DEVELOPMENT OF LAND SURFACE MONITORING ALGORITHMS FOR SATELLITE IMAGES

Abstract

Nowadays various satellite images are available which are used for different applications like: object detection and identification, land cover classification, soil moisture retrieval and vegetation mapping. Satellite images can be categorized according to the mode of acquiring the images as active sensor and passive sensor images. An active sensor which is having its own source of illumination, whereas the passive sensor depends on some other sources like sunlight or emerging radiation from target objects. Synthetic Aperture Radar (SAR) images are active sensor images as the SAR sensor possesses its own source. Optical images, which measure the sunlight radiation reflected from the earth in the visible and infrared band range of the electromagnetic spectrum, are called as passive sensors. Moderate-resolution Imaging Spectroradiometer (MODIS) sensed images are the passive sensor images as MODIS sensor does not emit their own radiation, but receive natural light from the earth’s surface. It is a need of current research to explore some techniques by which utilization of satellite images may be enhanced. But it is still a challenging task to use the different kind of freely or less expensive data for land surface applications such as land cover classification, agricultural areas monitoring, seasonal change monitoring, and hot spot monitoring, etc. Therefore, data applicability may be enhanced by using various techniques in which optimization based method is one of them. Researchers are using various techniques for land surface monitoring such as classification, change detection and parameter retrieval, etc. Image classification is a procedure of grouping pixels into meaningful classes and it can be referred as extracting useful information from satellite images. Various classification methods are presented to classify satellite images such as: supervised, unsupervised and optimization based method based classification, but all these techniques still need the attention once it is used for a low resolution satellite images like MODIS. Optimization based method is a combination of techniques, rather than a technology. There are various optimization based techniques like: genetic algorithm (GA), particle swarm optimization (PSO), neural network (NN), support vector machine (SVM), fuzzy logic, principal component analysis (PCA) etc., and their Abstract ii combinations which were used by many researchers for satellite image classification, image segmentation and other applications. During the land surface monitoring, the major factors arise, such as subsurface fire (hot spot) monitoring, agricultural areas monitoring especially the identification of the unimodal and bimodal agriculture areas. Land cover classification such as urban, vegetation and water is supportive in the management and planning of urban areas, although monitoring of vegetation can be important for several reasons like: (1) Agricultural areas monitoring and assessment of the state is an essential part in the process of adaptive management and (2) agricultural areas monitoring can be useful to identify the influence of threatening processes, such as climate change, disease and many others. Agricultural areas monitoring is conventionally done by fixed sites in the places where circumstantial variables are measured at consistent intervals to observe the changes (Briggs and Freudenberger, 2006). After monitoring the agricultural areas, changes have to be detected in it, but the main challenge is that there is no single technique to map the changes in a different kind of applications, whereas various change detection techniques have their own limitations and it is observed that still it is difficult to apply a single change detection approach for maximum types of change scenarios. The main aspect is to implement a suitable change detection technique that produces information about where changes are occurring in the agricultural areas. It is also important to know that the land surface is also used for area estimation and the extraction of unimodal (i.e. annual or yearly growth information) and bimodal (i.e. biannual or half yearly growth information) about the vegetation phenology in agricultural areas. 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. Studies have shown that the temporal domain of multispectral data often offers more information of land surface than the spectral, spatial, or radiometric domains. A regional-scale monitoring development procedure is challenging because it needs satellite images that have an adequate spatial resolution, high temporal resolution, wide geographic coverage and minimal cost. Satellite images with high temporal and moderate spatial resolution may provide the opportunity to extract the detailed, large area and land surface characteristics. Harmonic analysis with time series images is useful for analyzing the intra and inter-annual variations in land surface condition. This type of analysis may provide the possibility to quantify and Abstract iii classify some fundamental characteristics, related to the phenology of vegetation, water and other land covers. Subsurface fire (in this thesis, termed as hot spot) is presently a major problem in coal producing countries like India, South Africa, Indonesia, Australia, USA, Germany, and China. However, the nature and severity of the problem varies from country to country. These fires burn causes the destruction of the natural and man-made land structure, constitutes a hazard to human beings. Mostly fires take place due to spontaneous combustion of coal, which increases the temperature of the surface. Various optical/TM satellite images are used by researchers for hot spot detection and monitoring since last two to three decades, but still detection of hot spots with less false alarm and better accuracy is challenging task. Many researchers have used various techniques for hot spot monitoring with satellite images like principal component analysis (PCA) with information fusion, fuzzy logic, contextual thresholding model, image stacking, and wavelet transform based model. The application of low resolution satellite images with the fusion of some other high resolution images may be useful for hot spot monitoring. Therefore, present research work takes care of this issue and explores the applicability of such satellite data which are helpful to provide surface roughness information and can effectively be utilized for hot spot monitoring. Therefore, in this thesis, after observing the various challenges to use satellite data for land surface monitoring, the following tasks were formulated. 1) Critical analysis of some land cover classification techniques for satellite images. 2) A potential application of KLT tracker on satellite images for automatic change detection in agricultural areas. 3) Development of Harmonic analysis based technique for agricultural areas monitoring. 4) An efficient application of fusion approach for hot spot detection with MODIS and PALSAR-1 data. For this purpose, two types of satellite data have been used: (i) Advanced Land Observing Satellite Phased Array L-band Synthetic Aperture Radar (ALOS PALSAR), (ii) Moderate-resolution Imaging Spectroradiometer (MODIS) for land surface classification, agricultural areas monitoring, agriculture pattern monitoring (i.e., unimodal and bimodal pattern monitoring) and hot spot monitoring. ALOS PALSAR level 1.1 in the Committee on Earth Observation Satellites (CEOS) format, which is launched by Japan Aerospace Exploration Agency (JAXA) in 2006, is used for the land surface classification, agricultural Abstract iv area monitoring and hot spot monitoring. The fully polarimetric PALSAR data, which is the single look complex data has four different polarization modes such as horizontally transmitted horizontally received (HH), horizontally transmitted vertically received (HV), vertically transmitted horizontally received (VH) and vertically transmitted vertically received (VV) polarization. The ALOS PALSAR satellite is more suitable to land monitoring for various causes like: the long operating wavelength (23.6 cm), which is sensitive to land, structures and moisture appearances with 25 m resolution. MODIS data which is a payload scientific equipment launched into Earth orbit by NASA was taken as the optical data for land surface classification, time series analysis and for the fusion purpose with PALSAR to monitor the hot spot. MODIS images have good temporal coverage of morning & evening time daily images (i.e. known as Aqua and Terra), these images are freely available. MODIS offers excellent possibilities for land surface monitoring, as it has 36 different spectral bands in the spectral wavelength range from 0.4 μm to 14.4 μm, which allows for a full and automated atmospheric correction and detailed cloud masking. In this thesis, the surface reflectance product of MODIS (MOD09Q1) which comprises two spectral bands i.e., Red (620-670 nm) and Near-Infrared (841-876 nm) with 250 m resolution is used. This thesis includes seven chapters. Chapter 1 presents the introduction of the thesis which consists of motivation, scope and objectives of the thesis. Chapter 2 discusses the brief literature review of the land surface monitoring, agriculture area monitoring and hot spot monitoring with satellite images. Chapter 3 presents the critical analysis of different land surface classification techniques. Image classification is one of the techniques which can be used for monitoring purposes and it can be classified mainly into two categories: supervised and unsupervised classification. Supervised classification techniques on satellite images require the need for human interaction to determine the classes and training samples. Unsupervised classification techniques classify the image automatically, but has the limitation to accurately divide land cover into natural classes, whereas supervised classification techniques assign the sample pixels to the related classes accurately which are already labeled by virtue of ground truth, existing maps, or inference from experimental data. K-means and ISO-data based clustering are usually used for satellite image applications for unsupervised classification. Both techniques use an input metric like Euclidean distance for measuring likeness or similarity of pixels belonging to each cluster. There