Efficient Segmentation Algorithm for Estimation of Revised Reservoir Capacities in Google Earth Engine

The periodic assessment of the revised capacity is essential and conventional techniques through bathymetric surveys and inflow-outflow techniques are time and resource intensive. The application of optical remote sensing data is practiced for a long time to compute the revised capacity of the reservoir but has the limitation of selecting threshold values and inefficacy of the method during the cloudy season. This research proposed the application of the Otsu segmentation technique in Google Earth Engine (GEE) to determine revised water spread using microwave remote sensing data. The Otsu technique is efficient in classifying the image into two distinct classes using within and between-class variances. An application has been developed using Sentinel data in GEE, which has the capacity of cloud-based computing and applied to the four reservoirs of Mahanadi Reservoir Project (MRP) namely Ravishankar Sagar, Murumsilli, Dudhawa, and Sondur reservoir in the Chhattisgarh state of India. The original capacities of these reservoirs vary from 910.51 MCM of Ravishankar (RS) Sagar to 165.89 MCM of Murumsilli reservoir. The revised water spreads computed in the GEE were further used to compute revised volume and in turn the revised cumulative capacities at different levels. The analysis suggested that 17.65% of dead storage (118.26 MCM from 143.00 MCM) and 9.25% of gross storage (85.26 MCM from 910.52 MCM) of RS Sagar has been lost in 42 years (1978 to 2020). Similarly, 10.33% gross capacity of Murumsilli, 13.12% of Sondur, and 12.37% of Dudhawa reservoirs have been lost due to the deposit of sediments from the catchments, and results were found comparable with bathymetric survey results of three reservoirs. The proposed application has been developed in GEE so it can be used in any part of the world in cloudy weather with no human interference.


Introduction
Worldwide, 0.5-1% of reservoir capacities are lost annually, and estimates suggest that up to 25% of the world's total reservoir capacity could be lost in the next 25 years (Parera et al. 2023). In India, about 70% of reservoirs have an actual sediment rate more than twice the designed rate (Andredaki et al. 2015;Vorosmarty et al. 2003;Jaiswal et al. 2016;Rakhmatullaev et al. 2013;WCD 2000). Revised capacity can be estimated using conventional methods like the bathymetric survey or digital image processing of multi-date satellite data in optical bands (Singh et al. 2021). Digital water body classification has been used by many investigators through optical bands in the past (Goel et al. 2002;Rodrigues et al. 2011;Merina et al. 2016;Fallah et al. 2016;Tsolakidis and Vafiadis 2019;etc.). In optical bands, the normalized difference water index (Gao 1996) is frequently utilized to identify water pixels (Dadoria et al. 2017;Pandey et al. 2016;Avinash and Chandrmouli 2018). However, this method is prone to several limitations such as cloud interference, chlorophyll concentration, and total suspended solids (Pekel et al. 2016). Microwave data, on the other hand, are not affected by these issues and can be utilized for identifying water pixels via segmentation techniques (Goumehei et al. 2019). These data are also commonly used for identifying water spread and soil moisture due to their ability to penetrate clouds and resistance to atmospheric effects (Cheng et al. 2021). Thresholding techniques are frequently employed in microwave datasets to distinguish water under the assumption of low backscattering compared to other features (Li and Wang 2015). With advances in technology and computing techniques, these data are commonly utilized for various purposes. Condeça et al. (2022) utilized the Google Earth Engine to calculate water spreads and stored water volumes automatically using NDWI from Landsat 4 and 5 ETM and Landsat 8 Operational Land Imager images.
Image segmentation is a popular technique used in a variety of applications, including text detection and medical image processing (Sari et al. 2014;Esphtein et al. 2010;Shekhar et al. 2015;Francis and Sreenath 2020;Lin et al. 2017;Bindu 2009;Feng et al. 2017;Nyo et al. 2022;Nina et al. 2011;Mustafa et al. 2017;Karthika and James 2015;Pare et al. 2021;Srinivas et al. 2019;Biswas et al. 2015;Campos et al. 2012;Tulbure and Broich 2013). Object detection is a common use of image segmentation, and many techniques have been proposed, such as edge, region, or threshold-based detection (Pal and Pal 1993;Suresh and Srinivasa Rao 2019). Threshold segmentation is one of the simplest and most popular segmentation techniques, which splits the picture's grey scale information depending on multiple targets' varying grey values (Haralick and Shapiro 1985). The Otsu method is the most widely used threshold segmentation algorithm, where a globally optimal threshold is chosen by maximizing the variance between foreground and background (Otsu 1979;Liu 2011;Zhu et al. 2010;Huang et al. 2021). Different segmentation techniques can be categorized based on various criteria such as manual, semiautomatic, and automatic; pixel-based and region-based; classical, statistical, fuzzy, and neural network techniques; manual delineation, low-level segmentation, and model-based segmentation.
Remote sensing data processing requires specialized software and high-end hardware (Ranade et al. 2015), but the Google Earth Engine (GEE) is becoming a popular option due to its ability to handle large geospatial datasets on the cloud (Gorelick et al. 2017;Jain et al. 2021). GEE's capabilities include parallel processing, a fast computing platform, and powerful graphics processing units (Coleman et al. 2020;Martin et al. 2019;Gorelick et al. 2017). Additionally, GEE makes it easy to share results with policymakers and researchers and can be used for decision support (Kumar and Mutanga 2018). GEE has been widely applied in land cover mapping, floods, disaster management, vegetation, etc. (Wahap and Shafri 2020;Goldblatt et al. 2018;Bui et al. 2020;Rayner et al. 2021;Kumar et al. 2022;Agnihotri et al. 2019;Clemente et al. 2020;Amitrano et al. 2018;Ghaffarian et al. 2020;Yang and Wang 2021;Liu et al. 2021;Juntakut et al. 2021). Matunga and Kumar (2019) highlighted GEE's potential due to its user-friendly frontend and ease of interactive data management on the cloud. Nghia et al. (2022) used GEE to analyze floods downstream of the Mekong River delta and observed decreasing floods in the area. The present study first time applied the Otsu algorithm in GEE with the objectives of application of SAR data for the estimation of revised capacities of the reservoir. The SAR data is least affected by clouds making possibilities to determine revised water spreads during the rainy season where the application of optical RS is not possible. The paper presented an efficient algorithm (Otsu technique) for the assessment of revised capacity of any reservoirs of the world without using any specialized software. The comparison of Otsu based application in GEE with bathymetric survey showed the potential of proposed technique for application in other reservoirs.

