Potential of Using Machine Learning Regression Techniques to Utilize Sentinel Images for Bathymetry Mapping of Nile River

Abstract

This paper investigates the potential of using the Sentinel-2 images for deriving bathymetric maps for the Nile River in Egypt. Regression analysis technique with the aid of in situ measurements were used. Exploring the effects of the size of calibration data, and water depth is another aim of this paper. Linear and machine learning regression techniques (Fine Decision Tree (FDT), and Random Forest (RF) Algorithms) are investigated. The study area is about 23 km between Assiut and Delta barrages. Around 82,000 depth points, are available. Regression models are developed using the depth data and the corresponding digital values of the 13 Sentinel-2 imagery bands. The linear regression model was not applicable (R2 = 0.02). The FDT, and RF models, have (R2 = 0.86) when half of the depth points were used for calibration, and the RMSE of the testing data (the other half of the depth points) equaled to 2.86 and 2.09 m, respectively. When fewer points were used for calibration, the R2 = 0.83 and 0.84, and RMSE = 2.65 and 2.19 m, for FDT and RF, respectively. When only the shallow water areas were considered, the RMSE reached 1.51 and 1.30 m for the FDT and RF, respectively. The VNIR bans of Sentinel-2 images are not enough to estimate the water depth of the Nile river. Collecting in situ measurements for small areas, is adequate for estimating the water depth and produces bathymetry maps for the Nile River in Egypt, and larger in situ measurements are not required.

Keywords

• Remote sensing
• Sentinel-2
• Bathymetry mapping
• Nile River