are many other unsupervised classification algorithms for satellite images Abstract v that are based on techniques like K-nearest neighbor clustering, fuzzy models, Markov Random Field (MRF) models, particle swarm optimization, genetic algorithms and some others. Although, several supervised methods are available such as Maximum Likelihood, Minimum Distance, Parallelepiped, Mahalanobis Distance, Neural Networks and Support Vector Machine (SVM). Land cover classification techniques were critically analyzed by applying the supervised and unsupervised classification techniques and optimization based techniques onto the MODIS and PALSAR images of the Roorkee region to classify it into the four major classes (i.e. urban, water, bare land and agriculture). For the ALOS PALSAR data, the overall accuracy is obtained as 59%, 58%, 56% and 70% for K-means, ISO-data, Parallelepiped and Minimum Distance techniques, respectively, whereas for MODIS data, the overall accuracy is obtained as 44%, 48%, 53% and 59% for K-means, ISO-data, Parallelepiped and Minimum Distance techniques, respectively. optimization based techniques such as GA, SVM and NN are also critically analyzed for PALSAR and MODIS data and the overall accuracy for PALSAR data is obtained as 65%, 72% and 71%, respectively, whereas for MODIS data, the overall accuracy is obtained as 49%, 62% and 51%, respectively, for GA, SVM and NN. It is observed that although these techniques are good and some of them producing quite good results but these techniques are affected by various factors such as supervised classification techniques require precise a priori information for each class before classification. Further, they are based on the distance measure, which may lead a misclassification of pixels lying on the boundary of two classes. On the other hand, unsupervised classification methods do not require training samples, i.e., a priori information; however, they too can misclassify the pixels. So, it is observed that it may be difficult to use these techniques for change detection or monitoring purpose. Chapter 4 discusses the development of automatic change detection algorithm. In recent years, understanding and prediction of the mass of vegetation and productivity have become gradually significant, specifically with advances in the economic value of the environment due to land use and climate change. Fully polarimetric Synthetic Aperture Radar (SAR) data provide a good option for monitoring the vegetation, because it has the capability to provide appropriate and unaffected data without being strongly affected by atmospheric conditions and cloud cover. Therefore, in this chapter, the possibility of application of Kanade-Lucas-Tomasi (KLT) tracker to track agricultural areas with PALSAR satellite images has been explored, which is generally quite useful for monitoring or tracking the objects in the video Abstract vi images. KLT tracker has been used by many researchers for real-time applications like: optical flow measurement, multi-target tracking in real-time surveillance video, real-time people tracking system, aircraft tracking and an automated real-time image georeferencing/registration but its use of satellite images is very limited or not used. The study area is taken as the Roorkee city and its nearby places and it is located in the Haridwar district, Uttarakhand, India. The considered study area covers the main land covers like: water, urban and agricultural areas. Two PALSAR data sets were used, which were acquired on April 9th, 2010 and April 12th, 2011. After preprocessing both dated PALSAR images, the decision tree classification (DTC) technique has been applied to both the data sets to classify them into the three classes as water, urban and agricultural areas (which include tall and short vegetation) and overall accuracy (OA) for the Roorkee region was achieved 83.75% and 80.77%, respectively, where the total number of pixels belongs to the agricultural areas are 83254 and 89366. The classified pixels belongs to the agriculture class were further used as reference pixels to see the effect of KLT tracker for agricultural areas feature selection process. KLT tracker has been implemented on HH, HV and VV polarization images of the Roorkee region, but these images fail to select and track the agricultural areas features. Therefore, polarimetric indices ratio vegetation index (RVI) and cross polarization ratio (CPR) and the ratios of backscattering coefficients (HV/HH and HV/VV) were extracted by using the HH, HV and VV polarization images because these indices have the capability to distinguish different types of land cover. In KLT tracker, selection of appropriate parameters like window size, number of features and minimum eigenvalue and their optimization is a very critical task. For this purpose, a detailed critical analysis has been carried out to find suitable parametric values and after an exhaustive analysis, it is observed that KLT tracker selected the maximum agricultural area when window size, number of features and minimum eigenvalue are set as (4×4), 40,000 and 7000, respectively. Further, maximum agricultural areas have been selected and this process has been applied separately for all