Harvesting spatially dense legacy soil datasets for digital soil mapping of available water capacity in Southern France

doi:10.1016/j.geodrs.2020.e00353

Geoderma Regional

Volume 24, March 2021, e00353

https://doi.org/10.1016/j.geodrs.2020.e00353 Get rights and content

Highlights

•
We used dense samplings of legacy soil observations to map Soil Available Water Capacity (SAWC) over an irrigated perimeter.
•
The best results were obtained with Quantile Random Forest (QRF) using geographical locations as additional soil covariates
•
Feeding QRF with 2781 additional auger hole observations greatly increased the accuracy and resolution of the SAWC map
•
QRF performances were however limited by the observational uncertainties of the SAWCs determined from auger holes
•
Extending SAWC maps to the whole irrigated perimeter would need to automatize the digitalization of auger hole observations.

Abstract

Although considerable work has been conducted in recent decades to build soil databases, the legacy data from a lot of former soil survey campaigns still remain unused. The objective of this study was to determine the interest in harvesting such legacy data for mapping the soil available water capacities (SAWCs) at different rooting depths (30 cm, 60 cm, 100 cm) and to the maximal observation depth, over the commune of Bouillargues (16 km², Occitanie region, southern France).

An increasing number of available auger hole observations with SAWC estimations – from 0 to 2781 observations – were added to the existing soil profiles to calibrate quantile regression forests (QRFs) using the Euclidean buffer distances from the sites as soil covariates. The SAWC was first mapped separately for different soil layers, and the mapping outputs were pooled to estimate the required SAWC. The uncertainty of the SAWC prediction was estimated from the estimated mapping uncertainties of the individual soil layers by an error propagation model using a first-order Taylor analysis.

The performances of the SAWC predictions and their uncertainties were evaluated with a 10-fold cross validation that was iterated 20 times. The results showed that the use of a quantile regression forest that was fed with auger hole observations and that used the Euclidean buffer distances as soil covariates considerably augmented the performances of the SAWC predictions (percentages of explained variance from 0.39 to 0.70) compared to the performance of a classical DSM approach, i.e., a QRF that solely used soil profiles and only environmental covariates (percentages of explained variance from 0.04 to 0.51). The analysis of the results revealed that the performances were also dependent on the spatial patterns of the different examined SAWCs and was limited by the observational uncertainties of the SAWCs determined from auger holes. The best performance tended to also provide the best view of the uncertainty patterns with an overestimation of uncertainty.

Despite these gains in performance, the cost-efficiency analysis showed that the augmentation of soil observations was not cost efficient because of the highly time-consuming manual data harvesting protocol. However, this result did not account for the observed gain in map details. Furthermore, the cost efficiency could be further improved by automation.

Introduction

Digital soil mapping (DSM) has been recognized as the appropriate solution to provide spatial soil information for land users, scientist communities and policy and decision makers in agriculture and the environment (McBratney et al., 2003; Sanchez et al., 2009). The principle of DSM is to predict a soil property or soil classes and the associated prediction uncertainty by determining the quantitative relationships between the soil information available over a limited set of locations and the spatial data reflecting the state factors of soil formation (environmental covariates). DSM has now moved from a largely academic movement toward an operational activity (Minasny and McBratney, 2016, Arrouays et al., 2017a,b).

However, the performances of DSM predictions of soil properties often exhibit more uncertainty than initially expected. For example, the percentages of explained variances of less than 0.5 were observed for 95%, 76%, 100% and 86% of the tested soil properties for DSM applications at the catchment scale (Nussbaum et al., 2018), at the regional scale (Vaysse and Lagacherie, 2015), at the national scale (Mulder et al., 2016), and at the global scale (Hengl et al., 2014), respectively.

