STUDY OF DIP TECHNIQUES TO DETECT HOTSPOTS WITH LOW RESOLUTION SATELLITE DATA

Abstract

Subsurface coal mine fires (hotspots) have occurred in nearly all parts of the world like India, US, Indonesia, South Africa, Australia, China, Germany and many other countries. However, the nature and magnitude of the problem differs from country to country. India, where large scale mining began more than a century ago, accounts for the world's greatest concentration of subsurface fires which exist from last 100 years [134]. Jharia Coal Field (JCF) in Jharkhand (India) contains nearly half of subsurface mine fires in Indian coalfields [1]. Due to severe effects caused by these hotspots from economical and environmental aspects, this problem is being taken veryseriously by scientists throughout the world but still success needs to be achieved in developing a solution for the same which is cost effective, time and scene independent and offers significant performance. Hence, major effort needs to be done in the direction of detecting and monitoring these hotspots. Since ground based monitoring of hotspots is very expensive and cumbersome, research is required to be focused on exploiting the potential of the satellite image analysis for this purpose. Satellite images are available with different spatial resolutions. Most of the existing hotspot detection algorithms use high spatial resolution satellite images. These high spatial resolution satellite data are very expensive (e.g., Landsat TM, ASTER, IKONOS etc.) and have low revisiting frequency. These limitations reflect the urgent need of exploring the possibility of using less expensive or freely available satellite image but with low resolution, for hotspot detection problem. The aim of the present research work is to provide economic and promising solution to the problem of hotspot detection to the end user using freely available low resolution satellite data. The satellite images used in this research work are acquired from the Advanced Very High Resolution Radiometer (AVHRR) on board the National Oceanic Atmospheric Administration (NOAA) satellite series. NOAA/AVHRR data have spatial resolution of 1 km at nadir and these data are freely available to the end users. In this thesis, a broad methodology is proposed for the detection of hotspots using freely available NOAA/AVHRR images of Jharia coalfield region of India bounded by Latitude (N) 22°00'00" to 24°00'0" and Longitude (E) 85°00'00" to 87°00'00". In the present research work, a contextual model is developed for hotspot detection. Further, the performance of this model is refined by employing machine intelligent techniques. The proposed research then focuses on developing separate algorithms for hotspot detection by utilizing the advantages offered by soft-computing techniques. Later, after successful detection, the attention is given toward developing a technique for monitoring the spectral characteristics of these hotspots over a specified time period. Finally, this research ends with proposing a quite new approach, based on subpixel analysis, for estimating the area occupied by these hotspots. The thesis is divided in seven chapters. First chapter presents the introduction of the thesis which includes the motivation, major research gaps, anddetails about the study area and satellite dataused. A brief review of the existing techniques and their limitations for the problem of hotspot detection is presented in Chapter 2. Chapter 3 of the thesis explores the applicability of adaptive image processing techniques and presents a sensor independent classical approach to detect hotspots in NOAA/AVHRR images. After doing comprehensive analysis of several existing algorithms e.g., Otsu's method [152], entropy based thresholding [161] as well as contextual algorithm proposed by Flasse and Ceccato [35], it is found that these algorithms are not fully adaptive, do not make use of important multi-channel information and therefore, are not appropriate for hotspot detection. In this chapter, a contextual algorithm is presented which is generically adaptive in nature and is able to detect hotspots reliably and successfully in different observational conditions. After observing the radiometric response of known subsurface fire pixels for 10 years of 70 NOAA/AVHRR datasets of the region of interest, a set of test criteria, based only on the statistics of image pixels, is proposed to detect potential hotspots. Further, an automated

model is developed which, once developed for a particular region of interest, can successfully detect hotspots in future years of images of the same region. The performance of the proposed model is compared with other existing methods and the results indicate that the proposed contextual algorithm outperforms existing ones and provides successful detections of most of the hotspots [44]. The machine intelligent techniques i.e., Support Vector Machine (SVM) [155], Neural Networks (NNs) [39, 147], and Rough Set Theory [110] has been explored in Chapter 4 for refinement of contextual model developed in previous chapter. Since machine intelligent techniques provide the flexibility of enhancing the learned knowledge when new sets of input are presented to it, the performance of the model developed in previous chapter can also be improved immensely if this model can be made intelligent. In this way, the model is able to learn from the hotspot information of different regions of interest and can use that learned information for predicting unknown hotspots from the same region. The output of contextual model developed in previous chapter serves as training data to train these machine intelligent models. SVM works on structural risk minimization principle that minimizes an upper bound on the expectedrisk and therefore has greater ability to generalize [155]. The present chapter also implements well known Neural Networks classification in order to compare the classification performance with SVM. Rough set classification, which is also implemented, eliminates redundant information in training data and then generates rules to classify the test data. The proposed algorithm exploits aforementioned advantages of these machine intelligent techniques and employs them for refinement of proposed hotspot detection model. Obtained results show that the application of SVM and Rough sets immensely improves the performance of the hotspot detection model [42, 45]. Nowadays soft computing tools i.e. Fuzzy logic [165], Principal Component Analysis [2] are very emerging techniques as they utilize the fused information of multi-spectral image for different applications. Therefore, in Chapter 5, we have critically analyzed these tools for hotspot detection by incorporating multi-channel information obtained from the NOAA/AVHRR data. Since there is no perfect boundary in between the two classes i.e., hotspots and non-hotspots, the hotspot detection problem creates geometrical fuzziness. Proposed framework resolves this fuzziness by employing fuzzy logic vu approach for classifying AVHRR image in hotspots and non-hotspots. An efficient hotspot detection tool should extract minimum number of features from the image for hotspot detection while producing highest detection accuracy. Therefore proposed tool inherits the advantages of principal component analysis (PCA) for features selection. Using PCA, the most prominent components which contain almost total variance information of all input features are extracted and used for further analysis. Thus, the proposed tool works on minimal set of features and presents efficient approach for hotspot detection. The obtained results indicate that the fuzzy based approach is quite successful in handling uncertainty in the data and provides accurate information aboutthe spatial allocation of hotspots, whereas, PCA based approach works effectively with fewer numberof input features and presents a very simple and effective solution to the problem of hotspot detection [43, 46, 47]. Chapter 6 of the thesis provides the detailed time series analysis of 10 years of NOAA/AVHRR images (120 images) aiming to develop a technique which is able to monitor the spatial and spectral characteristics of hotspots overthe specified time period and also classifies the region of interest in different land features. To achieve the objective, in this chapter, harmonic analysis (based onFourier transformation) of 10-year time series (1995-2005) NOAA/AVHRR yearly composite images is performed to develop an innovative technique for hotspot detection and land-features classification based on temporal changes in the NDVI and various AVHRR band values. After determining the spatial allocation, the next step is to estimate the actual area occupied by the hotspots in the region of interest. Since, the size ofthe hotspots is less than the spatial resolution of the NOAA/AVHRR images, in order to determine the actual area, subpixel analysis ofAVHRR images needs to be performed. The chapter proposes a methodology to determine the hotspots area using constrained energy minimization based subpixel analysis technique. Using subpixel analysis, it is possible to break the resolution barrier and identify the individual entities contributing to the pixel's spectrum, and their respective proportions. The proposed subpixel analysis based approach successfully gives hotspot fraction images as output and provides the estimate of actual area occupied by these detected hotspots [48]. vin The contributions made in the thesis are summarized and scope of future work is outlined in final Chapter 7. The results were validated with the hotspots information provided by BCCL (Bharat Coking Coal Limited, India). These reports are based on ground observations (using GPS) and include the name of the place and location (Latitude, Longitude) of the hotspots in Jharia region [119, 123].