PLEASE SCROLL DOWN FOR ARTICLE

Transcription

1 This article was downloaded by:[rmit University] [RMIT University] On: 28 June 2007 Access Details: [subscription number ] Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: Registered office: Mortimer House, Mortimer Street, London W1T 3JH, UK International Journal of Remote Sensing Publication details, including instructions for authors and subscription information: Evaluation of radiometric resolution on land use/land cover mapping in an agricultural area Online Publication Date: 01 January 2007 To cite this Article: Rao, N. Rama, Garg, P. K. and Ghosh, S. K., (2007) 'Evaluation of radiometric resolution on land use/land cover mapping in an agricultural area', International Journal of Remote Sensing, 28:2, To link to this article: DOI: / URL: PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: This article maybe used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material. Taylor and Francis 2007

2 International Journal of Remote Sensing Vol. 28, No. 2, 20 January 2007, Evaluation of radiometric resolution on land use/land cover mapping in an agricultural area N. RAMA RAO*, P. K. GARG and S. K. GHOSH Geomatics Engineering Section, Department of Civil Engineering, Indian Institute of Technology Roorkee, Roorkee , India (Received 10 September 2005; in final form 3 April 2006 ) Motivated by the increasing availability and importance of high radiometric resolution remote sensing data, this study aims to determine whether current generation high radiometric resolution remote sensing data could be used accurately for land use/land cover classification in place of traditional moderate radiometric resolution multispectral data. A comparative study has been carried out to evaluate the utility of the simulated 12-bit LISS-III sensor compared with that of the original 7-bit LISS-III sensor for land use/land cover classification. It has been found that there is a small increase of 3% in overall accuracy by using the high radiometric resolution (12-bit) LISS-III data over the moderate radiometric resolution (7-bit) LISS-III data for land use/land cover classification. A 4 6% difference in classification accuracy of simulated LISS-III for cotton, chillies, black gram, and sugar cane indicates the usefulness of high radiometric resolution data for ground classes that are more heterogeneous within the same class. This study suggests that high radiometric resolution satellite data may not be a requisite for accurate land use/land cover mapping, whereas the spectral bandwidth, band placement, and method of classification parameters may be more important. 1. Introduction Land use/land cover mapping and analysis is one of the most successful applications of remote sensing. However, identification, discrimination, and mapping of various land use/land cover classes depends on the spatial, spectral, and radiometric resolution of the remote sensing satellite (Teillet et al. 1997). For some of these applications, relative radiometric accuracy of multispectral systems is of paramount important; while for others, the ability to delineate spatial features is more important (Bonan 1993). There have been extensive studies performed on the suitability of various kinds of spatial and spectral resolutions to various categories of application areas such as land use/land cover mapping, urban studies, vegetation monitoring, etc. (Curran et al. 1992). Incidentally, most of these studies used satellite data of 8-bit, 10-bit, or 12-bit radiometric resolution, and little attention has been paid to the potential impact of the radiometric resolution on the accuracy of land use/land cover mapping. While it is commonly thought that greater spatial resolution is the key to proper land use classification, finer spectral and radiometric resolution also have potential advantages that remain only partially explored. Radiometric resolution *Corresponding author. ramunr@yahoo.com International Journal of Remote Sensing ISSN print/issn online # 2007 Taylor & Francis DOI: /

