BAU researchers build an AI model to forecast river floods in advance

Researchers at Bangladesh Agricultural University (BAU) in Mymensingh have built a flood forecasting model that uses artificial intelligence (AI) to predict how far the water in a river will rise or fall, Jagonews24 reported on 28 April 2026. The team used machine learning and deep learning, two branches of AI, and trained the model on 26 years of weather and river records from the Old Brahmaputra, from 1999 to 2024.

Predicting river levels early and accurately has long been a major challenge in flood-prone Bangladesh, and it is harder still in places where measurements are few. The researchers said the work was aimed at coping with floods driven by climate change and at giving useful warnings in exactly those data-poor areas. According to the report, the new approach gave very accurate forecasts while using far less data than conventional methods.

The team collected data from four important points on the Old Brahmaputra: Islampur, Sarishabari, Dewanganj and Mymensingh. They used rainfall, temperature, river water level and the rate of flow to estimate how much the river would rise or fall, and tested several different AI models against one another.

The research was led by Dr Md Touhidul Islam, associate professor in BAU's Department of Irrigation and Water Management. The team included Professor Dr A. K. M. Adham of the same department, along with undergraduate and postgraduate students.

The Ministry of Science and Technology (MoST) of the Government of Bangladesh and the University Grants Commission (UGC) funded the project, and the Bangladesh Agricultural University Research System (BAURES) gave technical and administrative support. According to the report, the work began in the middle of 2025 and was published in 2026 in an international journal ranked in the Q1 to Q2 category.

The journal records agree with the report. A paper by Md Touhidul Islam, Sujan Chandra Roy, Nusrat Jahan, Nahidul Islam and A. K. M. Adham, titled "From climate signals to smart predictions: advanced machine learning and deep learning for hydrological modeling in data-scarce Bangladesh", appeared online in H2Open Journal on 20 February 2026 (volume 9, issue 2, pages 81 to 109). Its funding record names the University Grants Commission of Bangladesh, under award crop_science-16/2022-2023. An earlier paper by the same group on the same river, in the Journal of Water and Climate Change on 8 October 2025, names the Ministry of Science and Technology as its funder.

BAURES itself is not new. The national encyclopaedia Banglapedia records that the university set up the BAU Research System in 1984 to coordinate and manage its research projects, which are sponsored either by the university or by outside agencies.

The H2Open Journal paper describes the study as an evaluation of five machine learning and deep learning models for predicting daily water levels in the Old Brahmaputra: linear regression, random forest, XGBoost, the light gradient boosting machine (LightGBM) and long short-term memory (LSTM). It used 26 years (1999 to 2024) of hydrological and meteorological data.

The models were tested under four scenarios, each giving the model a different kind of input:

• Water level only: the river's own recent levels are used to predict the next day's level.
• Rainfall only: no river measurement is used, only rain.
• Temperature only.
• Rainfall and temperature together.

The scenarios matter because many small rivers and remote reaches in Bangladesh have no gauge that records the water level, while rain and temperature are measured more widely. A model that can work from weather data alone could, in principle, give a forecast where there is no river record.

To tune the models, the authors searched through combinations of settings (a "grid search") and checked each one with five-fold cross-validation, in which the data are split into five parts and the model is tested repeatedly on the part it did not learn from. They also fed the models "lagged" values from the previous one to five days, so that the model could see how the river and the weather had been changing. Performance was measured with six statistics, and principal component analysis was used to rank the models.

According to the paper's abstract, the results were:

• Water-level-only scenario: the random forest model had a mean absolute error of 0.1445 metres, a coefficient of determination (R squared) of 0.9916 and a Nash-Sutcliffe efficiency of 0.9916.
• Rainfall and temperature scenario: LSTM did best, with R squared of 0.8145.
• Rainfall-only scenario: LSTM reached R squared of 0.8064.
• Temperature-only scenario: predictive capability was limited, with R squared of 0.5346.
• Transfer to another station: tested at Sarishabari without recalibration, the LightGBM model reached a mean absolute error of 0.1205 metres and R squared of 0.9917 in the water-level scenario.

The authors concluded that the study provides operational frameworks for climate-driven forecasting in data-scarce deltaic areas, and highlighted LSTM's capability for catchments that have no river gauges.

In the Jagonews24 report, Professor Adham explained the results in terms of accuracy. When past water-level data were used, he said, the random forest model (written RFM in the report) gave results that were up to 99.16 percent accurate. Where data were scarce, the deep learning model LSTM, analysing only rainfall and temperature, forecast correctly up to 81.45 percent of the time. An SVM (support vector machine) model was also used in the research. Such accuracy despite limited data, he said, points to great potential for data-poor areas.

These are the same numbers that appear in the paper as R squared values: 0.9916 for the random forest in the water-level-only scenario and 0.8145 for LSTM with rainfall and temperature. R squared, the coefficient of determination, is a standard statistical measure of how much of the variation in the observed values a model can reproduce. A value of 1 means the model reproduces the variation fully. It is a measure of fit across the whole record rather than a count of right and wrong forecasts.

