Digital soil mapping using artificial neural networks

In the context of a growing demand of high‐resolution spatial soil information for environmental planning and modeling, fast and accurate prediction methods are needed to provide high‐quality digital soil maps. Thus, this study focuses on the development of a methodology based on artificial neural networks (ANN) that is able to spatially predict soil units. Within a test area in Rhineland‐Palatinate (Germany), covering an area of about 600 km², a digital soil map was predicted. Based on feed‐forward ANN with the resilient backpropagation learning algorithm, the optimal network topology was determined with one hidden layer and 15 to 30 cells depending on the soil unit to be predicted. To describe the occurrence of a soil unit and to train the ANN, 69 different terrain attributes, 53 geologic‐petrographic units, and 3 types of land use were extracted from existing maps and databases. 80% of the predicted soil units (n = 33) showed training errors (mean square error) of the ANN below 0.1, 43% were even below 0.05. Validation returned a mean accuracy of over 92% for the trained network outputs. Altogether, the presented methodology based on ANN and an extended digital terrain‐analysis approach is time‐saving and cost effective and provides remarkable results.     

Assessing and mapping topsoil organic carbon stock at regional scale: A scorpan kriging approach conditional on soil map delineations and land use

In order to assess the potential of soils as C reservoir at regional scale, accurate estimates of soil organic carbon (SOC) are required, and different approaches can be used. This study presents a method to assess and map topsoil organic carbon stock (Mg ha⁻¹) at regional scale for the whole Emilia Romagna plain in Northern Italy (about 12 000 km²). A Scorpan Kriging approach is proposed, which combines the trend component of soil properties as derived from the 1:50 000 soil map with geostatistical modeling of the stochastic, locally varying but spatially correlated component. The trend component is described in terms of varying local means, calculated taking into account soil type and dominant land use. The resulting values of SOC, sand, silt, and clay contents are retained for calculating topsoil SOC stocks, using a set of locally calibrated pedotransfer functions (PTFs) to estimate bulk density. The maps of each soil attribute are validated over a subset of 2000 independent and randomly selected observations. As compared to the standard approach based on the mean values for delineation, results show lower standard errors for all the variables used for SOC stock assessment, with a relative improvement (RI) ranging from 4 per cent for SOC per cent to 24 per cent for silt. The total C stock (0–30 cm) in the study area is assessed as 73·24 ± 6·67 M t, with an average stock of 62·30 ± 5·55 Mg ha⁻¹. The SOC stock estimates are used to infer possible SOC stock changes in terms of carbon sequestration potential and potential carbon loss (PCL). Copyright © 2010 John Wiley & Sons, Ltd.     

Potential of multi‐objective models for risk‐based mapping of the resilience characteristics of soils: demonstration at a national level

Policy makers rely on risk‐based maps to make informed decisions on soil protection. Producing the maps, however, can often be confounded by a lack of data or appropriate methods to extrapolate using pedotransfer functions. In this paper, we applied multi‐objective regression tree analysis to map the resistance and resilience characteristics of soils onto stress. The analysis used a machine learning technique of multiple regression tree induction that was applied to a data set on the resistance and resilience characteristics of a range of soils across Scotland. Data included both biological and physical perturbations. The response to biological stress was measured as changes in substrate mineralization over time following a transient (heat) or persistent (copper) stress. The response to physical stress was measured from the resistance and recovery of pore structure following either compaction or waterlogging. We first determined underlying relationships between soil properties and its resistance and resilience capacity. This showed that the explanatory power of such models with multiple dependent variables (multi‐objective models) for the simultaneous prediction of interdependent resilience and resistance variables was much better than a piecewise approach using multiple regression analysis. We then used GIS techniques coupled with an existing, extensive soil data set to up‐scale the results of the models with multiple dependent variables to a national level (Scotland). The resulting maps indicate areas with low, moderate and high resistance and resilience to a range of biological and physical perturbations applied to soil. More data would be required to validate the maps, but the modelling approach is shown to be extremely valuable for up‐scaling soil processes for national‐level mapping.     

利用人工神经网络以及相关地形属性绘制数字土壤地图

Detailed soil surveys involve costly and time-consuming work and require expert knowledge. Since soil surveys provide information to meet a wide range of needs, new methods are necessary to map soils quickly and accurately. In this study, multilayer perceptron artificial neural networks (ANNs) were developed to map soil units using digital elevation model (DEM) attributes. Several optimal ANNs were produced based on a number of input data and hidden units. The approach used test and validation areas to calculate the accuracy of interpolated and extrapolated data. The results showed that the system and level of soil classification employed had a direct effect on the accuracy of the results. At the lowest level, smaller errors were observed with the World Reference Base (WRB) classification criteria than the Soil Taxonomy (ST) system, but more soil classes could be predicted when using ST (7 soils in the case of ST vs. 5 with WRB). Training errors were below 11% for all the ANN models applied, while the test error (interpolation error) and validation error (extrapolation error) were as high as 50% and 70%, respectively. As expected, soil prediction using a higher level of classification presented a better overall level of accuracy. To obtain better predictions, in addition to DEM attributes, data related to landforms and/or lithology as soil-forming factors, should be used as ANN input data.     

