Finding Out Suitable Index for Wetland Mapping in Barind Plain of India and Predicting Dynamics of Its Area and Depth

10 Remote Sensing and GIS play an important role in mapping and monitoring natural resources 11 and their management. The present study attempts to delineate wetland in the lower Tangon 12 river basin in the Barind flood plain region using suitable water body extraction indices. The 13 main objectives of this present study are mapping and monitoring the flood plains wetlands 14 along with the future status of wetland areas of 2028 and 2038 using the advanced Artificial 15 Neural Network-based Cellular Automata (ANN-CA) model. Apart from wetland area 16 prediction, wetland depth simulation and prediction are also carried out using statistical 17 (Adaptive Exponential Smoothing) as well as advanced machine learning algorithms such as 18 Bagging, Random subspace, Random forest, Support vector machine, etc. for the year 2028. 19 The result shows a remarkable change in the overall wetland area in the upcoming two 20 decades. The small wetland patches away from the master stream are expected to dry out 21 during the forecast period, where the major wetland patches nearer to the master stream with 22 greater depth are rather sustainable but their depth of water may be reduced in the next 23 decades. All models show satisfactory performance for wetland depth mapping, but the 24 Random subspace model was identified as the best-suited depth predicting method and 25 machine learning models explored better results that adaptive exponential smoothing. This 26 recent study will definitely be very helpful for the policymakers for managing wetland 27 landscape as well as the natural environment.

and Selmi, 2018). Depth is a hydro-ecological aspect whose variation is linked to the species 117 richness, diversity, comfortability, particularly to the fish species (Hamza and Selmi, 2018; 118 Ouma, 2020). Wetland trend analysis (depth and area) is very essential for sustainable 119 management planning. Lack of such information to the administration is also one of the 120 primary causes of rapid wetland conversion around the world, particularly in floodplain 121 regions (Talukdar and . Therefore, predicting wetland area and depth of water is 122 another major focus of this work. Regarding the originality issue, it is to be mentioned that     175 Making a distinct difference between water and non-water pixels is very challenging using a 176 common threshold due to spectral nearness of some land use land cover components (Ngoc et 177 al., 2019). In recent times, numerous remote sensing spectral indices are available for 178 delineation of water bodies as well as wetland mapping (Paul and Pal, 2020;Saha et al., 2021). Spectral indices like NDWI (Mcfeeters, 1996) 261 The Bag algorithm is a very popular ML algorithm, which was widely used to develop Only the learners not trained on X are involved in this stage, the OBB estimated using the 280 following equation (Eq.11)

Wetland mapping using water indices and validation
where, X= vector, x=variables, y= output spaces, N= data sample, T=number of base learners,

284
The OOB error of bagging can be estimated using the following equation (Eq. 12).
where, x is the variables 287 288 The RF classification algorithm is the modification of the CART (classification and where, ik =conditioning variables, 1, 2,…n= input vectors x.

313
The error of RF algorithms can be estimated using the following equation (Eq. 14). where, x and y= conditioning variables, mg= margin function.

316
The margin function can be estimated as follows (Eq. 15). 319 The RS is a well-known random sampling-based ensemble machine learning algorithm 320 developed by Ho (1998 (Table. S1). These values were higher than other applied indices like 417 NDWI and MNDWI and therefore, RmNDWI could be treated as accepted. The RmNDWI 418 maps were further used for wetlands area simulation and prediction.  The simulated wetland map derived from ANN-CA techniques, reported that wetland area  (Table. 2). The detailed areal extent of simulated and predicted wetlands is 432 mentioned in Table. 2.

433
The results of kappa statistics, overall accuracy, AUC of ROC were mentioned in Table 5.

Predicted wetland depth and accuracy assessment
Due to the absence of spatiotemporal datasets, time series wetland depth mapping is very 448 difficult over a wide geographical area. The recent work calibrated NDWI map for wetland 449 depth mapping. The calibration process was carried out using 33 field-specific wetlands but their depth of water may be reduced.

467
To execute the best representative model from the applied five models, it is necessary to 468 measure the accuracy of the simulated wetland water depth from observed depth.

488
Wetland demarcation is a very important task for the implementation of proper management  It is also observed that the large core of wetlands is relatively safe, secure, and having a result (Paul and Pal, 2020). No studies so far attempted to do this at pixel level. Therefore, 542 the present work could be treated as a landmark in this approach since the study applied 543 multiple ML models for predicting water depth precisely at pixel scale. Here lies the novelty 544 of this work.

545
Whatever may the absolute difference in prediction result using different ML models, the 546 overall trend shows declining water depth which is of immense ecological concern. It may