the considered polarimetric indices images. KLT tracker selects the number of agricultural areas pixels on RVI, CPR, HV/HH and HV/VV as 68036, 67610, 67410 and 66209, respectively, and tracks these agricultural areas pixels on second dated images as 78884, 79436, 81408 and 78991, respectively. After tracking the selected agricultural area successfully for all the four input images, it is observed that KLT tracker produces the most effective results for accurate tracking with HV/HH image. A comparison has been made between the selected and tracked agriculture pixels with the decision tree classified images for 2010 and 2011 data Abstract vii sets of the Roorkee region. After KLT tracker implementation on Roorkee region images, it is observed that 67410 number of agricultural area pixels are identified from 2010 image and in the same way on the 2011 image 81408 numbers of agricultural areas pixels are tracked efficiently, which are almost close to the observed pixels from the decision classification technique for Roorkee region. Changes occurred in agricultural areas are observed by applying the image differencing technique onto the agricultural areas pixels identified image and agricultural areas pixels tracked image. It is observed that change pixels are observed as positive change pixels (number of pixels are 27308) and negative change pixels (number of pixels are 13310), while no change pixels are observed as 54100 pixels in agricultural areas. Thus, the KLT tracker may be useful for this type of applications. Chapter 5 explores the utilization of MODIS normalized difference vegetation index (NDVI) time series data for land use classification to identify the agricultural areas growth cycle (i.e. annual and biannual change information). Monitoring of agricultural land in particular regions is still a challenging task. Many researchers have performed land use classification on to the time series data by applying different classification techniques like: supervised and unsupervised classification, decision trees, support vector machine neural networks, contextual algorithms, sub-pixel analysis based algorithms and object-based algorithm, but these techniques have some limitations predominantly in moist tropical regions due to limitations of satellite data and complex biophysical environmental conditions. Therefore, in this chapter an attempt has been made to develop such a monitoring technique by which nature of agricultural variations may be obtained by utilizing the MODIS time series images. MODIS data has moderate resolution (250 m) and its temporal frequency is very high. The considered test area for this task was of western Utter Pradesh (Shahanpuer, Muzaffarnagar and Deoband) and part of Uttarakhand (Roorkee), India, and their nearby regions and these regions comprise the urban, water, bare grounds and agriculture bodies and at least two cultivation periods are there and major crops such as wheat and rice have a short cycle. For time series analysis, Fourier or harmonic analysis has been used which transforms the complex temporal profile into smaller repetitive signals, which is helpful to enhance the signal analysis with intra and inter-annual frequencies. Each term refers to the number of full cycles completed by the wave at a certain interval (i.e. the first harmonic amplitude term has one cycle, and the second harmonic amplitude term has two cycles). First harmonic amplitude term (A1) provides annual or yearly growth (Unimodal) information about vegetation phenology in agricultural Abstract viii areas and the second harmonic amplitude term (A2) provide biannual or half yearly (bimodal) growth information (Canisius et al. 2007). The additive term (A0) generally implies the arithmetic mean of NDVI over the time series. It is observed, in most of the studies first and second harmonic amplitude terms are used to interpret the periodic pattern changes in the time series data and three, four, five, and six are the higher order harmonic amplitude terms which have been discarded because these terms indicate the trivial amplitude. The main challenge with this type of study is an efficiently classified image is required, which should have the capability to segregate the agricultural areas. Many researchers are using different classification techniques which may be useful for single image classification, but with time series analysis, these algorithms are not effective. Some researchers have used the additive term (A0) image, first harmonic amplitude term (A1) and the second harmonic amplitude term (A2) images for land cover classification, but they are not able to produce better classification accuracy. Therefore, in this chapter, an attempt has been made to develop the pattern recognition based classification technique by using the additive term (A0) image which involved the probability density function (PDF) for classification in one hand and in the other hand the classified image is used as a reference image for the masking of water and urban regions in A1 and A2 images to identify the annual and biannual change pixels. The accuracy assessment has been carried out onto the classified images by using the confusion matrix and the estimated overall accuracy