OR/15/030 Introduction

Nicole Archer, Barry Rawlins, Stephen Grebby, Ben Marchant and Bridget Emmett. 2015. Identify the opportunities provided by developments in earth observation and remote sensing for national scale monitoring of soil quality. British Geological Survey Internal Report, OR/15/030.

Background

Defra last reviewed the opportunities for using remote sensing The UK Soil Indicators Consortium (UKSIC) identified the following eleven indicators for inclusion in national scale soil monitoring: pH, organic carbon, bulk density, phosphorus (Olsen P), nitrogen (total N), magnesium (extractable), potassium (extractable), copper (aqua regia extractable), cadmium (aqua regia extractable), zinc (aqua regia extractable) and nickel (aqua regia extractable). Since the UKSIC report (Black et al., 2008), soil depth and, in particular, peat depth has been identified as an important requirement for soil monitoring. There is also interest in the monitoring of soil erosion and compaction. However, we are not explicitly considering soil erosion monitoring, as a pilot project to establish a soil erosion network for England and Wales is currently underway. Combining the two sets of soil indicators above gives a total of fourteen that we consider in this report.

Soil monitoring activity is still at a relatively early stage of development. Optimal approaches to measuring state and change in soil may be quite different. The spatial variation of change in soil properties should determine the resampling strategy for monitoring, and this may be quite different from the baseline variation of that property’s state (Lark et al., 2006). There are many examples of using remotely sensed data as covariates for reducing the uncertainty in predictions of soil properties such as organic carbon in combination with traditional, ground-based measurements (Rawlins et al., 2009). However, this, and other exploitation of remote sensor data depends on statistical correlations between the remotely sensed measurement and the soil property of interest. Such a correlation may have a direct physical basis (e.g. if the soil property directly influences the radiative properties of the land surface), but may also arise through secondary relationships (e.g. soil organic carbon may be correlated with clay content, which in turn affects soil water content and so the radiometric measurement). When we are concerned with measuring changes in the soil, a remote sensor variable that proves useful for predicting landscape-scale variation of baseline values may prove rather less useful, if its value in the former context is due to correlations with indirect rather than direct physical effects on the radiometric properties of the landscape. Our review of technologies therefore considers evidence for the relative importance of direct physical effects and indirect relationships in determining correlations of predictive value.

The aims of this report are to address the following questions:

1. What are the current and potential future opportunities to monitor any of the listed set of indicators using remote sensing? Defra wish to understand the potential to monitor both state and change.
2. When will these be ready for use and what level of further development is required?
3. Could remote sensing of any of these indicators replace and/or complement traditional field based national scale soil monitoring?
4. Can meaningful measures of change be derived?
5. How could remote soil monitoring of individual indicators be incorporated into national scale soil monitoring schemes (such as the Countryside Survey?)

In addressing these questions we have considered the full range of earth observation, satellite navigation and telecommunications tools available.

Sensor platforms

In the context of remote sensing, the different sensor types can generally be mounted on either airborne or spaceborne platforms. Here, airborne platforms are differentiated according to whether the aircraft is manned (i.e. airplane or helicopter) or unmanned (i.e. UAV — un-manned aerial vehicle). The choice of sensor-platform combination is not only governed by the soil property of interest, but by several additional factors including the associated cost and practicality.

Irrespective of the application, the use of UAVs for remote sensing is somewhat restricted by the size and mass of the sensor payload they are capable of carrying. As a result, UAVs have traditionally been used as a platform for imaging systems for the acquisition of spectral reflectance data (Honkavaara et al., 2013; Torres-Sánchez et al., 2013), thermal infrared imagery (Berni et al., 2009; Zarco-Tejada et al., 2012), or for the derivation of photogrammetric digital elevation models (DEMs) (d'Oleire-Oltmanns et al., 2012). However, with recent advances in compact sensor technology, more potential UAV applications are emerging as some are now capable of carrying RADAR sensor payloads (Koo et al., 2012; Remy et al., 2012). After the initial outlay to purchase the UAV and any appropriate sensors, the cost of acquiring remotely sensed data from an unmanned airborne platform generally consists of the staff-time for only a single operator. Moreover, UAVs can be utilised to acquire very high spatial resolution data — typically on the order of centimetres — because they can be operated at lower altitudes than manned aircraft. However, imagery acquired at a very high spatial resolution usually covers a small spatial extent on the ground. As a consequence, achieving national-scale coverage using UAVs is likely to require considerable time and financial support. Accordingly, soil monitoring through repeat UAV surveys is arguably better suited to detailed local monitoring programmes.

