Published online 16 November 2005
Published in Vadose Zone J 4:1119-1122 (2005)
DOI: 10.2136/vzj2005.0023
© 2005 Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
NOTES
Estimating Surface Soil Moisture from Soil Color Using Image Analysis
Magnus Persson*
Dep. of Water Resources Engineering, Lund Univ., Box 118, SE-221 00 Lund, Sweden
* Corresponding author (magnus.persson{at}tvrl.lth.se)
Received 11 February 2005.
 |
ABSTRACT
|
|---|
In this technical note the ability to estimate surface soil moisture (
) from soil color using image analysis is evaluated. Four natural soils and uniform fine sand were used. Calibration soil samples with varying from 0 to 0.40 m3 m–3 in 0.05 m3 m–3 increments were prepared and photographed. The variations in soil color with were investigated in both the RGB (red, green, and blue) and HSV (hue, saturation, and value) color spaces. Generally, all tested soils became darker when wetted up to a certain limit (around 0.25 m3 m–3). However, many soils actually became lighter again at the highest levels. This was due to that some water was visible on the soil surface causing reflections. A simple linear regression model between S and V was selected to estimate from the soil color. The model performed excellent in the fine sand and in two natural soils with a root mean square error (RMSE) of 0.011 to 0.017 m3 m–3. In the two other soils the RMSE was about 0.025 m3 m–3. An independent validation data set was also collected for the sand. The calibrated model performed well also in the validation data set with a RMSE of 0.015 m3 m–3. From the limited data presented in this study, it seems that the relationship between soil color and is stronger in light colored soils with low organic matter content. Some examples of practical applications of the method are also suggested in the paper.
Abbreviations: HSV, hue, saturation, and value • RGB, red, green, and blue
|
INTRODUCTION
|
|---|
LABORATORY MEASUREMENTS of soil moisture are of great interest in soil science. The gravimetric methods are probably the most commonly used today. These methods include the standard method of oven drying of soil samples, but also putting entire soil columns on a scale and record the weight. Oven drying is, of cause, destructive and can only be performed once for the same soil volume. Measuring the weight of soil columns is more efficient, but will only give an average for the entire column. If the profile in a soil column or the small-scale variability in is of interest, there are few nondestructive options. Examples are various types of small-scale TDR probes or tensiometers (Nissen et al., 1998; Persson and Wraith, 2002; Persson and Haridy, 2003), but to get high spatial resolution, many probes have to be used. Furthermore, insertion of any probe in the soil will interrupt the flow paths. Thus, there is a need for a method that can measure in a nondestructive way at a high spatial and temporal resolution.
During recent years, image analysis has found many applications within soil science. Dye tracers have been used for many years by soil scientists investigating the effects of soil heterogeneity as they allow visualization of spatial flow patterns (see, e.g., Flury and Flühler, 1995). This method has proven very useful for detecting preferential flow paths in the soil. Traditionally, image analysis of the dye photographs has only involved separation between stained and nonstained soil. However, during the 1990s, image analysis improved to the extent that estimation of dye concentration from soil color was possible (Forrer, 1997; Ewing and Horton, 1999). The image analysis involves several corrections of, for example, variation in color temperature of illumination and spatial variation in illumination (see Forrer, 1997 for details). The color as expressed in a specific color space (e.g., RGB or HSV) of the corrected images can then be related to dye concentration. In most cases, an empirical polynomial equation is adopted to model the relationship between RGB or HSV values and dye concentration (Forrer, 1997; Ewing and Horton, 1999; Aeby et al., 2001). Other more complex models like neural networks have also been used (Persson, 2005).
There have also been studies where the soil color has been related to various physical soil properties. Chen et al. (2000) found an exponential relationship between the RGB values of aerial images and soil surface organic matter content. Post et al. (2000) used soil color information to predict soil albedo. It is also a well-known fact that soils become darker when wet, therefore soil color might also be useful for estimations.
