6.4.3.1 Prediction of mineral categories using spectroscopy
The overall accuracy of the MNL models using the original RS data were relative high (> 0.74) (Table 6.7). The smoothed RS data, regardless of scale, improved the overall accuracy little, except for mica using medium-scale RS data. Nevertheless, using a combination of medium and long-scaled RS data substantially improved the overall accuracy of the models. The probability of a sample belonging to a specific class also increased using a combination of medium and long-scaled RS data (Fig. 6.6). Many samples were positioned around the center of the diagram using the original RS data, indicating substantial confusion between the categories. This improved by using scaled RS data, because many samples moved towards the corners of the diagrams. The remaining samples located along the axes indicated confusion between two categories, which was largest for absence and presence in a calcite-poor environment. High confusion between the smectite classes remained, despite using scaled RS data (Fig. 6.6c).
Table 6.7: Overall classification accuracy of the predicted mineralogy by MNL using the sample data.
Mineral Pixel Medium Long Medium + Long
Mica 0.74 0.70 0.79 0.76
Smectite 0.78 0.79 0.74 0.89
Kaolinite 0.75 0.79 0.76 0.86
6.4.3.2 Prediction of mineral abundances based on x-ray diffraction analysis
Models using the original RS data resulted in low coefficients of determination and high RMSE values for all minerals and were deemed unsuitable for predicting mineral abundances (Table 6.8). The use of individual scaled RS data had mixed effect on the prediction models and most accurate models were derived for calcite and mica. We found that the prediction models resulted in highest accuracy using both the medium and long-range RS data. The mica and calcite models performed well for both the medium and the combined ranges; the model statistics showed high Radj
2
and lowest RMSE for predictions from the calibration and cross-validation. The goodness-of-fit for smectite indicated that 57% of the variation was explained by the model. The model for kaolinite was limited to explain 45% of the variation. These results indicated that the use of smoothed data improved model performance.
6.4.4 Digital soil mapping for mineral characterization
6.4.4.1 Application of prediction models to the study area
This section presents the spatial explicit prediction models for mica because of its high predictive accuracy (Table 6.8). The effect of using medium (Fig. 6.7b) and long-range variability (Fig. 6.7c) was clearly observed. Employing the smoothed data, areas of high and low abundances were well differentiated whereas the original data produced a map that hardly showed spatial patterns. Model predictions in areas interpolated with FRK seamlessly joined most of the spatial patterns of areas where RS data were available. The mineralogy was not well characterized in the nodata- area in the East, running from North to South using medium-scaled RS data. In this area, the accumulated FRK prediction errors (Fig. 6.4) were too high and the
Table 6.8: Model accuracy for the predicted minerals by MLR
Mineral Scale Radj2 RMSEcal RMSEval
Calcite Pixel 0.42 12.8 14.2 Medium 0.54 10.6 17.3 Long 0.34 12.7 22.9 Medium + long 0.71 8.9 12.0 Mica Pixel 0.46 3.4 3.7 Medium 0.61 4.9 9.0 Long 0.55 5.2 8.9 Medium + long 0.70 4.6 6 Smectite Pixel 0.53 7.4 10.2 Medium 0.29 9.2 25.8 Long 0.36 7.8 19.5 Medium + long 0.57 6.0 8.3 Kaolinite Pixel 0.26 2.9 3.2 Medium 0.10 3.6 7.9 Long 0.24 2.9 4.5 Medium + long 0.45 2.5 3.8
predicted abundances disrupted the smooth spatial patterns observed in the map (Fig. 6.7d). The mapped calcite abundances were overestimated in the 30 wt. % and 50 wt. %-interval (Fig. 6.8a) causing a shift in the median of 4.5 wt. % (p=0.02). Similarly, mica was overestimated in the 15 wt. %-interval (Fig. 6.8b) causing a shift in the median of 1.5 wt. % (p=0.02). The distribution of modelled and measured kaolinite abundances were similar (p=0.05) (Fig. 6.8c). Smectite abundances were predicted outside the range of measured abundances (Fig. 6.6d), yet, they originated from the same distribution (p=0.53). The modelled and measured mineral categories did not have a similar distribution (p < 0.01) using the initial thresholds (Table 6.9). Increasing the threshold of mica to 5 wt. % resulted in a similar distribution (p = 0.51). 22 wt. % of the mapped kaolinite falsely represents the presence of kaolinite in either a calcite-rich or -poor environment (p < 0.01). Smectite required a lower calcite threshold (3 wt. %, p = 0.63).
Table 6.9: Proportional distribution of the sampled and modelled mineral categories.
Mineral class Mica Sample Sample* Kaolinite Smectite
(mica 5 wt. %) Map Sample Map Sample Sample* (calcite 3 wt. %) Map Presence Ca-rich 0.35 0.24 0.20 0.23 0.29 0.21 0.34 0.40 Presence 0.49 0.40 0.45 0.40 0.56 0.27 0.14 0.13 Absence 0.15 0.36 0.35 0.37 0.15 0.52 0.52 0.47 Figure 6.7: Predicted mica based on XRD data and using the (a) original RS data, (b) RS data representing medium-range variability, (c) RS data representing long-range variability and (d) combination of scaled RS data.
