Aquaculture water dissolved oxygen concentration remote sensing estimation method based on measured data and machine learning
By combining empirical modal decomposition and machine learning algorithms, the problems of insufficient time and space coverage and low accuracy of dissolved oxygen monitoring in aquaculture fish ponds are solved, and high-precision dissolved oxygen concentration prediction is achieved, providing real-time and reliable data support for aquaculture water quality management.
Patent Information
- Application Number
- CN202510548070.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-28
- Publication Date
- 2025-09-02
AI Technical Summary
Traditional dissolved oxygen monitoring methods have limited space-time coverage and high cost in aquaculture fish ponds, making it difficult to capture dynamic changes. In addition, traditional spectral analysis has poor feature extraction effect in dynamic breeding environments, resulting in insufficient model accuracy.
Combining empirical modal decomposition and machine learning algorithm, through adaptive decomposition and feature reconstruction of water spectral signals, a high-precision prediction model for dissolved oxygen concentration suitable for farmed fish ponds is established, and multi-scale eigenmodal functions are extracted using EMD and dynamic modeling and training is combined with machine learning models.
It realizes large-scale high-precision monitoring of dissolved oxygen concentration in aquaculture fish ponds, breaks through the bottlenecks of traditional methods in spatiotemporal resolution and environmental adaptability, and provides low-cost and efficient water quality management data support.
Smart Images
Figure CN120578950A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hydrological remote sensing technology, and in particular to a remote sensing estimation method for dissolved oxygen concentration in aquaculture water based on measured data and machine learning. Background Art
[0002] Since the 1990s, China has gradually become a major aquaculture producer, reaching the highest global production in 2021, accounting for 57% of global production (FAO-Food and Nations 2022). Pond aquaculture is one of the most important aquaculture methods, accounting for approximately 47% of all aquaculture production in China (Bureau of Fisheries, Ministry of Agriculture and Rural Affairs et al. 2021); globally, pond aquaculture accounts for approximately 38.1% of the global aquaculture area (Xu et al. 2024). With increasing stocking density and outdated technical and management methods, pond aquaculture faces increasingly prominent environmental and product quality issues. The pursuit of economic benefits, coupled with a lack of the necessary technology and management experience for intensive aquaculture, has led to deteriorating water quality, frequent diseases, and declining production quality in pond aquaculture, posing significant challenges to the development of the aquaculture industry (Liu et al. 2021). Dissolved oxygen is a key indicator of water health, particularly in aquaculture environments, where its level directly affects the growth rate and health of aquatic products. Studies have shown that only when dissolved oxygen concentration is maintained within an appropriate range can the healthy growth of aquatic products be promoted (Summerfelt 2000; Boyd and Tucker 2012; Boyd 2017). Too low dissolved oxygen concentration will seriously affect water quality and the sustainable development of aquaculture (Mallya 2007).
