Assessing the capability of UAV-borne multispectral and LiDAR sensors for mapping water levels in small water bodies.

Abstract

Efficient and spatially explicit quantification of water levels in small water bodies is vital for sustainable water resource management, particularly in arid and data-scarce regions. While satellite remote sensing provides regional-scale observations, its coarse spatial resolution limits applicability to small water bodies. Unmanned Aerial Vehicles (UAVs) bridge this gap by offering high-resolution, site-specific spatial and temporal changes. The first objective of this study was to systematically review UAV-based studies on water level mapping, assessing their progress, challenges, and opportunities. The reviewed literature shows that UAVs achieve accuracies up to R² = 0.85 relative to in situ and airborne data, utilising photogrammetry, LiDAR, and radar sensors-mostly integrated with DJI platforms. Advanced modelling techniques such as linear regression, convolutional neural networks, and support vector machine regression achieved mean RMSEs of 0.65 m, 0.035 m, and 0.39 m, respectively. Literature indicated a potential gap of UAV-derived data for mapping water levels in small water bodies. Hence, the second objective was to assess the spatial and temporal resolution capabilities of drones in detecting and mapping water levels in small water bodies. The DJI Matrice 300, equipped with a Mica Sense multispectral camera and a Zenmuse L1 LiDAR sensor, was used to monitor water level variations at High Down Stud Dam, Pietermaritzburg. Multispectral images were used to assess the changes in water surface extent through water spectral-based indexing, such as MNDWI. UAV LiDAR data enabled the precise quantification of water level fluctuations through elevation differencing between water surfaces and land boundaries. To address the limitations of using only UAV-derived data, water depth measurements obtained with a dipstick were used for validation. Rando Forest (RF) and Extreme Gradient Boosting (XGBoost) models were implemented to assess the predictive accuracy of the developed approach. The Random Forest (RF) model demonstrated strong predictive performance, achieving coefficients of determination (R2 ) of 0.96 and 0.97 for May and October, respectively, with corresponding RMSE values of 0.12 m and 0.11 m, and relative RMSE (rRMSE) of 1.36% and 1.14%. Similarly, the XGBoost model performed well, attaining R² values of 0.95 and 0.96 in May and October, respectively, with RMSE values of 0.13 m for both months and rRMSE of 1.57% (May) and 1.39% (October). The study will contribute to a broader perspective of combining remote sensing techniques with in-situ validation and machine learning techniques in assessing the capabilities of drones in mapping water levels in small water bodies.