retrieved as 76.78% from proposed classification technique and 70.49% from minimum distance classification. To identify the annual and biannual change pixels, a pixel wise comparison has been made between each and every pixel of agriculture and bare land classes, whoever be the dominant pixel among A1 and A2 pixels is assigned as the annual change pixel otherwise it is marked as the biannual change pixel. Chapter 6 presents the hot spot monitoring, which is also known as subsurface fires. Subsurface fires pose a serious threat to natural resources and cause significant societal and environmental degradation. Therefore, we have considered a hot spot monitoring as one of the part of land surface monitoring. Several researchers have used TM/optical data for identification of hot spot, but the main challenge with optical images is that they have some limitations due to cloud coverage, while SAR images are unaffected by clouds and haze and these images may provide valuable information over burned areas but the use of SAR data is very limited for this type of application. SAR images have not only been used for several ecological applications like vegetation monitoring and estimating biomass, but also used for Abstract ix burned area monitoring. It is also observed that NIR is the most suitable spectral band in optical images for hot spot monitoring because hot spot regions are generally characterized by a low reflectance in the NIR band in correspondence with the subsurface fire due to the loss of vegetation and the presence of ash and charcoal. The other problem associated with Optical images is its resolution; therefore, there is a need of research through which the Optical image can be efficiently used for hot spot and non-hot spot region’s classification. Therefore, this work explores the possibility of fusion of SAR (i.e., PALSAR data) and Optical (i.e., MODIS) data which may provide the complementary information and for this purpose, vegetation greenness and roughness information which are obtained from MODIS and PALSAR satellite images respectively are used for fusion by using the genetic algorithm (GA) based pixel level image fusion method, which is efficiently optimized the weights of source images for fusion (Mumtaz et al. 2008). The Jharia coal field (JCF) of India contains most of the subsurface mine fires in the Indian coal-fields and JCF has been chosen as the study area. A contextual thresholding algorithm is also developed, which consists rule based tests to detect the hot spots. It takes the fused or unfused image as the input image and by employing the rule based tests; it provides the output image as the hot spot detected image. False alarm rate (FAR) and hot spot detection accuracy (HDA) are taken as performance parameters. By using MODIS band 1 and band 2 images, GEMI and MSAVI indices have been obtained. The extracted Vegetation indices (Vis) may provide some useful information over the hot spot because GEMI index reflects the photosynthesis capability of healthy vegetation and it is insensitive to soil, ground and is capable of mapping vegetation activity and detecting hotspots in sparsely vegetated lands and MSAVI index minimizes the soil effects and supposed to give the vegetation information. By using the different PALSAR polarization images (i.e. HH, HV and VV), cross polarization ratio (CPR), HV/HH and HV/VV indices images have been obtained because CPR is used to distinguish between sparsely vegetated fields, forested areas and bare grounds, while HV/HH is used to maximize the variation of surface and volume scattering and have the ability to distinguish vegetation and bare grounds and HV/VV gives the low values for bare ground and gives a better idea of the bare field. The developed contextual thresholding algorithm has been applied on unfused MODIS and PALSAR images for the detection of hot spots. By calculating HDA and FAR, their performance has been evaluated. For GEMI, MSAVI, CPR, HV/HH and HV/VV, the detection accuracy (i.e. HDA) is retrieved as 60.67%, 37.89, 1.56%, 1.33% and 0.22%, respectively, while FAR is retrieved as 3.45%, 4.14%, 3.62%, 2.51% and 1.27%, Abstract x respectively, and it is observed that no single image is capable to produce high detection accuracy with low FAR. Therefore, to obtain the fusion coefficient, fusion is planned in a systematic manner to critically analyze the importance of fusion of indices obtained from PALSAR and MODIS data. One PALSAR data for the month of April, 2011 and four MODIS data for the same month of April, 2011 have been considered and assuming that conditions are approximately same during the whole month and this PALSAR data is also fused with each MODIS image separately. For the development purpose, only one MODIS data is taken with the fusion of PALSAR data and the developed model produces the HDA as 81.59% and FAR as 1.08%. The obtained coefficient is validated with the fused image of the MODIS and PALSAR data of another year and produces the HDA as 74.39% and FAR as 2.27%. Finally, Chapter 7 provides the summary of attained results and enlists the major contributions made in the thesis. The perspective of future study using current results is also discussed in this chapter.