* adjusted threshold in abundance (wt. %), indicating the detection threshold of MICA.
6.4.4.2 Evaluation of predicted soil mineral composition using spectroscopy or x-ray diffraction
The overall similarity between the maps with reclassified abundances and mineral categories were relative low (between 0.34 and 0.44, Table 6.10). Nevertheless, substantial agreement between to two approaches was found using a more qualitative comparison using the error matrices and observed spatial patterns. Areas with high abundances (Fig. 6.9) coincide with the minerals classified as present either in a calcite- rich or poor environment (Fig. 6.10 and Table 6.10). High concentrations of calcite (Fig. 6.9d) showed similar patterns as mica (Fig. 6.10a) and kaolinite (Fig. 6.10b) in a calcite-rich environment, although, the error matrices indicated relative high misclassifications. The absence of smectite had high similarity (Table 6.10c) but the other two classes deviated strongly, which agrees with the findings in section 6.3.2.1. High user accuracy was found for the presence of kaolinite, but substantial spatial differences were observed, especially in the North (Fig. 6.9b and Fig. 6.10b). As a result, the error matrix shows low agreement between the mapped absence of kaolinite. Finally, the effect of the FRK prediction errors were less pronounced in the mapped mineral categories.
Figure 6.8: Histograms of the sampled (right axis) and mapped (left axis) abundances (wt. %) of (a) calcite, (b) mica, (c) kaolinite and (d) smectite.
Figure 6.9: Predicted soil mineral abundances, based on XRD data and scaled RS data.
Table 6.10a: Error matrix (%) mica characterization. Reclassified
abundances Presence Ca-rich Presence Absence User accuracy Presence Ca-rich 0.13 0.28 0.14 0.23 Presence 0.04 0.13 0.12 0.43 Absence 0.03 0.04 0.07 0.50 Producer’s Accuracy 0.64 0.28 0.22 0.34*
*Overall accuracy
Table 6.10b: Error matrix (%) kaolinite characterization. Reclassified
abundances Presence Ca-rich Presence Absence User accuracy Presence Ca-rich 0.16 0.24 0.08 0.34 Presence 0.04 0.12 0.02 0.66 Absence 0.08 0.19 0.05 0.16 Producer’s Accuracy 0.57 0.22 0.34 0.34*
*Overall accuracy
Table 6.10c: Error matrix (%) smectite characterization. Reclassified
abundances Presence Ca-rich Presence Absence User accuracy Presence Ca-rich 0.16 0.06 0.17 0.40 Presence 0.10 0.04 0.07 0.20 Absence 0.13 0.02 0.22 0.60 Producer’s Accuracy 0.40 0.34 0.47 0.43* *Overall accuracy
6.5 Discussion
6.5.1 Main findings 6.5.1.1 Data collectionThe MICA analysis applied to the samples had difficulties in matching some minerals present in the sample with the spectral library (Mulder et al., 2012b). The MICA analysis used in this study was adapted from an approach developed at the remote sensing level of spectroscopy (Kokaly et al 2011a) rather than the laboratory level. To improve the analysis, the expert-decision rules need to be further refined to match the level of detail of high-resolution spectral data. More specifically, this involves specific adjustments of the characteristics describing the diagnostic absorption features. Despite that, MICA delivers insightful information for large- scale studies, consisting of different mineral categories. This provides important information on parent material and soil formation. The highest interest for regional
scale studies, are the coarse differences and general spatial patterns of minerals. Therefore, spectroscopy is considered to be an important and cost efficient way to obtain mineral information. In addition, mineral compositions derived using the MICA algorithm are based on absorption features in reflectance signatures, thus the basis of mineral identification is similar to absorption processes causing variations in the satellite remote sensing data. By contrast, XRD results are based on different types of processes arising in the interaction of electromagnetic energy with the sample.
In this work, x-ray analysis was used to obtain mineral abundances. This involved the use of costly equipment while the analysis was labour-intensive. Alternatively, mineral abundances could be estimated using other spectral modelling approaches, such as proposed by Mulder et al. (2013).
6.5.1.2 Variogram analysis and Fixed Rank Kriging
The different scales of spatial variability within the area were determined by variogram analysis (Fig. 6.3). The high nugget values (20% to 43% of the total semivariance) suggested the presence of a substantially amount of noise in the original RS data. Also, for some RS data more variance was captured by the long- range structure than the medium-scale structure. These findings were further used in the FRK. Here, we found lower prediction errors for RS data having lower sill values at the medium-scale. This indicated that the basis functions had more difficulty in capturing the spatial process for strongly spatially correlated RS variables. The prediction errors might be further reduced by increasing the number of basis functions. At the long-scale the opposite was found, this indicated that the increased nugget reduced the prediction accuracies for those RS data having lower spatial correlation. Overall, the resulting FRK prediction errors and the relation with the variogram were confirm the expectations (Isaaks and Srivastava, 1989).