Study Area and Data Used
Mahanadi reservoir project situated in the Chhattisgarh state of India consists of four reservoirs namely Ravishankar Sagar, Murumsilli, Dudhawa, and Sondur reservoirs. The location map and salient features of these reservoirs are depicted in Fig. 1

Methodology
The estimation of revised capacity with the help of digital image analysis is an indirect method that works on the principle that the sediment deposited in the reservoir reduces the water spread from its original spread at the same level. The revised water spread areas computed through image analysis are used to compute revised estimates of capacities and differences between original and revised capacities can be considered as loss in volume due to sediment. The flow chart for the computation of revised capacities through GEE is presented in Fig. 2.

SAR Microwave Data
The microwave data used in the study were obtained from the constellation of Sentinel satellites 1 A and 1B launched on April 10, 2014, and April 22, 2016, respectively by European Space Agency (ESA) having a combined revisit of six days. The Sentinel satellite's C-band synthetic aperture radar (5.4 GHz) has dual-polarization capabilities, including HH & VV, HV, and VH, allowing for image acquisition regardless of weather and at any time of day and night for various applications such as surface water monitoring (Pham Duc et al. 2017). The same transmitted and received (HH & VV) polarization is found most useful in the case of water delineation (Bourgeau-Chavez et al. 2001). SAR has four acquisition modes, including an extra wide swath of 400 km for oceans and coastal monitoring, strip mode of 80 km on special order, wave mode of 20 km for the ocean, and interferometric wide (IW) swath of 250 km for land features. Preprocessing with the Sentinel 1 toolbox involves using the terrain height supplied in the product general annotation to correct the ellipsoid projection of the GRD products, and applying a median filter as a speckle filter to reduce granular noise caused by interference at the transmitter and receiver antenna.

Application of Otsu Algorithm in GEE
The date and area filters were applied to select all the scenes within the selected period to get the data set for the period of interest in GEE through Java scripting. In this study, the "VV" mode of polarization and the "IW" instrument mode were used to select SAR data. The image processing of remote sensing data for the identification of water pixels from the rest of the image is essentially a binarization where thresholding is an efficient technique. The Otsu method is one of the most popular and successful methods due to its simplicity and effectiveness for image thresholding (Liu and Yu 2009;Cao et al. 2021) proposed by Nobuyuki Otsu (Otsu 1979) used for automatic image thresholding for image processing and computer vision (Sezgin and Sankur 2004).
The Otsu method is a powerful tool for segmenting images by distinguishing between foreground and background pixels based on their brightness. This method calculates the interclass variance among different classes, with the key to perfect segmentation being the difference in inter-class variation among different clusters (Bangare et al. 2015). The Otsu method converts the image into a binary image by selecting a threshold value between 0 and 255 on a gray-level image. This threshold is found by either minimizing the variance of with-in-class density or maximizing variance between classes. The Otsu method requires the histogram of the image as input for the binarization process. An initial threshold (t) is selected which divides the whole histogram into two clusters. The within-class variance can be computed using the following equation: Where, w 1 (t) and w 2 (t) are the probabilities of two clusters divided by the threshold value (t), 2 1 (t) and 2 2 (t) are the variance of these two clusters. The variance of these clusters was determined using the following equation: Where X i is the pixel value of the i th pixel, is the mean value of pixels and n is the total number of pixels. Using the above equation, variance is computed for both clusters and interclass variance for different threshold values and the threshold value gives the minimum within-class variance and is considered the Otsu threshold for the division of the image into two classes (black and white). On the other hand, Otsu suggested the other option of maximizing the between-class variance ( 2 b (t)) using the following equation: Where, 1 (t) and 2 (t) are the mean of the first and second clusters. The developed GEE program resulted from a threshold value having a maximum within-class variance, and a chart showing water pixels, and water spread on different dates of pass of Sentinel 1 over the area of interest.

