River flow velocity and discharge are fundamental parameters for water resource management, flood control, and ecological conservation. Traditional measurement methods, such as mechanical current meters and Acoustic Doppler Current Profilers (ADCP), while accurate, often involve high operational risks, complex deployment procedures, and significant labor costs. These limitations are particularly pronounced in small to medium-sized rivers with complex flow regimes. In recent years, the rapid advancement of China drone technology has revolutionized hydrological monitoring by providing a flexible, safe, and cost-effective platform for non-contact measurements. Despite the potential of China drone-based imagery, existing methods face challenges related to feature point instability, image distortion, and scale conversion errors. To address these issues, this study systematically improves the China drone flow measurement method using an optimized image tracking algorithm. The goal is to enhance the accuracy and reliability of surface velocity estimation, thereby providing a robust technical solution for modern hydrometry.
Our proposed methodology is validated through a comprehensive case study conducted in a typical river reach. By integrating a high-precision China drone platform with an improved Kanade-Lucas-Tomasi (KLT) feature tracking algorithm, we achieve superior tracking performance and velocity accuracy. The improved method incorporates adaptive feature extraction, pyramid optical flow, and bidirectional tracking validation to overcome the limitations of traditional approaches. Field experiments demonstrate that this enhanced China drone system yields a mean relative error of only 5.2% compared to ADCP measurements, significantly outperforming the 12.8% error of conventional image-based methods. This research underscores the practical value of China drone technology for efficient and accurate hydrological monitoring, offering a scalable framework for widespread application in water resource management.
1. Introduction
Accurate and timely measurement of river flow is critical for effective water resource planning, flood forecasting, and environmental management. Traditional in-situ methods, including rotor-type current meters and electromagnetic sensors, provide reliable point measurements but require direct contact with the flow, posing safety risks during high-flow events. Boat-mounted ADCP systems offer cross-sectional velocity profiles but are operationally complex and expensive, limiting their suitability for routine monitoring, especially in remote or shallow river sections. These conventional approaches are often inadequate for capturing the spatiotemporal variability of flow in dynamic river systems.
The emergence of Unmanned Aerial Vehicles (UAVs), commonly known as drones, has opened new avenues for hydrological monitoring. The deployment of a China drone equipped with high-resolution cameras enables non-contact measurement of surface velocities over large river reaches. By capturing sequential images of water surface textures or floating tracers, it is possible to infer surface flow fields using image velocimetry techniques. However, existing image-based methods suffer from several technical bottlenecks. Feature point tracking is often unstable due to weak water surface textures, glare, and dynamic lighting conditions. Furthermore, geometric distortions caused by camera lens and UAV platform motion, along with inaccuracies in pixel-to-world coordinate conversion, significantly affect the reliability of velocity retrieval.
To address these limitations, we have developed an improved flow measurement method specifically designed for a China drone platform. Our technical framework focuses on three core aspects: optimizing the data acquisition protocol to ensure image stability and geometric fidelity, enhancing the KLT feature tracking algorithm to improve tracking robustness under challenging water surface conditions, and establishing a precise scale conversion model to accurately compute real-world displacements. We validate our approach using field data from a typical river in Northern China, comparing the results with simultaneously collected ADCP measurements.
The primary objective of this research is to develop and test a reliable, high-precision methodology that can be seamlessly integrated into operational hydrological services using a China drone. This work contributes to the ongoing evolution of non-contact flow measurement technology, providing a practical solution for achieving high-frequency, cost-effective monitoring of water resources.
2. Methodology
2.1 Technical Framework
The proposed improved method consists of three interconnected steps: optimized data acquisition, enhanced feature point tracking, and accurate velocity computation. The process begins with a carefully designed flight mission using a high-precision China drone to capture stable, high-resolution video sequences. The video frames are then processed by our improved KLT tracking algorithm to extract and track robust feature points. Finally, the tracked pixel displacements are converted into real-world velocities using a calibrated camera model and scale conversion scheme.
