ESTIMATION OF WATER SPREAD AREA OF A RESERVOIR USING SATELLITE DATA AND DEEP LEARNING.

Abstract

Water, being a basic human need and a vital natural resource available on the earth, has always been a prime concern for human civilizations. In most of the parts on earth, the exponentially increasing fresh water demand-supply gap has posed a serious water scarcity threat and made this issue of global importance. In the 21st century, most of the nations are struggling with this upcoming fresh water availability challenge, which is also reflected in their national policies. The water demand of many countries is continuously rising year by year and to cope up any future disastrous scarcity, their national policies are emphasizing on appropriate management of water resources. This requires a serious focus on scientific assessment and periodic analysis of surface water area with reservoir variables. The reservoir water resource management mainly depends on outflow of the reservoir and climatic variables, which needs to be properly correlated with estimation of surface water. Conventional methods are very time-consuming, need a periodic estimation of surface water and provide a tedious method to set relationships with meteorological parameters and reservoir variables. Advanced remote sensing techniques are handy and cost-effective tool for collecting a huge data of water resources. In the last two decades, coarse to medium resolution data have been widely used for water assessment at a local and global scale. The Landsat-8 (OLI) and recently launched Sentinel-2 have extended new opportunities for water resource assessment. However, there is a serious lack of a framework to manage the water resources of a reservoir. In the last few years, artificial intelligence (AI) techniques have become prevalent in reservoir operation planning and management approaches of water resource scheduling. The recent techniques based on Deep learning (DL) of AI are able to accurately predict and forecast different time-series data. Various deep learning algorithms as the Convolution Neural Network (CNN) and Recurrent Neural Network (RNN) are preferred to develop a relevant model for a relationship between massive water resource time-series data. However, for the derivation of appropriate modeling and conclusions based on different data sources, types of problems, space and time complexity, the selection of an appropriate DL algorithm has become a prominent challenge for research communities. i The research work presented in this thesis evaluates the accurate and fast assessing data of surface water area which is acquired using temporal Landsat-8 OLI and Sentinel 2A satellite images. Subsequently, the work shows the relationship between surface water area and the actual outflow of a reservoir with meteorological parameters. This study also acquired the relationship of actual reservoir outflow and the surface water area of monthly time series data of years 2014 2020. One of the essential parts of this research is the prediction of outflow variable of a reservoir using the analysis of input parameters of the created model using the Long Short Term Memory model of Deep Learning. This study has used temporal satellite data and compared the classification output of images generated by different water indices from satellite data of LANDSAT-8 OLI and Sentinel 2A MSI and calculated the average surface water area using digital image processing techniques. The results obtained in this study have matched with the original resource data. The efficacy of different existing water indices, such as Normalized Difference Water Index (NDWI), Water Ratio Index (WRI), Modified Normalized Difference Water Index (MNDWI), Automated Water Extraction Index (AWEI_Sh and AWEI_Nsh), Further Normalized Difference Vegetation Index (NDVI) using Landsat-8 and Sentinel-2A satellite images have been carried out and compared. A semi-automatic Double-Window Flexible Pace Search (DFPS) thresh-holding method has been used to determine an optimal threshold to separate water and non-water features from the generated ratio images using water indices. Since none of the data of surface water area is maintained by the concerned department, the data of actual water area has been obtained using Google Earth to compare with the observed area using satellite images and water indices. Using Google Earth as the reference, the water surface area was extracted and computed by applying the respective threshold values for all the indices. The area has been calculated for all months of the years 2014 to 2020. The trend variation in computed surface water area for a year is contributed by the factors like temperature, rainfall, inflow by a canal, water outflow, and dry summer season. Therefore, using an appropriate and optimum threshold, the proposed method could extract water bodies from Landsat TM, Landsat ETM+, Landsat OLI, and Sentinel-2A imagery with high accuracy. ii The variations in the surface water area and rainfall of all months of 2014-2020, as depicted in the graphs, show that the rainfall is low, and the surface water area remains average in the month of June. In the monsoon period from July to September, the rainfall increases initially and further starts decreasing. The variation in surface water area in the monsoon also follows the same pattern as the rainfall variation. In the post-monsoon and winter period from October to February, rainfall decreases to a minimum. The surface water area reaches a maximum in February due to low water consumption. After that, in consecutive three months i.e. March, April, and May, rainfall increases slightly, whereas surface water area reduces because water consumption becomes higher in the dry and summer season. The graph, depicting varying results of the surface water area and temperature of all months of 2014-2020 shows that the temperature moves on higher side, and the surface water area remains at an average level in June. In the monsoon period from July to September, the temperature decreases, and the surface water area moves down after initial increment. The temperature continually keeps on decreasing in the post-monsoon period and touches the minimum temperature level during the winter period from October to February. The surface water area also touches its maximum level in the month of February due to lower water consumption. After that, in the subsequent months of March, April, and May, the temperature increases and surface water area gets reduced because of higher water demand and maximised evaporation because of dry and summer season. Historical sources such as rainfall, temperature, and outflow are highly essential for the appropriate prediction of reservoir outflow time series analysis. Based on different climatic variables like rainfall, rainfall intensity, runoff rate, temperature and surface water area, a novel model is developed using Long Short-Term Memory of Deep Learning to predict the reservoir's outflow. The discussion on parameter setting effect on model performance has been summarized and prominent factors affecting reservoir operation are analysed. The multipurpose reservoir prediction model using the climatic parameters, analyze the monthly and yearly water demand of the Kaylana lake reservoir and confirms the reciprocal relationship between predicted outflow of the model and the predicted surface water area in all seasons of the year. iii iv

Our study in this thesis proposes a novel water resource management framework which has been developed by combining techniques to extract surface water area using satellite data and reservoir water outflow prediction used by modern deep learning techniques. The proposed resource management framework reduces the risk of flooding downstream as well ensures sufficient water storage for monthly utilization. The model has been trained and tested using the data obtained and performed well with better accuracy and precision. This can predict all parameters entirely accurately and efficiently and the output result that combined iterations and neurons of a hidden layer mainly impact manipulating the model precision, computation speed is primarily affected by the batch size.