| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Agronomy Dep., Maringá State Univ., 87020-900, Maringá, Paraná, Brazil
b Soil Science and Plant Nutrition Dep., Univ. of São Paulo, Escola Superior de Agricultura "Luiz de Queiroz", C.P. 9, 13418-900, Piracicaba, São Paulo, Brazil
* Corresponding author (jamdemat{at}carpa.ciagri.usp.br)
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Abbreviations: CEC, cation exchange capacity • OM, organic matter • RS, remote sensing • SR, spectral reflectance • TM, thematic mapper
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Most agricultural and environmental plans require soil analysis, or at least should require analysis to better implement any change. Furthermore, better practical methods that can rapidly estimate soil properties are needed to improve quantitative assessments of land management problems (Shepherd and Walsh, 2002). The system of agricultural planning has become a series of steps starting with soil surveys that lead to fertilizer recommendations. Typically, soil samples are analyzed, application rates calculated, and the fertility interpretations are made. Soil samples are normally composites consisting of 15 to 20 individual samples for an area of 12 to 20 ha (van Raij et al., 1996). In some cases where precision agriculture is conducted, recommendations call for one soil sample per hectare (Wolkowski and Wollenhaupt, 1994). Additionally, the soil properties analyzed also help establish if more or less samples are needed per area. Finally, while it is clear that soil analyses are necessary, they are still expensive, time-consuming, and often create undesirable environmental waste products.
According to Demattê et al. (2001), the costs of soil analyses with precision agriculture systems are very expensive when compared with more traditional methods. Thus, more affordable alternatives are required and these can be achieved through the use of new technologies to estimate soil attributes.
Remote sensing (RS) has been explored as an alternative method for determining soil attributes (Galvão et al., 1997). Correlations between the different bands of the electromagnetic spectra and SR data have led to a better understanding of complex soil components. These studies are the basis for new paradigms of nondestructive methods to quantify soil attributes (Coleman et al., 1991; Ben-Dor and Banin, 1995; and Madeira Netto, 1996). According to Coleman et al. (1991), "estimation of soil variables from spectral data used in soil mapping is an important potential application of multispectral remote sensing." Shepherd and Walsh (2002) developed a scheme to use a spectral library as a fast and nondestructive estimation of soil attributes based on analyses of diffusion reflectance spectroscopy. These spectral data banks are essential and facilitate soil RS analysis.
The airborne and satellite RS soils data have to be analyzed so that parameters such as atmospheric interference, sensor resolution, and surface condition do not interfere with the results (Guyot et al., 1996; Huete, 1996). Furthermore, Ben-Dor (2002) detected problems when using RS data to analyze tropical soils, while Coleman et al. (1993) obtained low correlation for estimating soils attributes from space platforms. These authors stated the following: "The low coefficients observed could be attributed to atmospheric particles that are known to affect the electromagnetic energy that is sensed from satellite and aircraft platforms." The successful correlation of statistical data and Landsat TM imaging with sand, clay, Fe, and OM content of different soils was developed by Ben-Dor and Banin (1995) in Madison County, Alabama, USA.
The use of precision agriculture, which required rapid assessment of soil attributes, began in the 1980s (Searcy et al., 1989) and has become increasingly more popular in the 1990s (Schueller, 2000). Nevertheless, soil attributes are routinely processed and quantified by traditional methods in the laboratory. With precision agriculture, more soil samples are required, and they use many chemical products, which are harmful to the environment. In Brazil, there are legislative regulations that limit the use of chemical products in the laboratory especially for methods that produce environmental pollutants (Afonso et al., 2003). For example, the determination of the organic matter in the soil is based on the Walkley–Black method (WB) (Jackson, 1982), due to its higher speed and simple instrumentation. However, the disadvantages of the WB method are the use of concentrated H2SO4 and the presence of Cr6+ (Miyazawa et al., 2000), which are potential pollutants.
It is important to develop low cost, highly efficient, new methods of soil analyses that produce less environmental pollutants. In fact, Hummel et al. (1996) emphasized the importance of RS to estimate CEC, moisture, and soil nutrients as the basis for site-specific management, even though satellite results were still not readily available.