The objective of the present study is to investigate if it would be possible to estimate from soil color using image analysis. Soil samples were photographed and the relationship between soil color and was modeled using various empirical models. Several different soil types were tested to examine the effect of soil texture.
|
MATERIALS AND METHODS
|
|---|
Experimental Set-Up and Camera Settings
Four natural soils, named after their origin, and pure silica fine sand were used. The natural soils were all sampled in farm lands at the 0- to 0.1-m depth except the Värpinge soil which was collected at the 0.4-m depth. Selected soil properties are listed in Table 1. Known amounts of distilled water and oven-dry soil were packed to their natural bulk density in standard soil sample retaining rings (Soilmoisture Equipment Corp., Goleta, CA) commonly used in pressure extractors. The rings are 0.01 m high and 0.051 m in diam. The varied from 0 to 0.40 m3 m–3 in increments of 0.05 m3 m–3. However, in some soils it was not possible to pack the soils into the rings at the highest . A knife was used to compact and smooth the soil surface to make the soil surface free from shadows. Three samples were prepared for each , thus up to 27 samples were prepared for each soil. Since the Värpinge soil contained a fair amount of clay, this soil was not oven dried to prevent cementation of clay particles. Instead this soil was air dried for several months. A subsample was then oven dried to calculate the residual . The air-dried soil ( = 0.04 m3 m–3) was then mixed with water in 0.05 m3 m–3 increments as described above.
The soil samples were put on a table and photographed. All photographs were taken using a Nikon D100 digital camera (Nikon Corp., Tokyo, Japan) with a 50-mm lens (AF Nikkor 50 f/1.8D). The camera was placed on a tripod above the table, at approximately 1-m height. Care was taken so that both the camera and the table were aligned horizontally. To reduce vibration a cable release was used to take the pictures. The exposure and white balance was set using a gray card placed on the table (see Bockaert, 2003 and Persson, 2005 for details). To avoid color casts it is important to use a light source with a constant color temperature. In this study ambient light was used (fluorescent tubes). A flat field image of the gray card was used to investigate the need for corrections of the heterogeneity of the illumination (see, e.g., Forrer, 1997; Aeby et al., 2001; Persson, 2005). However, it was found that this was not necessary.
Images were captured in the RGB color space, where R, G, and B are the brightness of the individual red, green, and blue signals. All colors between white (R = G = B = 255 for an eight-bit image) and black (R = G = B = 0) are represented as a mixture of these three colors. The RGB color space have been successfully used to estimate dye concentration (e.g., Forrer, 1997), but also the hue, saturation, and value (HSV) color space have been used (Ewing and Horton, 1999). The RGB color spaces can be transformed into the HSV color space using the following relationships (Hanbury, 2002)
 | [1a] |
 | [1b] |
 | [1c] |
For all soil samples, a representative area of approximately 150 by 150 pixels was selected, representing a physical size of several square centimeters. The median RGB values of this area were calculated. The corresponding HSV values were calculated from the median RGB values according to Eq. [1a] through [1c]. When calculating dye concentration from soil color, an empirical polynomial equation is normally adopted to model the relationship between RGB (or HSV) values and dye concentration (Ewing and Horton, 1999; Forrer, 1997; Aeby et al., 2001). A similar approach was also used in the present study to find a model for calculating from soil color.
Validation
To validate the model, an independent data set was also collected for the fine sand. This was achieved by placing saturated soil samples in a pressure extractor and applying various pressures to drain the samples. The same soil sample retaining rings was used. Immediately after opening the pressure extractor after each run the samples were placed on the table and photographed. In total 18 soil samples were used. Again, a representative area of approximately 150 by 150 pixels was selected of each sample and the median RGB and HSV values of this area was calculated. The of the samples were then determined by drying them in the oven for 24 h at 105°C. The validation data covered a range of 0 to 0.39 m3 m–3.
|
RESULTS AND DISCUSSION
|
|---|
All soils tested became darker when wet, at least initially. However, many soils actually became lighter again at the highest levels. In the soils tested R was highest followed by G and B. There was a high correlation between the RGB values (r > 0.97). The standard deviation of the RGB values for the selected 150 by 150 pixel area of the soil samples varied a lot between the different soils and , between 2 and 15. The highest standard deviations were found when the soil was dry or close to saturation. When the soils were dry it was difficult to achieve a smooth soil surface. Due to the microtopography areas between soil particles or soil aggregates were shaded and thus darker. As the soil was wetted these shadows disappeared to a large extent. When close to saturation, a water film covered the soil surface causing reflections of the light sources. These reflections were very bright, almost completely white. By choosing the median RGB values and not the mean did overcome these problems to a certain degree, but still this resulted in that some soils appeared lighter close to saturation.
