1- Faculty of Agriculture - Tabriz University
2- Robotics and Soft Technologies Research Center, Tabriz Branch, Islamic Azad University, Tabriz, Iran
2- Robotics and Soft Technologies Research Center, Tabriz Branch, Islamic Azad University, Tabriz, Iran
Abstract: (427 Views)
Extended Abstract
Introduction and Objectives: Studying and monitoring the water level of rivers and canals by the Ministry of Energy and related organizations is an important part of water resources management in the catchment area. On the other hand, water level measurement is a vital task in hydrological monitoring, but it often faces limitations such as lack of resources, high costs, and high time requirements. These limitations often lead to delays in measurements and potential inaccuracies, especially in remote or harsh environments. In general, most of the traditional methods have significant errors and costs and make continuous monitoring and control almost impossible. Recent advances in technology have led to a paradigm shift towards image processing systems for water level monitoring. These non-contact methods have attracted attention due to their high potential in terms of accuracy, reliability, cost-effectiveness, and reduced time required. In this research, presenting a new approach of combining image processing and machine learning, it has been tried to reduce costs and increase performance to extract water level indicators and measure the water level instantly and automatically. This approach is based on creating a set of gauge images recorded by a smartphone (which is a common device with easy access) for clear and turbid water states, to train machine learning-based models. This method, unlike the traditional methods with instantaneous and continuous measurement of water level, makes water supply systems work better and manage critical situations such as floods, river overflows and erosion.
Methodology: This paper contributes to this growing research by evaluating an image-based water level detection system using a standard smartphone camera. The RGB image processing algorithms in this research include filtering, noise reduction, color detection, resizing, grayscale conversion, Hough detection and transformation, and projection to obtain digital characters and watermarks that only include the area of scale lines. Also, all the mentioned steps and modeling steps have been done in Mathematica software. The experimental data includes 244 observation data, which were randomly considered to be 201 training data images, and 43 test data images, which were captured by a mobile phone camera with a fixed position. Considering the capabilities of various machine learning models, including artificial neural networks (ANN) and deep learning (DL) in image processing and analysis, we focused on these models for accurate water level estimation. Our study involves taking water level images, identifying the water edge in a gauge, and using these models to estimate the water level. Machine learning, a branch of artificial intelligence, aims to develop computer systems with the capacity to learn from data. This process involves computer learning through hands-on experience, starting with organizing data, choosing a machine learning algorithm, entering data, and enabling the model to independently learn patterns or generate predictions, and gain self-programming capability. In general, the comparative analysis of the performance of these models aims to show the potential of combining image processing and machine learning in overcoming the traditional obstacles of water level measurement in hydrological studies.
Findings: The results of the model were evaluated using three evaluation criteria: coefficients of determination (R), root mean square error (RMSE), and Nash-Sutcliffe agreement coefficient (NSE) and visual charts. In this research, two artificial neural network (ANN) and deep learning (DL) models, which are subsets of machine learning models, were used to estimate the water level in muddy and clear conditions by image processing. The results showed that the deep learning model is acceptable in both conditions of performance and also according to the evaluation indicators of the deep learning model with the lowest error of 28.39 mm and the highest coefficient of determination of 0.973, it was chosen as the best model for water level estimation. Therefore, according to the performance of the deep learning model, it is possible to automatically and continuously check the process of checking and monitoring the water level, in addition to laboratories, in hard-to-reach places and without high costs, and by making the right decisions from Prevent possible accidents.
Conclusion: Due to the problems of manual measurement and field monitoring, automatic and continuous monitoring of water level by humans becomes difficult and even impossible. Despite these obstacles, today researchers' interest in image processing systems has increased with the advancement of technology. According to the work done to detect and estimate the water level, today most of the methods go towards achieving the maximum efficiency with the minimum facilities. Finally, two methods for extracting information from digital images for water level monitoring systems are compared. Combined techniques for water level detection in clear and muddy images were compared based on visual evaluation and statistical accuracy. Based on the experimental results, these techniques and models were all able to extract water level information from the image. The image processing technique and deep learning model for detecting and estimating water surface features from images that include two turbid and clear states at three levels of low, medium, and high altitude had acceptable results and high efficiency. Due to the increasing progress in the field of image processing and machine learning, future research can add different states of water and create models based on new algorithms of machine learning and artificial intelligence.
