Abstract:
River meandering causes severe environmental and socio-economic risks such as erosion, habitat loss, and damage to infrastructure. Traditional methods for studying river meandering often rely on retrospective data analysis and are usually limited to specific regions. Due to that, it is challenging to scale these methods or use them for proactive risk management. This research aims to overcome these limitations by creating a more flexible predictive model. With the help of satellite imagery and machine learning, the study focuses on the early identification of meandering patterns and risk assessment.
This study adopts a time-series forecasting approach to predict river centreline deviations and identify risk areas. Annual satellite images (2000–present) are retrieved via Google Earth Engine and pre-processed through image filtering, cloud removal, and spectral band selection. The main river area is then extracted using OpenCV-based contour detection, followed by skeletonisation to derive yearly centrelines. A breadth-first search algorithm is applied to remove additional branches, preserving only the longest path. These centrelines are resampled to a uniform number of points to ensure consistent analysis. Then, pointwise deviations between consecutive years are used as input to the ARIMA model, which forecasts future deviations. Additionally, explainability is incorporated into the system through the visualisation of each processing stage.
The system demonstrated strong performance with low error metrics, achieving an MAE of 0.6899, RMSE of 0.8845, and MAPE of 0.52% for row deviations, and 1.4227, 1.7308, and 0.86% respectively for column deviations. The predicted centreline trends closely aligned with actual deviations, affirming the model's accuracy in forecasting river meandering patterns