The correlation between the RGB values and was high (r < –0.70). The R values had the highest correlation with in all soils, followed by G and B. A more interesting behavior was found when looking at the data in the HSV color space. In Fig. 1
the HSV values are plotted against for the different soils. Since the behavior of the Odarslöv soil was virtually identical to the Lund soil, only the latter soil is presented in Fig. 1. It should be noted that since R always was higher than G and B, the V values was always R/255 and thus the behavior of R and V was identical. The V value always had the largest correlation with . In the fine sand and Värpinge soils, H had a very low correlation with . In the other soils, however, H had a higher correlation with than S. The S value typically increased with at low but then decreased again at high .
Modeling the Soil Color– Relationship
Since R, G, and B had virtually identical behavior as increased, the HSV color space was chosen for the model. Several different models with both first and second-order terms in H, S, and V. Finally, the following empirical model was chosen,
 | [2] |
where a, b, and c are best fit parameters. The results are presented in Table 2. When more terms were used in the model, the fit improved slightly; however, Eq. [2] was chosen due to its simplicity. A second-order polynomial equation between and V was also tested since V had the highest correlation with . This model did not perform as well as Eq. [2]. Clearly S contributed with some information of soil color that was related to that was not found in V. A model similar to Eq. [2] using R and G was also tested, but this was also inferior to Eq. [2].
The model (Eq. [2]) performed excellent in the fine sand, Löddeköpinge and Värpinge soils. It should be noted, however, the limited range for the two latter soils. In the Lund and Odarslöv soils the model performed well in the 0.05 to 0.25 m3 m–3 range. Still the overall RMSE in these soils was acceptable. Probably much of the errors can be related to the difficulties in packing the samples to a homogeneous and bulk density. The fine sand was by far the easiest soil to work with. It should also be noted that the soil color was calculated as the median values of 150 by 150 pixels. If Eq. [2] would be used on individual pixel color values, the RMSE would have been considerably higher.
The validation data set showed a slightly higher variability in estimated using Eq. [2] and the parameters achieved from the calibration, see Fig. 2
. However, the estimation can still be considered acceptable with an RMSE of 0.015 m3 m–3 and an r2 of 0.9859.
From the limited data presented in this study, it seems that the relationship between soil color and is stronger in light-colored soils with low organic matter content. Therefore, the method presented in this paper might be applicable to a limited range of soils. For the pure fine sand, however, the method performed excellent. This kind of material is commonly used in laboratory studies, thus estimating from soil color might be most applicable for laboratory studies. Some examples of potential use are to estimate the profile in transparent soil columns or to study how decrease due to evaporation from soil surfaces. Another interesting application would be in pressure extractors. Normally soil samples are removed from the extractor and weighted when equilibrium has been reached for the current pressure. Thus, each soil sample can only be used once for each pressure. If the lid of the pressure extractor was transparent, the pressure could instead be increased stepwise and the soil samples could be photographed to determine without removing them. Alternatively, a camera chip could be placed inside the extractor.
It should be noted that a photograph of soil only shows the uppermost soil particles. Using the method presented in this study only the surface soil moisture will affect the estimated . When handling soil samples care should be taken to avoid evaporation as the soil surface can dry out in a few minutes.
Since different camera chips (and films) render colors in different ways, the obtained relationship between soil color and is camera specific. If photographs are taken under different lightning conditions, corrections for color cast and differences in illumination has to be considered. This will introduce further uncertainty in the model (see Persson, 2005 for details).
