FUSION APPROACH ON LOW RESOLUTION SATELLITE IMAGES FOR LAND COVER APPLICATION

Abstract

Since the launch of Landsat-1, the first Earthresource satellite in 1972, satellite image processing has become an increasingly important tool for the inventory, monitoring, management of earth resources and many other applications (Draeger et al. 1997). The increasing availability of information products generated from satellite images has added greatly to our ability to understand the patterns and dynamics of the earth resource systems at all scales of inquiry (Lambin et al. 2001). Satellite remote sensors can be divided into two major types of imaging systems: optical (optical and thermal) and radar imaging systems. Optical imaging systems operate in the visible and IR (Infra Red) regions of the spectrum. Their operational use is weather dependent, since clouds are not transparent at visible/IR wavelengths (0.4-14 urn). Some of the satellite images working on optical and thermal images can be listed as AVHRR (Advanced Very High Resolution Radiometer), MODIS (Moderate Resolution Imaging Spectroradiometer), Landsat (land Satellite), LISS (Linear Imaging Self Scanner), SPOT (Satellite Pour L'Observation de la Terra or earth observing satellites) and many others. On the other hand, radar imaging systems works in microwave region (1 GHz to 30 GHz), and are very much atmosphere and weather independent. ERS (European remote sensing satellite), JERS (Japanese earth remote sensing), ENVISAT (Environmental Satellite Advanced Synthetic Aperture Radar), RADARSAT (Radar Satellite), and PALSAR

(Phased Array L-band Synthetic Aperture Radar) are some of the radar satellite sensors available for various applications. One of the important application of satellite image processing is the generation of landuse/ land-cover maps (Anderson et al. 1976)'in comparison to more traditional mapping approaches such as terrestrial survey and basic aerial photo interpretation which is quite cumbersome. Land-use mapping/ classification using satellite imagery has the advantages of low cost, large area coverage, repetitively, and computitivity. Eventually maximum high resolution satellite images is expensive with some extent. Therefore, it is need of current research to explore some techniques by which utilization of freely available satellite image may be enhanced. The increasing availability of satellite imagery with significantly improved spectral and spatial resolution has offered greater potential for more detailed land cover classification. The availability of MODIS image with greatly improved spectral, spatial, geometric, and radiometric attributes provides significant new opportunities and challenges for remote sensing-based land cover classification (Friedl et al. 2002) as well as other applications. MODIS has several spectral bands with are useful for various application in one hand and in other hand its spatial resolution varies from 250 m to 1000 m. This spatial resolution is not so enough to get good classification accuracy on MODIS images. Therefore, there is a need to explore the possibility of the use of techniques like fusion that may be helpful to increase the utilization of MODIS image for land cover classification/ land use maps. Time series analysis is another important aspect by which changes may be identified in a particular class of the land cover. Alparone et al. 2004 have demonstrated the benefit of combining optical and radar image for improved land cover mapping in several studies. With the availability of multifrequency and high-resolution space borne radar data, such as Advanced Land Observation Satellite (ALOS) Phase Array L-type Synthetic Aperture Radar (PALSAR), an increased interest in tools to exploit the full information content of both image types is arising. Unsupervised clustering is a fundamental tool in image processing for geosciences and satellite imaging applications (Stuckens et al. 2000). A well review of clustering method is reported by Jain et al. in 1999. For example, unsupervised

clustering is often used to obtain vegetation maps of an area of interest. This approach is useful when reliable training image are either scarce or expensive, and when relatively little a priori information about the image is available. Unsupervised clustering methods play a significant role in the pursuit of unsupervised classification (Richards and Jia 1999). Eventually this unsupervised clustering can be used for hotspot and non-hotspot region classification. Recent advances in satellite image processing have expanded opportunities to characterize the seasonal and inter-annual dynamics of natural and managed Land use/ land cover communities. The development of a regional-scale monitoring procedure is challenging because it requires remotely sensed image that have wide geographic coverage, high temporal resolution, adequate spatial resolution and minimal cost. The MODIS offers an opportunity for detailed, large-area Land use/ land cover characterization by providing global coverage of science quality image with high temporal resolution (1-2 days) and intermediate spatial resolution (Justice and Townshend 2002). The spatial, spectral, and temporal components of the MODIS may be appropriate for multitemporal harmonic analysis. Harmonic analysis is useful for analyzing seasonal and inter-annual variation in land surface condition (Wan et al. 2004). This type of analysis may develop the possibility to quantify and classify some fundamental characteristics, related to the phenology of vegetation, water and others. The main aim of this thesis was to study and to maximize the utilization of low resolution freely available satellite image for various land cover classification. For this purpose, freely available MODIS image has been used and it is attempted to develop suitable algorithm for fusion technique for land cover classification and harmonic analysis. This thesis is organized in seven chapters. The first chapter presents the introduction of the thesis which includes motivation, major research gaps, and details about the study area and satellite image used. Brief review of related work is presented in chapter II. We have explored the fusion technique for enhancing the land cover classification of low resolution satellite image especially freely available satellite image like MODIS in the chapter III. One of the aim of this chapter is to analyze the