These authors converged toward the conclusion that the density of soil observations used for calibrating the DSM models was the main factor that limited the DSM performances. Most of the soil information used as input in DSM applications has been either soil maps or the spatial sampling of sites with soil property measurements. The average densities used in most operational DSM applications have been low, e.g., 4–12 sites/km² (several study areas in Nussbaum et al., 2018), 0.07 sites/km² (Vaysse and Lagacherie, 2015), 0.03 sites/km² (Mulder et al., 2016), and 0.001 sites/km² (Hengl et al., 2014), which limits the performances of soil prediction, especially when the pattern of variation in the soil property is largely below the spacing of soil profiles (Vaysse and Lagacherie, 2015; Gomez and Coulouma, 2018). In addition, further experiments that consisted of varying the spatial density of soil input confirmed this analysis (Somarathna et al., 2017; Wadoux et al., 2019; Lagacherie et al., 2020). Consequently, it is of paramount importance to increase the density of soil inputs to improve the performance of DSM models in predicting soil properties (Voltz et al., 2020).

The most straightforward way to increase the density of DSM model soil inputs involves harvesting the legacy soil data that have not yet been stored in the existing soil databases. Arrouays et al. (2017a),b) showed that during the period 2009–2015, the numbers of legacy soil profiles stored in global and national soil databases increased by 1046% and 45%, respectively. However, they estimated that a large amount of soil legacy data can still be harvested. This is even more true in some areas across the world where soil surveying has been particularly active in the past.

For example, in southern France, the BRL irrigation company conducted detailed soil surveys over its irrigation perimeter between 1957 and 1992, which resulted in detailed soil maps, 25,000 soil profiles (5/km²) and 203,000 auger hole observations (31/km²). At this stage, such soil data have not yet been harvested and therefore cannot be used as input for DSM applications. However, this data has great potential for improving DSM performance and should be thoroughly examined.

In this paper, a spatially dense set of soil observations harvested from soil survey documents was tested for improving the performances of DSM models in mapping soil available water capacities for different rooting depths (0–30 cm, 0–60 cm, 0–100 cm) and at maximum observation depth, and the associated uncertainties. Our aim was to evaluate the cost-efficiency ratio of using such soil observations and to evaluate the added value of using euclidian buffer distances as additional inputs of DSM models as proposed by Hengl et al. (2018). The study is conducted in the commune of Bouillargues, which is one of the communes included in the BRL irrigation perimeter.

Section snippets

The study area

This study took place in the administrative commune of Bouillargues in the Occitanie administrative French region (Fig. 1). Located in southern France, Bouillargues covers 16 km² and is mainly devoted to vineyards, agricultural lands, forests, and scrublands. Bouillargues has a Mediterranean climate characterized by a moderate average annual rainfall (600 mm) and dry and hot summers.

The study area is topographically split into two subregions with the large flat valley of the Vistrenque in the

DSM models for soil profiles

In this study, we used several mapping models derived from the random forest algorithm. Hereafter, we provide a general description of random forest and its derivatives used in this study.

Preliminary results

In Fig. 6, we present the distributions of SAWC30, SAWC60, SAW100 and SAWCmax for the soil profiles (left panel of Fig. 6) and auger holes (right panel of Fig. 6). We first observed that the distributions of SAWC regardless of the considered soil depth were bimodal for both the soil profiles and auger holes, with i) a higher peak for higher values of SAWC30 and SAWC60 and with ii) a higher peak for lower values of SAWC100 and SAWCmax. Additionally, it is worth noting that both the SAWC ranges

Soil available water capacity

The selected case study considered the soil available water capacity, which is among the most highly demanded properties of end users, as the targeted soil property (Richer-de-forges et al., 2019). This paper completes the small set of papers that were devoted to the digital mapping of SAWC (Hong et al., 2013; Malone et al., 2009; Padarian et al., 2014; (Poggio et al., 2010); Román Dobarco et al., 2019; Ugbaje and Reuter, 2013, Amirian-Chakan et al., 2019) and the even smaller set of papers

Conclusion

In this study, the main lessons were as follows:

•
A QRF approach using euclidian buffer distances outperformed a classical QRF approach in predicting SAWC with a dense set of profiles and auger holes
•
The addition of a dense spatial sampling of auger hole observations dramatically increased the performance in predicting SAWCs and increased the spatial resolutions of the SAWC pattern delineations, but there were limitations due to the uncertainty of the auger hole observations.
•
The performances in

Declaration of Competing Interest

None.