This study was based on the following question: "Can the content of soil components, obtained by traditional physical-chemical analyses, be substituted by a remote sensing methodology? Can a sensor located 800 km from the target be utilized by scientists for soil quantification"? Thus, the objective of this study was to develop a methodology to determine physical and chemical soil attribute contents by sensors in the laboratory and by satellite. The results were compared with traditional analyses with the main goal of correlating these different methods. The hypothesis tested was that soil contents estimated by sensors could be well correlated with traditional soil analyses. Background literature shows that the incident energy interacts with soil components. Thus, the reflected energy acquired by a sensor could indicate the soil sample constituents. An adequate atmosphere correction and detection of bare soil in images facilitated this goal. Results achieved in this study suggest that this methodology can present an efficient cost-effective quantification of soil attributes.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
|
Geologically, the area studied is situated in the Itararé formation of the Tubarão group (Instituto de Pesquisas Tecnológicas [IPT], 1981). This formation is a complex association of lithofacies, mostly sedimentary, that appears in fairly fast vertical and horizontal successions. The predominant lithologies consist of immature psammite with a heterogeneous granulation, leading to feldspathic psammite and even arkosic sandstone. Concomitantly with this lithology of the Itararé formation, there are eruptive dike elements of the Serra Geral formation, comprising intrusive bodies of tholeitic basalts.
Methodology Procedure
The procedure to determine the methodology had the following phases: (1) Data bank: collection of soil samples from the field; collection of the pure spectral data of these samples using a laboratory sensor; collection of spectral data from satellite sensors; determination of soil attributes using laboratory analyses; (2) Calibration: determination of statistical models that correlated spectral data with each soil attribute from both, laboratory and satellite sensors; and (3) Validation: comparison between the determined and estimated contents of soil attributes by both sensors. The specific methodologies used are described in the following sections.
Soil Sampling
Within the studied area of 184 ha, regular 100 x 100 m grids were used to divide the field. All areas sampled were bare soil. Soil samples were collected from each point of the grid, resulting in one sample per hectare (Fig. 1). Each point was georeferenced using the differential global positioning system (DGPS) with a maximum error of 3 m. At each point, the soil was sampled at 0- to 20- and 80- to 100-cm depths, which corresponded to the surface (Epipedon) and subsurface horizons (Nanni et al., 2004), respectively, with a total of 368 samples.
The collected soil samples were sent for physical, chemical, and mineralogical analysis in the laboratory. The texture groups of the soils were defined according to Empresa Brasileira de Pesquisa Agropecuária (EMBRAPA, 1999a). The contents of total sand, silt, and clay were determined by the densimeter method (Camargo et al., 1986). The texture ratio was defined by dividing the mean values of the clay contents found in the A and A/B horizons (when present) by the contents of clay from the B horizon, with the exception of the BC horizon (EMBRAPA, 1999). The contents of Ca, Mg, K, and the sum of the bases (S) were determined according to van Raij and Quaggio (1986). Organic matter, total and effective acidity, pH in water and in KCl, CEC, the values of base saturation (V%) and Al (m%) were determined according to EMBRAPA (1996b). The contents of total Fe, Si, and Ti were determined by using the sulfuric acid based methodology from EMBRAPA (1996b). The color obtained was established with the Munsell color panels. Soil samples were taken to laboratory where they were dried at 45°C for 48 h and sieved to 2 mm. The following analyses were performed: particle-size distribution (PSD); total Fe2O3, SiO2, and TiO2 extracted by sulfuric acid digestion (Camargo et al., 1986); sum of cations (SC); CEC; cation saturation, CS (CS = SC x 100/CEC); aluminum saturation, AS (AS = Aluminum/SC+Aluminum); clay activity, T (T = CEC x 100/clay), and OM (van Raij and Quaggio, 1986). Munsell colors were used to characterize the dry and moist soil samples.
A detailed soil map (EMBRAPA, 1996) was made and the following soils were detected: Typic Haplorthox, Typic Paleudult, Arenic Abruptic Paleudalf, Typic Udorthent, Typic Argiudoll, Rhodic Paleudalf, Typic Dystrochrept, and Typic Udifluvent (Soil Survey Staff, 1998).
