversity over large areas [12,13]. For example, we now have access to relatively fine-resolution (decameters) satellite-borne spectral data at the global scale [1,14]. Using these data, many studies have attempted to predict ecological factors like plant species richness, plant phenology, plant traits and habitat characteristics [4,12,15,16]. However, most of these studies have attempted this only at local scale or involving only one or a few species, traits, or habitat types, preventing broad-scale generalization. At the same time, no study has succeeded in providing sufficiently detailed insight into a given area based solely on remotely sensed information to use remote sensing as a stand-alone tool for the monitoring and management of nature. This is likely related to the fact that retrieving some of the most important abiotic factors for plant species distribution—namely soil chemical factors such as soil fertility, soil moisture, and pH [17]—with remote sensing is notoriously hard, as most remote sensing signals do not penetrate into the soil and hence do not gather data directly from the soil.

Predicting soil moisture, fertility and pH using remotely sensed data in nature areas has been attempted several times, but often at too coarse a spatial resolution to be relevant for practical nature management [18] or across fairly small spatial extent, preventing the generalization of results [19]. The hypothesis behind using multispectral satellite imagery to generate such predictions is that soil chemistry is thought to affect the plants’ structure and leaf chemistry, causing differences in the reflectance of plants depending on the underlying soil chemistry and moisture [20,21] and hence the plants can act as indicators of soil chemical properties. For example, Schmidtlein (2005) and Möckel et al. (2016) [22,23] used imaging spectroscopy to map these factors at a high spatial resolution, but the study areas were relatively small. On the other hand, studies combining a relatively fine resolution, and large areas are starting to improve our ability to predict these important soil factors at relevant scales and extents for nature monitoring and management [24].

Here, we take the next major step to bridge this gap by using plant community data from >50,000 small (5 m radius circular) vegetation plots with indicator values for soil moisture, soil fertility, and pH, combined with current multispectral satellite data to show how this can be used to predict some of the most important abiotic soil factors for local plant diversity and distribution—soil moisture, soil fertility, and pH—with sufficient detail to be relevant for nature monitoring and management. More specifically, we investigate (1) if nine bands of Sentinel-2 data are linked to plants’ preference for soil moisture, fertility, and pH, and if so (2) how well we can predict these plant-indicated soil factors using current state-of-the-art remote sensing and modeling techniques.

2. Materials and Methods

2.1. Vegetation Data

In this study, we used vegetation data from the Danish national program for the monitoring of nature (NOVANA, freely available from https://naturdata.miljoeportal.dk [9,10,11], accessed on 16 September 2016). We extracted higher-plant species lists for 97,334 vegetation plots (5 m diameter radius). The plots were randomly placed within sites belonging to 32 different habitat types (Figure S1), all of them listed in the EU Habitats Directive’s Annex I [25]. These natural and semi-natural habitats range from grasslands, over heathlands, dunes, and shrublands to wetlands, but in this analysis, we excluded forests (Figure S1). The plot data were collected by trained field botanists from 2004 to 2015 and cover all parts of Denmark (Figure 1, but here only the plots used for this study are shown). For analysis, we used 58,071 plots of which 31,001 were unique, meaning that some plots were revisited several times over the years (See Figure S2). However, since the vegetation plots were georeferenced in the field using traditional GPS having a vertical uncertainty of 5–10 m, revisits were not exactly at the same place at every revisit. The exact GPS models used differed depending on inventory year and region.

To represent the three main drivers of local plant diversity, soil moisture, fertility, and pH, in each plot, we calculated—based on the plant species list for each plot—the mean Ellenberg indicator values (EIVs [21]) for soil moisture (EIV_F), soil nutrients (EIV_N), and soil reaction (EIV_R). Most European plant species are assigned an indicator value for each of the soil factors mentioned above, see [21]. These indicator values represent the plants’ preferences for soil moisture, the fertility of the soil (productivity of the plants), and how calcareous the soil is (pH), respectively [21,26]. The soil moisture EIVs are in the interval 1 (very drought-tolerant species) to 12 (submersed water plants). The soil nutrient indicator values are in the interval from 1 (indicating plants with low productivity, typically adapted to growing in very nutrient-poor sites) to 9 (plants that are highly productive, typically found in very nutrient-rich sites). The soil reaction indicator values are in the same interval, with 1 indicating plants adapted to growing in strongly acidic soil (actual pH 3.5–4) and 9 indicating plants living almost exclusively in calcareous soils (actual pH 7.5–8) [21,26]. Earlier studies have shown that these EIVs are typically rather tightly correlated with corresponding measured abiotic variables [8,26]. Soil nutrients and pH are often highly correlated, so to represent the nutrient availability independently from pH, we also calculated the EIV_N/EIV_R ratio [27], referred to as the N/R ratio in the rest of this paper. The N/R ratio is thought to represent the adaptation of plants to grow in nutrient-poor or rich sites independently from what the soil pH is [27]. Not all species had an Ellenberg indicator value—in this case, they were not included in the mean calculation.