3 444 N. R. Rao et al. refers to the sensitivity of a remote sensor to variations in the reflectance levels. The higher the radiometric resolution of a remote sensor, the more sensitive it is to detecting small differences in reflectance values. The advantages of high radiometric resolution are well documented in domains such as mineralogical mapping (e.g. Smailbegovic et al. 2000). Smailbegovic et al. (2000) have shown that an area within a scene that was identified from Landsat-TM data as predominantly kaolinite rich was found to consist of not only kaolinite, but also illite- and jarosite-abundant zones when using 10-bit Landsat-TM data simulated from AVIRIS. Upon our extensive literature search and review, it has been observed that the potential advantages of higher radiometric resolution have been only partially explored and are confined to studies related to the measurements of aerosol optical depth (Devara et al. 2001), leaf pigments such as chlorophyll (Joseph et al. 1997), and water quality parameters (Nanda et al. 1998). With few exceptions, such studies used multispectral data like Landsat TM, MODIS, VEGETATION, or ETM +, rather than moderate radiometric resolution multispectral sensors, such as the IRS LISS-III. A very limited number of studies have reported on the impact of radiometric resolution in land use/land cover classification (Platt 2001). Platt (2001) has reported the comparative performance of synthetic Landsat data (8-bit) simulated from AVIRIS hyperspectral sensor with a subset of AVIRIS bands (10-bit), which are similar to Landsat-TM spectral bands, for a location on the urban fringe of Colorado, USA. In this study, the author reported that AVIRIS data (10-bit) allowed for improved urban land use classification over synthetic Landsat-TM data (8-bit). However, this study has not taken into consideration the impact of the narrow spectral bands of AVIRIS (as AVIRIS has 220 bands with 10 nm bandwidth) on classification, besides having only one broad class, i.e. urban land use. It is difficult to infer from existing studies whether higher radiometric resolution multispectral sensors offer an improved sensitivity for land use/land cover classification. With advances in sensor technology, the spate of space-borne imaging systems is rapidly changing, and in the near future, higher (12-bit or more) spectral and radiometric resolution imagery will be routinely available to the user community. The potential advantages and compatibility (spatial, spectral, and radiometric resolution) of using high-end imagery with the now available satellite imagery has to be considered in many studies that involve historic and temporal datasets (e.g. change detection, climate change, etc.). Owing to the importance of land use/land cover information in an agriculture-dominated area, a study has been carried out to evaluate the impact of high radiometric resolution (12-bit) over moderate radiometric resolution multispectral imagery (7-bit) in land use/land cover classification. 1.1 Characteristics of LISS-III and Hyperion The LISS-III sensor carried by IRS-IC, IRS-ID, and Resourcesat-1 has been in wide use by various national and international remote sensing communities for a range of natural resource monitoring and inventory. LISS-III is a multispectral sensor of high spatial resolution with moderate spectral (4 bands) and radiometric resolution (table 1). The increasing availability of hyperspectral satellite data (e.g. Hyperion) has opened vistas to evaluate the spectral and radiometric quantization of broadband multispectral sensors, like LISS-III, TM, or ETM +. The Hyperion sensor carried by the NASA Earth Observing 1 (EO-1) satellite (Pearlman et al. 2001) is the first space-borne hyperspectral instrument to acquire both visible

4 Remote Sensing Letters 445 Table 1. Characteristics of LISS-III and Hyperion. LISS-III Hyperion Sensor Type Push broom Push broom Spatial resolution (m) Swath (km) Repetitivity (days) Spectral bands (nm) Discrete (4 bands) Continuous (242 bands) Radiometric quantization (bit) 7 12 SNR SNR decreases as wavelength increases near-infrared [(VNIR) gm] and shortwave infrared [(SWIR) gm] spectra (table 1). 2. Materials and methods 2.1 Study site The study site for the present work lies in the Guntur district in the Andhra Pradesh state of India, situated between 16u u N latitudes and 79u u E longitudes. A location map of the study area is shown in figure 1. The annual rainfall of the district is 689 mm. The climate is semi-arid and generally warm in summer. The soils, in general, are very fertile, and are broadly classified as black cotton, red loamy, and sandy loamy. The black cotton area is 70%, red loamy is 24%, and sandy loamy is about 6% of the area in the district. The predominant crops grown in the district are paddy, jowar, and bajra among cereals, black gram, green gram, and red gram among pulses, cotton, chillies, sugar cane, turmeric, and tobacco among non-food and commercial crops. A single image each of Hyperion (date of pass: 28 September, 2003 and path/row: 143/48) and IRS LISS-III (date of pass: 27 September, 2003 and path/row: 100/61) Figure 1. Location map of the study area.

5 446 N. R. Rao et al. has been procured for the kharif season (June November) of Since the width of the Hyperion image is 7.5 km, the common area from the LISS-III image has been extracted. Both the Hyperion and LISS-III datasets have been georeferenced with rms error of 0.31 pixel size after selecting 30 ground control points (GCPs) from Survey of India (SOI) toposheets and ensuring proper distribution throughout the image. Both the images have been resampled with 23.5 m pixel size using the nearest neighbourhood method, using ERDAS IMAGINE software. The standard atmospheric corrections on LISS-III data have been performed using the 6S radiative transfer code. The Hyperion sensor collects hyperspectral image data over the continuous spectrum from 356 to 2577 nm at a spatial resolution of 30 m. The delivered USGS Hyperion product will contain 242 bands, out of which 44 bands are not calibrated. The main reason for not calibrating all bands is due to decreased sensitivity of the detectors within the non-calibrated spectral regions. All bands that are not calibrated are set to zero (null values) in the Level 1R product. The 198 calibrated channels will cover a complete spectrum from 426 to 2395 nm. Also, due to an overlap between the VNIR and SWIR channels (VNIR bands and SWIR bands 77 78), there will be 196 unique channels in the final data product. Atmospheric water vapour bands that absorb almost the entire incident and reflected solar radiation are easily identified by visual inspection of the image data. Accepting this as a good criterion for band elimination for land surface applications has yielded the subset of 161 bands used in this study. The wavelengths removed correspond to strong atmospheric water vapour absorption bands between 1356 and 1417 nm, 1820 and 1932 nm, and above 2395 nm. Since Hyperion operates from a space platform with consequently modest surface signal levels and full-column atmospheric effects, its data demands careful processing. Atmospheric water vapour is a key factor in the atmospheric correction of remote sensing data. The atmospheric correction of hyperspectral data has a clear advantage over multispectral data since the magnitude of water vapour effects in every pixel can be assessed directly from a few spectral channels of the data themselves. The striping has been minimized using the Fast Fourier Transform technique (Datt et al. 2003). This de-striped Hyperion image is further processed to perform standard atmospheric corrections using Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) atmospheric correction software module. The desired properties (reflectance, column water vapour, etc.) are derived from the spectral radiance at each image pixel using look-up tables that are generated from these simulations. FLAASH is available as an optional module in the ENVI software package. The 12-bit radiometric resolution LISS-III image has been derived from the atmospheric corrected Hyperion image. Simulation is achieved by weighting the appropriate narrow Hyperion spectral bands within the LISS-III spectrum (table 2) according to the spectral response function of the LISS-III instrument. This Table 2. List of Hyperion bands used to simulate LISS-III image. LISS-III Hyperion bands nm nm nm nm