effect of classification accuracy on major type of land cover types like agriculture, water and urban bodies with fusion of ASTER image to MODIS image and enhance the classification accuracy of MODIS image at spatial level. For this purpose, we have considered to fuse, high resolution i.e., like 15m resolution ASTER image with moderate resolution i.e., like 250 m MODIS satellite data. MODIS band 1 and band 2 are used as a moderate resolution data, where as ASTER band 2 and band 3 are considered as high resolution data. Curvelet transformation has been applied for fusion of these two satellite images and Minimum Distance classification technique has been applied on the resultant fused image for classifying the fused image in major land cover classes (i.e., agriculture, urban and water). The fuzzy based fusion is also applied for fusion of these two satellite images. After the fusion by fuzzy based approach, the minimum distance classification technique is used to classify the resultant fused image in major land cover classes (i.e., agriculture, urban and water). It is quantitatively observed that the overall classification accuracy of MODIS image after fusion is enhanced at spatial level. The quality of fused image is assessed by quality indicators. Another important point which one should consider while doing the time-series analysis, where every MODIS image may require the same number of high resolution image for fusion in one hand and in another hand every time one has to carry out the complex computation of fusion. Generally high resolution image i.e., ASTER is not freely available. So, it is another point of research to develop such a methodology or coefficients by which use of high spatial resolution image(i.e., ASTER in our case) for fusion for time series analysis may be minimized. Therefore, in this chapter, we have attempted to explore to find the possible methodology to search some fusion coefficient by which ASTER image use may be minimized, while analyzing time series image for observing the land cover classification. For this purpose, firstly we have carried out fusion of MODIS (band 1 and band 2) image with ASTER (band 2 and band 3) image using curvelet transform and fuzzy based technique. After this, we have obtained a fusion coefficient that may be minimizing the use of ASTER image(i.e., every time of fusion of MODIS image with ASTER data, we may use this derived coefficient). Another advantage of this fusion coefficient is that every time we do not have to carry out curvelet transform on ASTER image by which it reduces the computation complexity up to a certain extent. We have obtained the fusion coefficient with three MODIS and one ASTER image of Page | vi month March 2001. The obtained coefficient is validated with the MODIS image of another year and gives the satisfactory result. In the chapter IV, another important aspect of fusion of different sensors image like optical and radar images (where both can provide the complimentary information) is carried out and the quality of fused image is assessed by various assessment indicators. For this purpose an attempt has been made to fuse the PALSAR image with MODIS image using curvelet based fusion and quality assessment of fused image has been done. PALSAR image has a advantage of availability of image in four different channels. These four channels are HH (Transmitted horizontal polarization and received also in horizontal polarization), HV(Transmitted horizontal polarization and received vertical polarization), VH (Transmitted vertical polarization and received horizontal polarization) and VV (Transmitted vertical polarization and received vertical polarization)(www2 2009), which provides various landcover information. We have used the curvelet and fuzzy based technique for fusing the PALSAR (HH, HV and VV) image to MODIS (band 1 and 2) image in spatial resolution. Each band of PALSAR (i.e., HH, HV and VV) is individually fused with MODIS band 1 and Band 2 separately in one hand and in other hand fused image of MODIS band 1 and 2 is individually fused PALSAR (HH, HV and VV) bands separately. The quality of fused image is assessed by assessment indicator like Correlation, RMSE (root mean squared error), Relative Mean Difference, Relative Variation Difference, Deviation Index, PSNR (Peak signal-tonoise ratio) and Universal Image Quality Index. These indicators are applied to measure and compare the performance of the fused images. The results are quite encouraging and we get quite a good overall classification accuracy after fusion. Thereby in near future it may provide a better platform for the maximize the use of MODIS images. Classification of various class is one aspect whereas focus for certain class is another aspect, therefore in this thesis, we have considered for both type of objective where in case 1, we have classified image into three major land cover classes (Agriculture, water and urban) and in second case we have attempted to focus the subsurface fire (i.e., hotspots) in the image and considered it as one class. We have considered Jharia, region of India as a test area for this purpose. In the chapter V, we

have explored the application of MODIS and LISS-III image for hotspot and nonhotspot regions. Although MODIS provides a special product MOD14A2 for fire product classification. But this special product is not only sufficient for hotspot and non-hotspot regions. Therefore objective of this chapter is to use various information of different corresponding bands to focus the hotspot in the images. For this purpose, in this chapter, an approach based on Binary Division Algorithm is used for hotspot and non-hotspot regions using band 1 and band 2 of MODIS and band 2 and band 3 of LISS - III for the Jharia (India) Region. Results are compared with the MOD14A2, the exclusive MODIS product for thermal anomalies and fire, and it is found that the proposed approach gives quite satisfactory results in comparison to MOD14A2 products. It is important to analyze the changes in the classes from year to year therefore harmonic analyses is another aspect to see the respective changes in different classes, hence in the next chapter, i.e., chapter VI, we have tried to characterized the changes for agricultural and water land use/land cover in Western Utter Pradesh and part of Uttarakhand of India form the year 2001 to 2008. In this perspective, we have considered the MODIS NDVI (Normalized difference vegetation index) and NDWI (Normalized difference water index) images for agricultural and water regions respectively. Harmonic analysis, also known as Fourier analysis, decomposes a timedependent periodic phenomenon into a series of sinusoidal functions in which each defined by unique additive and amplitude values. In consequence, the additive image Ao and the amplitude images Ai were produced, respectively. With these images we have checked the changes in agricultural and water land use/land cover in the test area, in one hand we have analyzed the changes or variation in agriculture and water for the whole image, whereas in the other hand, we have analyzed the changes or variation in agriculture and water for the selected region of interest. Such type of study is very helpful in near future to optimize the use of MODIS image in one hand and in another hand to develop monitoring system by which changes during the particular month may be observed. Finally in chapter VII, the contributions made in the thesis are summarized and scope of future work is outlined.