Academia.eduAcademia.edu

Figure 13 – uploaded by Mogens Humlekrog Greve

See full PDFdownloadDownload figure
Figure 4. Prediction errors in dry and remoistened samples within different sand classes. From Stenberg (2010).

Figure 4 Prediction errors in dry and remoistened samples within different sand classes. From Stenberg (2010).

Related Figures (33)

The system also acquires soil electrical conductivity (EC) for two soil depths: shallow conductivity EC-SH (0-31cm), deep conductivity EC-DP (0-91cm). The conductivity is recorded together with GPS coordinates and spectral data.
The system also acquires soil electrical conductivity (EC) for two soil depths: shallow conductivity EC-SH (0-31cm), deep conductivity EC-DP (0-91cm). The conductivity is recorded together with GPS coordinates and spectral data.
Figure 3. Maps showing the TOC distribution across the field, based on manual grid sampling and mobile VIS/NIR measurements.
Figure 3. Maps showing the TOC distribution across the field, based on manual grid sampling and mobile VIS/NIR measurements.
Table 1. Statistical results for calibration, cross validation and prediction of chemical and biochemical characteristics, proportions of EOM organic C mineralized after 3 and 91 days of incubation (C3q and Co1q), Indicator of Residual Organic C (Iroc) and RothC model initial pool (DPM), Values expressed as % of Dry Matter (DM), as % of Organic Matter (OM), as % of added C and as % of Organic Carbon (OC). nl: number of samples after elimination of calibration outliers, n2: number of samples in the validation set.
Table 1. Statistical results for calibration, cross validation and prediction of chemical and biochemical characteristics, proportions of EOM organic C mineralized after 3 and 91 days of incubation (C3q and Co1q), Indicator of Residual Organic C (Iroc) and RothC model initial pool (DPM), Values expressed as % of Dry Matter (DM), as % of Organic Matter (OM), as % of added C and as % of Organic Carbon (OC). nl: number of samples after elimination of calibration outliers, n2: number of samples in the validation set.
Figure. 1. Measured values of the Iroc indicator (i.e. calculated from measured SOL, CEL and LIC biochemical fractions and C3q) and fitted values of the DPM pool of the RothC model (i.e. fitted to C mineralization kinetics of EOM during laboratory incubations) vs. corresponding values predicted using NIR spectra. Black circles: calibration set; white circles: validation set.
Figure. 1. Measured values of the Iroc indicator (i.e. calculated from measured SOL, CEL and LIC biochemical fractions and C3q) and fitted values of the DPM pool of the RothC model (i.e. fitted to C mineralization kinetics of EOM during laboratory incubations) vs. corresponding values predicted using NIR spectra. Black circles: calibration set; white circles: validation set.
“ Number of PLS components used in the model. n is models using NIR measurements, v is models using VisNIR measurements ** Including 75% of all reference samples (91 samples) in the calibration and using the remaining 25% as validation samples.
“ Number of PLS components used in the model. n is models using NIR measurements, v is models using VisNIR measurements ** Including 75% of all reference samples (91 samples) in the calibration and using the remaining 25% as validation samples.
SEcy = Standard error of cross validation, calculated in the calibration step Table 1. NIRS calibration statistics, Bioforsk Loken
SEcy = Standard error of cross validation, calculated in the calibration step Table 1. NIRS calibration statistics, Bioforsk Loken
Figure 2. NIRS versus reference values of clover content (%) in grass mixtures (+/- SEcy ¢ lines). Samples artificially mixed by dry matter weight.
Figure 2. NIRS versus reference values of clover content (%) in grass mixtures (+/- SEcy ¢ lines). Samples artificially mixed by dry matter weight.
= forage constituent combination not available. = determined traditional, quality of the NIR-results does not meet the quality standards. Table 1. Available forage constituent combinations within Samplinq®, marked by the RMSEPfyixea values.
= forage constituent combination not available. = determined traditional, quality of the NIR-results does not meet the quality standards. Table 1. Available forage constituent combinations within Samplinq®, marked by the RMSEPfyixea values.
Table 3. Results produced for the running validation on grass silage in 2008 for the different sites.
Table 3. Results produced for the running validation on grass silage in 2008 for the different sites.
Table 2. Summary of the results produced for the Daily Reference sample in March at the different sites.
