Project
title / New methods of soil mapping
/ MAFF
project code / SR0120

ministry of agriculture, fisheries and food CSG 15

Research and Development

Final Project Report

(Not to be used for LINK projects)

Two hard copies of this form should be returned to:
Research Policy and International Division, Final Reports Unit
MAFF, Area 6/01
1A Page Street, London SW1P 4PQ
An electronic version should be e-mailed to
Project title / New methods of soil mapping
MAFF project code / SR0120
Contractor organisation and location / National Soil Resources Institute
Cranfield University, Silsoe, Bedford MK45 4DT
Total MAFF project costs / £ 179,002.00
Project start date / 01/09/99 / Project end date / 30/08/01
Executive summary (maximum 2 sides A4)
To tab in this section press the tab key and the Control key together
Press the DOWN arrow once to move to the next question.

CSG 15 (1/00) 3

Project
title / New methods of soil mapping
/ MAFF
project code / SR0120

In order that the true significance of soil can be incorporated into environmental assessments and environmental models, national 1:50,000 scale map coverage of soils is required. This is unlikely to happen in short- or medium-term timescales because, using conventional techniques, field survey is labour intensive and hence expensive. The necessary skill-base is also rapidly disappearing as the last generation of trained soil surveyors retire. A new, rapid and cost effective method of soil map preparation is therefore urgently required. Promising results have been obtained in this project for the application of mathematical techniques, in the form of a predictive soil mapping model, to a range of existing environmental and remotely sensed data.

The model has been developed in trial areas around Melbourne, in Derbyshire and Harold Hill, in Essex. The areas were selected to make use of relevant, but not necessarily universally available data, in a wide range of soilscapes. Once developed the model was used as a predictive tool to generate soil maps in nearby test areas where soilscapes were broadly similar to those of the trial areas.

Soil should be considered as a continuum as it changes in some aspects from place to place. It can also change dramatically over short distances and exhibit complex patterns of contrasting soil properties. A soil map at 1:50,000 scale can only represent an approximation of the true soil pattern. Within a mapped area of a particular soil series, the unit of mapping at this scale, soils exactly matching that series profile class might only occupy about 70% of the mapped area. Many of the remaining aberrant soils will be closely related to the named series but nevertheless will belong to a different class.

The results obtained from the predictive model at Melbourne are very encouraging and of the six soilscapes studied, soil series predictions were at least 70% correct in three, with one soilscape having an almost 80% success rate. The other three soilscapes had success rates between 55 and 60%. It is felt that further progress can be made in developing a successful model by concentrating on a few important features;

·  The Ordnance Survey Digital Terrain Model (DTM) was unexpectedly found to be inadequate in dealing with the hydrological aspects of the landscape. It identified many hundreds of false sinks in the landscape because it had insufficient altitude information in areas of low relief. A Centre for Ecology and Hydrology (CEH) DTM has recently been identified which has been developed for hydrological applications and will also allow a closer grid spacing for data and the incorporation of improved landform classification techniques.

·  Soilscapes are currently used to extrapolate the model results out from the trial areas into the test sites. Further stratification is required if results are to improve and NATMAP soil associations, which are largely based on landscape, should be tried.

·  High-resolution airborne geophysical data (HiRES) has been shown to be particularly valuable in refining the prediction of soil parent material. However, HiRES coverage is very limited and is only available for part of the Melbourne study area. HiRES data, collected at 200 m line spacing should be incorporated into any new study.

·  Geologists are not usually asked to predict the precise lithology of materials at 1 m depth and this type of data is not routinely stored in BGS databases. Soil parent materials are strictly defined and many of the concepts are new to geologists. However, field survey geologists regularly encounter these materials and it should be possible to infer detailed information from published maps, especially when reviewed by experienced field survey staff. This is an iterative process accurately linking the geological map polygons to soil parent materials and further local refinement should benefit the model results.

·  Results of the predictive model should be critically evaluated by soil surveyors and geologists with expert knowledge of the relevant landscapes. In this way it may be possible, where necessary, to refine the way the model interprets the input data. Selective ground truthing of the model output should be incorporated into any new study.