Each plot also holds information on which habitat type was present. This information is based on a preliminary assignment in the field and a subsequent verification process involving the species found in the plot [11].

2.2. Remote Sensing Data

In this study, we used multispectral imagery from the European Space Agency Sentinel-2 mission [14]. To match the plants’ growing season, we retrieved all images overlapping the study area from 1 June to 31 August 2016 through the Copernicus Open Access Hub (https://scihub.copernicus.eu, accessed on 18 May 2018). The total dataset consisted of 38 image products (see Appendix A for specific product names). To obtain surface reflectance, all images were atmospherically corrected using the MAJA ver. 1.0 software tool [28] with default settings. Unlike most other software tools for satellite imagery processing, MAJA takes advantage of using imagery time series to estimate cloud mask and aerosol optical thickness. Thanks to the use of this multi-temporal information, MAJA cloud masks are highly reliable [29] and the accuracy of the surface reflectance estimates computed is among the best when applied to Sentinel-2 [30,31]. This processing was conducted in a Docker installation of CentOS version 7.4.

Table 1. Overview of the input data used for modelling. Res.: resolution, Quant.: quantity.

	Type	Res.	Source	Period	Processing	Quant.	Link
Vegetation plot data	Field-based 5 m radius circular plots with plant species and habitat data	Point data	The Danish monitoring program for habitats in Annex 1 of the EU Habitats Directive	2004–2015	For each plot: calculation of mean Ellenberg indicator values based on the plant species	58,071 plots	https://naturdata.miljoeportal.dk/, accessed on 16 September 2016
Sentinel-2 data	Satellite-based remote sensing	10–60 m	European Space Agency	Summer 2016	Atmospheric correction (see main text)	38 scenes	https://scihub.copernicus.eu/, accessed on 7–18 May 2018

gives an overview of both the vegetation and the remote sensing data used for modelling (see Section 2.3 and Section 2.4).

2.3. Preprocessing

To provide a basis for testing how well Sentinel-2 imagery can predict local characteristics (i.e., the EIVs), we extracted the raw values for the band numbers 2, 3, 4, 5, 6, 8, 8a, 11, and 12 (after atmospheric correction) corresponding to the position of each vegetation plot. Due to missing coverage or clouds that overlapped the vegetation plots, we decided to use the satellite data from one single cloud-free day (16 August 2016) as this resulted in the largest possible dataset with 58,071 vegetation plots having band data from the Sentinel-2 images and the above-mentioned average Ellenberg indicator values. For this day, satellite data were not available for the whole country and therefore our study is only based on plots from approximately two-thirds of the country (Jutland, Figure 1). However, the data do cover all main soil and habitat types despite this (except for rock habitats that constitute a tiny part of the Danish landscape), so the model is probably quite broadly applicable both to the rest of Denmark and to other parts of the north-European temperate region.

2.4. Statistical Analysis

To test if the Sentinel-2 data were linked to the abiotic environment in the plots (the EIVs were used as response variables), we used supervised learning AI algorithms (see below) on a training data set consisting of 30,000 randomly selected vegetation plots of the total 58,071 plots covering the 32 different habitat types, found in Figure S1. For validation, we predict model estimates in the remaining vegetation plots (n = 28,017) and report the predictive power (R² and standard deviation) based on that in

Table 2. Modeling results and characteristics for each response (columns) and each selected model (rows). “Satellite” denotes that all 9 bands (see main text) from the satellite data were used as explanatory variables. “Habitat” denotes that only habitat was used as a categorical explanatory variable. F, N, and R: EIVs for plants’ preferences for soil moisture, nutrients, and reaction (pH), respectively. Std.: standard deviation, GBT: gradient-boosted trees, RF: random forest, DT: decision tree, NN: nearest neighbors, LR: Linear regression.