Table 2. Summary of the results produced for the Daily Reference sample in March at the different sites.
Table 4. Results produced for the proficiency test in the Sampling® network at 2008 for both the grass silage samples. Only for the Polish site z-scores above 3 were found. These high values can be caused by sample shipment. The shipment of the samples took more than a week due a strike of the Polish mail service. In result some damage may be presumed especially the grass silage samples. The overall results for the proficiency test are found sufficient, although some improvements can be made. Next time also a couple of prepared samples will be included in the proficiency test to evaluate the results without sample preparation and the risk of deterioration of samples.
Table 4. Results produced for the proficiency test in the Sampling® network at 2008 for both the grass silage samples. Only for the Polish site z-scores above 3 were found. These high values can be caused by sample shipment. The shipment of the samples took more than a week due a strike of the Polish mail service. In result some damage may be presumed especially the grass silage samples. The overall results for the proficiency test are found sufficient, although some improvements can be made. Next time also a couple of prepared samples will be included in the proficiency test to evaluate the results without sample preparation and the risk of deterioration of samples.
Clay mineral absorptions are mostly due to OH, H2O and CO; overtones and combinations. Kaolinite absorb characteristically at doublets near 2200 and 1400 nm (Fig. 2). Illite and montmorillonite has single bands near 1400 and 2200 nm, and also near 1900 nm. Illite has weaker bands also near 2300 and 2400 nm. In illite and montmorillonite the 1400 and 1900 bands are due to combination vibrations of water bound in the interlayer lattices. Such water is present at larger abundance in montmorillonite and not present at all in kaolinite. Therefore absorption bands are weaker in illite and the 1900 band not at all in kaoline (Fig. 2). The 1400 absorption in kaolinite is due to overtones of OH vibrations of structural water. The 2200 absorptions are due to AI-OH plus OH combinations. For these and other minerals see e.g. Clark (1999) and Hunt and Salisbury (1970).
Clay mineral absorptions are mostly due to OH, H2O and CO; overtones and combinations. Kaolinite absorb characteristically at doublets near 2200 and 1400 nm (Fig. 2). Illite and montmorillonite has single bands near 1400 and 2200 nm, and also near 1900 nm. Illite has weaker bands also near 2300 and 2400 nm. In illite and montmorillonite the 1400 and 1900 bands are due to combination vibrations of water bound in the interlayer lattices. Such water is present at larger abundance in montmorillonite and not present at all in kaolinite. Therefore absorption bands are weaker in illite and the 1900 band not at all in kaoline (Fig. 2). The 1400 absorption in kaolinite is due to overtones of OH vibrations of structural water. The 2200 absorptions are due to AI-OH plus OH combinations. For these and other minerals see e.g. Clark (1999) and Hunt and Salisbury (1970).
Although SOC and clay calibrations usually can be regarded as robust, the variation betweer prediction results is not insignificant, especially regarding the root mean squared error (RMSE). It is frequently suggested that a smaller soil variation would support better calibrations indicating a scale effect (e.g., Udelhoven et al., 2003). However, reviewing a large number of published prediction results it was found that the variation in the predicted soil property itself could explain a very large part of the variation in prediction performance (Fig. 3) (Stenberg and Viscarra Rossel, 2010). For SOC, field scale calibrations could be separated, but there was only a tendency that they performed better than large scale calibrations (Fig. 3). Regarding SOC it has also been observed that very sandy soil may be worse predicted than more clayey ones (Fystro, 2002; Stenberg et al., 2002). Recently it was shown that the sandiest soils in a wide data set were over predicted (Fig. 4) (Stenberg, 2010). This could be explained by the dark humus dominating the spectra of relatively featureless quartz (Clark, 1999). Interestingly and largely unexplained, this effect was eliminated by volumetrically standardized remoistening (Fig. 4).