Spectral Data Acquisition
Laboratory Spectral Data Acquisition
Soil samples (of approximately 80 g) were placed on 9-cm diam. Petri plates and radiometric analysis was done in a controlled environment using the Infra-Red Intelligent Spectroradiometer sensor (IRIS), with a spectral resolution of 2 nm (from 400 to 1000 nm) and 4 nm (from 1000 to 2500 nm) (Geophysical Environmental Research Corp. [GER], 1996). According to Demattê and Garcia (1999) and Demattê et al. (2004), the geometry of the data acquisition indicates a bidirectional SR factor, which expresses a ratio between the sample's reflected spectral flux and a standard spectral flux (a BaSO4 plate under the same conditions) (Nicodemus et al., 1977).
Laboratory Spectral Data Acquisition Converted to Satellite Spectral Range
A simulation was established between satellite and terrestrial data for comparison. The mean SR data obtained by the IRIS sensor was calculated in the wavelength ranges corresponding to satellite sensor bands (convolution), for example, 450 to 520, 520 to 600, 630 to 690, 760 to 900, 1550 to 1750, and 2080 to 2350 nm. This is a very common procedure that is used when researchers need to simulate satellite sensor data (Latz et al., 1984).
Satellite Spectral Data Acquisition
A TM-Landsat 5 image, bands 1, 2, 3, 4, 5, and 7, was used to obtain satellite level data (WRS 220_076 quadrant B; passed overhead on 27 Aug. 1997). The Spring program developed by the Brazilian National Space Research Institute, INPE (2003), was utilized to set up and manipulate the database. Spring is the newest generation of geoprocessing programs conceived for object-oriented programming with multiple functions and algorithms for processing georeferenced databases (Câmara et al., 1996).
Digital number values (DN) obtained from the TM Landsat images were transformed into SR and properly adjusted for atmospheric effects (Thome et al., 1997; Demattê and Nanni, 2003). Processing the interactions between solar radiation, the atmosphere and soil compounds, the image was corrected, band-to-band, where the atmospheric effects were eliminated, and then digital numbers were converted into "real" SR values (Zullo, 1994). Digital number values were first normalized to "top of atmosphere" apparent reflectance and then corrected for Rayleigh scattering and ozone absorption by using the 5S radiate transfer code simulation (Tanré et al., 1992; Vermote et al., 1995) with the following equations:
 |
 |
 |
) (mw cm–2 sr–1 µm) is the radiance detected by the sensor; DN() is the digital number per TM band; n, m calibration coefficients of the TM; 
app is the apparent reflectance "top of the atmosphere"; E0, the solar irradiance exoatmospheric per TM band; 
z is the zenithal solar angle; d, the earth-sun distance (astronomical units); * corrected reflectance Rayleigh/Ozone; To3 ozone transmittance (absorption); a,r atmospheric reflectance (Rayleigh), and Tr atmospheric transmittance totals (Rayleigh total). The input 8-bit image value was transformed by the Spring program through the implementation of a specific algorithm language. After the conversion and correction process, the output 8-bit image values were scaled to the range of 0 to 100% reflectance. The zero image digital number corresponded to 0% reflectance, whereas the DN 255 corresponded to 100% reflectance.
After transformation, the image was georeferenced with the Spring program using georeferenced points from the field, which were identical to the Landsat image, using the equation of the zero order (nearest-neighbor interpolation) obtained by SR values (Jensen, 1986).
The identification of bare soils, in the image, was performed by a methodology proposed by Demattê et al. (2000). The method consists of vegetation indices, soil line concepts, image composition and soil genesis information regarding certain locations of the field in the image. When all indices demonstrated bare soil, the spectral data could be collected. The Normalized Difference Vegetation Index (NDVI) refers to the band ratio between red (R) and the near-infrared reflectance (Jackson, 1983). In the Spring program, the NDVI index was calculated by the equation C = G [(A – B)/(A + B)] + O, where A refers to near infrared band; B = red band; G = image gain, and O = image offset. Furthermore, to increase SR contrast between vegetation and soil, NDVI was partially compensated for illumination, surface declivity, and geometry (Lillesand and Kiefer, 1999). Some vegetation indexes are denominated of the "soil line." The soil line is a linear relationship between the near infrared (NIR) and R reflectance of bare soil as characterized by slope and intercept parameters (Garey et al., 2004). The data can be observed on a bidimensional graph formed by both, visible and NIR bands (Huete, 1989). Thus, vegetation present in a studied image will be, theoretically, proportional to the Euclidean orthogonal distance on this "soil line." Reciprocally, Euclidean distance, based on vegetation index, has a complementary orthogonal index, which relates to soil optical properties with less vegetation (Fukuhara et al., 1979). Garey and Sabbagh (2002) describe that remotely sensed estimations of soil surface properties can lead to improved representation of spatial heterogeneity.