The acquisition of various remotely sensed datasets (e.g., multi- and hyperspectral imagery, RADAR, LiDAR, radiometric, thermal) from manned airborne platforms is well established. Although having a higher operating altitude than their unmanned counterparts, manned airborne platforms can still be utilised to acquire high spatial resolution data — typically on the order of metres. Data acquired from manned airborne platforms also has the added benefit of it covering a larger spatial extent than that of UAV-acquired data, therefore providing a more practical means of acquiring data at regional or even national-scale. However, despite the advantages, the commissioning of manned airborne surveys is generally costly, particularly if considered for repeat surveys for national-scale monitoring purposes.

With many relevant sensors already in operation, spaceborne platforms offer an attractive means of monitoring soils at a national-scale, because image scenes typically cover large spatial extents and can be acquired at either no cost or for a relatively small fee. Although only a handful of scenes may be required for national coverage, spaceborne data do generally have a coarser spatial resolution than airborne platforms — typically on the order of tens to hundreds of metres. Nevertheless, with frequent revisit times, spaceborne platforms provide access to temporal datasets that can be readily used for monitoring purposes.

A recent review by Croft et al. (2012) highlights the major challenges to remote sensing and modelling of soil properties: 'One of the greatest challenges facing the broad-scale adoption of remote sensing methods in soil science and soil organic carbon [SOC] studies is the site-specific nature of relationships between RS-measured variables and SOC.'

Site-specific relationships between remotely sensed variables and soil properties can occur for various reasons:

1. Reported models are empirical in nature. These models are often only relevant for a particular instrument at a point in time and space, as the complex relationship between soil constituents and soil reflectance is not taken into account (Bartholomeus et al., 2011)
2. Sensor characteristics vary between sensors. The transfer of prediction models between one sensor to another can be complex, due to differences in spectral resolution, sampled wavelengths, location of spectral bands and the number of bands used (Bartholomeus et al., 2011).
3. The use of different numerical methods and data pools can also cause difficulties when comparing the statistical quality of mapped soil parameters (Selige et al., 2006).
4. Soil reflectance responds to temporally invariant factors (e.g. soil type, mineralogy, geology) and temporally variant factors (e.g. tillage, moisture, soil roughness, crop residue cover (Ladoni et al., 2010).
5. Inherent data accuracy of airborne and satellite data, which can be due to variations in illumination, changes in terrain and atmospheric attenuation (Ben-Dor et al., 2002). This can also cause problems particularly when, for example, SOC has low concentrations or a small range of SOC values, and measurement uncertainly can exceed spatial and temporal differences in SOC content (Stevens et al., 2008).

Challenges in northern temperate regions

In regions of northern latitudes, such as the UK (50–60 ºN), there are particular challenges for remote sensing of non-vegetated soils to measure their properties. The window of opportunity for remote sensors to measure soil surfaces is generally reduced because: 1) bare, or partially bare, ground exists mainly during the autumn and winter months, due to tillage practices, where soils are cultivated in the autumn and sown later that autumn; and in the case of grassland are seldom bare. 2) Cloud cover during the winter is common in northern latitudes, preventing clear-day skies, which are necessary for remote sensors to provide spectral reflectance data of soil properties. 3) Soil moisture can cause variation to spectral response, which during the winter can be very variable, including aspects of freeze/thaw and snow, which have a large effect on spectral reflectance. 4) During the winter, the angle of the sun is low, which also affects spectral response of remote sensors. To reduce spectral response variability in the visible and near infra-red, it is best to have a small azimuth angle.

Mulder et al. (2011) reviewed the use of remote sensing in soil and terrain mapping and summarised the various ways remotely sensed data offered possibilities for extending existing soil survey datasets. The suggested uses of remote sensing were:

1. Soil composition can be assessed by using remote sensing to segment the landscape into approximately homogenous soil-landscape units, which then aids the assessment of soil composition using classical or more advanced methods,
2. The spectral data measured by remote sensors can be analysed using physically-based or empirical methods to derive soil properties,
3. Remotely sensed imagery can be used as a data source to support soil mapping as described by (Grunwald et al., 2011), and (Minasny et al., 2013)
4. Remote sensing methods facilitate mapping of inaccessible areas, reducing the need for extensive time-consuming and costly field surveys.

Measuring state ands change of soil indicators

The monitoring of the status of soil indicators and the monitoring of change in these indicators are two quite different challenges. Lark (2009) notes that the status and change of a soil indicator are different variables and their variability is likely to differ. For example, Lark (2009) considered the example of metal concentrations in the soil. The baseline status of a particular metal is likely to be primarily related to the underlying geology, whereas the change might be related to anthropogenic processes such as land use change and pollution. This example illustrates that a soil monitoring network that is suitable to monitor the status of a soil indicator might not be suitable to monitor the change in the property in two regards. First, the statistical design of the network might not be able to estimate the change in the indicator with the same precision as the status can be estimated. Second, a measurement method (e.g. a remote sensing technology) that is suitable to infer the status of the indicator might not be suitable to infer change in that property.