Although SOC and clay calibrations usually can be regarded as robust, the variation betweer prediction results is not insignificant, especially regarding the root mean squared error (RMSE). It is frequently suggested that a smaller soil variation would support better calibrations indicating a scale effect (e.g., Udelhoven et al., 2003). However, reviewing a large number of published prediction results it was found that the variation in the predicted soil property itself could explain a very large part of the variation in prediction performance (Fig. 3) (Stenberg and Viscarra Rossel, 2010). For SOC, field scale calibrations could be separated, but there was only a tendency that they performed better than large scale calibrations (Fig. 3). Regarding SOC it has also been observed that very sandy soil may be worse predicted than more clayey ones (Fystro, 2002; Stenberg et al., 2002). Recently it was shown that the sandiest soils in a wide data set were over predicted (Fig. 4) (Stenberg, 2010). This could be explained by the dark humus dominating the spectra of relatively featureless quartz (Clark, 1999). Interestingly and largely unexplained, this effect was eliminated by volumetrically standardized remoistening (Fig. 4).
Figure 1. On-line fitting of reflectance probe to 120 litre chicken manure digester. into the reactor wall (Figure 1). Digestate samples for GC VFA analysis were collected from ¢ sample port situated at a similar level but on the opposite side of the reactor. The reactor contents were mixed thoroughly by a gas recycling mixing system before and during sample collection and stopped just before NIRS measurement to prevent gas bubbles causing a false measurement. The 100 ml samples drawn off for GC analysis were part of bulk digestate samples of approximately six litres. The remainder of each bulk sample was carefully returned to the digester feed inlet during NIR analysis to cause further agitation within the digester contents and thus a more representative measurement. The same 256 scan averaging used in the previous experiment was employed.
Figure 1. On-line fitting of reflectance probe to 120 litre chicken manure digester. into the reactor wall (Figure 1). Digestate samples for GC VFA analysis were collected from ¢ sample port situated at a similar level but on the opposite side of the reactor. The reactor contents were mixed thoroughly by a gas recycling mixing system before and during sample collection and stopped just before NIRS measurement to prevent gas bubbles causing a false measurement. The 100 ml samples drawn off for GC analysis were part of bulk digestate samples of approximately six litres. The remainder of each bulk sample was carefully returned to the digester feed inlet during NIR analysis to cause further agitation within the digester contents and thus a more representative measurement. The same 256 scan averaging used in the previous experiment was employed.
Figure 3. Total VFA calibration model using data from all four pilot reactors treating pig slurry and maize silage.
Figure 3. Total VFA calibration model using data from all four pilot reactors treating pig slurry and maize silage.
Figure 5. Estimation of chicken manure total VFA using NIRS model created from pig slurry and maize silage data.
Figure 5. Estimation of chicken manure total VFA using NIRS model created from pig slurry and maize silage data.
A study was performed in 2006 in two field trials on Radde experimental farm. The ley trials consisted of a mixture of different grass species, red clover (Trifolium pratense), white clove (Trifolium repens) and in one of the trials chickory (Cichorium intybus). In the study NIR- analysis was performed on the same samples as the regular botanical analysis were performe on. Thus sampling error was minimized. Figure 1 show a good relationship between NIR- spectra and total clover content. The differences between the manually sorted clover content and the NIR predicted content were on average 4.5 % clover. This must be considered to be good. However, this can not always be expected as the trial conditions regarding 1.e. varieties and harvest time were the same in each trial representing both calibration and validation samples. Thus, the distinction between clover and grass would be well defined in the PLS calibration by means which can not be guaranteed for extrapolation to other conditions. This was indicated by using the clover content calibration for one of the trials for prediction of the other trial. This trippled the average deviation between measured and predicted clover content. However, this type of validation, comparing one set of conditions with another, is very rigorous and would overestimate errors compared to a representative calibration set.
A study was performed in 2006 in two field trials on Radde experimental farm. The ley trials consisted of a mixture of different grass species, red clover (Trifolium pratense), white clove (Trifolium repens) and in one of the trials chickory (Cichorium intybus). In the study NIR- analysis was performed on the same samples as the regular botanical analysis were performe on. Thus sampling error was minimized. Figure 1 show a good relationship between NIR- spectra and total clover content. The differences between the manually sorted clover content and the NIR predicted content were on average 4.5 % clover. This must be considered to be good. However, this can not always be expected as the trial conditions regarding 1.e. varieties and harvest time were the same in each trial representing both calibration and validation samples. Thus, the distinction between clover and grass would be well defined in the PLS calibration by means which can not be guaranteed for extrapolation to other conditions. This was indicated by using the clover content calibration for one of the trials for prediction of the other trial. This trippled the average deviation between measured and predicted clover content. However, this type of validation, comparing one set of conditions with another, is very rigorous and would overestimate errors compared to a representative calibration set.
