Integrating geospatial and multi‐depth laboratory spectral data for mapping soil classes in a geologically complex area in southeastern Brazil

Soil mapping across large areas can be enhanced by integrating different methods and data sources. This study merges laboratory, field and remote sensing data to create digital maps of soil suborders based on the Brazilian Soil Classification System, with and without additional textural classification, in an area of 13 000 ha in the state of São Paulo, southeastern Brazil. Data from 289 visited soil profiles were used in multinomial logistic regression to predict soil suborders from geospatial data (geology, topography, emissivity and vegetation index) and visible–near infrared (400–2500 nm) reflectance of soil samples collected at three depths (0–20, 40–60 and 80–100 cm). The derived maps were validated with 47 external observations, and compared with two conventional soil maps at scales of 1:100 000 and 1:20 000. Soil suborders with and without textural classification were predicted correctly for 44 and 52% of the soil profiles, respectively. The derived suborder maps agreed with the 1:100 000 and 1:20 000 conventional maps in 20 and 23% (with textural classification) and 41 and 46% (without textural classification) of the area, respectively. Soils that were well defined along relief gradients (Latosols and Argisols) were predicted with up to 91% agreement, whereas soils in complex areas (Cambisols and Neosols) were poorly predicted. Adding textural classification to suborders considerably degraded classification accuracy; thus modelling at the suborder level alone is recommended. Stream density and laboratory soil reflectance improved all classification models, showing their potential to aid digital soil mapping in complex tropical environments.     

Fluctuations in method‐of‐moments variograms caused by clustered sampling and their elimination by declustering and residual maximum likelihood estimation

Soil data accumulated in national and regional archives derive from many sources and tend to be concentrated in zones of particular interest. Experimental variograms computed from such data by the usual method of moments can appear highly erratic, and therefore models fitted to them are likely to be unreliable. We have explored two methods of avoiding the effects, one by computing declustering weights and incorporating them into the method of moments, the other using residual maximum likelihood. The methods are illustrated with data on bulk density, exchangeable magnesium, cation exchange capacity and organic carbon of 4182 samples of soil from numerous soil surveys in the whole of Australia and stored in the CSIRO's national archive. The experimental variograms of all four variables are erratic. Cell declustering produced much smoother sequences of estimates to which plausible models could be fitted with confidence. The residual maximum likelihood models closely matched those models over several hundred km. Finally values were simulated at the same sampling points from the residual maximum likelihood models, reproducing ‘spiky’ experimental variograms such as those computed from the data. The simulation showed that clustered design of sampling causes spiky artefacts. We conclude that where data are clustered experimental variograms should be computed with declustered weighting or variogram models be fitted by residual maximum likelihood.     

A Two-Factor Empirical Deterministic Response Surface Calibration Model for Site-Specific Predictions of Lime Requirement

This paper describes the development of an empirical deterministic two-factor response surface model for the Woodruff lime-requirement buffer (WRF). The model may be used to produce variable-rate lime requirement maps, or to predict lime requirements in real-time. Hence it may be suitable as a component of a decision support system (DSS) for the site-specific management of acid soil. The models' predictions were compared to those of a one-factor response surface, and those of a linear regression. The models tested were validated against soil-CaCO₃ incubations using a statistical jackknifing procedure for error and bias estimations. The Akaike Information Criterion (AIC) was used to ascertain the best model in terms of goodness of fit and parsimony. The two-factor response surface model produced the best lime requirement estimates, followed by the single-factor model, then the conventional linear regression. The advantages of the response surface models are their improved prediction accuracy, and their flexibility in the choice of any target pH (from pH 5.5 to 7) without the need for excessive calibrations. The uncertainty of the model was assessed using data from an agricultural field in Kelso, New South Wales, Australia. Block kriged maps of soil pH measured in 0.01 M CaCl₂ (pH_CaCl2), WRF buffer pH (pH_buffer) and lime requirements to a target pH of 7 were produced, to compare their spatial distributions. Finally the economic and agronomic benefits of site-specific liming were considered.     

Improved estimates of organic carbon using proximally sensed vis–NIR spectra corrected by piecewise direct standardization

We investigated the use of piecewise direct standardization (PDS) to remove the effects of water and other environmental factors from proximally sensed (field) visible–near infrared (vis–NIR) spectra. Our hypothesis was that the PDS‐standardized field spectra can be used to predict soil carbon effectively with calibrations derived from existing spectroscopic databases of spectra recorded in the laboratory on dried, ground and sieved samples. In our experiments we used field spectra recorded in situ with a portable spectrometer at 124 sites in 11 paddy fields in Zhejiang Province, China. We sampled the soil at these same sites, recorded their spectra in the laboratory and measured their soil organic carbon (SOC) contents with a conventional laboratory technique. Two‐thirds of the samples were used to relate the laboratory spectra to SOC by partial least squares regression (PLSR), and the remaining one‐third was used as an independent validation dataset. We selected a representative set of samples from corresponding field and laboratory spectra that we could use as the PDS transfer set. Piecewise direct standardization was used to relate each wavelength in the laboratory spectra to the corresponding wavelength and its neighbours in the field spectra. The field spectra of the validation samples were then corrected with PDS so that they acquired the characteristics of the spectra measured under laboratory conditions. The approach was evaluated by (i) quantifying the similarity between the PDS‐standardized spectra and their corresponding laboratory spectra, (ii) measuring the accuracy of their SOC predictions on the independent validation dataset and (iii) comparing these results with those of direct standardization (DS). Both PDS and DS led to considerable improvements in the predictions of SOC (R² = 0.71, R² = 0.60, respectively), compared with those with original field spectra (R² = 0.03). However, fewer transfer samples were needed with PDS to obtain similar results.     

Prediction of soil organic matter using a spatially constrained local partial least squares regression and the Chinese vis–NIR spectral library

We need to determine the best use of soil vis–NIR spectral libraries that are being developed at regional, national and global scales to predict soil properties from new spectral readings. To reduce the complexity of a calibration dataset derived from the Chinese vis–NIR soil spectral library (CSSL), we tested a local regression method that combined geographical sub‐setting with a local partial least squares regression (local‐PLSR) that uses a limited number of similar vis–NIR spectra (k‐nearest neighbours). The central idea of the local regression, and of other local statistical approaches, is to derive a local prediction model by identifying samples in the calibration dataset that are similar, in spectral variable space, to the samples used for prediction. Here, to derive our local regressions we used Euclidean distance in spectral space between the calibration dataset and prediction samples, and we also used soil geographical zoning to account for similarities in soil‐forming conditions. We tested this approach with the CSSL, which comprised 2732 soil samples collected from 20 provinces in the People's Republic of China to predict soil organic matter (SOM). Results showed that the prediction accuracy of our spatially constrained local‐PLSR method (R² = 0.74, RPIQ = 2.6) was better than that from local‐PLSR (R² = 0.69, RPIQ = 2.3) and PLSR alone (R² = 0.50, RPIQ = 1.5). The coupling of a local‐PLSR regression with soil geographical zoning can improve the accuracy of local SOM predictions using large, complex soil spectral libraries. The approach might be embedded into vis–NIR sensors for laboratory analysis or field estimation.