One of the most important aspects of this work is that it is crucial that care is taken when collecting soil data from the image. Many parameters were analyzed from the soil sampling location before a decision was made to evaluate it and use it in the model. More than one remote sensing method was used to evaluate a point in the image, such as: the soil line parameter; image composition; landscape; and evaluation of the spectral curve directly by the software. Only when all of these parameters were conclusive and indicated that the point was bare soil, would the point be collected and those data inserted into the model.
The soil sampling grids were overlapped on the TM-Landsat image with the Spring program. DN for each point, which coincided with the soil sampling point of the field, was extracted by a "pixel reading" algorithm (INPE, 2003). Each DN was multiplied by 100 and then divided by 256 to obtain the reflectance value of the satellite orbital image.
The resulting data obtained from reflectance for each band (SR by IRIS sensor and simulated TM-Landsat bands) were tabulated for statistical analysis.
Statistical Analysis
Laboratory Spectral Data Statistics
The statistics used soil analyses from surface and subsurface samples collected in the field and its respective spectral data. The objective was to develop a model that could estimate a specific soil attribute using spectral data.
Spectral data collected in the laboratory presented a total of 674 bands according to the sensor specification. The evaluation of the entire spectrum would form too many bands, thus making the statistics difficult. Due to this, we selected specific wavelength ranges. Three concepts were used to select the bands from the laboratory spectral curves: 1) empirical observation of the analyzed spectrum, which showed SR curve inflections, convex and concave portions, and variation of reflectance intensity in all spectrum; 2) literature observations that depicted the correct wavelengths that have relationships with soil attributes (Henderson et al., 1992; Madeira Netto, 1993, 1996; Demattê and Garcia, 1999; Demattê and Nanni, 2003), and 3) wavelengths characterized by a strong inflection, such as iron oxides (481 nm), water and OH groups (1417 and 1927 nm), kaolinite (2206 nm), and gibbsite (2265 nm).
Based on these observations, twenty-two specific wavelengths (or ranges of wavelengths) "bands" were chosen. When the band is specific, for example, 480 nm, the reflectance data are exactly from this wavelength. If the band was in a wavelength-range, for example, 480 to 580, we used the mean spectral reflectance from this range (Fig. 2a ). On the other hand, some of the wavelengths of the spectrum showed inflection (absorption features), which have been previously reported in the literature (Henderson et al., 1992; Madeira Netto et al., 1993, 1996; Demattê and Garcia, 1999; Demattê and Nanni, 2003). In these cases, we modified the methodology and called these bands "Reflectance Inflection Difference (RID)." The RID is the difference between reflectance value at the highest and lowest points of inflection (or Amplitude of spectral data at this range), represented by unique data (Fig. 2b). In this study, 13 RID values were selected (demonstrating the height of the curve between peak and valley).
|
The data matrix was analyzed based on the 13 physical and chemical attributes of each soil sample (e.g., PSD and FeO3). This matrix was submitted to correlation analysis using the SAS system (SAS Institute, 1992). Thirteen soil attributes (OM, SC, CEC, T, CS%, AS%, total Fe, Si, Ti, silt, clay, sand, and the silt/clay ratio) were correlated with their respective SR characteristics (predictor variables).
Correlations between soil attributes and the obtained SR characteristics were determined using the SAS regression function. To avoid biases in the analysis, independent variables or SR characteristics were evaluated for colinearity, ensuring that two or more variables would not overlap.
Using the SAS multivariate analysis component (STEPWISE), multiple linear regression equations were developed to estimate soil attributes from the SR data. According to Glantz and Slinker (1990), the stepwise procedure allows the user to establish a secure regression model. Stepwise is a procedure for sequentially entering independent variables one at a time into a regression equation in an order that improves the regression equation's predictive ability. This method is particularly useful for screening data sets in which there are many independent variables that identify a smaller subset of variables that determine the value of a dependent variable.