A large investment has been made in compiling the datasets for this project. Further work is required to extract the full value from the assembled datasets. Successful results have been obtained in certain landscapes but important issues remain unanswered. Until these issues are addressed it is impossible to give clear answers to questions about the feasibility of producing reliable soil maps by these new mathematical techniques.

Recommendations are made for further development of this methodology for cost-effective predictive soil mapping.

CSG 15 (1/00) 3

Project
title / New methods of soil mapping
/ MAFF
project code / SR0120
Scientific report (maximum 20 sides A4)
To tab in this section press the tab key and the Control key together
Press the DOWN arrow once to move to the next question.

CSG 15 (1/00) 3

Project
title / New methods of soil mapping
/ MAFF
project code / SR0120

Introduction

Detailed soil maps at 1:25,000 and 1:50,000 scale are available for approximately 25% of England and Wales (Fig. 1). For the remaining 75% of land the only published map is the reconnaissance 1:250,000 scale National Soil Map (NATMAP). There is, however, continual and increasing pressure to resolve national environmental questions using soil data at 1:50,000 scale or better. Many environmental models and decision support systems for land management issues require detailed spatial and attribute information for soils in order to match other detailed environmental data obtained from remote sensing and digital terrain analysis (Band et al., 1991, 1993). Currently, data derived from reconnaissance soil surveys are the major source of soil information for a variety of land analysis techniques, ecological modelling, and management applications.

Fig 1: Distribution of large and medium scale maps for England and Wales.

Field survey is labour intensive and, therefore, it is becoming increasingly unlikely that a detailed soil mapping programme will be reinstated using traditional methods, especially as the necessary mapping skills are disappearing as experienced staff retire. In the short or medium term, if detailed soil maps are to be produced for those areas where coverage currently does not exist, less expensive methods of mapping need to be developed.

Soil distribution is intimately linked to climate, surface geology, landform and vegetation. Large-scale digital datasets exist nationally for these variables, and there is evidence that reliable soil maps can be derived from these by appropriate mathematical techniques, in conjunction with expert integration of the data.

The main objective of this research was to investigate the application of mathematical techniques to a range of environmental and remotely sensed data as a means of producing 1:50,000 scale soil maps. This is particularly relevant to areas for which soil distribution maps are only currently available at a reconnaissance scale of 1:250,000 scale.

Environmental Modelling

The origins of the functional approach to the soil system can be traced back to the end of the 19th century, when Dokuchaev suggested the following equation to represent soil formation:

where soil (s) is a function (f) of parent material (pm), climate (c), the biosphere (b) and the age of the land (a) (time factor).

Jenny (1941) takes this concept further and states that soil is a function of five soil forming factors. He postulated that:

where cl is climate; o is organisms; r is relief; p is parent material; and t is time. There is then space for additional, as yet unidentified, soil-forming factors.

This concept was again updated by Jenny (1961) when he developed what might now be regarded as a systems framework.

Since 1970 there have been many attempts to characterise the meso-scale spatial variability of measured soil attributes (Beckett & Webster, 1971; Webster, 1985; Yates & Warick, 1987; Loague & Gander, 1990). These attempts have concentrated on the characterisation of patterns rather than on the linking of pattern to process.

Two techniques are commonly used for spatial predictions:

(i)  geostatistics, a quantitative interpolation approach (e.g. kriging) that relates the spatial covariance function to the spatial separation of the data, and

(ii)  soil-landscape analysis, which relates soil attributes to qualitative measures of landscape position such as toe-slopes and interfluves (as an attempt to account for the ranges of processes involved).

Both techniques require large databases and their results are not transferable. Interpolation techniques ignore pedogenesis, whilst methods based on landscape position have lacked a consistent quantitative framework.

Methodology

This section outlines the methodology adopted in this study to produce 1:50,000 scale soil maps. The modelling approach is based on Jenny’s functional approach to soil system, using small-scale soil maps, medium scale lithology maps as well as terrain and remotely sensed data as input layers. Within this framework, the statistical approach to model construction is based on Bayesian belief networks. The resulting models are extrapolated to other areas based on soilscapes, which are areas of common lithology, landform and soil genesis.

