By Titia Mulder.
Summary
Soil iron is an important indicator in the soil science; Iron is an indicator for soil fertility, the usability of an area to cultivate specific crops and it can indicate the age of the deposits (Bartholomeus et al, 2006). Determining the spatial distribution of different types of iron with traditional fieldwork and laboratory analysis is time-consuming and expensive. Remote sensing has proven to be a useful tool to determine the presence of iron in extended areas and various research fields (Bartholomeus et al, 2006).
There are chemical and physical properties that influence the reflectance. They originate from the chemical and physical properties of the surface, the soil and from the spectrometer. These properties are so called spectral chromophores. For the soil reflectance the most important chemical chromophores are minerals, organic matter content and water. The most important physical chromophores are texture, surface roughness, viewing angle, radiation intensity, incident angle and azimuth angle of the source (Ben-Dor et al., 1999). With airborne image spectroscopy some characteristics of the terrain, like slope and aspect, do influence the reflectance. With PARGE/ATCOR the influence of these characteristics can be corrected for but the effects are still present because the PARGE/ATCOR model is not performing well for rugged terrain (Schläpfer and Richter, 2002).
The main objective of this research is to quantify the effect of slope on the prediction of soil iron content, using spectral reflectance based iron indices.
The laboratory experiment will give the data to asses the influence of slope on the prediction of soil iron content. An experimental set up is designed where the slope of the sample can be varied, resulting in different illumination angles.
From these reflectance data the iron indices are calculated and a model is developed to predict the iron content. Three different indices are used in order to estimate the soil iron content; the ratio based Redness Index and the area and standard deviation of the continuum-removal. With multiple regression and general statistics the relation between slope and the prediction of the iron content is quantified. From these relations a correction model is created with multiple linear regression to correct for the error caused by slope. In order to make the link with the “real world”, the laboratory results are compared with airborne image spectroscopy data. An assessment is made on the performance of a ROSIS- image, taken in the same area as the soil samples, on the predictability of the iron content (Appendix A).
Paired-Samples T-tests are used to address the significant difference in the estimated iron content between measurements done under a certain slope and measurements from nadir. The ratio based Redness Index is not influenced by slope, the other indices are influenced by slope. The slope has little influence on the D550_area_lab and minimum and maximum errors are present up to 10% for slopes of 100 or smaller. With increasing slope, the influence of the slope increases as well but there is not a clear trend.
D550_S.D._lab is influenced by slope. On average, the index does significantly deviate from the nadir measurements up to an error of 2.5% and the minimum and maximum errors are up to 5%. The influence of slope is larger for steeper slopes and a so called “V-shape” is formed when results are plotted.
For both the area and the standard deviation of D880_lab the slope is clearly influencing the indices. The iron content is significantly deviating from the nadir measurements and the average deviates up to 5% with minimum and maximum error deviating around this average of 5%. The effect is stronger for the negative facing slopes, this might be due to the retroreflectance properties of the material and the surface roughness.
A correction slope model is made with a linear regression for each slope and for each index.
The capability of the prediction for both the area and the standard deviation for D550 and D880 are improved and more accurate. The RPD values for D550_area_lab are on average 1.93 and the R2 is 0.72 and the RPD values for D550_S.D._lab are on average 1.90 and the R2 is 0.71, this means that these correction models for the estimation of the iron content can be classified as a class B model according to Chang et al. (2001). RPD values for D880_area_lab are on average 2.91 with a R2 of 0.77 and the RPD values for D880_S.D._lab are on average 2.09 with a R2 of 0.76, this means that these correction can be classified as a class A model (Chang et al., 2001).
Finally, it is concluded that the ratio based Redness Index is not influenced by slope and the other indices are indeed influenced by slope. The error correction model for the slope is performing well for the area and standard deviation of D880 and the performance of the error correction model is less for the area and standard deviation of D550.
The estimation of soil iron content with ROSIS data is not accurate. There are several reasons why the ROSIS image has a bad performance in the estimation of soil iron content:
-
The ROSIS-image has a strong noise in the data. Due to the applied MNF some signal is lost.
-
Due to the rugged terrain the atmospheric and topographic corrections are not accurate and the reflectance is deviating strongly from the laboratory measurements.
-
The coarse resolution makes a good SMA difficult and the influence of mixed pixels is large. The mixed pixels consist of bare soil and olive trees.
-
The interpolation of the masked pixels is not the most accurate method.
-
The texture map is based on too few samples and therefore inaccurate.
Taken al these possible errors into account, the correlation of iron calculated with RI_rosis with the original iron content is 0.26. When the masked pixels and correct texture is taken into account than the correlation increases up to 0.40, which is still very low.
If the expected error of the slope for D550_S.D._rosis is compared with the overall error than it turns out that the error caused by the data quality and the field conditions is 8.96 times as big as the error caused by the slope under laboratory conditions.
The effect of slope is negligible when the ROSIS-image is being used for the quantification of soil iron content because the field conditions are of more influence than slope and the indices are too sensitive to be calculated with ROSIS.
For improvement of the error correction model for slope, it is most important that the standard deviation of the error caused by slope is minimized.
It would be recommended to use an other spectrometer with a smaller band width and a finer resolution. With better data, the reflectance will probably better correspond to the reflectance measured under laboratory conditions. The SMA will discriminate the vegetation better from bare soil and due to the finer resolution, the amount of mixed pixels will be lower.
The texture map must be of good quality, the best way to collect this data is to go into the field and take samples which are than analyzed in the laboratory.
More field samples of the iron content are needed in order to get insight into the spatial variability of the iron. All the calculated statistics are based on a relative small data set; the reliability of these statistics can be enhanced when more samples are taken into account (Isaaks and Srivastava, 1989).