	F			N			R			N/R
	Std.	R2	Model	Std.	R2	Model	Std.	R2	Model	Std.	R2	Model
Satellite	1.34	0.26	GBT	1.09	0.59	RF	0.93	0.54	RF	0.16	0.29	RF
Habitat	1.23	0.36	GBT	1.40	0.05	DT	1.29	-	NN	0.18	-	LR
Satellite + habitat	0.81	0.73	RF	0.92	0.70	DT	0.71	0.73	RF	0.16	0.23	GBT

.

The nine intensity values from Sentinel-2 were treated as a vector with 9 numeric elements. Using the raw intensity values instead of the traditional indices like NDVI (see e.g., reference [32]) has previously been shown to work well for this kind of modeling [23]. To investigate if the links between the imagery and the EIVs were influenced by habitat type, this information was also included as a predictor (categorical variable).

To make sure we obtained the best model, we let the function “Predict” in Wolfram Mathematica version 14 automatically select (default settings) which machine learning algorithm provided the best predictions for each EIV and for satellite data only, habitats only, and these two data sources combined, respectively (Table 2). The “random forest” algorithm was the most frequently selected method, but also the algorithms “decision trees” and “nearest neighbor” were found to provide the best predictions in some cases (Table 2).

To investigate the within-plot, among-year variation in the calculated Ellenberg indicator values, the indicator values were analyzed in a mixed linear model with habitat type as a fixed factor and plot as a random factor, assuming that the residual variation was normally distributed. The R (ver. 4.3.2) procedure lmer (ver. 1.1-35.1) in the package lme4 [33] was used for this part of the analysis.

3. Results

Overall, we found that the Sentinel-2 imagery explained a significant part of the variation in the plot-level abiotic environment (Figure 2, Table 2), with >26, 59 and 54% of the variation explained in the soil moisture, fertility, and pH levels indicated by the plant species composition, respectively. This means that satellite data are linked to and can indeed be used to predict plant-indicated local-scale soil moisture, nutrient levels, and pH across numerous habitat types with some uncertainty. We also found that combining satellite data with information about the plots’ habitat type considerably improved this predictive power (over 70% of variation explained in all cases mentioned above), but also that habitat—when used alone as a predictor—consistently had no or lower predictive power compared to using satellite data alone (Table 2). We found significant within-plot, among-year variation in the calculated Ellenberg indicator values (Table S1). The result details can be seen in Figure 2 and Table 2. For examples showing the actual vs. the predicted values in selected sites, see Figure 3.

4. Discussion

4.1. Soil Fertility and pH

Earlier studies have been able to explain around 35–40% of the variation in plant-indicated soil fertility using remotely sensed spectral data [32,34]. However, we were able to explain almost 60%, resembling the findings by Möckel et al. 2016 [23], who reached a similar result. Both Möckel et al. (2016) [23] and we tried to utilize the information in all available spectral bands, whereas previous studies have typically used NDVI—which is only based on two bands—as the only spectral predictor. This suggests that important information for predicting soil fertility is present not only in the red and infrared bands but also in other bands. We therefore recommend that this full suite of information (i.e., all spectral bands) is used in future studies and applications of multi- and hyperspectral data when studying soil fertility indicated by plants. In our study, the nutrient ratio was predicted less well. This parameter integrates both nutrient and soil reaction and hence indicates the actual soil fertility independently from soil reaction, and this is often important in ecological questions as it mirrors the actual nutrients available for plant growth [27]. Consequently, this result means that it can be quite complicated to use remotely sensed spectral data to gain insight into plant-relevant soil fertility dynamics at local scale. That said, our relatively good ability to predict soil nutrient status (i.e., only EIV_N) is indeed encouraging in this respect, and we do believe there is a potential for further development of methods that could improve our results (see “Perspectives”). When adding habitat information to our models, the predictive power for both soil fertility and pH rose to 70% or more, indicating that knowledge of habitats further refines our models and should be used whenever available.