Modelling approach:

The modelling approach adopted in this study has a form similar to Jenny’s (1941) functional model but differs in several terms:

where for each site, soil (Si) is a function (f) of the spatial coverage of the National Soil Map (Sin), derived lithology at 1 m depth (Li), and other variables derived from terrain analysis (Ti) and remote sensing (RSi).

In the context of this study the predicted 1:50,000 scale soil maps for Melbourne and Harold Hill were modelled on the following input data:

·  National Soil Map (NATMAP)

NATMAP is an important input layer as it is the only source of soil information covering the whole of England and Wales. In each training area the NATMAP soil associations (defined as a group of soil series soil types that occur together in predictable patterns in distinct soil landscapes) were overlaid and cross-tabulated with the soil series identified on the detailed soil map and the surveyors individual auger borings. In this way relationships were established between individual NATMAP soil associations and their constituent soil series

·  Lithology

Surface geology maps should describe materials at about 1 m depth (ie at the base of the active soil layers) and they are available for nearly all of England and Wales at either 1:50,000 or locally at 1:10,000 scale. If geologists record sufficiently carefully at these shallow depths, spatial surface geology data could, potentially, provide more detailed information on the distribution of soil parent material than is available from the broad 1:250,000 scale NATMAP. By cross-validating the detailed soil map or individual auger bore data relating to that survey with the surface geology map the relevance of the geology-derived lithology with respect to soil development was established.

·  Terrain analysis

Landform is an important factor determining the spatial distribution of soils. Soil-landscape relationships are well documented in the national Soil Bulletins that accompany NATMAP in the form of traditional block diagrams. In this project these relationships are described by both landform parameterisation and landform classification.

·  Remote sensing

Remotely sensed information is quantitative in character and provides spatial coverage. Relationships between radiometric and magnetic airborne data and soil series / soil lithology were established.

Statistical analysis:

A probabilistic approach has been chosen because there is likely to be a wide range in the quality of input data as the methodology is required to be applicable across the whole of England and Wales. The Bayesian rule-based method is analogous to conventional soil survey in that it can exploit a wide range of available data and offers a flexible approach to reducing the uncertainty associated with soil survey. It uses the best available evidence of soil variation for an area, and adopts the best current interpretation for mapping. The ‘knowledge base’ of the model can be derived from training areas, expert knowledge or a combination of the two. However, in contrast to the conventional survey methods, the Bayesian rule-based method can provide a numerical estimate of the probability of occurrence of a particular soil type, or for a given set of attributes, as these vary across the study area.

The statistical analysis was based on point observations, the density of the points being determined for practical reasons by the OS digital terrain model (50 m grid). This is determined by the fact that the terrain derivatives provided the bulk of the numerical input data. Sampling to a smaller grid size would be inappropriate. In addition, the airborne geophysical data were gridded to the same 50 m grid. Relevant attribute data from other input layers were also attached to these point observations.

Extrapolation:

Environmental modelling relies on training areas where models are constructed and validated. The ‘trained’ models are then used for extrapolation across unmapped areas. Stratification is required to distinguish ‘soilscapes’ that are representative of the training area, from those that are unrepresentative (Bell et al., 1992). In this context, a "soilscape" is defined as an area of similar soil-landscape relationships. "Soilscapes" are therefore based on common lithology, landform and soil genesis.

Study sites

It has long been recognised by soil surveyors that patterns of soil distribution in non-glaciated terrain relate more strongly to topography, landscape and, in particular, surface solid geology, than is the case in landscapes formed in glacial deposits. The initial choice of trial areas was severely constrained by the incomplete digital coverage of geological maps (now resolved) and the limited extent of remotely sensed high-resolution airborne geophysical (HiRES) data. With these constraints in mind, the new mapping techniques have been tested in one glaciated and one largely non-glaciated landscape, namely Harold Hill and Melbourne respectively. HiRes data were only available for part of the Melbourne study area.