Acknowledgments

This research was granted by the French National Research Institute for Agriculture, food and environment (INRAE), the French Research and Technology Agency (ANRT) and BRL Exploitation. Philippe Lagacherie is a collaborator of the GLADSOILMAP Consortium supported by the LE STUDIUM Loire Valley Institute for Advanced Research Studies.

References (54)

Alireza Amirian-Chakan et al.
Some practical aspects of predicting texture data in digital soil mapping
Soil Tillage Res.
(2019)
I. Cousin et al.
Influence of rock fragments on the water retention and water percolation in a calcareous soil
Catena
(2003)
T. Hengl et al.
A generic framework for spatial prediction of soil variables based on regression-kriging.
Geoderma
(2004)
P. Lagacherie et al.
Analysing the impact of soil spatial sampling on the performances of Digital Soil Mapping models and their evaluation: a numerical experiment using clay contents obtained from Vis-NIR-SWIR hyperspectral imagery
Geoderma
(2020)
B. Minasny et al.
Digital soil mapping: a brief history and some lessons
Geoderma
(2016)
V.L. Mulder et al.
GlobalSoilMap France: high-resolution spatial modelling the soils of France up to two meter depth
Sci. Total Environ.
(2016)
J. Padarian et al.
Predicting and mapping the soil available water capacity of Australian wheatbelt
Geoderma Reg
(2014)
E.J. Pebesma
Multivariable geostatistics in S: the gstat package
Comput. Geosci.
(2004)
L. Poggio et al.
Soil available water capacity interpolation and spatial uncertainty modelling at multiple geographical extents
geoderma
(2010)
M. Román Dobarco et al.
Uncertainty assessment of GlobalSoilMap soil available water capacity products: a French case study
Geoderma
(2019)

D.L. Shrestha et al.

Machine learning approaches for estimation of prediction interval for the model output

Neural Netw

(2006)

K. Vaysse et al.

Evaluating digital soil mapping approaches for mapping GlobalSoilMap soil properties from legacy data in Languedoc-Roussillon (France)

Geoderma Reg.

(2015)

Kévin Vaysse et al.

Using quantile regression forest to estimate uncertainty of digital soil mapping products

Geoderma

(2017)

D.J.J. Walvoort et al.

An R package for spatial coverage sampling and random sampling from compact geographical strata by k-means

Comput. Geosci.

(2010)

A. Alfons

cvTools: Cross-Validation Tools for Regression Models

(2012)

D. Arrouays et al.

Digital soil mapping across the globe

Geoderma Reg.

(2017)

D. Arrouays et al.

Soil legacy data rescue via GlobalSoilMap and other international and national initiatives

GeoResJ.

(2017)

D. Arrouays et al.

The GlobalSoilMap project specifications

GlobalSoilMap: Basis of the Global Spatial Soil Information System - Proceedings of the 1st GlobalSoilMap Conference

(2014)

D. Baize et al.

Guide des sols

(1995)

J. Böhner et al.

SAGA — analysis and modelling applications

J. Bourrier

La mesure des caractéristiques hydrodynamiques des sols par la méthode Vergière.

Bull.Tech. du Génie Rural

(1965)

L. Breiman

Random forests

Mach. Learn.

(2001)

D.J. Brus et al.

Sampling for validation of digital soil maps

Eur. J. Soil Sci.

(2011)

E. Dominati et al.

A soil change-based methodology for the quantification and valuation of ecosystem services from agro-ecosystems: a case study of pastoral agriculture in New Zealand

Ecol. Econ.