The paper also reports the mean absolute error, which is easier to picture: on average, the random forest's next-day prediction differed from the measured level by 0.1445 metres in the water-level scenario. The Nash-Sutcliffe efficiency, the third figure, is a measure developed for judging hydrological models; a value of 1 again means a perfect match with the observations.

The gap between the two scenarios is the central finding. With the river's own history, the models were very close to the observed levels. With weather data alone, accuracy fell, but the deep learning model still explained most of the variation, which is why the researchers see it as a way to forecast where no gauge exists.

The study compared several families of machine learning. Each learns patterns from past data rather than from the physical equations of water flow that traditional hydrological models solve.

• Random forest: an ensemble method that builds a large number of decision trees during training. For a numerical prediction such as a water level, the output is the average of the trees' predictions. Combining many trees corrects the tendency of a single decision tree to overfit its training data.
• Long short-term memory (LSTM): a type of recurrent neural network, the family of deep learning models designed for sequences such as daily time series. It was designed to overcome the "vanishing gradient" problem of earlier recurrent networks and can keep information over long sequences, which suits rivers whose level depends on what happened days before.
• Support vector machine (SVM): a supervised learning model developed at AT&T Bell Laboratories, used for classification and regression. Its regression form, support vector regression, is the version used for numerical forecasts.
• LightGBM and XGBoost: gradient boosting methods built on decision trees. LightGBM is a free and open-source framework originally developed by Microsoft, with a focus on performance and scalability.
• Gated recurrent unit (GRU): a simpler relative of LSTM introduced in 2014, with fewer parameters, which the team tested in its other studies.

The report stressed that the innovation is not a commercial product but a digital, software-based technology. That is also why the researchers said farmers would not have to spend any extra money to use it.

The 2026 paper is one of a series the BAU group published on the same river in the months before the report. Each tested a different part of the forecasting problem.

In July 2025, in Progress in Disaster Science, the team compared nine machine learning models for predicting monthly maximum and minimum water levels at three key stations on the Old Brahmaputra, using a 34-year record from 1990 to 2024. Random forest regression performed best, with R squared of 0.87 to 0.92 for maximum levels and 0.82 to 0.92 for minimum levels. Midstream stations were predicted more accurately than downstream ones, which the authors linked to tidal effects.

In October 2025, in the Journal of Water and Climate Change, the group evaluated five algorithms (linear regression, random forest, XGBoost, support vector regression and LightGBM) on daily data from 1999 to 2024 from three hydrological stations and one meteorological station, using the same four input scenarios. Random forest reached R squared above 0.98 and mean absolute errors below 0.15 metres at all stations when it used water levels. The combined rainfall and temperature scenario gave R squared of 0.72 to 0.78, while temperature alone gave only 0.11 to 0.46. The authors wrote that the findings showed the feasibility of climate-driven forecasting for ungauged river basins.

In November 2025, in Results in Engineering, the team tested six algorithms for forecasts from 1 to 30 days ahead, using 26 years of daily data from Jaganathganj station and testing the models' transferability at Islampur, upstream. LightGBM was best for short-term forecasts, with R squared of 0.985 one day ahead and 0.939 seven days ahead, while the GRU model did best 30 days ahead (R squared 0.841). Adding rainfall and temperature reduced errors for long-term predictions by 13.5 to 17.8 percent, and the models kept a Nash-Sutcliffe efficiency above 0.90 for short-term forecasts when moved to another station.

In December 2025, in Water Practice and Technology, the group turned from water level to discharge, the volume of water flowing past a point, using 22 years of data (2000 to 2021) from the Mymensingh station. All models exceeded R squared of 0.98 for one-day predictions, but accuracy fell sharply with longer lead times, and the models systematically underestimated peak monsoon discharges.

In February 2026, in H2Open Journal, a further paper from the team examined how many past days of data a model should look at. Using 9,497 daily observations from Jaganathganj, it found that a GRU model looking back 30 days gave the best next-day forecasts (R squared 0.9903, mean absolute error 0.1255 metres), and that all six models tested exceeded a Nash-Sutcliffe efficiency of 0.97.

Read together, the papers show the same pattern as the 28 April report: river-level history gives very accurate short-range forecasts, weather data alone give useful but less accurate ones, and models trained at one station can be applied at another.

The Old Brahmaputra leaves the left bank of the Brahmaputra north of Bahadurabad, flows roughly southeast past Jamalpur and Mymensingh towns, and joins the Meghna at Bhairab Bazar, according to Banglapedia. It was once the main course of the Brahmaputra: James Rennell's maps of the 18th century show the great river flowing east of the Madhupur Tract, along what is now the Old Brahmaputra. In 1787, after a major flood, the main flow shifted west into the channel now called the Jamuna. An earthquake reported from the Mymensingh region in 1782 may also have contributed, and the shift seems to have taken about 30 years.