[0003] The concentration of dissolved oxygen in water is influenced by a combination of physical, chemical, and biological factors. Physical and chemical factors such as temperature and salinity directly affect the solubility of oxygen in water. Increased water temperature reduces oxygen solubility, while high salinity further reduces the water's saturated oxygen solubility capacity (Sherwood et al. 1991). Furthermore, electrical conductivity (EC) and total dissolved solids (TDS) are closely related to dissolved oxygen, as they reflect the concentration of dissolved ions in water and influence oxygen transport and dissolution. pH indirectly influences dissolved oxygen concentrations by influencing carbonate systems and biological activity. Wind speed and atmospheric pressure are physical factors that influence oxygen exchange between water and the atmosphere. Photosynthesis by aquatic plants and respiration by aquatic animals are important biological regulators of dissolved oxygen concentration. Photosynthesis by aquatic plants, particularly algae and phytoplankton, is a significant source of dissolved oxygen in water (Hargreaves and Tucker 2002). Chlorophyll-a (Chl-a) is a measure of algae and phytoplankton photosynthesis, and changes in its concentration can significantly affect water oxygen levels. Light intensity (unit: Lumen) can partially represent the intensity of photosynthesis. Nitrate, nitrite, and ammonia nitrogen concentrations in the nitrogen cycle also affect dissolved oxygen. Excessive ammonia nitrogen and nitrate can lead to excessive algal growth, causing eutrophication, which indirectly affects dissolved oxygen levels (Qiu et al. 2024). Furthermore, suspended particulate matter (TSM) can affect algal photosynthesis by blocking light, thereby altering dissolved oxygen levels. Fluctuations in dissolved oxygen concentrations are particularly pronounced in aquaculture ponds. Due to high stocking densities, the decomposition of large amounts of organic matter at the bottom of the ponds and the vigorous respiration of aquacultured organisms continuously deplete dissolved oxygen in the water. Therefore, artificial interventions, such as the use of aeration equipment and regulation of water circulation systems, have become key measures to maintain dissolved oxygen levels. The diurnal fluctuations in dissolved oxygen concentrations in fish ponds are much greater than those in natural water bodies. This not only affects the growth of aquatic organisms but can also lead to eutrophication, further damaging aquatic ecosystems (Oberle et al. 2019). Traditional methods for monitoring dissolved oxygen concentrations rely primarily on field sampling. While accurate, these methods are often time-consuming, costly, and difficult to implement for large-scale monitoring. Furthermore, traditional monitoring methods generally only capture data from specific sites, lacking a comprehensive understanding of temporal and spatial variations across the entire water body (Topcu and Brockmann 2015).
[0004] In recent years, with the rapid development of remote sensing technology, water quality parameter inversion based on remote sensing imagery has become a more convenient and rapid method for monitoring dissolved oxygen. This method has been tried for environmental monitoring in open waters of varying scales and for monitoring natural water bodies such as lakes and rivers. However, attempts in fish ponds are still in the early stages. In a previous study, Cui Wenjun et al. (2017) performed first-order differential processing on spectral data measured at the Pearl River Estuary and conducted correlation analysis with simultaneously observed dissolved oxygen data. They found that the third and fourth bands of Landsat 8 OLI data showed the strongest correlation with dissolved oxygen concentrations. Based on this, they established a linear inversion model for dissolved oxygen concentrations in the Pearl River Estuary's coastal waters. However, using only simple linear methods for modeling is ineffective. Furthermore, because dissolved oxygen is a non-optically active factor, there is no direct physical process linking spectral information to changes in dissolved oxygen concentration (Gholizadeh et al. 2016). Therefore, some researchers have incorporated other water quality information, such as water temperature, into the model to improve inversion accuracy. Dong et al. (2024) used the MLR method to model dissolved oxygen concentrations in the Zhejiang Sea, incorporating water temperature information, enhancing the model's accuracy. Kim et al. (2020) employed a stepwise multiple regression approach, incorporating not only water temperature but also chlorophyll concentration, which is closely related to dissolved oxygen concentration, to predict dissolved oxygen concentrations in the Yellow Sea within the Korean border.
[0005] With the development of remote sensing technology and machine learning methods, researchers have leveraged remote sensing data to further improve the predictive performance of machine learning models. Salas et al. (2022) used Sentinel-2 satellite data to model dissolved oxygen concentrations in the Little Miami River using the Support Vector Machine (SVM) and Random Forest (RF) algorithms, respectively, achieving excellent prediction results. Andromachi Chatziantoniou et al. (2022) used CMEMS data and the SVR algorithm to develop a daily dissolved oxygen concentration estimation model for the Agrilia Ocean Ranch. Guo et al. (2021) used Landsat and MODIS satellite data and, based on the SVR algorithm, introduced water temperature as an input feature to invert dissolved oxygen concentrations for four inland lakes, including Lake Huron. They derived seasonal trends in dissolved oxygen concentrations in Lake Huron with good generalizability. However, for aquaculture ponds, the smaller water area and strong human disturbances make the inversion of dissolved oxygen concentrations more complex. Moreover, the factors affecting dissolved oxygen concentration in different types of fish ponds are different.