4.2. Soil Moisture

In contrast to our soil fertility and pH (EIV_N and EIV_R) results, we achieved poorer results when it comes to predicting soil moisture than earlier studies (R² = 29%). For example, both Möckel et al. (2016) [23], Weber et al. (2018) [34], and Löfgren et al. (2018) [32] were able to explain >50% of the variation in soil moisture indicated by plants. However, their studies were all conducted solely in dry grasslands, and this could explain this difference: in dry grasslands the soil moisture gradient is quite short compared to when all the different habitat types are considered, from bogs and fens to fresh- and saltwater meadows to moist and dry grasslands, as we do here. We believe that this longer moisture gradient causes the spectral signals to vary much more independently of the soil moisture. For example, it seems intuitive that in dry grasslands, the plants in moister areas are greener than those in drier areas. In other habitats, there may not be this difference in the appearance of the plants, which could explain our lower explanatory power. Hence, for endeavors to remotely sense soil moisture it may be necessary to include other techniques, such as lidar-based topographical wetness indices [35]. Also, our modeling highlighted the importance of having access to habitat type data to make good predictions of soil moisture from spectral data.

4.3. Limitations and Uncertainties

Our models had lower predictive power than what would be acceptable for everyday nature management and monitoring. Following from that, our approach does have limitations when it comes to practical applications, as our models may not be precise enough for such uses. On the other hand, plant indicator values themselves are also uncertain and do not necessarily mirror abiotic soil conditions one to one [36,37,38], and naturally, predicting uncertain factors will inevitably introduce some uncertainty into the results. That said, plant indicators uniquely integrate abiotic conditions over time and space, unlike any other data source [8], and therefore our ability to predict them using remotely sensed data is crucial for developing future digital nature mapping and monitoring solutions.

Our models were constructed using satellite imagery captured on a single day. As a result, we are missing information on how the spectral signal changes throughout the year, and the spectral data may not necessarily correspond to the recording year of the floristic data from the vegetation plots. Therefore, there is a risk that our models are overfitted to the data. For this reason, we recommend that future studies test these models on data collected over multiple days across several years before practical applications, to ensure they are not overfitted to data from a specific day. This will pinpoint issues with generality and ensure that models are sufficiently general for applications under day-to-day and year-to-year varying phenology and long-term weather conditions (e.g., moist and dry summers) that are known to strongly influence soil characteristics.

The within-plot, among-year variation in the EIVs is probably partly due to the GPS uncertainty, rendering it impossible to find the same position when revisiting vegetation plots. It could also be due to changes in plant community composition over the years. Considered together with the fact that our satellite data only capture a glimpse of conditions in time (see above), this means that we cannot expect perfect accuracy of the local predictions of our models.

Notably for soil moisture, but also for pH (EIV_F and EIV_R), our models’ predictive power strongly improved by including habitat type as an explanatory variable. This probably means that the abiotic and biotic information represented by the habitat type is important to make good models of soil abiotic conditions, and at least for some habitat types obtaining this information from remote sensing is notoriously hard [39].

4.4. Perspectives

For soil nutrients (EIV_N) and pH (EIV_R), we achieved considerable predictive power (R²: 0.5–0.6) using satellite data alone. This means that our method could have potential for further development, e.g., by using finer resolution data (both spectrally and spatially), for example from drones or airplanes [40,41]. Several of our models were machine learning-based (random forests or gradient-boosted regression trees). Another possible future improvement could be to try implementing deep learning-based methods, as they may be better at capturing patterns in the data [42].

Including habitat type improved our models significantly. So, being able to remotely sense this information and feed it into, for example, models like ours in the future would be clearly desirable. Without knowledge of habitat type, models like the ones we developed here will have poorer performance and their results will be less useful in practical applications. Currently, researchers and technicians from several different parts of the World are working on such a remote-sensing-based classification of habitat types over large areas [43,44,45]. While these models are still far from perfect, they are slowly improving, and this could possibly enhance our ability to improve models like the ones we developed here.

While satellite data often suffer from poor resolution compared to drone or lidar data, they have a clear strength in that they offer time series data (i.e., repeated data capture of the same area). This enables developing models that can predict changes in soil characteristics over time. Such change detection models may encounter similar errors to those we have demonstrated here, but we see significant potential in these models and strongly encourage researchers to test and further explore this research direction in the future.