(2014)

C. Gomez et al.

Importance of the spatial extent for using soil properties estimated by laboratory VNIR/SWIR spectroscopy: examples of the clay and calcium carbonate content

Geoderma

(2018)

T. Hengl

GSIF: Global Soil Information Facilities

T. Hengl et al.

SoilGrids1km - Global soil information based on automated mapping

PLoS ONE

(2014)

Cited by (5)

Countrywide mapping and assessment of organic carbon saturation in the topsoil using machine learning-based pedotransfer function with uncertainty propagation
2023, Catena
Stakeholders and policymakers have been becoming more and more interested not just in the potential organic carbon (SOC) saturation level of soils but also in spatially explicit information on the degree of SOC deficit, which can support future policy and sustainable management strategies, and carbon sequestration-associated spatial planning. Thus the objective of our study was to develop a cubist-based pedotransfer function (PTF) for predicting and mapping the saturated SOC content of the topsoils (0–30 cm) in Hungary and then compare the resulting map with the actual SOC map to determine and assess the degree of SOC deficit. It was assumed that topsoils covered by permanent forests can be practically considered as saturated in SOC. Using the monitoring points of the Hungarian Soil Information and Monitoring System located in forests as reference soil profiles, we developed a cubist-based PTF. The transparent model structure provided by cubist allowed to show that not just the physicochemical properties of soils (e.g., texture, and pH) but also environmental conditions, such as topography (e.g., slope, altitude, and topographical position) and climate (e.g., long-term mean annual temperature, and evaporation), characterizing landscape are important factors in predicting the level of SOC saturation. Our results also pointed out that there is SOC deficit on large part of the country (∼80%) showing high spatial variability. It was also revealed that the most considerable potential for additional SOC sequestration can be found related to soils with medium to high actual SOC content.
Complex hydrological knowledge to support digital soil mapping
2022, Geoderma
Citation Excerpt :
We used random forest (RF) and Cubist machine learning algorithms to predict soil attributes in the study area. Both techniques are common in DSM works, producing good results on the prediction of soil texture (da Chagas et al., 2016; Fongaro et al., 2018; Lagacherie et al., 2019; Poppiel et al., 2019b), SOC (John et al., 2021; Li et al., 2021; Moura-Bueno et al., 2021), soil classes (Flynn et al., 2019; Lamichhane et al., 2021; Odgers et al., 2014; Vincent et al., 2018; Zeraatpisheh et al., 2019), and other environmental factors related to soil (Hengl et al., 2018; Styc et al., 2021; Zhang et al., 2021). We applied the k-fold cross-validation (CV) method to the soil data to construct the prediction models using the R package “caret” (Kuhn, 2008).
Drainage network (DN) is the representation of all the stream channels developed over the landscape. The morphometry of DN describes the relationship between channel characteristics and basin geometry, which is regulated by a series of processes, such as weathering, geomorphology, sediment erosion/deposition. The interaction between these factors impacts soil formation, resulting in a relationship between DN morphometry and soil characteristics. Unstable surfaces produced shallow soils, enhancing surface runoff and high drainage density, while stable landscapes favor vertical infiltration through old and weathered soils. Digital soil mapping has benefited from multiple environmental variables, such as relief and satellite data, but DN information can offer great contributions for the prediction of soil attributes. In this work, we applied a set of complex DN variables to perform digital maps of clay, sand, and soil organic carbon (SOC), for a 1378 km² site in the São Paulo state, Brazil. We analyzed the relationship between the drainage density (DD), drainage frequency (DF), channel sinuosity, and confluence angle with soil classes and attributes. The Cubist and Random Forest algorithms were tested to predict the soil information, and to evaluate the impact of the new hydrological variables. The results showed that landscapes predominated by clayey soils favor surface runoff and increase channel formation, with higher channel sinuosity and acute confluence angles. These new drainage variables contributed 35 to 55% for SOC prediction. DD and DF were the most important drainage variables on the models, ranging from 65 to 70%. The external validation reached R² of 0.72 and 0.56 for the prediction of clay and sand, respectively. The impact of DN information on the model performance suggests that more work is needed to better explore and understand the relationship between DN and soil information.
Combining laboratory measurements and proximal soil sensing data in digital soil mapping approaches
2021, Catena
Citation Excerpt :
The comparisons of the EMI integration approaches showed strong differences in performances across the three approaches with only a significative improvement of ECe prediction performances when EMI measurements are integrated through the regression co-kriging approach. Although a larger number of sites were used for calibrating the RF algorithms in approach 1, we did not observe an improvement of the results as observed by Somarathna et al. (2017), Wadoux et al. (2019a), Lagacherie et al. (2020) and Styc et al. (2021). Conversely, the introduction of pseudo values of ECe derived from EMI measurements decreased the performances, which revealed the sensitivity of RF calibration to the uncertainty of data inputs.