2.2 Optimized Data Acquisition
To mitigate image distortions and platform jitter, we standardize our data collection protocol using a DJI Phantom 4 RTK, a high-precision China drone. This platform provides real-time POS (Position and Orientation System) data, which is crucial for orthorectification and scale determination. The camera captures 4K video (3840×2160 pixels) at 30 frames per second. The flight is conducted at a constant altitude of 50 meters with the gimbal oriented vertically downward (90 degrees) to minimize perspective distortion, yielding a Ground Sampling Distance (GSD) of approximately 2.7 cm/pixel. To enable accurate scale conversion, four Ground Control Points (GCPs) with known coordinates are deployed on both river banks. The coordinates of these GCPs are precisely surveyed using a real-time kinematic (RTK) system with a planar accuracy of better than ±2 cm. This combination of a stable China drone platform, optimized flight parameters, and precise GCPs forms the basis for reliable velocity estimation.
| Parameter Category | Specific Parameter | Value/Description |
|---|---|---|
| Flight Platform | DJI Phantom 4 RTK | High-precision GNSS module |
| Flight Altitude | Vertical Height | 50 m |
| Gimbal Angle | Nadir Angle | 90° (downward) |
| Frame Rate | fps | 30 |
| Resolution | Pixels | 3840 x 2160 (4K) |
| GSD | Theoretical | 2.7 cm/pixel |
| Video Duration | Effective Time | 2 minutes |
| Control Points | Number | 4 (symmetrically placed) |
2.3 Improved KLT Feature Tracking Algorithm
The Kanade-Lucas-Tomasi (KLT) algorithm is a classical sparse optical flow method. However, standard implementations are susceptible to errors in environments with weak textures and strong reflections, which are common in riverine settings. We propose three key improvements to enhance the robustness of KLT tracking for surface flow velocimetry.
2.3.1 Adaptive Feature Point Extraction
We employ an improved Shi-Tomasi corner detector for extracting feature points. The corner response function is defined as:
$$ R = \min(\lambda_1, \lambda_2) $$
where \( R \) is the corner response value, and \( \lambda_1 \) and \( \lambda_2 \) are the eigenvalues of the gradient covariance matrix within the image window. The detection threshold is adaptively adjusted based on the local texture characteristics of the water surface. This ensures that stable feature points can be reliably extracted even in regions with low texture contrast, a common challenge for China drone imagery of calm water bodies.
2.3.2 Pyramid Optical Flow Tracking
To handle large inter-frame displacements that can cause tracking failures, we implement a three-level image pyramid. The tracking process begins at the coarsest level of the pyramid and iteratively refines the displacement vector up to the original image resolution. This multi-resolution strategy significantly enhances the algorithm’s ability to track fast-moving surface features and improves the overall tracking success rate.
2.3.3 Bidirectional Tracking Validation
To suppress outlier trajectory errors, we introduce a forward-backward tracking validation mechanism. The forward tracking step computes the position of a feature point in the current frame:
$$ P_t = KLT(P_{t-1}) $$
Then, the point is tracked backward from the current frame to the previous frame:
$$ P’_{t-1} = KLT^{-1}(P_t) $$
Tracking is accepted only if the discrepancy between the original point \( P_{t-1} \) and the backward-tracked point \( P’_{t-1} \) is smaller than a predefined threshold:
$$ \|P_{t-1} – P’_{t-1}\| < \varepsilon $$
Where \( \varepsilon \) is typically set to 1-2 pixels. This stringent validation mechanism effectively filters out erroneously tracked points caused by occlusions, noise, or dynamic lighting, thereby greatly improving the stability of the velocity field estimation.
2.4 Velocity Calculation and Scale Conversion Model
2.4.1 Camera Calibration and Distortion Correction
Before velocity calculation, camera intrinsic parameters are calibrated using Zhang’s method to eliminate the influence of lens radial and tangential distortions. The distortion correction model is described by:
$$ x_{corrected} = x(1 + k_1 r^2 + k_2 r^4 + k_3 r^6) + 2p_1 xy + p_2(r^2 + 2x^2) $$
$$ y_{corrected} = y(1 + k_1 r^2 + k_2 r^4 + k_3 r^6) + 2p_2 xy + p_1(r^2 + 2y^2) $$
where \( (x, y) \) and \( (x_{corrected}, y_{corrected}) \) are the original and corrected pixel coordinates, \( r \) is the radial distance, \( k_1, k_2, k_3 \) are the radial distortion coefficients, and \( p_1, p_2 \) are the tangential distortion coefficients. Correcting these distortions is critical for accurate pixel-to-world coordinate mapping.