[0006] Given the complexity of dissolved oxygen monitoring in fish ponds, in-depth exploration of water spectral information is an important method for improving the accuracy of dissolved oxygen prediction models. The water reflectance spectrum contains a wealth of information about water quality parameters, but due to the dynamic changes in the aquaculture environment and the interplay of multiple factors, traditional spectral analysis has difficulty accurately extracting effective features.
[0007] Existing dissolved oxygen monitoring technology suffers from three core flaws: First, traditional field sampling methods rely on fixed-site data collection, which has limited temporal and spatial coverage and is costly, making it difficult to capture the dynamic patterns of dissolved oxygen changes in small water bodies and under strong artificial disturbances in fish ponds. Second, feature extraction methods based on spectral analysis struggle to separate effective features directly related to dissolved oxygen concentration from nonlinear and non-stationary water reflectance spectra due to dynamic changes in the aquaculture environment (such as water temperature fluctuations and interference from bait placement) and the coupling effects of multiple factors (such as algae growth and fish activity), resulting in insufficient model accuracy. Third, although empirical mode decomposition has demonstrated the advantages of adaptive signal processing in fields such as turbulence analysis and mechanical monitoring, its application in water spectral analysis still faces technical gaps, especially in the lack of mature solutions for spectral noise suppression and feature reconstruction in the complex environments of fish ponds. The root cause of these problems lies in the lack of adaptability of traditional technologies to dynamic disturbances and the lack of cross-domain technology integration. Summary of the Invention
[0008] The purpose of the present invention is to provide a remote sensing estimation method for dissolved oxygen concentration in aquaculture water based on measured data and machine learning, so as to solve the technical problems existing in the background technology.
[0009] By combining empirical mode decomposition with machine learning algorithms, and through adaptive decomposition and feature reconstruction of water spectral signals, a high-precision prediction model for dissolved oxygen concentration in fish ponds in the Greater Bay Area has been established, enabling large-scale monitoring of dissolved oxygen concentrations in fish ponds. Furthermore, based on measured data, an in-depth analysis of the spatiotemporal distribution of dissolved oxygen concentrations in fish ponds and their key influencing factors is conducted, providing a basis for regional-scale aquaculture water quality monitoring and scientific management.
[0010] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0011] like Figure 1-8 As shown, a remote sensing estimation method for dissolved oxygen concentration in aquaculture water based on measured data and machine learning comprises the following steps:
[0012] Step 1: Establish a variable dataset for dissolved oxygen modeling, which includes remote sensing reflectance data and empirical mode decomposition data;
[0013] Step 2: Construct three machine learning models for estimating dissolved oxygen concentration based on RF, SVR, and XGBoost, and train them on the variable dataset and the measured dissolved oxygen concentration data on the day of the satellite pass;
[0014] Step 3: Compare and evaluate the prediction accuracy of the RF model, the SVR dissolved oxygen concentration estimation machine learning model, and the XGBoost dissolved oxygen concentration estimation machine learning model based on the R2, RMSE, and MAPE indices. Select the model with the highest accuracy as the final concentration estimation model, and make predictions on the measured dissolved oxygen concentration data and variable data sets obtained on the day of the satellite transit.
[0015] Furthermore, in step 1, the remote sensing reflectivity data is the remote sensing reflectivity data of the relevant bands after the remote sensing data is corrected by Rayleigh scattering. The remote sensing reflectivity data includes reflectivity data of the bands of 443nm, 482nm, 561nm, 655nm, 865nm, 1609nm and 2201nm. Then, the reflectivity of each pixel of the remote sensing reflectivity data is decomposed by empirical mode decomposition. By identifying the local extreme points of the signal sequence, the upper and lower envelopes are constructed using linear interpolation, the envelope mean is calculated, and the intrinsic mode function is extracted from the original signal. Components are calculated repeatedly until the termination condition is met. Then, the energy contribution rate of each intrinsic mode function component is calculated to evaluate its importance. A 3×3 window is established to extract the local statistical features of the intrinsic mode function, and a feature validity check mechanism is established. The number of valid non-zero pixels in the window is greater than 5. At the same time, the residual term features are analyzed. Finally, the features of all bands are combined to form a unified feature matrix. The decomposed features are standardized for subsequent model establishment. Through adaptive decomposition and multi-dimensional feature extraction, effective characterization of multispectral reflectance data is achieved.