The Old Brahmaputra floodplain, stretching from the southwestern corner of the Garo Hills along the eastern rim of the Madhupur Tract to the Meghna, has a gentle landscape of broad ridges and depressions. Its soils are more oxidised and darker than those formed over the last 200 years on the younger Brahmaputra floodplain, Banglapedia notes.

Three of the four stations named in the report, Islampur, Dewanganj and Sarishabari, are upazilas of Jamalpur district. Banglapedia, citing the 2011 census, lists Jamalpur as a district of 2,115.16 square kilometres with seven upazilas, whose main rivers are the Jamuna, the Old Brahmaputra, the Banar and the Jhinai. In that census Islampur had 298,429 people, Dewanganj 258,133 and Sarishabari 325,320. Banglapedia gives agriculture as the main source of income at 65.50 percent, far ahead of commerce at 11.88 percent.

Mymensingh, the fourth station, is where the research was done. BAU was established in 1961, and its campus of about 1,200 acres lies on the western bank of the Old Brahmaputra, 3 kilometres south of Mymensingh town, according to Banglapedia.

Floods are a recurring part of life in Bangladesh. Banglapedia records that about 26,000 square kilometres, 18 percent of the country, is flooded each year, and that in severe floods the affected area can exceed 55 percent. In an average year about 844,000 million cubic metres of water flows into the country from May to October through the three main rivers, the Ganges, the Brahmaputra-Jamuna and the Meghna, which is 95 percent of the annual inflow. Only about 187,000 million cubic metres is generated by rainfall inside the country in the same period.

The encyclopaedia divides the country's floods into three kinds: monsoon floods, which rise and fall slowly and cover vast areas; flash floods, which rise and fall suddenly, generally in the valleys of hilly areas; and tidal floods. It describes the Old Brahmaputra valley as one of the distinct floodplain areas of the northeast region, separated from the Brahmaputra by the Madhupur Tract.

The researchers said their framework could be linked directly to the forecasting centre of the national river network. That centre is the Flood Forecasting and Warning Centre (FFWC) of the Bangladesh Water Development Board (BWDB). According to a BWDB description of its service, the FFWC was established in 1972 and is fully operational in the flood season, from 1 April to 31 October, as directed by the government's Standing Orders on Disaster. Its forecasts rest on a real-time hydrological monitoring network, a data exchange agreement with India for rainfall and water levels from outside the border, weather data from the Bangladesh Meteorological Department, and satellite and radar images.

The same document traces how the system grew. In 1972 the centre had 10 real-time flood monitoring stations and forecast by gauge-to-gauge correlation. The MIKE11-FF numerical model was introduced in 1992 with 16 forecasting points, a larger model with GIS followed in 1995 to 1996 with 30 points, and by 2009 forecasting covered 52 locations on 23 rivers, along with inundation mapping. Its products include daily bulletins, forecast hydrographs, warning messages and flood maps. The BAU team presents its AI models as a possible addition to this kind of physical modelling, especially for rivers and reaches the gauge network does not cover.

The researcher said in the report that farmers would benefit directly if the technology were used in the field. With an accurate advance warning of a flood, farmers could bring in ripe rice and other crops early, move livestock to safe places, and plan irrigation so that water is used properly, reducing the risk of large financial losses.

The districts along the study river know such losses. In early July 2024, Dhaka Tribune reported from Jamalpur that about 50,000 people were marooned in 18 unions of Dewanganj, Islampur and Madarganj upazilas, with several thousand hectares of cropland under water, as the Jamuna at Bahadurabad Ghat in Dewanganj rose 92 centimetres above the danger level. A day later, on 6 July 2024, the paper reported that the number had risen to 150,000 people in 45 unions of six upazilas, including Islampur, Dewanganj and Sarishabari, according to the executive engineer of the Jamalpur office of the Water Development Board.

The researchers' aim, the report said, is to apply the model to different rivers of the country and build an effective early warning system. Their research showed that a model built for one area could also work well in another, which the Sarishabari and Islampur transferability tests in their papers were designed to check.

The innovation was ready as an operational framework that could be connected directly to the national river network's forecasting centre, according to the report. With government sponsorship and technical support, the researchers said, the AI-based technology could be developed further into a full operational system and expanded easily to the field. They also hoped that, through a government initiative, the forecasts could reach farmers cheaply as a mobile app or a phone message.

Photo: Ashis Saha / Wikimedia Commons (CC BY-SA 3.0)

Sources: Jagonews24, H2Open Journal (IWA Publishing), Journal of Water and Climate Change, Results in Engineering, Progress in Disaster Science, Water Practice and Technology, Bangladesh Water Development Board (FFWC), Banglapedia, Dhaka Tribune, Wikipedia