Computation of Revised Capacity and Sediment
Revised water spread areas at nearly equal intervals obtained from the Otsu segmentation technique were used to determine the revised bed level by fitting an appropriate curve. It is evident that the storage up to the revised bed level is filled with sediment and above it, the cone formula (Ninija Merina et al. 2016;Dadoria et al. 2017;etc.) was used to determine revised water volumes between two consecutive levels.
where, V R is the revised volume between two consecutive levels h 1 and h 2 , h = h 2 -h 1 is the difference between these two levels, and A 1R and A 2R are the revised water spread on these levels. The cumulative revised capacities were computed by adding revised volume between these levels and compared with the original cumulative capacities to ascertain the loss of capacities at different levels of the reservoir. The revised capacities obtained from Otsu-based image classification were compared with bathymetric survey results conducted in the year 2021 for three reservoirs in MRP.

Analysis of Results
The present study suggested Otsu, an efficient technique for the segmentation of SAR images that can be automated through GEE and applied to Ravishankar Sagar, Murumsilli, Sondur, and Dudhawa reservoirs.

Revised Capacity of Murumsilli Reservoir
Murumsilli reservoir started functioning in the year 1922 having a gross storage capacity of 164.54 MCM, water spread of 25 sq km, and live storage capacity of 161.91 MCM. The ten different scenes were during the period of 2020-22 (mean 2021). The Otsu segmentation technique yielded a threshold value of -15.03 to segregate water pixels from the rest of the image. The revised water spread at different levels was used to fit a non-linear curve to determine the new zero elevation for the reservoir and found that sediment had filled all the space up to 359.95 m in Murumsilli reservoir. The revised water spreads on few dates and revised capacities are presented in Fig. 4. From the analysis, it has been observed that nearly 94.1% of the dead and 10.33% of gross storage of the Murumsilli reservoir was lost in 99 years with an average rate of 0.17 MCM/year.

Revised Capacity of Sondur Reservoir
To assess the revised capacities of the Sondur reservoir ten different dates were selected to cover the whole range of live storage from 455.19 to 471.07 m. The threshold value to distinguish water pixels from the rest of the image was found as -14.93 and used to determine revised water spread areas on ten different dates to cover the whole live storage zone. The revised capacities based on revised water spreads are depicted in Fig. 5. From the analysis, it can be concluded that 13.    (Fig. 7).

Discussion
In the study, the Otsu-based segmentation technique is proposed and applied to four reservoirs in India, and the results were compared with the bathymetric survey. These reservoirs were classified into different classes as insignificant (up to 0.1%), significant (0.1 to 0.5%), or serious (more than 0.5%) suggested in IS 12,182 by the Bureau of Indian Standard (BIS). Also, the same standard devised a way to select a reservoir catchment for soil conservation by a concept of prescribed limit which can be expressed by the following equation: The revised capacity and the rate of sedimentation in different reservoirs computed with the proposed technology and its comparison with the bathymetric survey were presented in Table 2. From the analysis, all reservoirs have an annual loss in the range of 0.1 to 0.23%, which can be categorized in the significant class. Also based on the 100-year life period, the Sondur and Dudhawa reservoirs need soil conservation measures to arrest excessive sediment in the reservoirs.

Limitations and Benefits of the Otsu Technique
The Otsu method has difficulties when an image does not have a bimodal histogram, noise in the images, and availability of more than two classes (Kumar and Tiwari 2019). The Otsu technique proposed with SAR data in this research is useful to work in all weather and GEE API-based script can be applied to any reservoir in the world (Zhou et al. 2020). Therefore, the region of interest should be near the water body and not consist of a large area outside the water spread. SAR data are most suitable for determination in water spread through Otsu segmentation in the rainy season and cloudy days where optical remote sensing does not work well.

Conclusions
Reservoir sediment is an important aspect of planning water resources and management of releases for different purposes. The present paper introduced and presented a noble technique of Otsu segmentation for the classification of water spreads using SAR data. Automatic classification reduces human interference and leads to the automation of processes. The Otsu-based assessment showed that the Ravishanker reservoir lost nearly 84.26 MCM of gross capacity in 41 years (1979 to 2019) whereas the hydrographic survey conducted in the year 2021 found this loss as 84.62 MCM. Similarly, other reservoirs also confirmed a close resemblance of loss of reservoir capacity at different levels and cumulative capacities for these reservoirs with the hydrographic results. The advantage of the proposed methodology lies in its application in the google earth engine where no need for storing data, no software requirement, application of a robust Otsu technique for segmentation, and application for any reservoir in any part of the world without any human interference.