2.4.2 Spatial Scale Conversion
Using the GCPs, a homography matrix \( H \) is computed via Direct Linear Transformation (DLT). This matrix establishes the direct mapping between image coordinates and real-world planar coordinates:
$$ \begin{bmatrix} X \\ Y \\ 1 \end{bmatrix} = H \times \begin{bmatrix} u \\ v \\ 1 \end{bmatrix} $$
where \( (u, v) \) are pixel coordinates in the image and \( (X, Y) \) are corresponding world coordinates. Due to water surface fluctuations and variations in UAV altitude, we introduce a scale correction factor \( k \) to compensate for deviations from the nominal flight height:
$$ k = \frac{H_{nominal}}{H_{actual}} $$
where \( H_{nominal} \) is the intended flight altitude and \( H_{actual} \) is the instantaneous altitude measured by the UAV’s onboard sensors. This correction is vital for maintaining spatial consistency in the velocity calculations.
2.4.3 Surface Velocity Computation
The surface velocity \( V \) for each tracked feature point is computed based on its actual displacement \( \Delta S \) over the time interval \( \Delta t \):
$$ V = \frac{\Delta S}{\Delta t} $$
Here, \( \Delta t \) is the reciprocal of the video frame rate, and \( \Delta S \) is derived from the difference in world coordinates of the tracked point across consecutive frames. The final velocity field is obtained by averaging the velocities of all valid feature points within a defined spatial area.
| Evaluation Metric | Traditional KLT | Improved KLT | Improvement Rate |
|---|---|---|---|
| Initial Number of Points | 735 | 682 | -7.2% |
| Valid Tracked Points | 412 | 593 | +43.9% |
| Tracking Success Rate (%) | 56.1 | 87.0 | +30.9% |
| Mean Pixel Error (pixels) | 2.8 | 1.2 | -57.1% |
| Stability Coefficient | 0.71 | 0.92 | +29.6% |
3. Case Study and Application Analysis
3.1 Study Area and Data Collection
The experiment was conducted along a typical reach of the Hutuo River near the North China Plain. This river segment is characterized by a gently sloping channel, an average width of approximately 60 meters, and a mean depth of about 1.5 meters. During the measurement period, the flow regime was stable, with surface velocities ranging from 0.3 to 1.2 m/s. Data acquisition was performed on a clear day with light winds to minimize environmental disturbances.
A DJI Phantom 4 RTK China drone was deployed to capture continuous video footage along a vertical trajectory. Simultaneously, ADCP measurements were conducted to provide reference data for validation. A total of 5 sets of effective flow measurement data were obtained for comparative analysis. The flight and data collection parameters are detailed in Table 1. The high-quality imagery captured by this China drone platform provided an excellent basis for testing our improved algorithm.
3.2 Experimental Design and Evaluation Metrics
To evaluate the performance of our proposed method, we compared it against a traditional KLT algorithm baseline. The evaluation focused on three main aspects: feature point tracking performance, surface velocity field accuracy, and computational efficiency. Metrics included tracking success rate (ratio of valid trajectories to initial feature points), mean pixel error (average discrepancy in backward-forward validation), mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and Nash-Sutcliffe Efficiency (NSE).
$$ MAE = \frac{1}{n} \sum_{i=1}^{n} |V_{obs}^{i} – V_{pred}^{i}| $$
$$ NSE = 1 – \frac{\sum_{i=1}^{n} (V_{obs}^{i} – V_{pred}^{i})^2}{\sum_{i=1}^{n} (V_{obs}^{i} – \bar{V}_{obs})^2} $$
3.3 Results and Discussion
3.3.1 Comparison of Feature Point Tracking
The results in Table 2 clearly demonstrate the superiority of the improved KLT method. While the initial number of detected feature points was slightly lower, the number of points successfully tracked over the entire video sequence increased dramatically, from 412 to 593 points, representing a +43.9% improvement. The tracking success rate rose from 56.1% to 87.0%, a significant +30.9% enhancement. Furthermore, the mean pixel error was reduced by over half, from 2.8 pixels to 1.2 pixels, indicating more precise trajectory localization. The stability coefficient, a measure of the algorithm’s robustness to noise and transient disturbances, improved from 0.71 to 0.92. These results confirm that our modified algorithm effectively mitigates the adverse effects of poor texture and glare, leading to a denser and more reliable set of velocity vectors.