Digital soil mapping (DSM) products are limited in accuracy because of the lack of soil inputs. Soil sensing is a promising alternative to direct soil measurements that could provide much denser spatial samplings. Although using relevant and detailed soil sensing input in DSM is considered as vital to increase the prediction performances, there has been no studies in the literature that compare and develop the methods for integrating new sources of soil data that can be applied as inputs of DSM. This paper fills this gap on the example of mapping electrical conductivities from sites with laboratory measurements, in-field EM38MK2 measurements and spatially exhaustive covariates. Three different approaches are tested for putting in synergy real measurements and EM38MK2 measurements: (i) EM38MK2 measurement considered as measured points, (ii) EM38MK2 measurement used for building a new soil covariate and (iii) EM38MK2 measurement considered as a soft data in a regression co-kriging approach. According to soil analysis's financial expenditure, choosing an optimal sample size to merge laboratory analysis and in-field EM38MK2 measurements as surrogate data was done on the best method. The results showed (i) the utility of EM38MK2 data in DSM as a surrogate input data for mapping soil salinity (ii) Regression co-kriging was the best method for integration and (iii) The impact of EM38MK2 data on the gains of performance becomes greater and greater as the sizes of real measurements of soil salinity decrease. Hence, in other areas worldwide that soil sensing as alternative data is accessible, this research's future utilization could be possible as a promising way to tackle one of the essential constraints of DSM.
Soil parent material prediction through satellite multispectral analysis on a regional scale at the Western Paulista Plateau, Brazil
2021, Geoderma Regional
Citation Excerpt :
Despite the role of soil and geological information for economical and sustainable activities (Maltman, 2012; Prokopovich, 1984), most available databases are limited and with coarse spatial resolution, hampering its application for local and regional planning (Adhikari and Hartemink, 2016; Dobos et al., 2013; McBratney et al., 2014; Nolasco de Carvalho et al., 2015). Digital soil mapping (DSM) and remote sensing (RS), emerged as important tools to fill this gap and produce more detailed maps of natural resources (Dharumarajan et al., 2021; Searle et al., 2021; Styc et al., 2021; Wadoux et al., 2019) and to improve legacy databases (Lamichhane et al., 2021; Nauman and Thompson, 2014; Odgers et al., 2014; Pelegrino et al., 2016; Vincent et al., 2018). The use of RS coupled with geographical information systems (SIG) has resulted in new representative environmental data, which assisted on the prediction of soil attributes, landscape processes, and geological information (Grimm et al., 2008; Minasny and Hartemink, 2011; Poppiel et al., 2019b).
Parent material is the main source for soil textural, mineralogical, and other physical attributes. The knowledge over this factor is explored generally in low scale geology maps, insufficient for most users. Remote sensing can offer assistance in this regard, since it allows the evaluation of soil properties, as largely indicated in literature, being a potential tool to delineate parent material. Thus, we explored a multi temporal Landsat image composition with bare soil reflectance to extract soil properties and distinguish discrepant lithological classes at the western plateau, São Paulo State, Brazil. The area is 247,737 ha large, where 981 soil samples were collected at 0–20 cm depth. We acquired the synthetic soil image and linked the pixel's spectra with soil attributes. We performed a digital soil mapping procedure to generate maps of attributes related to parent material. The soil maps offered a great input on identifying the transitions between sandstone and basalt as soils from these formations have significant differences in clay, sand, Fe₂O₃ and TiO₂ contents. Therefore, the use of remote sensing coupled with digital soil mapping is a strong alternative to conventional methods to improve low scale PM maps to enhance detail on regional and local scales.
Data supporting the soil salinity evolution appraisals in the Flumen irrigation district, NE Spain
2021, Data in Brief
The dataset presented comprises (raw data) scans of the marked paper contact prints from a dedicated photogrammetric flight and a diagram showing the location of each of these photograms. The flight was commissioned specifically for the soil survey presented herein. The scanned paper prints are those used in the field to characterize the soil salinity of 27,500 ha within the Flumen irrigation district, in the semi-arid Central Ebro Basin, in Spain. On these prints, the soil surveyors marked the locations of the sampling sites. IRYDA, the extinct Spanish Ministry of Agriculture agency, in charge of designing and implementing new irrigation districts, commissioned the flight in 1975. These paper prints enabled us to resample the soils years later, to: (i) determine the soil salinity evolution from 1975 to 1985 [1], and from 1975 to 1999 [2]; (ii) apply electromagnetic induction (EMI) [3] for the same purpose; (iii) use multivariate analysis to discriminate the salinity trends from 1975 to 1999 in the different soil units [4]; and (iv) for land evaluation [5]. The report and two volumes of annexes [6], [7], [8], prepared by the contractor INYPSA for IRYDA, contain data on the soluble salts and other soil components sampled in 1975 as well as soil descriptions and agronomical data. The aerial photographs presented herein allow the sites sampled in 1975 to be located. This is the first step in exploiting the legacy data to appraise [9] the effects on the salinity and other soil properties recorded in 1975. The irrigation and the change from basin and border flooding to pressurized techniques, with the merging of many plots, govern these effects.

