Quick Search:

uzh logo
Browse by:

Zurich Open Repository and Archive

Maintenance: Tuesday, 5.7.2016, 07:00-08:00

Maintenance work on ZORA and JDB on Tuesday, 5th July, 07h00-08h00. During this time there will be a brief unavailability for about 1 hour. Please be patient.

Permanent URL to this publication: http://dx.doi.org/10.5167/uzh-30621

Bartholomeus, H. The influence of vegetation cover on the spectroscopic estimation of soil properties. 2009, Wageningen University, The Netherlands, Faculty of Science.

[img] PDF - Registered users only


Soils and its resources are of major importance for the production of fiber and food. Recently, the soil functioning as carbon pool has been added to the list of soil functions. To determine soil quality as a resource there is a need for regular monitoring of its chemical and physical properties, both in time and space. Quantitative estimation of the exact amount, spatial distribution and temporal change of soil properties is still challenging. Conventionally, soil samples are analyzed by means of soil extraction procedures. Spatially, soil samples are collected according to a specific sampling scheme. Spatial interpolation techniques are used to prepare continuous maps, but for accurate interpolation intensive sampling is required. To achieve the high sampling density required to map the high spatial variability and slow temporal changes in soil properties new techniques are required. Visible and Near Infrared (VNIR) spectroscopy is a promising technique for the quantitative estimation of soil properties and is used frequently for laboratory and field studies. However, the developed models are usually location specific and bridging the gap to imaging spectroscopy introduces spectral mixing problems. The topic of this thesis is how to link spectral reflectance information to soil properties. This is done by the development of statistical models, which can be divided in two classes: multivariate models (PLSR) and univariate models (spectral indices). The objectives were to investigate the robustness of VNIR spectroscopy based soil property prediction models and the influence of vegetation on these models. Furthermore, methods were developed for quantitative mapping of soil properties in fractionally vegetation covered agricultural fields. First, the scaling issues that play a role in VNIR spectroscopy are described. Spatial up-scaling needs to be dealt with when the step from point spectroscopy to imaging spectroscopy is made. Mixed pixels, in our case because of fractional vegetation cover, limit the direct transition of techniques and models from point spectroscopy to imaging spectroscopy. The use of different sensors to acquire spectral information introduces spectral scaling problems, which complicates model transfer from one sensor to another. Finally, directional scaling issues should be considered, including fixed protocols for the measurement setup. Next, the accuracy and robustness of Partial Least Squares Regression (PLSR) and indices based models to estimate soil properties are investigated. Both with PLSR and indices it is possible to get good model calibrations and predictions for SOC. For the soil iron content, indices in the VIS and beginning of the NIR can be used to get qualitative information, but the full spectrum should be used to get quantitative information about iron content. Spectral indices become inaccurate when they are used to estimate SOC beyond the range of values that were used in the calibration phase, but PLSR is less sensitive to extrapolation. In general, model calibrations are location-specific. Including soil-types that are not used for model calibration leads to large errors in prediction. Transfer of the calibration of one year to another causes a decrease in accuracy, which indicates that practical implementation of spectroscopy requires stable measurement conditions, good calibrations and standard sampling protocols. If we want to get rid of the local calibrations, we will have to accept lower accuracies, but we gain in the ease of use. Users willing to choose one of the techniques for a particular application have to weigh the accuracy against the area to be covered. While laboratory spectroscopy has the advantage to allow stable calibration trough time, it still requires sample preparation (drying, sieving, grinding). Field spectroscopy can reach an accuracy comparable to laboratory spectroscopy, but requires calibration before each campaign. Imaging spectroscopy might be a practical way to spatially evaluate soil properties on large scales. The large number of samples compensates for the lower accuracy. Also for imaging spectroscopy, a calibration before each campaign will be necessary. Imaging spectroscopy based models lack behind in prediction accuracy when the results are studied at the pixel level. The lower prediction accuracies are the result of 1) instrumentation specifications (spectral resolution, low instrument signal-to-noise ratios), 2) disturbing external factors (atmospheric attenuation, geometric and optical distortions, mixed pixels) and 3) internal factors (soil moisture, structure). Remote sensing studies often use the per pixel accuracies as a measure of the quality of imaging spectroscopy derived products. However, for most applications (e.g. precision agriculture, regional carbon stock estimates) a per pixel knowledge of the soil properties is not required. The power of imaging spectroscopy lies in the ease with which a high sampling density can be reached and the spatial distribution of soil properties can be determined, which offers great potential in comparison to traditional soil sampling techniques. The influence of vegetation on the predicted soil properties is large. Already, with low amounts of fractional vegetation cover the predictions becomes inaccurate. How much the predicted soil property deviates from the measured value depends on the amount of fractional vegetation cover, the index used and the amount of the soil property. Despite the strong influence of fractional vegetation cover on soil reflectance, it is possible to determine soil properties in partially vegetated areas. This can be done by processing of the output of multiple indices, or by removing the vegetation influence from the spectral measurements in an early stage of the processing chain. The fact that two iron indices react differently on the influence of vegetation can be used to minimize the vegetation influence. Averaging the results of two indices gives a more accurate iron prediction than using one of these indices individually. How large the improvements are depends on the vegetation type, (local) calibration of the model and soil type. A method that can be applied more general is Residual Spectral Unmixing (RSU). With this approach it is possible to filter the influence of vegetation from the mixed spectra, and the residual soil spectra contain enough information to map the SOC distribution within agricultural fields with a good accuracy. Because RSU deals with the vegetation influence in an early stage of the processing chain, it can easily be implemented for other areas. The vegetation influence is removed before the soil property prediction model is calibrated, which gives more flexibility in choice for the prediction models to be used. Finally, it was concluded that VNIR spectroscopy can be used to estimate soil properties on laboratory, field and image scale. The achieved accuracy for the laboratory and field techniques are good enough for monitoring of the small temporal changes in carbon stock. The high sampling density of imaging spectroscopy offers great potential in comparison to traditional soil sampling techniques. However, high prediction accuracies can only be achieved when models are locally calibrated. In the case of field spectroscopy and imaging spectroscopy calibration for every campaign will be necessary. To get models that are robust to variation in soil types, PLSR should be used. Imaging spectroscopy derived products have low prediction accuracies when evaluated at the pixel level, but can be used to evaluate changes in soil properties over time, or to support the division of fields in management zones for precision agriculture applications. Vegetation has a large influence on soil reflectance and cannot be ignored when soil properties are estimated from mixed reflectance spectra. With a combination of indices or by using advanced pre-processing of the spectral information it is possible to derive spatially continuous information about soil properties from imaging spectroscopy data in partially vegetated areas.


3 downloads since deposited on 19 Feb 2010
0 downloads since 12 months

Detailed statistics

Additional indexing

Item Type:Dissertation
Referees:Schaepman M E, Kooistra L
Communities & Collections:07 Faculty of Science > Institute of Geography
Dewey Decimal Classification:910 Geography & travel
Deposited On:19 Feb 2010 08:33
Last Modified:05 Apr 2016 13:55
Number of Pages:146
Additional Information:This object is not available until: 2010-09-30 - see url below
Official URL:http://www.grs.wur.nl/UK/Publications/phd/2009/

Users (please log in): suggest update or correction for this item

Repository Staff Only: item control page