The regression (REG) program of SAS was used with the SELECTION = STEPWISE option, in which the system established the dependent variables that had higher coefficient of determination with independent variables. The means of the model were further improved by a guided analysis procedure, and checked to ensure that no assumptions had been violated. Besides the R2 coefficient, other statistical parameters, such as Cp, can also indicate if the model is suitable (Glantz and Slinker, 1990), that is, if the dependent variables were appropriately chosen for the fitness of the model. The Cp parameter estimates the magnitude of the biases introduced into the estimates of the dependent variables by leaving variables out of the regression model. Moreover, Cp is particularly useful for screening large numbers of a regression obtained in all possible subset regressions. This coefficient was also used by Coleman et al. (1993), Coleman and Tadesse (1995), and Ben-Dor and Banin (1995).
Soil attributes were used to determine correlation equations with the respective reflectance in each band. Thus, each band had a factor in the equation according to its contribution. However, bands without correlation were not included in the final equation.
Multiple regression equations were established and divided between the surface and subsurface soil layers, thus creating one model for each attribute and for each layer. To predict soil attributes, the SR equations were then tested with soil samples that were not used in the model's development. The correlation coefficients were computed between the estimated values (SR analysis) and measured (routine laboratory analysis values) to evaluate the accuracy (Fig. 3 ). No samples were omitted from analysis with the multiple-regression equations. A successful prediction was evaluated for estimated values and determined values, using the coefficient of correlation (R2) and root mean square error (RMSE). The RMSE, also defined as the standard error of the estimate, established an actual variability for the line of means in the underlying population (Glantz and Slinker, 1990).
|
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
|
Gerbermann and Neher (1979), through field radiometer, reported lower R2 for sand content than this work, although it was very close to the coefficient found by Shepherd and Walsh (2002). The R2 for sand reached 0.796, where bands 22 and 18 corresponded to 80% of the model.
The clay attribute had the best R2 values (0.915) and the model was composed of 13 variables with a Cp of 8.1, which indicated an unbiased model. Similar to what occurred with sand, the prediction for clay found bands 18 and 22 to be essential. This fact is related to weathered tropical soils, where increasing contents of clay are generally associated with a reduction of sand. Coleman et al. (1991) determined a R2 value of 0.63 with their model, where part of the spectrum agrees with our model, such as: 481–600 nm (B3, 481–596 nm), 1150–1300 nm (B7, 975–1350 nm) and the range between bands 18 to 22 (infra red, 2080–2350 nm). Compared with Al-Abbas et al. (1972), there were some bands in common such as B4 (596–710 nm), B7 (975–1350 nm), and the spectral ranges that correspond to B18 to B22. According to these authors, the statistical analyses of their data showed that the variable that contributed the most to R2 was band B10 (1000–1400 nm), while in our study the variable that contributed the most to R2 (0.58 of 0.796) was band B22 (2389–2498 nm).
Multiple regression equations used to evaluate the soil samples of the subsurface layer attributes (Table 2) showed lower values, although the R2 was still significant. The coefficients for soil attributes, such as sand and clay, remained high at 0.82 for sand and 0.88 for clay. Therefore, these data agree with those reported by Shepherd and Walsh (2002). However, the R2 value found for OM was lower and correlated to that observed by Ben-Dor and Banin (1995). Due to the importance of OM in fertility and soil quality, the authors believed that more attention should be given to improve the OM estimation. Chemical attributes, such as CEC and SC, had similar coefficients for the surface and subsurface layers. The R2 for the chemical attribute CEC was 0.85 for the subsurface and 0.91 for the surface, while for the SC chemical attribute, the R2 was 0.87 and 0.88 for subsurface and surface, respectively. Cation saturation (CS%), Al saturation (AS%), and T had higher subsurface R2 values when compared with surface values. The multiple regression equations showed high correlation because SR is derived from the interaction between radiant energy and soil attributes (Sabins, 1997).