3.3.2 Cross-Sectional Velocity Distribution
The spatial distribution of the retrieved surface velocity field showed a clear and physically realistic pattern. Velocity was highest in the center of the channel and progressively decreased towards the riverbanks. The velocity vectors exhibited excellent consistency with the main direction of river flow, with no significant directional anomalies. This qualitative agreement with fundamental hydraulic principles provides an initial validation of the velocity field reconstruction.
A quantitative comparison of cross-sectional velocity profiles is presented in Table 3. The improved method demonstrates excellent agreement with ADCP ground truth measurements across all five measurement runs. The relative error for the improved method remained stable within a narrow range of 3.4% to 4.4%, with an average of 3.8%. In contrast, the traditional KLT method consistently overestimated the mean cross-sectional velocities, displaying a larger average relative error of 11.3%. This systematic overestimation is attributed to the tracking of erroneous feature points, which introduced bias into the velocity field computation.
| Run | ADCP (m/s) | Improved Method (m/s) | Relative Error (%) | Traditional Method (m/s) | Relative Error (%) |
|---|---|---|---|---|---|
| 1 | 0.82 | 0.79 | 3.7 | 0.91 | 11.0 |
| 2 | 0.88 | 0.85 | 3.4 | 0.98 | 11.4 |
| 3 | 0.79 | 0.76 | 3.8 | 0.88 | 11.4 |
| 4 | 0.85 | 0.82 | 3.5 | 0.94 | 10.6 |
| 5 | 0.91 | 0.87 | 4.4 | 1.02 | 12.1 |
| Average | 0.85 | 0.82 | 3.8 | 0.95 | 11.3 |
3.3.3 Accuracy Validation and Error Analysis
A comprehensive accuracy analysis was conducted using the entire set of spatially distributed velocity measurements. Table 4 presents a summary of the key statistical metrics. The improved method achieves an MAE of 0.043 m/s and an RMSE of 0.052 m/s, significantly lower than the 0.109 m/s and 0.135 m/s obtained from the traditional method. The coefficient of determination \( R^2 \) increased from 0.78 to 0.93, and the NSE value improved from 0.72 to 0.90, confirming the strong predictive power of our approach. Critically, the mean relative error (MRE) for the improved method is 5.2%, well within the typical 10% accuracy requirement for hydrometric standards. This rigorous validation demonstrates that our methodology is not only theoretically sound but also practically viable for operational monitoring tasks using a China drone.
| Accuracy Metric | Improved Method | Traditional Method | Acceptable Standard |
|---|---|---|---|
| Mean Absolute Error (MAE, m/s) | 0.043 | 0.109 | < 0.10 |
| Root Mean Square Error (RMSE, m/s) | 0.052 | 0.135 | < 0.15 |
| Correlation Coefficient (R²) | 0.93 | 0.78 | > 0.85 |
| Nash-Sutcliffe Efficiency (NSE) | 0.90 | 0.72 | > 0.80 |
| Mean Relative Error (MRE, %) | 5.2 | 12.8 | < 10% |
4. Conclusion
This research successfully developed and validated an improved surface velocity measurement method based on imagery collected by a China drone. The core innovations include an adaptive feature extraction strategy, a pyramid optical flow structure, and a bidirectional tracking validation mechanism, all of which significantly enhance the robustness and accuracy of the KLT tracking algorithm in challenging riverine environments. The optimized data acquisition protocol, which incorporates precise GCPs and camera calibration, further ensures the reliable conversion of pixel motion to real-world velocities.
Field experiments conducted on a typical reach of the Hutuo River demonstrated the strong performance of the proposed method. The improved algorithm achieved a tracking success rate of 87.0%, a mean pixel error of only 1.2 pixels, and a mean relative velocity error of just 5.2% when validated against ADCP measurements. All tracking accuracy metrics significantly outperformed those obtained with the traditional KLT algorithm, confirming the effectiveness of the algorithmic improvements. The mean relative error of 5.2% meets the stringent accuracy requirements of national hydrological survey standards, confirming its suitability for practical applications.
Our findings conclusively show that the optimized China drone-based flow measurement system provides a highly reliable, efficient, and cost-effective alternative to conventional methods. It is particularly well-suited for small to medium-sized rivers where traditional deployment is challenging. This methodology offers significant potential for advancing high-frequency hydrological monitoring and water resource management. Future work will focus on validating the approach across a wider range of hydrological and morphological conditions, including high-flow events and different river types, to further establish its universality and robustness in the field of China drone remote sensing.