The design of a soil monitoring network refers to the configuration of locations and times at which the soil indicator is measured. The precision of an estimate of the baseline status of a soil indicator should improve with the sampling effort in space. This precision depends on the design and the spatial variability of the measurements of the soil indicator. If the variability of a particular indicator is well understood it is possible to estimate the precision with which a particular design will estimate the baseline status. Such a process was conducted by Black et al. (2008) when they considered the design of a UK soil monitoring network that would use conventional measurement methods. They used previous surveys of soil properties (the National Soil Inventory and Countryside Survey) to establish models of the spatial variability of key soil indicators, such as soil organic carbon, pH, copper and zinc. They then tested the precision with which different designs could estimate the means of these indicators. If the same process is to be used to determine the precision with which a soil monitoring network could estimate the change of a soil indicator, then it is necessary to quantify both the spatial and temporal variation of the indicator. Information about the temporal variation of soil properties tends to be less plentiful than spatial information, because of the expense of conducting a survey at multiple times and the time that must elapse before meaningful changes can be observed. Therefore, it tends to be more difficult to establish whether a monitoring network is suitable to monitor change with a specified precision. Black et al. (2008) only considered the precision with which changes in soil organic carbon could be estimated. Where temporal information about soil properties is not available it is necessary to conduct reconnaissance surveys prior to designing a soil monitoring network.

The second point about whether a particular measurement method is suitable to infer both status and change in a soil indicator is particularly pertinent for remote sensing technologies. Often these technologies do not directly measure the soil indicator of interest. Instead, they measure a property that is correlated to the indicator of interest and a statistical model is used to infer the indicator. If we return to the example of soil metal concentrations, it might be possible to use a radiometric sensor to identify variations in parent material and a statistical model calibrated to relate these variations to the metal concentrations. However, if the radiometric sensor is used to re-measure the soil at regular intervals then the information it gathers will still primarily relate to the variation in geology. It will say less about pollution or the changes in land use which might have caused changes to the soil metal concentrations.

One key question is 'What is meant by a meaningful change in a soil property?' Meaningful could refer to a statistically significant change i.e. a change that is sufficiently large that we would not have expected it to occur by random chance. However, this is not always the most useful definition. The magnitude required for a change to be statistically significant decreases with sampling effort. If very intensive soil sampling is employed then it is possible that tiny changes to soil indicators might be statistically significant. However, these changes might be too small to cause a noticeable change to soil functionality. It is preferable to consider meaningful changes in terms of soil functionality. For instance, there might be critical thresholds in the concentrations of soil nutrients below which the soil is unsuitable to maintain production or heavy metal concentrations above which plant toxicity may occur. A meaningful change in the property could be said to have occurred if soil nutrient or heavy metal concentrations cross a threshold. Black et al. (2008) reviewed critical thresholds or ‘action levels’ for 13 soil indicators. A coupled question is how precisely the soil indicator and these action levels need to be monitored. Black et al. (2008) suggested that the development of such quality measures was urgently required, but to our knowledge, these have not been determined to date.

National scale soil monitoring and incorporation of remote methods

To understand the potential for remote methods to enhance soil monitoring it is necessary to clarify some terminology. We refer to ground-based soil sampling followed by laboratory measurements of soil indicators at national survey sites as primary observation. If a remote sensor makes a direct measurement of a soil indictor this is a secondary observation. Where we can establish a statistical relationship between a series of remotely sensed observations at a large number of sites where a series of primary observations have also been made, we refer to the former as a secondary covariate. Such a secondary covariate may have an indirect relationship with the soil indicator. For example, in their study on mapping soil organic carbon across Northern Ireland, Rawlins et al. (2009) stated that the two dominant factors which explained the negative correlation (indirect relationship) between SOC and gamma-radiation derived from potassium (40K) decay were: i) the variation in mineral-K content which decreased with increasing quantities of soil organic matter, and ii) increased soil moisture resulting in greater attenuation of the gamma signal from the soil. The secondary covariate can be included in statistical approaches to predicting soil properties as a fixed effect.

To enhance soil monitoring, by contributing to the detection of a meaningful change in soil indicators, preliminary approaches are likely to utilise remote sensing as a secondary covariate to improve predictions of soil properties at unsampled locations. For example, in areas of cultivated soils of England, it may be possible to use either airborne (Selige et al., 2006) or satellite (Jaber et al., 2011) hyperspectral data as a fixed effect to predict SOC concentrations in topsoil across large regions.

The CEH Countryside Survey soil sampled a total of 396 sites across England and Wales in 2007 (Emmett et al., 2010), but it would only be possible to capture vegetation-free, remotely sensed data on soil reflectance for a subset of cultivated sites predominantly in southern and eastern England. If it was possible to establish a strong statistical relationship between the hyperspectral data for these sites and the primary measurements, this could be used as a fixed effect to make predictions (with associated uncertainties) for other areas of cultivated land for which hyperspectral data were available as a way of wider mapping of the status of soil indicators.