Digital mapping of soil ecosystem services in Scotland using neural networks and relationship modelling. Part 2: Mapping of soil ecosystem services

A framework for estimating the distribution of soil ecosystem service (ES) supply is described that is based on the concept of matrix multiplication. This approach enables relationships between fundamental soil variables and associated environmental characteristics to be linked to soil processes, and hence to ecosystem functions and ecosystem services. The parameterization of these relationships was achieved using a combination of data from the Scottish Soils Database and expert knowledge. Baseline data to allow mapping of processes, functions and services across Scotland is given by digital maps of soil classes. The matrix multiplication approach constrains the relationship linkages to linear relationships and ignores potential synergies between factors at each stage, but does provide a mechanism for relating fundamental soil characteristics to ecosystem services. The approach has been tested by developing maps of selected ecosystem services in Scotland and comparing these with existing maps of the same or similar ESs. While the values and their ranges differ in each case, the spatial distribution of services is similar. The proposed mechanism is extensible at every level and can also be used to explore the impacts of land management options on environmental characteristics. This is demonstrated by using the model to estimate impacts of liming on three ecosystem services: Agricultural Capability, Carbon Sequestration and Drinking Water Provision. The model is shown to produce reasonable estimation of the impacts of this management option. Further discussion of improvements to the system and its potential applications is given.     

自动土壤图基于知识的分类

A machine-learning approach was developed for automated building of knowledge bases for soil resources mapping by using a classification tree to generate knowledge from training data. With this method, building a knowledge base for automated soil mapping was easier than using the conventional knowledge acquisition approach. The knowledge base built by classification tree was used by the knowledge classifier to perform the soil type classification of Longyou County, Zhejiang Province, China using Landsat TM bi-temporal images and GIS data. To evaluate the performance of the resultant knowledge bases, the classification results were compared to existing soil map based on a field survey. The accuracy assessment mad maalysis of the resultant soil maps suggested that the knowledge bases built by the machine-learning method was of good quality for mapping distribution model of soll classes over the study area.     

Digital mapping of cation exchange capacity using genetic programming and soil depth functions in Baneh region,Iran

This study evaluates the performances of a combination of genetic programming and soil depth functions to map the three-dimensional distribution of cation exchange capacity (CEC) in a semiarid region located in Baneh region, Iran. Using the conditioned Latin hypercube sampling method, the locations of 188 soil profiles were selected, which were then sampled and analyzed. In general, results showed that equal-area quadratic splines had the highest R², 89%, in fitting the vertical CEC distribution compared to power and logarithmic functions with R² of 81% and 84%, respectively. Our findings indicated some auxiliary variables had more influence on the prediction of CEC. Normalized difference vegetation index (NDVI) had the highest correlation with CEC in the upper two layers. However, the most important auxiliary data for prediction of CEC in 30–60 cm and 60–100 cm were topographic wetness index and profile curvature, respectively. Validation of the predictive models at each depth interval resulted in R² values ranging from 66% (0–15 cm) to 19% (60–100 cm). Overall, results indicated the topsoil can be reasonably well predicted; however, the subsoil prediction needs to be improved. We can recommend the use of the developed methodology in mapping CEC in other parts in Iran.     

Analysis on pedodiversity and spatial subset representativity—the German soil map 1:1,000,000

Knowledge about spatial soil variation in terms of measured pedodiversity, as well as the spatial distribution of soils in terms of spatial subset representativity, offers the possibility to estimate the quality and variance within a soil map. Additionally, it can help to identify representative sample locations. Demonstrated at the German soil map at a scale of 1:1,000,000, this study describes a methodology to analyze the distribution of taxonomical pedodiversity using the Simpson index and a new approach to derive representative spatial subsets based on a modified χ²‐test (χ_m²), which can be used as monitoring areas. To analyze the spatial composition of the soil map and to detect differences in the underlying mapping schemes of the German soil map 1:1,000,000, three different spatial data structures were studied: (1) the entire soil map, (2) the soil map segmented into geomorphological regions, and (3) the soil map segmented into the Federal States of Germany. Representative patches of varying sizes were statistically derived for all spatial subsets as well as the entire soil map ranging from 20 km × 20 km up to 70 km × 70 km. The results show that the measured pedodiversity is linked to both the geomorphology as well as the political borders of the Federal States. On the one hand, this reveals the uncertainty of measuring pedodiversity on the basis of soil‐class maps as the spatial representation of pedodiversity is influenced by the different mapping traditions and methods applied in the 16 Federal States of Germany. On the other hand, it allows the analysis of the aggregation schemes of different landscapes. The presented approach helps to understand large soilscapes and to compare different soil maps of different states and countries as well as to enhance the soil map with additional information. Furthermore, the representative patches can be used to select soil‐monitoring areas.     