6 Remote Sensing Letters 447 simulated image had the same radiometric resolution as the Hyperion data with 12-bit grey levels, from which it has been derived, but essentially the same band placement and bandwidth as the original or base LISS-III data. 2.2 Method of image classification To find out if the simulated LISS-III image contains more information than the original LISS-III image for land use/land cover mapping, classification has been carried out on both the original images. This study used a variety of supervised classification algorithms, but focused on a single one: the maximum likelihood (ML) classifier. MLC is a widely accepted classifier due to its robustness and simplicity. MLC is applied to both the original LISS-III image (4 bands) and the simulated LISS-III image (4 bands) for detailed land use/land cover classification. The training samples for different land use classes have been interactively defined based on the homogeneity of the samples and information derived from topographic maps and field visits. The training pixels were extracted from the same areas from both the images. To enable a meaningful calculation of statistics, the minimum training samples of a class are chosen based on the criteria of maintaining a minimum 10 n training pixels for each class (table 3), n being number of spectral bands. Accuracy assessment of the classified images has been carried out by generating classification error matrix by using different sets of testing pixels for both of the classified images. These testing pixels were collected during field visits performed prior to image classification and after image classification. Care has been taken to ensure that testing pixels are different from those pixels that have been used as training pixels. 3. Results and discussion The Kappa coefficient from the original LISS-III is 0.831, while it is for the simulated LISS-III. It is observed that the difference in classification accuracy is only 3%. This indicated higher radiometric resolution (12-bit) data does not bring about an increase in classification accuracy in comparison to the original LISS-III (7-bit). In order to statistically test the significance of the accuracy differences, t-test has been conducted between the Kappa coefficients of the classification result. The Z-value obtained (1.41) indicates that there is no difference between the two images Table 3. Training and testing samples for each class. LU/LC Class No. of training pixels No. of testing pixels (holdout) Water body Fallow land Bare soil Rice Cotton Sugar cane Chillies Built-up Tobacco Shrub land Red gram Black gram

7 448 N. R. Rao et al. (a) (b) Figure 2. (a) LU/LC map derived from original LISS-III data. (b) LU/LC map derived from simulated LISS-III data. obtained by using 12-bit and 7-bit data for land use/land cover classification. The resulting maps are shown in figure 2. Another aspect to be examined relates to per-class accuracy when using satellite images of different radiometric resolutions. This has been examined by plotting conditional kappa coefficients of each class for both datasets (figure 3). It has been observed that for water, rice, tobacco, and red gram, both the simulated LISS-III (12-bit) and the original LISS-III (7-bit) images give the same accuracy, and that their kappa values are close to 0.9. For sugar cane, cotton, chillies, black gram, tobacco, shrub land, bare soil, and agricultural fallows, the simulated LISS-III (12- bit) data gives 4 6% higher classification accuracy than the original LISS-III (7-bit) data. Thus, it can be inferred that for land use classes that are spectrally pure with no background effect, moderate radiometric resolution data, like LISS-III, gives the same classification accuracy in comparison to high radiometric resolution data, like simulated LISS-III (12-bit) in this case. In case of settlements (built-up), there is a marginal decrease (2.6%) in classification accuracy when using simulated LISS-III (12- bit). This decrease can be explained by the high spectral confusion that arises between settlements and other spectrally similar classes, such as sandy soil, bare soil, etc. Further, for fallow lands, cotton, chillies, and shrub land, the difference in classification accuracy between two datasets is 12%, 13%, 10%, and 12%, respectively (figure 3). This increase in classification accuracy is a remarkable feature, indicating enhancement in classification accuracy for ground classes that are more heterogeneous within the same class. In particular, the increase in classification accuracy of shrub land and agricultural fallows by the simulated LISS-III is interesting, and difficult to explain. It can be explained, partly, by the

