Reconstruction of the 2003 Daya River Flood, Using Multi-resolution and Multi-temporal Satellite Imagery

Floods are a common disaster in many parts of the world. It is considered to be the most common, costly and deadly of all natural hazards. Flooding is not just confined to certain region of the world but is a globally pervasive hazard. India experiences one of the highest incidences of Flood, and the area subjected to it is estimated to be one-eight of the geographical area (0.410 M km2) and have been occurring almost regularly each and every year.

Among the several states prone to flooding, Orissa state is one of them, which is mainly due to the existence of Mahanadi delta system. Studies have been carried out related to understand dynamic processes and the flooding problem in Mahanadi river system in the past. In this study, Daya river system, which is a part of Mahanadi Delta and which is one of the most flood prone areas in Orissa is selected. Heavy monsoon rainfall and congestion of drainage pattern in the river system are considered to be the main cause of flooding. Moreover, one of the main causes of flooding in this area is due to low and high tide (tidal effect), from the Chilka Lake.

With advances in the field of Remote sensing and GIS technology makes possible to monitor and study various natural and environmental hazards in a large spatial extent using a long time-series data, which in turn provide a means for quick response, recovery and mitigation activities during and after any natural calamities.

In this study, an attempt has been made to reconstruct the Daya Flooding event, 2003, which help in understanding the dynamic of Daya river system. Since flood is a dynamic phenomenon, the period of submergence vary greatly spatially. To capture and analyse the 2003-flooding phenomenon, which took place during 28th August 2003 to 20th September 2003, multi-sensor (optical and microwave data) and multi-temporal satellite data before and during the flooding period were acquired. Two broad approaches are adopted and implemented which integrates the information extracted using remote sensing techniques, for inundation mapping and GIS approach to determine the maximum extent of flooding and to overcome the constraint of temporal resolution in the application of satellite images in flood inundation mapping, with a strong historic and geomorphic data that in turn helps in understanding the dynamic and the flooding pattern prevalent in this region.

From historical data analysis, it is concluded that regarding the necessity of having proper and accurate historical information’s, which will provide a complete picture in understanding the real phenomenon of the event under consideration. In this study, past 40 years flood event maximum discharge data at various locations are analyzed and it is concluded that Daya River flooding event of 2003, is the third highest flood in magnitude with a return period of 14 years. It’s also found that a large magnitude flooding has a low exceedence probability when compared with event having less discharge. Moreover, the trend in flooding pattern is also studied and observed that whether there is any deviation in the trend observed.

In Remote sensing approach, various techniques of extraction of flooded extent have been performed on multi-temporal RADARSAT imagery of 4th, 11th, 13th and 20th September 2003 and various optical datasets i.e. IRS-Pan, LISS-III and Aster of different dates which help to analyse the inundation pattern extracted from various datasets i.e. visual interpretation and automated digital techniques. A comparative analysis of inundation extent extracted is made. Various digital classification techniques such as supervised, unsupervised, thresholding/density slicing, textural analysis based classification and Principal component analysis based classification were explored which help to find a quick, accurate method for flood mapping which can be made operational in future. Analysis from optical dataset by visual interpretation gives quite a reliable and stable inundation extent. The extent of inundation from visual interpretation of dB RADARSAT image (50 m resolution) is taken as a reference extent for determining the variation in spatial extent; due to its high contrast which inturn help in clear identification of land-water boundary extent. Percentage inundation in the reference images of 4th, 11th, 13th and 20th September 2003 are 42.78 %, 50.31%, 38.86% and 36.03%, where maximum and minimum flood extent is observed on 11th September and 20th September 2003 respectively.

Considering the maximum flood extent of 11th September, the variation in spatial extent observed by different digital techniques is presented here.
By thresholding, for the dB, DN images are about 7.71% (in case of 50m,dB); 2.89% (50 m, DN) and -11.01% (100 m, DN). This indicates low variation in extent for 50 m dB and DN as compared with that of 100 m DN. Similarly, variation in extent by unsupervised technique for the three dataset on 11th September gives about 7.29 %( in 50m,dB); 2.89% (in 50 m,DN) and 3.45% (in 100 m,DN). Inundation extent given by this technique is more than that of extent by visual interpretation for all the three dataset. Less deviation of the extent is observed from 100 m, DN as compared with the 50 m images. In case of supervised classification, variation in extent is observed when two different classifiers are applied i.e. maximum likelihood classifier (MLC) and minimum distance (MD) to mean classifier. The inundated map from the supervised classification varies to a large extent. For 50 m, dB variation of 0.48% and 18.45 % is observed from the resulted map by Maximum likelihood and Minimum distance to mean classifier. In case of the other two dataset, percentage deviation in extent observed is 50% and 26.36% (in case of 50 m, DN); 0.577% and 28.47% (in case of 100 m, DN). This clearly indicates the accuracy of extent generated by MLC as compared with MD, considering the deviation observed in 50 m, dB and 100 m, DN images. In Principal component analysis, the extent given by principal component 1(dB image) is deviated by 5.35% and that of 50 m, dB and DN Principal component 1 images by 15.52% and 35.26% respectively. It’s also highlighted in this study regarding the importance of colour composition for accurate feature identification and extraction. A colour composite of PC1:PC3:PC1; PC3, PC3, PC1 gives a distinct land-water boundary which help to visualise and extract the inundation extent accurately. Simple thresholding technique comes out to be the best reliable and suitable automatic techniques for quick inundation mapping accurately. Applicability of textural based classification for automatic flood inundation extraction is also attempted here. It’s observed that textural measures such as Homogeneity, Contrast and Second Moment come out to be the best suitable measures for extracting flood extent which is comparable with that of reference visually interpreted results of 50m, dB dataset. In 50 m db image, a deviation of 0.498%,-3.32% and 0.102 % could be observed when comparing the homogeneity, contrast and second moment derived inundation extent with that of reference extent of September 11th 2003. Similarly, in case of 50 m DN deviation observed was 0.626 %, -3.24% and 0.201%; and in 100 m DN the deviation is -34.906%, -49.49% and -26.56% respectively. It could be inferred form this study that dataset of 50 m dB and DN gives more accurate extent as compared with that of 100 m DN.

The percentage inundation obtained is about 49.57%, 48.68% and 49.75% from IRS-Pan (5.8 m), Pan-sharpened LISS-III (5.8 m), IRS-1C (LISS-III, 23.5 m) respectively for same date i.e.08-9-2003, which shows less variation of extent within the optical dataset. Hence, variation in extent using multi-sensor dataset is highlighted in this study.

In GIS based approach, attempt was made to generate accurate DEM from Aster Epi-polar images, field contour map to developed Cost-distance grid i.e. least accumulation cost- distance matrix, which is then used to integrate with the inundation map derived from RADARSAT to get the inundation extent map that correspond with the peak flood discharge. The existing DEMs i.e. public domain Aster DEM, DEM generated using filed contour map could not generate cost-distance matrix in Arc GIS environment. Hence maximum inundated extent could not be generated in this study. This reflects the necessity of an accurate DEM for flood related studies.