半干旱沙区土类/亚类的遥感调查制图方法

传统土壤调查制图存在低时效性、低精度等问题。为了解决半干旱沙区土壤遥感调查制图问题,该文以科尔沁左翼后旗为例,基于野外实地调查和专家知识分析了半干旱沙区土壤类型分布特征与环境因素之间的关系,并探讨了基于多时相Landsat8 OLI影像数据的半干旱沙区土类/亚类遥感调查制图方法。结果表明:利用多时相Landsat8 OLI影像数据提取的归一化差异水体指数(modified normalized difference water index,MNDWI)、盐分指数(salt index,SI)、归一化差异湿度指数(normalized difference moisture index,NDMI)、归一化差异植被指数(normalized difference vegetation index,NDVI)等环境信息,可实现对沼泽土、盐碱土、草甸土、风沙土及其亚类等半干旱沙区主要土壤类型的遥感调查制图。应用本文提出的半干旱区土类/亚类遥感调查制图方法对科左后旗进行土壤遥感调查制图和精度验证,总体精度约为72.84%,Kappa系数为0.667 8。该方法可为半干旱沙区数字土壤调查制图提供思路和参考。     

Small scale digital soil mapping in Southeastern Kenya

Digital soil mapping techniques appear to be an interesting alternative for traditional soil survey techniques. However, most applications deal with (semi-)detailed soil surveys where soil variability is determined by a limited number of soil forming factors. The question that remains is whether digital soil mapping techniques are equally suitable for exploratory or reconnaissance soil surveys in more extensive areas with limited data availability. We applied digital soil mapping in a 13,500 km² study area in Kenya with the main aim to create a reconnaissance soil map to assess clay and soil organic carbon contents in terraced maize fields. Soil spatial variability prediction was based on environmental correlation using the concepts of the soil forming factors equation. During field work, 95 composite soil samples were collected. Auxiliary spatially exhaustive data provided insight on the spatial variation of climate, land cover, topography and parent material. The final digital soil maps were elaborated using regression kriging. The variance explained by the regression kriging models was estimated as 13% and 37% for soil organic carbon and clay respectively. These results were confirmed by cross-validation and provide a significant improvement compared to the existing soil survey.     

Characterization of some Middle European soil textures by gamma‐spectrometry

Information about the variability of different soil attributes within a field is essential for sustainable land management and precision agriculture. Mobile proximal gamma‐ray spectrometry can map soil characteristics of vast areas at different scales rapidly and cost‐effectively. This study aims at investigating reliability and capability of mobile‐gamma‐spectrometry (radiometrics) data to map typical soils of Middle Europe. In this paper, we investigate relationships between the radioelement concentrations (K, U, Th, and dose rate) and soil parameters (texture, CEC, pH, and organic‐C content) at four different field sites and soil textures. The data reliability is confirmed at the survey start. Mobile data have an excellent linear correlation (nearly 1:1) with the stationary readings (of identical devices, acquisition setups, and soil conditions) but moderate correlation with laboratory data (of different devices, setups, and sample conditions). Dried lab samples have systematically higher radioelement concentrations than the field soils (normally wet). Consequently, the mobile‐gamma‐spectrometric data is sufficiently accurate for soil mapping, and its calibration by laboratory data is less useful due to the varying environmental conditions. Single absolute radioelement concentrations show only moderate correlations with the different soil parameters, particularly clay content and CEC. This may be related to varying environmental conditions (soil moisture, soil structure, vegetation, land use, etc.) between the study sites. Investigations of the ratios of radioelement concentrations yield a clear improvement of their correlations to soil parameters, especially for sand and clay contents, CEC, and organic C. Additionally, multiple‐linear‐regression models were established using the element concentrations of potassium and thorium to predict silt content and pH. The results of the highly correlated models were confirmed by comparing with clay and silt content and pH value, respectively, to six additional independent field samples. Briefly, applications of gamma‐ray data for soil mapping offers the possibility of the development of quantitative relationships regarding soil parameters like sand and clay contents, CEC, and organic C. Classification of soil textures by gamma‐ray data seems to be promising, though a broader database of soils is needed for further research. We recommend gamma‐ray mapping as a complementary or even an alternative to common mapping techniques.     