View full text

Harvesting spatially dense legacy soil datasets for digital soil mapping of available water capacity in Southern France

Highlights

Abstract

Introduction

Section snippets

The study area

DSM models for soil profiles

Preliminary results

Soil available water capacity

Conclusion

Declaration of Competing Interest

Acknowledgments

Soil Tillage Res.

Catena

Geoderma

Geoderma

Geoderma

Sci. Total Environ.

Geoderma Reg

Comput. Geosci.

geoderma

Geoderma

Neural Netw

Geoderma Reg.

Geoderma

Comput. Geosci.

cvTools: Cross-Validation Tools for Regression Models

Digital soil mapping across the globe

Geoderma Reg.

Soil legacy data rescue via GlobalSoilMap and other international and national initiatives

GeoResJ.

The GlobalSoilMap project specifications

GlobalSoilMap: Basis of the Global Spatial Soil Information System - Proceedings of the 1st GlobalSoilMap Conference

Guide des sols

SAGA — analysis and modelling applications

La mesure des caractéristiques hydrodynamiques des sols par la méthode Vergière.

Bull.Tech. du Génie Rural

Random forests

Mach. Learn.

Sampling for validation of digital soil maps

Eur. J. Soil Sci.

A soil change-based methodology for the quantification and valuation of ecosystem services from agro-ecosystems: a case study of pastoral agriculture in New Zealand

Ecol. Econ.

Importance of the spatial extent for using soil properties estimated by laboratory VNIR/SWIR spectroscopy: examples of the clay and calcium carbonate content

Geoderma

GSIF: Global Soil Information Facilities

SoilGrids1km - Global soil information based on automated mapping

PLoS ONE