|
The equations developed in the present study were tested on soil samples that were not used in the development of the spectral model. Figure 4 shows the correlation between the values estimated by spectral reflectance and the ones determined by chemical analysis for the subsurface soil layer. The SR estimated clay content showed values of R2 = 0.81 in laboratory analysis (Fig. 4), and RMSE = 91.0. The same high results were observed for SC (R2 = 0.81, RMSE = 55.1) and CEC (R2 = 0.72, RMSE = 89.9), which agree with Shepherd and Walsh (2002). The sand contents also had a significant R2 = 0.63 and RMSE = 92.9 and had low variability.
|
Similar results were obtained when using equations for the surface soil layer and unknown samples (Fig. 5 ). Clay and sand contents showed good correlations with R2 of 0.46 and 0.68, respectively, with RSME values of 118.7 and 86.5, respectively.
|
One potential explanation for the differences reported in these studies is laboratory conditions, which have less environmental interference compared with field methods used in prior studies.
These findings permitted us to understand the variations observed in the spectral data from the laboratory and allowed us to go to the next step: satellite sensing (Demattê, 1999).
The Conversion of Laboratory SR Soil Attribute Estimates to the TM System
The SR laboratory sensor data were used to simulate Landsat TM data (6 bands) (Table 3). Such data enabled the comparison between models with diverse sets of explanatory spectra bands for a specific soil attribute. The simulation had also allowed us to compare laboratory and simulated satellite data without atmospheric corrections (Latz et al., 1984).
|
Coleman et al. (1991) evaluated soil attributes using field spectral data, and obtained R2 coefficients ranging from 0.21 to 0.68 for silt, clay, iron oxides, and OM. The difference in results between that study and the present one demonstrates the importance of the spectral data technique. The methodology used in this work resulted in a significant gain in the R2 values (Table 3). The R2 values obtained for clay and Fe2O3 were 0.83 and 0.88, respectively (Table 3). Even the CEC, which is dependent on chemical elements (Ca, K, Mg, H, and Al), showed meaningful results with an R2 of 0.70. An adequate silt equation could not be established since it was not a significant dependent variable as found in laboratory (Table 3).
A field spectroradiometer was used by Coleman et al. (1991) to simulate TM-Landsat bands and establish a stepwise regression equation to predict important soil attributes. The R2 value found by these authors was 0.108, which is lower than that found in the present work (0.527). On the other hand, our model presented a 13.42 Cp index, making the model biased (Glantz and Slinker, 1990), while the model obtained by Coleman et al. (1991) had only three bands that covered almost the entire spectrum. Since organic matter influences all spectra, it is expected that more variables are needed to compose the model. In fact, it was necessary to use 22 bands and 13 RID in this study.
The key bands for clay were B5 (1150–1300 nm) and B7 (2080–2350 nm) and gave high R2 value (0.837), thus corroborating with Coleman et al. (1991). All simulated bands were used to construct the model, with the exception of B1 and B2, thus demonstrating that the activity of the clay attribute above 600 nm for the evaluated spectrum and conditions tested.
The determination of the main specific bands, within the spectrum, is important for developing any successful model. The spectral data are a function of all soil attributes and their interaction with incident energy. Soil attributes have specific bands, which have more significant interactions, but they also influence, to a small degree, other parts of the spectrum. An example of this point is the OM. The spectral curve of OM material has very low reflectance through the entire spectrum (Madeira Netto, 1993; Demattê et al., 2003). The iron oxides influence certain bands with absorption features, such as 450 and 850 nm, (White et al., 1997). Table 3 depicts that almost all of the soil attributes had reflectance in the B3, B4, and B5 bands.
Therefore, if we are interested in estimating iron oxide content, all the spectral bands that show alterations should be included in the model. These considerations lead us to conclude that all bands are potentially important to explain the content of a certain soil attribute, thus making stepwise regression an important tool in this task.
Soil Attribute Estimates by Spectral Reflectance Obtained at the Satellite Level
Great advances in RS research have been reported in the last decade. Through its evolution, the challenge of estimating soil contents, by a sensor positioned at approximately 438 miles or 705 km (Landsat) from the target, has increased. To date, research to quantify soil attributes, using satellite data, has had limited success. Therefore, our purpose was to expand the work in this field.
The TM-Landsat spectral bands were first simulated using laboratory data, as discussed earlier. The purpose was to verify the soil line pattern from the sampled soils (Fig. 6 ).
|
|
Multiple regression equations obtained by TM-Landsat bands (Table 4) had lower R2 values for all soil attributes than those simulated in the laboratory (Table 3). Although, both models used six bands of TM, one was a simulation using laboratory data with no atmospheric interference. Thus, the simulated TM data presented better results than the original image data (Table 4). Such results could be due to factors inherent to the soil that cause different intensity and forms in the spectral response such as size, particles distribution, soil moisture, and crust (biogenic and physical) occurrences.
|
Frazier and Cheng (1989) obtained R2 value of 0.98 for predictive organic C equations, using the ratio between the TM's fifth and fourth bands. These values were comparatively higher than those obtained in the present study (Table 4, 0.5 R2 for OM). However, these authors did not use SR values from TM-Landsat-image to detect bare soil. Instead, they used DN, which does not establish a relation between the real physical reflectance of the OM (Bittick et al., 1994; Demattê and Nanni, 2003).
The OM model showed a similar behavior with laboratory conditions (simulated data). Cp (0.9) indicated an unbiased equation (Table 4) and only two bands were included in the model. On the contrary, Coleman et al. (1993) obtained a model with almost all bands of Landsat, and they considered the model inadequate to predict OM. In the present study OM showed an R2 of 0.50.
The model to predict sand utilized band 7 (2080–2350 nm) and was similar to the model of Coleman et al. (1993). This spectral range appears in all equation used in the present study for the surface conditions. Thus, more studies should be conducted to further evaluate this attribute.
The clay R2 value was superior to those of OM and sand. The model obtained was similar to the model of Coleman et al. (1993), and showed coincidence for the following bands: B1, B3, B4, and B5.
The method described in this study can have several applications, directly in the laboratory and/or in the field. For both cases, there is a need of a vast data bank with spectral soil response from several regions. The need of standard patterns is essential. In the case of the laboratory, unknown soil samples would have their spectral reflectance measured, and with such models, the soil attribute contents could be determined.
The same method could have other application in precision agriculture. The data bank would be in a tractor mounted onboard computer that could be utilized in the field. Modern sensors could be developed to determine the real-time reflectance values directly on the tractor. Information on the reflected data would go through to the tractor-mounted computer and be used to determine the models to quantify soil attributes. The information from the evaluated soil attributes would then be sent to an application-sensor (at this point you can have several types of applied products, such as fertilizer or lime) that would, in real time, apply a product(s). These products are influenced by soil texture and organic matter contents. Thus, the tractor-sensors could determine the soil attributes and indicate the best corrective measure, that is, dosage of a specific product.
Data obtained from space, at first view, would have application to assist multiple agricultural and environmental activities. The development of soil surveys would be faster and have quantified soil information. Considering that new hyperspectral sensors are being developed, soil evaluation, in all aspects, could be improved.
The method chosen, laboratory or spectral image data, will depend on the user's objective, and the applications are unlimited.
|
CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
It is possible to estimate clay, Fe2O3, and TiO2 contents with stepwise models developed from satellite spectral data. These attributes showed R2 values as high as 0.72. On the other hand, chemical elements such as cations and Al had lower values and were not predictable with the same effectiveness, therefore, more studies should be conducted.
The success of the process of estimating soil attributes by satellite imaging depends on a number of factors. These factors consist of atmospheric image correction; geometric corrections; transforming the data in real reflectance, and the use of a detailed methodology for the identification of bare soil, which was developed in this study. The mentioned attributes could be determined by spectral reflectance as an alternative to traditional chemical methods, which are expensive, time-consuming, and produce environmentally hazardous by products.
The results reported here help validate this new method and correlate with data from several previous studies regarding the usefulness of RS to determine some soil attributes. Therefore, if we consider all the knowledge obtained to date related to this subject, it is clear that the scientific community should take the next step. This step would be to discuss the methods developed, determine standards and patterns (for each region) for the procedures, and begin implementing these into routine soil analyses.
|
ACKNOWLEDGMENTS |
|---|
Received for publication November 11, 2003.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  E. Ben-Dor, D. Heller, and A. Chudnovsky A Novel Method of Classifying Soil Profiles in the Field using Optical Means Soil Sci. Soc. Am. J., June 18, 2008; 72(4): 1113 - 1123. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||
from laboratory IRIS sensor data.