FRK generated a full coverage by interpolating missing areas with low FRK prediction errors up to a range of 6 kilometres. The missing data in the Eastern part of the area, running from North to South (120 km) having a width of 10 kilometre, appeared to be too large. The FRK positioned the medium-range basis functions at a resolution of 5 and 9.5 kilometre. As a result, there were too few observations available to accurately interpolate. The resulting prediction errors accumulated in the mineral prediction models, increased the average prediction errors (Table 6.6) and resulted in unrealistic predictions. As such, the accuracy of interpolation by FRK is mainly defined by the extent of the missing data and the desired map scale. Discrepancies between the latter substantially increase the prediction errors. The MNL was more robust for this phenomenon because the RS data was related to three categories rather than a continuous range of abundances.
6.5.1.3 Prediction models for soil mineralogy
Initially, the effect of smoothing RS data seemed relatively small in relation to mineralogy (Fig. 6.5). However, the prediction models clearly indicated that using original RS data resulted in lower model performance (Table 6.6 and 6.8). Noise in the original data (i.e. short scale variability) was removed by the FRK, leading to improved relations of smoothed RS data with mineralogy. In addition, the model accuracy and sample sizes of kaolinite were smaller than for calcite and mica. This may indicate that the MLR might have been hampered by sample size rather than by the soil-landscape relationships. Alternative to MLR, different analysis might improve the prediction of mineralogy. MLR assumes a linear relationship of mineralogy towards the predictor variables but this relation might be more complex. Recursive partitioning of the data, for example, can deal with nonlinearity and interactions between predictor variables (Breiman et al., 1984).
The most accurate models utilized a combination of scaled RS data. This indicates that variability of environmental factors explain soil variability better at a specific scale. In addition, the resulting spatial patterns of the soil system’s variability at regional scale were more representative. Vasques et al. (2012) also found that the use of medium and long-range variability observed in spatial data were needed to model major soil variability at large scales. They found that soil carbon was controlled by ecological processes that evolved and interacted over a range of spatial scales across the landscape. They suggested there were two appropriate scales to observe soil carbon and their results supported our use of multi-scale soil-landscape relationships.
6.5.1.4 Evaluation of maps
The soil mineral maps showed similarity between both approaches (Fig. 6.9 and 6.10). Unfortunately, the overall accuracy was rather low (between 0.34 and 0.44, Table 6.10). The modelled calcite and mica abundances were overestimated. As a result, the thresholds which were based on the measured abundances, were too low (Fig. 6.8). This resulted in higher misclassification of the presence of mica or kaolinite in a calcite-rich or -poor environment (Table 6.10a, b). Besides, comparison of kaolinite abundances (Fig. 6.9b) and categories (Fig. 6.10b) may not be that informative. Typically, kaolinite can be well discriminated by spectral analysis and the MNL model resulted in accurate models, while the MLR model was weak (R2 = 0.45) (Table 6.7 and Fig. 6.6). This may imply that the MICA approach was more reliable for kaolinite. Reflecting on the methods and findings of this paper several sources of errors could be identified that contributed to the disagreement, (1) the reclassification of abundances into mineral categories was subjective due to the unknown spectral detection thresholds and (2) soil mineral abundances were less accurately modelled due to the accumulation of the FRK predictions errors. As such, major differences resulted from the classification of absence in areas where low or unrealistic abundances were mapped. Overall, the spectroscopy-based maps
characterized regional soil mineral composition and defined areas with a high likelihood of occurrence of minerals.
The evaluation in section 6.4.4 was based on the distribution of the collected sample (Mulder et al., 2012a). Here, two assumptions were made, first, the spatial distribution of mineralogy was assumed to be fully sampled. Second, the overrepresented and missing intervals of the LHS were not considered to influence the distribution of mineralogy. The first assumption is inherent to the choice of using cLHS (Minasny and McBratney, 2006). The RS data was assumed to be representative of the landscape and its environmental variables (CLORPT, Jenny, 1941). By sampling the first three principal components of ASTER data and elevation, the sampling aimed to optimally capture the environmental variability which was thus the proxy for soil variability. As such, reflectances are a result of the environmental variables but also an expression of interactions. The precise interactions are unknown but they might influence soil variability, e.g. soil- vegetation feedbacks (Milcu et al., 2012) and, therefore, the sample may not be an optimal representation of the actual soil variability. The second assumption should be further investigated because the realized sample did not fully cover the LH covariate space. Therefore, careful interpretation of the statistical tests employed in section 6.4.4 is in place. Additional field data may be acquired to complete the coverage of the LH, and independent data is deemed necessary to validate the final prediction models and maps.