Introduction and Objectives: Studying and monitoring the water level of rivers and canals by the Ministry of Energy and related organizations is an important part of water resources management in the catchment area. On the other hand, water level measurement is a vital task in hydrological monitoring, but it often faces limitations such as lack of resources, high costs, and high time requirements. These limitations often lead to delays in measurements and potential inaccuracies, especially in remote or harsh environments. In general, most of the traditional methods have significant errors and costs and make continuous monitoring and control almost impossible. Recent advances in technology have led to a paradigm shift towards image processing systems for water level monitoring. These non-contact methods have attracted attention due to their high potential in terms of accuracy, reliability, cost-effectiveness, and reduced time required. In this research, presenting a new approach of combining image processing and machine learning, it has been tried to reduce costs and increase performance to extract water level indicators and measure the water level instantly and automatically. This approach is based on creating a set of gauge images recorded by a smartphone (which is a common device with easy access) for clear and turbid water states, to train machine learning-based models. This method, unlike the traditional methods with instantaneous and continuous measurement of water level, makes water supply systems work better and manage critical situations such as floods, river overflows and erosion.
Methodology: This paper contributes to this growing research by evaluating an image-based water level detection system using a standard smartphone camera. The RGB image processing algorithms in this research include filtering, noise reduction, color detection, resizing, grayscale conversion, Hough detection and transformation, and projection to obtain digital characters and watermarks that only include the area of scale lines. Also, all the mentioned steps and modeling steps have been done in Mathematica software. The experimental data includes 244 observation data, which were randomly considered to be 201 training data images, and 43 test data images, which were captured by a mobile phone camera with a fixed position. Considering the capabilities of various machine learning models, including artificial neural networks (ANN) and deep learning (DL) in image processing and analysis, we focused on these models for accurate water level estimation. Our study involves taking water level images, identifying the water edge in a gauge, and using these models to estimate the water level. Machine learning, a branch of artificial intelligence, aims to develop computer systems with the capacity to learn from data. This process involves computer learning through hands-on experience, starting with organizing data, choosing a machine learning algorithm, entering data, and enabling the model to independently learn patterns or generate predictions, and gain self-programming capability. In general, the comparative analysis of the performance of these models aims to show the potential of combining image processing and machine learning in overcoming the traditional obstacles of water level measurement in hydrological studies.
Findings: The results of the model were evaluated using three evaluation criteria: coefficients of determination (R), root mean square error (RMSE), and Nash-Sutcliffe agreement coefficient (NSE) and visual charts. In this research, two artificial neural network (ANN) and deep learning (DL) models, which are subsets of machine learning models, were used to estimate the water level in muddy and clear conditions by image processing. The results showed that the deep learning model is acceptable in both conditions of performance and also according to the evaluation indicators of the deep learning model with the lowest error of 28.39 mm and the highest coefficient of determination of 0.973, it was chosen as the best model for water level estimation. Therefore, according to the performance of the deep learning model, it is possible to automatically and continuously check the process of checking and monitoring the water level, in addition to laboratories, in hard-to-reach places and without high costs, and by making the right decisions from Prevent possible accidents.
Conclusion: Due to the problems of manual measurement and field monitoring, automatic and continuous monitoring of water level by humans becomes difficult and even impossible. Despite these obstacles, today researchers' interest in image processing systems has increased with the advancement of technology. According to the work done to detect and estimate the water level, today most of the methods go towards achieving the maximum efficiency with the minimum facilities. Finally, two methods for extracting information from digital images for water level monitoring systems are compared. Combined techniques for water level detection in clear and muddy images were compared based on visual evaluation and statistical accuracy. Based on the experimental results, these techniques and models were all able to extract water level information from the image. The image processing technique and deep learning model for detecting and estimating water surface features from images that include two turbid and clear states at three levels of low, medium, and high altitude had acceptable results and high efficiency. Due to the increasing progress in the field of image processing and machine learning, future research can add different states of water and create models based on new algorithms of machine learning and artificial intelligence.
Type of Study: Applicable |
Subject:
ساير موضوعات وابسته به مديريت حوزه آبخيز
Received: 2024/01/29 | Accepted: 2024/06/10
Received: 2024/01/29 | Accepted: 2024/06/10
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
