"A Framework for Long-range Ensemble Streamflow Forecasting in the Brahmaputra Basin"
This project aims to prepare reliable long-range streamflow forecasts at the Bahadurabad outlet of the Brahmaputra basin, which will provide sufficient duration for flood preparedness in the flood-prone regions of north-western Bangladesh situated downstream in the great Brahmaputra basin. The core element of the framework is a lumped hydrologic model, calibrated using CPC precipitation data as an alternative to station observed rainfall. The model provides streamflow forecast with up to 30-days lead time, and accounts for uncertainty of forecasts leveraging the ECMWF ensemble and extended meteorologic forecasts. Residual error from the model is then reduced through a Machine Learning based error correction model. The model has been developed by RIMES, with Mr. Nazmul Ahasan being the lead contributor. My involvements in the project include developing the error correction model using machine learning increase the accuracy of the model, integrating Muskingum-Cunge routing into the model, and providing assistance with the writeup. The model is being experimentally monitored by the Bangladesh Water Development Board, and I am charged with the task of operating the model on a regular basis.
Developing a Sustainable Flood Forecast Model for the Teesta River in Bangladesh
The goal of this project is to develop a flood forecast model for the flash flood-prone Teesta River in Bangladesh. The model has been calibrated using CPC rainfall data for 2015–2021 using a Hydrologic model. This model is expected to be used to predict the timing and magnitude of floods in the Teesta River, facilitating decision-making about flood preparedness and mitigation in the region. My involvement in the project includes data processing, model development, and the preparation of writeups.
"Flood Forecasting in the Kushiyara River Using Supervised Machine Learning"
Abstract: The northeastern region of Bangladesh is highly susceptible to flooding almost every year mainly due to its geographical location. Heavy flooding in this region causes immense damage to lives, property and mostly, agricultural crops. This region suffers from both flash flood and monsoon flood. However, the lack of rainfall information from the upstream catchment areas outside of Bangladesh makes flood forecasting in this region difficult. In addition, inaccuracies in rainfall forecasts and a lack of high-resolution bathymetry and topography data limit the applicability of hydrologic and hydrodynamic models to forecast floods in this region. Currently, the Flood Forecasting and Warning Centre (FFWC) of Bangladesh Water Development Board (BWDB) is producing short-range deterministic flood forecasts with a lead time of up to three days. However, medium-range (3 to 7 days) forecasts are critical for minimizing flood-related losses because they allow greater time for decision-making and planning. With a view to addressing these issues, four machine learning (ML) algorithms have been chosen in this study for flood forecasting in the Kushiyara river, which is one of the major rivers of the northeastern region of Bangladesh. Random Forest Regressor (RF), Categorical boosting regressor (CB), Deep Neural Network (DNN) and Long Short-Term Memory neural network (LSTM) have been chosen as the ML algorithms. Daily total precipitation (TP), volumetric soil water up to 100cm depth (SWVL), total precipitable water (TCRW), and total column water vapor (TCWV) from the fifth generation European Centre for Medium-Range Weather Forecasts Reanalysis (ERA5) has been used as inputs to the model in different combinations with historical water level (WL) data collected from BWDB. The models have been trained and tested using data from 2007 to 2020 to forecast floods with a lead time of 1, 3, 5 and 7 days, taking four different combinations of the input variables. When all the variables are taken as input together, the RMSE test scores for a forecast with a lead time of 1 day are found to be 0.31m, 0.38m, 0.52m and 0.39m for DNN, RF, CB and LSTM respectively. For a 7-day lead time, the values are 1.09m, 1.35m, 1.29m and 1.29m respectively for DNN, RF, CB and LSTM. The R2 values for the models are found to be 0.90, 0.84, 0.86, and 0.86, respectively, for DNN, RF, CB and LSTM for 7-day lead time. The study finds that the deep learning algorithms perform with better consistency regardless of lead times and input combinations. The LSTM models show more consistent results for a shorter lead time (1-day and 3-day) with only precipitation as input. However, for multiple variables, and when the lead time increases, the DNN models show a better overall performance considering training, validation and testing scenarios. The study suggests that despite the scarcity of rainfall data from the upstream areas, using ERA5 products in conjunction with deep learning algorithms may allow us to forecast and monitor high-volume flood occurrences in this region.
"A Spatiotemporal Analysis of Land-Use Changes in Rohingya Refugee Camps Using Multi-Temporal Satellite Image Analysis"
Abstract: The Rohingya Refugee Crisis is now one of the massive humanitarian catastrophes witnessed around the globe. About a million people of the Rohingyas took shelter in the Refugee camps of Teknaf and the Ukhia Upazilas of Bangladesh since they fled from Myanmar on 25 August 2017, making it one of the largest refugee camps in the world. Owing to the sudden change in the number of inhabitants and rapid development of human settlement in the region, drastic changes in land use patterns need to be observed. But coarse resolution satellite images (≥ 30 m) often fail to depict the actual human settlements in the region accurately. In this study, high-resolution Google Earth Images were manually digitized to demarcate the Rohingya inhabited regions to generate more accurate Land-use patterns by the Rohingya refugees between 2017 and 2021. The obtained result was then compared with Land Cover Maps generated by Maximum Likelihood Supervised Classification technique (MLSC) in GIS software using Landsat 8 images. The study also shows that the camp settlement areas are gradually changing and have increased by 7 times in 5 years from May 2017, and the vegetative cover has shown a steep reduction (54%). The research also established a correlation between the numbers of refugees with the increasing human settlement area.