Table 1. Composition of solid manures and cross validation errors obtained from calibration. Figure | shows that dry matter, total N, ammonium-N, MIN-N and total P in the solid manures could be well predicted with the used NIRS methodology with 1° values of 0.96, 0.93, 0.88, 0.92 and 0.90, respectively. Total K was less well predicted. The ratios between analyte variation range standard deviation and the root mean square error of cross validation (SD/RMSECYV) were 2.9-5.3 for dry matter, total N, ammonium-N, MIN-N and total P, but only 1.4 for total K (Table 1).
Table 1. Composition of solid manures and cross validation errors obtained from calibration. Figure | shows that dry matter, total N, ammonium-N, MIN-N and total P in the solid manures could be well predicted with the used NIRS methodology with 1° values of 0.96, 0.93, 0.88, 0.92 and 0.90, respectively. Total K was less well predicted. The ratios between analyte variation range standard deviation and the root mean square error of cross validation (SD/RMSECYV) were 2.9-5.3 for dry matter, total N, ammonium-N, MIN-N and total P, but only 1.4 for total K (Table 1).
LOI = loss-on-ignition, TOC = total organic carbon, TIC = total inorganic carbon, TN = total nitrogen, TS = total sulphur (TS), clay = particle size <2 um. Table 1. Quality standard set based for the SEP-values collected at this soil application. In both cases the z-scores can be evaluated equally in for instance a Shewhart chart (7) regardless the results are below or above the threshold.
LOI = loss-on-ignition, TOC = total organic carbon, TIC = total inorganic carbon, TN = total nitrogen, TS = total sulphur (TS), clay = particle size <2 um. Table 1. Quality standard set based for the SEP-values collected at this soil application. In both cases the z-scores can be evaluated equally in for instance a Shewhart chart (7) regardless the results are below or above the threshold.
Table 2. Composition of the calibration database and the validation set.
Table 2. Composition of the calibration database and the validation set.
Figure 1. NIRS application at Blgg using the Q-Interline FT-NIRS instruments.
Figure 1. NIRS application at Blgg using the Q-Interline FT-NIRS instruments.
Table 3. The number of outliers and accuracy found for validation sets from 2004/2005 (firs limited data set) and 2007/2008 (second, extended data set). 4) UuUulel woo The most used outlier test is the Mahalanobis distance. Although this parameter has no direct relation with prediction uncertainty a better alternative has still to be found. For the second outlier test relations between general soil characteristics and soil constituents like TOC, TN, TS and TIC were investigated. These relations turned out to be valuable to mark potential outliers. In case the predicted NIR-result is marked to be an outlier the sample is sent to the laboratory for analysis by the particular reference method.
Table 3. The number of outliers and accuracy found for validation sets from 2004/2005 (firs limited data set) and 2007/2008 (second, extended data set). 4) UuUulel woo The most used outlier test is the Mahalanobis distance. Although this parameter has no direct relation with prediction uncertainty a better alternative has still to be found. For the second outlier test relations between general soil characteristics and soil constituents like TOC, TN, TS and TIC were investigated. These relations turned out to be valuable to mark potential outliers. In case the predicted NIR-result is marked to be an outlier the sample is sent to the laboratory for analysis by the particular reference method.
Table 4. Second data set (2007/8): additional results calculated from the validation samples. _ SEP = prediction error. > Mean difference between NIRS and wet-chemical analysis; assumed to be acceptable if > -1 and < 1. > Measure for the accuracy (low values are highly accurate); assumed to be acceptable if < 1.4. * Correlation coefficient between NIRS and wet-chemical analysis. > Warning limit. ® Action limit.
Table 4. Second data set (2007/8): additional results calculated from the validation samples. _ SEP = prediction error. > Mean difference between NIRS and wet-chemical analysis; assumed to be acceptable if > -1 and < 1. > Measure for the accuracy (low values are highly accurate); assumed to be acceptable if < 1.4. * Correlation coefficient between NIRS and wet-chemical analysis. > Warning limit. ® Action limit.
Related topics:
NIR spectroscopyORGANIC SOILDiffuse reflectanceCarbon StockFuzzy C-Means Clustering AlgorithmSurface MeasurementPeat SoilElectric ConductivityField capacity
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved