Home Articles An Approach for Estimating Soil Organic Matter Content Using Synthetic IRS Satellite...

An Approach for Estimating Soil Organic Matter Content Using Synthetic IRS Satellite Data in Tropical Soils of Lop Buri, Thailand

K.W. Daniel
Email: [email protected]

N.K. Tripathi and K. Honda
Space Technology Applications and Research (STAR)
Asian Institute of Technology (AIT)
P. Box 4 Klong Luang, Pathumthani 12120
Fax: 66-2-524-5597

Timely information on the content and distribution of key soil nutrients is vital to support precision agriculture. This paper describes the newly developed approach, “Spectral Band Cloning” (here after, SBC). The purpose is to enhance the IRS satellite data for soil nutrient estimation, which otherwise is unattainable. The idea emanated from sensors integration, where the intraspectral relationships of spectrometer channels are mimicked by the corresponding IRS bands. New and synthetic bands were generated, which are competent to estimate soil organic matter (SOM). Forty-two samples from topmost soil layer, collected during a satellite-synchronized field survey in Lop Buri, Thailand, were analyzed chemically and spectrally in a laboratory. From raw spectrometer-driven spectra, SOM was successfully modeled from bands R410, R460, and R480 (R2 = 0.85), which are unobtainable in IRS. The SBC enabled decent modeling of SOM from synthetic IRS bands (R2 = 0.72). The model was implemented and verified on a GIS platform and generated a predicted SOM surface with a reasonable degree of accuracy. SBC is a promising to estimate other indiscernible biophysical parameters, which enhance precision farming, and could be employed to other satellite sensors.

1. Introduction
Satellite sensors are faster, cheaper, and objective data providers than conventional field-based surveys. This is due to wide area coverage and possibility of concurrent information recording represented in a digital energy pattern. Most satellite sensors, however, have a small number of satellite channels and are limited to extract satisfactory reflectance from objects, which vary with location, time, geometry of observation and waveband (Curran and Kupiec 1995). Soils show both spatial and temporal variability, and dictate a fine-tuned management at farm level, leading to precision agriculture (Bouma 1995). Nutrients of soils are not easy-to-detect variables from remote sensors (Palacios-Orueta and Ustin 1998, Clark 1999). However, several researchers have made significant contributions for the detection from remote sensing reflectance spectra (in situ measurements) and from radiance data recorded by satellite sensors. Recently, precision agriculture has gained interest from satellite data with a special preference for higher spatial and spectral resolutions (Mulla 1995, Strachan et al. 2002), where most existing satellites are handicapped.

The hyperspectral remote sensing data using AVIRIS sensor has been widely used for soil mineral detection (Hoffbeck and Landgrebe 1996, Palacios-Orueta and Ustin 1998, Price 1998, Galvao et al.. 2001). This wide application is due to extreme potential to detect at numerous narrow bands in a wide spectral range. However, the platform of AVIRIS is airborne, and requires special flight arrangement, as opposed to the regular observation by satellite sensors. Field and laboratory based spectrometers are endowed with high-resolution reflectance spectra, and have wider opportunity to capture the reflectance response of many objects (Milton et al. 1995, Palacios-Orueta and Ustin 1998, Clark 1999, Leone and Sommer 2000). Satellite sensors, which are composed of spectral bands with broad bandwidth, can benefit from these satellite sensors through sensor integration (Hoffbeck and Landgrebe 1996, Palacios-Orueta and Ustin 1998, Clark 1999)

Remote sensing of soils using optical satellite sensors, at spatial resolution of 30 meters, like Landsat-ETM, SPOT, and ASTER are suitable for small-scale farming system. However, their detection are restricted to the VNIR spectral regions (Plummer et al. 1995, Jensen 2000). They might have limited bands at SWIR region, but at a very lower spatial resolution ( 60 meters). Secondly, The spectral features of soils are monotonous and featureless within the VNIR region (Daniel et al. 2002). Hence, applying absorption features for the identification of soil nutrient purpose (Valeriano et al. 1995, Palacios-Orueta and Ustin 1996) is extremely difficult in VNIR. Some studies have addressed this problem by conducting spectrometric-based soil nutrients estimation from a wide range of spectral bands (Galvao and Vetorello 1998, Palacios-Orueta and Ustin 1998, Price 1998). However, the most commonly used spectrophotometers are limited to VNIR spectral range (Milton et al. 1995). SOM is considered a test nutrient. It is highly variable and react quickly to external changes and its decomposition rates show spatial variability (Palacios-Orueta and Ustin 1996). Therefore, the prime objective of this study is to integrate the spectroradiometer and satellite sensor (IRS-1C) through a newly developed method called, SBC, and estimate and classify SOM.

2. Methodology

2.1. The Data Set
The study area is found in Thailand, Lop Buri District, at geographical coordinates of 14°45′ – 15°00′ N and 100°50′ – 101°10′ E. Soils were developed over the parent material of sandstone and limestone. An intensive satellite-synchronized field survey was undertaken on April 2001. It was during “pedo-window” period: where there was cloud-free and uncovered soil condition, before the growing season. Forty-two soil samples from the top most layers were collected and geo-referenced with GPS. Soil samples (in three replicates) were brought to the laboratory for SOM determination through conventional dry ash method. The samples were also spectrally assessed using StellarNet spectroradiometer, over a 400 nm to 1290 nm spectrum at 10 nm intervals. Descriptions of the study area and laboratory procedures are found in Daniel et al. (2002, 2001). The satellite data used is IRS 1C scene (path 121, row 63), dated 08 April 2001. Both geometric the correction and conversion of original DN measures to the surface reflectance values was carried out in conjunction with the atmospheric correction.

2.2. Spectral Band Cloning
The Encyclopedia of Britannica (2002 edition) defines cloning as a process of generating identical genes enough for further study. In this study, cloning is tailored to the duplication of unknown spectral data from the known ones. Hence “spectral band cloning” is defined as, the process of duplicating satellite bands, assisted by intraspectral relationships from spectrometer channels. Daniel et al. (2002) reported the successful estimation of SOM (R2 = 0.85) from the laboratory-based spectral bands, using Artificial Neural Network (ANN). The SOM-sensitive bands, which were chosen from the spectrometer-driven models, were different from IRS-driven bands, which make the prediction of SOM (Eq-1) from the real IRS channels (R550, R650, and R815) impossible. Figure 1 shows the overview of SBS. From the 80 narrow channels of the spectroradiometer-driven spectral data, the IRS compatible bands (R550, R650, and R815) were chosen as predictors of the SOM-sensitive bands, which are identified by ANN. This is called the intra-spectral relationship of spectroradiometer bands.

Figure 1. Hypothetical Structure of SBS (Spectral Band Cloning) for IRS data
Yi = f (Xij)             (Eq-1)

Where Y is SOM-sensitive band. i is ANN chosen bands of R460, R460, and R480. j refers to the IRS channels of R550, R650, and R815.

From each sample points of the satellite image, the pixel values of G, R, and NIR were extracted to form the database for off-the-image modeling purpose. Then the intraspectral relationship developed through the spectrometer-driven SBC were imported and mimicked by the IRS. This helped to generate the synthetic IRS bands for latter modeling of SOM (Figure 2). This is through stepwise regression of several indexes developed from synthetic bands. To test the performance of the SOM estimating model, both measured and predicted layers of SOM were interpolated on IDW. The model from measured SOM refers to “what should have been produced”, and makes the “terrain nominal” (Vauglin 1999) that serves as a reference SOM surface of higher accuracy. The predicted layer, generated from interpolation of synthetically estimated SOM, refers to “what has actually been produced”. The accuracy of the prediction was undertaken by measuring the discrepancy between what should have been produced and what has actually been produced.

Figure 2. Structural overview of SOM modeling from Synthetic IRS bands
3. Results
In prior studies, Daniel et al.. (2002, 2001) reported the mean distribution of SOM was 12.53%, with a standard deviation of 3.33, and considered “too much” concentration. The spectral reflectance measured at the laboratory ranged from 20 to 50%. Sites of higher SOM concentration have correlated with low reflectance measures.

With respect to SBS, the ANN chosen SOM-sensitive bands from laboratory-based spectrometer data (Daniel et al. 2002) were R410, R460, and R480. Equations 2, 3, and 4 were derived through multiple regression technique (where R2 > 0.90) from the available spectral (i.e., IRS compatible) data of 550 nm (Green-band), 650 nm (Red-band) and 815 nm (NIR-band). The relationships between the compatible and non-compatible IRS bands of Eq-2,3, and 4 were mimicked by the existing IRS bands to produce the synthetic IRS bands. From those synthetic bands, several indexes were quantified. The stepwise multiple regression of those indexes gave the SOM model (Eq-5).

R410 nm = 1.1 + 0.2 (R815) +0.05 (R650) + 0.7 (R550)       Eq-2
R460 nm = -0.5 + 0.21 (R815) -0.02 (R650) + 0.7 (R550)      Eq-3
R480 nm = 0.04 + 0.1 (R815) -0.26 (R650) +1.08 (R550)      Eq-4
SOM (%) = 89.2 – 4.9(R460) +0.08 (R410)2 + 0.05 (R480)2 – 43 (R410/(R460* R480)) + 475.5 (R460/ (R410* R480)) – 4.3 ((R480* R480)/R460)) –0.89 ((R410*R460)/ R480) + 29 – 405.2* ÖR460/(R410 * R 480)       Eq-5

The SOM-sensitive synthetic IRS bands, R410 nm, R460 nm and R480 nm are implemented on a GIS platform, and are shown on Figure 3. Those cloned bands were first modeled from point data of each sampling site, and latter interpolated (with IDW) on the grid-formatted surfaces. The SOM predictive model (Eq-5) is implemented on the layer “predicted) which has more or less similar pattern with the “observed” layer of measured SOM layer. The predicted and the expected layers of SOM are in good agreement (R2 = 0.72).

Figure 3. SOM modeling from Synthetic IRS data and its comparison with measured point SOM layer

4. Discussion and Conclusions
Currently, very few soil nutrients are identifiable (mappable) from satellite sensors. Satellites with very good spatial resolutions are usually composed of few and broad spectral bands, which significantly limits the identification of soil nutrients. Integrating satellite and spectrometers are the current practices for mapping indiscernible earth’s objects. The relationship between surface reflectance values and SOM constituents of samples representative of Lop Buri district, Thailand, were analyzed at cross-platforms

This research has applied quite the inverse approach with the most currently employed methods from AVIRIS data. Due to the high number of spectral bands in AVIRIS, researches (Hoffbeck and Landgrebe 1996, Palacios-Orueta and Ustin 1998, Price 1998, Galvao et al. 2001) employed a mechanism to find out the few, but information-rich bands for nutrient modeling. However, in this study, the approach focused to maximize the existing few spectral bands to the level of adequate prediction of soil nutrients.

This research followed an empirical approach, which involves the derivation of quantified relationship between the remotely sensed data (through synthetic bands) and a measured SOM at laboratory (conventionally). The result does not provide the physics or surface interaction mechanisms. However, the developed method emanated from the knowledge and theory that the chemical components of soils behave differently at certain wavelengths. However, this study might have the limitation that the derived SBS models are for specific conditions at the time the measurements were obtained. Due to the absence of studies in this respect, comparing the obtained results with other studies is not materialized.

In a nut shell, however, it is safe to conclude from results of this study that:

  • The intra-spectral relationship existing in spectral pattern of spectrometers could be modeled and help to generate new and synthetic channels from the known ones. Consequently, the inter-spectral relationship of spectrometers could be mimicked by the satellite data and help to expand the IRS capability through filling the deficient bands through the band cloning technique
  • Through the added information harnessed from the “spectral band cloning”, it is possible to model other soil nutrients and biophysical parameters, which are not easily discernable from satellite sensors. The new technique could be implemented on Landsat TM sensors, which has more or less comparable spatial and spectral resolutions with IRS.
  • Precision farming could be benefited from such enhanced capacity where the potential of remote sensing will be substantially harnessed.


  • Curran PJ, Kupiec JA (1995) Imaging spectrometry: a new tool for ecology. In Advances in Environmental Remote Sensing. (Eds Danson, F.M. and Plummer, S.E.) Wiley, Chichester.
  • Daniel KW, Tripathi, NK, Honda K (2002) Artificial neural network analysis of laboratory and in situ spectra for the estimation of macronutrients in soils of Lop Buri (Thailand). Aust. J. Soil Res. Vol. 40, Issue 8 (In press).
  • Daniel KW, Tripathi, NK, Honda K, Apisit E (2001) Analysis of spectral reflectance and absorption patterns of soil organic matter. In Proceedings, ACRS- 22nd Asian Conference on Remote Sensing, 5-9 November 2001, Singapore. Vol. 1, pp. 742-747 (CRISP).
  • Escadafal R (1994) Soil spectral properties and their relationships with environmental parameters-Examples from arid regions. In imaging Spectrometry-A tool for Environmental Observations (Eds. J. Hill and J. Megier), Kluwer Academic Publishers, Amsterdam, pp. 71-87.
  • Galvao LS, Pizarro MA, Epiphanio JCN (2001) Variations in reflectance of tropical soils: spectral-chemical composition relationships from AVIRIS data. Rem. Sens. Env. 75: 245-255.
  • Galvao LS, Vitorello I. (1998) Role of organic matter in obliterating the effects of iron on spectral reflectance and color of Brazilian tropical soils. Int. J. Rem. Sens 19: 1969-1979.
  • Hoffbeck JP, Landgrebe DA (1996) Classification of remote sensing images having high spectral resolution. Rem. Sens. Env. 57: 119-126.
  • Jensen JR (2000) ‘Remote Sensing of the Environment: an Earth Perspective.’ (Prentice- Hall, Inc: NY).
  • Leone AP, Sommer S (2000) Multivariate analysis of laboratory spectra for the assessment of soil development and soil degradation in the Southern Apennines (Italy). Rem. Sens. Env. 72: 346-359.
  • Milton EJ, Rollin EM, Emery DR (1995) Advances in spectrometry. In: Advances in Environmental Remote Sensing” (Eds Danson, FM and Plummer, SE) John Wiley & Sons Ltd, Chichester, pp 9-32.
  • Mulla DJ (1995) Geostatistics, remote sensing and precision farming. In ‘Precision Agriculture: Spatial and Temporal Variability of Environmental Quality’. (Eds Lake JV and Bock GR) pp 100-119. (John Wiley & Sons)
  • Palacios-Orueta A, Ustin SL (1998) Remote sensing of soil properties in the Santa Monica Mountians 1. Spectral analysis. Rem. Sens. Env. 65: 170-183.
  • Plummer SE, Danson FM, Wilson AK (1995) Advances in remote sensing technology. In ‘Advances in Environmental Remote Sensing’. (Eds Danson, FM and Plummer, SE) pp 1-7. (John Wiley and Sons Ltd, Baffins Lane, Chichester)
  • Price JC (1998) An approach for analysis of reflectance spectra. Rem. Sens. Env. 64: 316-330.
  • Sreachan IB, Patley E, Boisvert B (2002) Impact of nitrogen and environmental conditions on corn as detected by hyperspectral reflectance. Rem. Sens. Env. 80: 213-224
  • Valeriano MM, Epiphanio JCN, Formaggio AR, Oliveira JB (1995) Bi-directional reflectance factor of 14 soil classes from Brazil. Int. J. Rem. Sens 1: 113-128.
  • Vauglin F (1999) Use of variograms to represent spatial uncertainty of geographic linear features. In ‘Spatial Accuracy Assessment: Land Information Uncertainty in Natural Resources’. (Eds Kim Lowell and Annick Jaton) pp 157-163. (Ann Arbor Press: Chelsea, Michigan)
  • K.W. Daniel received BA (Distinction), in Geography (1991) and served as graduate-assistant and assistant-lecturer at Alemaya University, Ethiopia. He has done M. Sc. in Soil and Water from Wageningen University, The Netherlands (1998), and worked as lecturer and researcher in former University. Since 2000, he is a Doctoral Candidate in the field of Remote Sensing and GIS, Asian Institute of Technology, Thailand. His research interests include Ground- and Satellite-based Soil Surveys, Utilization of Remote Sensing/GIS for assessment of Natural Resources and Precision Farming.
  • Nitin Kumar Tripathi, has done B. Tech. (1984) in Civil Engineering from Regional Engineering College, Warangal, India. He did M. Tech. (1987) and PhD (1995) in Geoinformatics from Indian Institute of Technology Kanpur, India. He received DAE Young Scientist Award in 1994 and All India Council of Technical Education’s Career Award for Young Teacher in 1995. He has worked in Geoinformatics Division, I.I.T., India from 1989 till 1999. Since 2000, he is working in Asian Institute of Technology, Thailand. His research interest includes development of interfaces for applications of GIS, GPS and Remote Sensing in costal zone, environmental pollution, disaster monitoring etc. Dr. Nitin is teaching fundamental and several specialized courses in GIS at AIT. He is also involved in conducting training in above topics for the Asian Region. He is the Editor-in-Chief of the Asian Journal of Geoinformatics.
  • Kyoshi Honda received the Doctor of Engineering (1992) from University of Tokyo in Civil Engineering. He worked as Landslide Engineer, Nippon Koei Co., Ltd., Japan (1982-1985), and as Assistant Professor in Mie University, Japan, 1985-1995. He is serving in Asian Institute of Technology (AIT) as a Visiting Faculty, seconded by NASDA, since 1995. He established Asian Center of Remote Sensing (ACRoRS) in 1997 and served as Center Director until mid 2002. His research interests are application of Remote Sensing and GIS for erosion control engineering, flood and sediment discharge models, 3D photogrammetry and image processing, forest feature extractions using high resolution remote sensing data, developing free software for image processing. He is actively involved in Digital ASIA and several other projects.