|
SUMMARY AND CONCLUSIONS
|
|---|
The ability of using image analysis for estimating surface soil moisture was investigated for four different natural soils and fine sand. Soil samples with known was prepared and photographed. The relationship between soil color expressed in both the RGB and HSV color spaces and was investigated. It was shown that all tested soils became darker when increased up to a certain limit. However, many soils became lighter again at the highest levels. This was mainly explained by that some water was visible on the soil surface causing reflections seen as bright spots on the soil surface. Furthermore, as the soil was wetter, it was easier to achieve a smooth soil surface avoiding shadows due to the microtopography. The R values had the highest correlation with in all tested soils, followed by G and B. In the HSV color space, V had the highest correlation with followed by S in the fine sand and Värpinge soils and H in the other soils. Several empirical models were tested to predict the relationship between the color information and . In the end, a simple linear regression model between S and V was selected to estimate . The model performed excellent in the fine sand and in two natural soils with a RMSE of 0.011 to 0.017 m3 m–3. In the two other soils the RMSE was about 0.025 m3 m–3. The calibrated model performed well also in the validation data set with a RMSE of 0.015 m3 m–3. The errors are probably mostly related to problems in packing the soil samples to a uniform and bulk density. From the limited data presented in this study, it seems that the relationship between soil color and is stronger in light-colored soils with low organic matter content.
|
ACKNOWLEDGMENTS
|
|---|
This study was funded by the Swedish Research Council.
|
REFERENCES
|
|---|
- Aeby, P., U. Schultze, D. Braichotte, M. Bundt, H. Wydler, and H. Flühler. 2001. Fluorescence imaging tracer distributions in soil profiles. Environ. Sci. Technol. 35:753–760.[Medline]
- Bockaert, V. 2003. The 123 of digital imaging interactive e-book. Version 2.0. Asimex Pte Ltd., Singapore.
- Chen, F., D.E. Kissel, L.T. West, and W. Adkins. 2000. Field-scale mapping of surface soil organic carbon using remotely sensed imagery. Soil Sci. Soc. Am. J. 64:746–753.[Abstract/Free Full Text]
- Ewing, R.P., and R. Horton. 1999. Discriminating dyes in soil with color image analysis. Soil Sci. Soc. Am. J. 63:18–24.[Abstract/Free Full Text]
- Flury, M., and H. Flühler. 1995. Tracer characteristics of Brilliant Blue FCF. Soil Sci. Soc. Am. J. 59:22–27.[Abstract/Free Full Text]
- Forrer, I. 1997. Solute transport in an unsaturated field soil: Visualization and quantification of flow patterns using image analysis. Ph.D. thesis ETH no. 12476. Eidgenössische Technische Hochschule Zürich, Switzerland.
- Hanbury, A. 2002. The taming of the hue, saturation and brightness colour. In H. Wildenauer and W. Kropatsch (ed.) Space, 7th Computer Vision Winter Workshop, Bad Aussee, Austria. 4–7 Feb. 2002. Inst. of Computer Aided Automation, Vienna Univ. of Technology, Austria.
- Nissen, H.H., P. Moldrup, and K. Henriksen. 1998. High-resolution time domain reflectometry coil probe for measuring soil water content. Soil Sci. Soc. Am. J. 62:1203–1211.[Abstract/Free Full Text]
- Persson, M., and J.M. Wraith. 2002. Shaft-mounted time domain reflectometry probe for water content and electrical conductivity measurements. Available at www.vadosezonejournal.org. Vadose Zone J. 1:316–319.[Abstract/Free Full Text]
- Persson, M., and S. Haridy. 2003. Estimating water content from electrical conductivity measurements with short time-domain reflectometry probes. Soil Sci. Soc. Am. J. 67:478–482.[Abstract/Free Full Text]
- Persson, M. 2005. Accurate dye tracer concentration estimations using image analysis. Soil Sci. Soc. Am. J. 69:967–975.[Abstract/Free Full Text]
- Post, D.F., A. Fimbres, A.D. Matthias, E.E. Sano, L. Accioly, A.K. Batchily, and L.G. Ferreira. 2000. Predicting soil albedo from soil color and spectral reflectance data. Soil Sci. Soc. Am. J. 64:1027–1034.[Abstract/Free Full Text]
This article has been cited by other articles:
|
|
|
|
 
S. R. Evett and G. W. Parkin
Advances in Soil Water Content Sensing: The Continuing Maturation of Technology and Theory
Vadose Zone J.,
November 11, 2005;
4(4):
986 - 991.
[Abstract]
[Full Text]
[PDF]
|
|
|
TOP
ABSTRACT
INTRODUCTION
), saturation (
), and value (
) of the soil samples plotted against water content 