Digital mapping of soil carbon sequestration potential with enhanced vegetation cover over New South Wales,Australia

Digital soil maps of soil organic carbon (SOC) sequestration potential resulting from a hypothetical 10% relative increase in long-term vegetation cover are presented at 100-m resolution across the state of New South Wales (NSW) in southeast Australia. This land management outcome is considered realistically achievable for many land managers, using strategies such as revegetation, grazing management or crop residue management. A mean state-wide potential increase of 5.4 Mg ha⁻¹ over the 0- to 30-cm depth interval was derived. Assuming a 20-year period of re-equilibration, this equates to an average SOC increase of 0.27 Mg ha⁻¹ year⁻¹. Sequestration potential is systematically influenced by a combination of climate, soil parent material and current vegetation cover, for example only 1.6 Mg ha⁻¹ SOC under dry conditions in sandy, infertile soil material with sparse vegetation cover, compared with 15.9 Mg ha⁻¹ under wet conditions in clay-rich, fertile soil material with moderate–high vegetation cover. The outputs could be used to identify locations of highest sequestration potential and thereby help prioritize areas and inform decisions on sequestration programmes. Future application of the method at field scale with high levels of accuracy, together with strategic sampling, may provide statistically reliable estimates of carbon sequestration, for application in carbon trading schemes such as Australia's Emissions Reduction Fund. The modelling involved a conceptually transparent ‘space-for-time substitution’ process. Multiple linear regression (MLR) and random forest (RF) modelling techniques were applied, but only MLR gave consistently meaningful results. The apparent failing of RF in this application warrants further examination.     

Filling the European blank spot—Swiss soil erodibility assessment with topsoil samples

Soil erodibility, commonly expressed as the K‐factor in USLE‐type erosion models, is a crucial parameter for determining soil loss rates. However, a national soil erodibility map based on measured soil properties did so far not exist for Switzerland. As an EU non‐member state, Switzerland was not included in previous soil mapping programs such as the Land Use/Cover Area frame Survey (LUCAS). However, in 2015 Switzerland joined the LUCAS soil sampling program and extended the topsoil sampling to mountainous regions higher 1500 m asl for the first time in Europe. Based on this soil property dataset we developed a K‐factor map for Switzerland to close the gap in soil erodibility mapping in Central Europe. The K‐factor calculation is based on a nomograph that relates soil erodibility to data of soil texture, organic matter content, soil structure, and permeability. We used 160 Swiss LUCAS topsoil samples below 1500 m asl and added in an additional campaign 39 samples above 1500 m asl. In order to allow for a smooth interpolation in context of the neighboring regions, additional 1638 LUCAS samples of adjacent countries were considered. Point calculations of K‐factors were spatially interpolated by Cubist Regression and Multilevel B‐Splines. Environmental features (vegetation index, reflectance data, terrain, and location features) that explain the spatial distribution of soil erodibility were included as covariates. The Cubist Regression approach performed well with an RMSE of 0.0048 t ha h ha^?1 MJ^?1 mm^?1. Mean soil erodibility for Switzerland was calculated as 0.0327 t ha h ha^?1 MJ^?1 mm^?1 with a standard deviation of 0.0044 t ha h ha^?1 MJ^?1 mm^?1. The incorporation of stone cover reduces soil erodibility by 8.2%. The proposed Swiss erodibility map based on measured soil data including mountain soils was compared to an extrapolated map without measured soil data, the latter overestimating erodibility in mountain regions (by 6.3%) and underestimating in valleys (by 2.5%). The K‐factor map is of high relevance not only for the soil erosion risk of Switzerland with a particular emphasis on the mountainous regions but also has an intrinsic value of its own for specific land use decisions, soil and land suitability and soil protection.     

Soil–landscape variability: mapping and building detail information for soil management

The study of soil–landscape relationships at a detailed scale (1:10 000) and its use for soil management was less common in developing countries. The study was conducted in western Ethiopia with the aim to explain the soil variability across landscapes, classify soils into mapping units and produce a map of these soils. This study was performed based on a discrete model of spatial variation. Five soil reference groups: Vertisols, Cambisols, Fluvisols, Luvisols and Leptosols were identified in the study site. Distribution of the soil reference groups was determined by landscape position. Variation in soil texture, colour, pH , exchangeable acidity, organic carbon, total nitrogen, available phosphorus (av. P), carbon to nitrogen ratio (C/N), exchangeable potassium (K), calcium (Ca), magnesium (Mg), sodium (Na) and cation exchange capacity (CEC ) was observed within and among soil mapping units (SMU s). Variability was considerably high for exchangeable Ca and CEC . Factor analysis result indicated that variation in soil properties within land unit was comparatively highest in Leptosols of SMU 9 (88.87%) and lowest in Vertisols of SMU 1 (60.82%). Moderate‐to‐fine scale mapping of soil properties helps to build detail information for soil management. Grouping fields into mapping units that require more or less similar management measure would be an important soil–landscape concept. As a result, mapping units could be used as cost‐effective means of treating variable field so as to optimize the forecasted benefits.