8 Remote Sensing Letters 449 Figure 3. The conditional kappa coefficients of individual classes for the best classification using the LISS-III data and the simulated LISS-III data. 1. Water. 2. Bare soil. 3. Agricultural fallows. 4. Rice. 5. Sugar cane. 6. Red gram. 7. Tobacco. 8. Cotton. 9. Chillies. 10. Settlements. 11. Shrub land. 12. Black gram. fact that the shrub land in the study area is associated with vegetation (e.g. grass, bushes), in many places creating mixed pixels in the LISS-III image, which mix with agricultural fallows, thereby reducing the accuracy of classification. However, in the simulated LISS-III, upon verification of the raw digital counts, it was found that there has been a clear (though moderate) difference in the digital numbers range, which might have helped the simulated LISS-III to increase classification accuracy of shrub land and agricultural fallows. This observation supports the inference drawn that heterogeneous classes with a soil background can be classified with improved accuracy with an increase in radiometric resolution. 4. Conclusions Several general inferences can be made from this study. A supervised classification with a simulated LISS-III image (12-bit) has shown a marginal increase in overall accuracy than with a LISS-III (7-bit) synthetic image for land use/land cover in an agriculture-dominated area. Based on the marginal increase (3%) in overall accuracy, it can be inferred that there is a small increase in classification accuracy by using the high radiometric resolution (12-bit) LISS-III data over the moderate radiometric resolution (7-bit) LISS-III data for land use/land cover classification. The comparatively increased classification accuracy of simulated LISS-III for sugar cane, cotton, chillies, and black gram indicates the usefulness of high radiometric

9 450 Remote Sensing Letters resolution data for ground classes that are more heterogeneous within the same class. However, further studies are recommended to confirm this observation. The classification accuracy of water, rice, red gram, and tobacco (almost similar with both images) makes clear that when the ground category is homogeneous, an LISS- III image (7-bit) can offer the same accuracy as that when using high radiometric resolution satellite data. Our results suggest higher spatial and spectral resolution data may be more useful to achieve high accuracy land use/land cover maps. Since classification accuracy is dependent on a number of factors besides resolution, caution should be used in extending the conclusions of this study to other research. Acknowledgements The first author wishes to acknowledge the All India Council for Technical Education, New Delhi for the financial support received in the form of a National Doctoral Fellowship to carry out this research work. References BONAN, G., 1993, Importance of leaf area index and forest type when estimating photosynthesis in boreal forests. Remote Sensing of Environment, 43, pp CURRAN, P.J., DUNGAN, J. and GHOLZ, H.L., 1992, Seasonal LAI measurements in slash pine using Landsat TM. Remote Sensing of Environment, 39, pp DATT, B., MCVICAR, T.R., NIEL, T.G. and JUPP, D.L.B., 2003, Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indexes. IEEE Transactions on Geoscience and Remote Sensing, 41, pp DEVARA, P.C.S., RAMKUMAR, M. and MAHESH KUMAR, R.S., 2001, High spectral resolution radiometric measurements of aerosol extinction over an urban region. Measurement Science and Technology Journal, 12, pp JOSEPH, G.S., RAJAN, S.K. and TANDON, B.Y., 1997, The use of high spectral resolution data for detection of chlorophyll concentration in Chilka lake. Indian Journal of Applied Physics, 12, pp NANDA, M.V., SUMAN, J. and SUJANA, R.K., 1998, Extraction of water quality parameters using OCM data. Indian Journal of Applied Physics, 13, pp PEARLMAN, J.S., BARRY, P.S., SEGAL, C.C., SHEPANSKI, J., BESIO, D. and CARMAN, S.L., 2003, Hyperion, a space-based imaging spectrometer. IEEE Transactions on Geoscience and Remote Sensing, 41, pp PLATT, R.V., 2001, A Comparison of AVIRIS and Synthetic Landsat Data for Land Use Classification at the Urban Fringe, Interim report (Austria: International Institute for Applied Systems Analysis). SMAILBEGOVIC, A., TARANIK, J.V. and KRUSE, F., 2000, Importance of Spatial and Radiometric Resolution of AVIRIS Data for Recognition of Mineral Endmembers in the Geiger Grade Area, Nevada, USA. Proceedings of the Nirth JPh airborne earth science workshop (Pasadena, CA: Jet Propulsion Laboratory). JPL Publication No: TEILLET, P.M., STAENZ, K. and WILLIAMS, D.J., 1997, Effects of spectral, spatial, and radiometric characteristics on remote sensing vegetation indices of forested regions. Remote Sensing of Environment, 61, pp