[0016] Furthermore, in step 1, the remote sensing data used Landsat OLI level-1 image data. In SeaDAS8.4.0, atmospheric correction was performed on the remote sensing image level 1 product to achieve Rayleigh-corrected reflectance. After correction, the image showed excellent results in identifying water bodies. The reflectance data of the first six bands of Landsat OLI were subtracted to 2201 nm, and some aerosol signals in the reflectance data were removed. The quality band of the Landsat satellite image was used as a reference to exclude pixels affected by clouds.
[0017] Furthermore, in step 1, the remote sensing data is preprocessed by performing radiometric calibration, atmospheric correction, and image mosaicking on the images covering the study area. Then, the boundary information of the aquaculture ponds in the study area is extracted by combining the SVM object-oriented classification and visual interpretation methods, and the normalized difference water index is used as a threshold filter to extract the area water body.
[0018] Furthermore, in step 1, the remote sensing data were first screened, including: 1. excluding any sampling points covered by clouds in the RGB image; 2. adding concentration samples outside the study area that were higher than the set value; 3. screening out valid sample points based on conditions such as image transit time and absence of cloud cover; 4. When obtaining reflectance data, the reflectance value was extracted from a 3 × 3 pixel window centered on each sampling point and averaged to ensure consistency among surrounding pixels.
[0019] Furthermore, the RF model uses the features after empirical mode decomposition as input variables, including four intrinsic mode functions, one residual, and the corresponding local mean and local standard deviation. The RF model achieves model robustness and prevents overfitting by adjusting hyperparameters such as the number of decision trees, the maximum depth of the tree, and the number of features considered at each split. The GridSearch method in Python is used in combination with 5-fold cross-validation to finally determine a combination of 200 trees, a maximum tree depth of 10, and tuned parameters.
[0020] Furthermore, in the model training data in step 2, 5-fold cross validation is used to ensure that the training dataset is randomly divided into different parts, and 70% of the data is selected as the model training dataset and 30% as the validation dataset.
[0021] The core purpose of the present invention is to solve the two major problems of dynamic spectral feature extraction and regional adaptability modeling through the innovative combination of empirical mode decomposition and machine learning algorithms: on the one hand, empirical mode decomposition is used to perform multi-scale decomposition (IMF extraction) of water spectral signals, remove noise interference and reconstruct key dissolved oxygen response characteristics; on the other hand, the decomposed characteristics are deeply correlated with regional environmental parameters (fish pond type, disturbance intensity) through machine learning models to construct a prediction model that takes into account both high precision and large-scale monitoring capabilities. Ultimately, it breaks through the technical bottlenecks of traditional methods in terms of spatiotemporal resolution, environmental adaptability and prediction reliability, and provides a low-cost, wide-area real-time monitoring method for aquaculture water quality management.
[0022] The present invention has the following beneficial effects due to the adoption of the above technical solution:
[0023] (1) This paper combines empirical mode decomposition with a machine learning algorithm. By adaptively decomposing and reconstructing the characteristics of water spectral signals, it establishes a high-precision prediction model for dissolved oxygen concentration in fish ponds, enabling large-scale monitoring of dissolved oxygen concentration in fish ponds. Simultaneously, based on measured data, it conducts an in-depth analysis of the spatiotemporal distribution characteristics of dissolved oxygen concentration in fish ponds and its key influencing factors, providing a basis for regional-scale aquaculture water quality monitoring and scientific management.
[0024] (2) Compared with traditional dissolved oxygen monitoring technology, the core advantage of the present invention is that it systematically solves the problems of insufficient spatiotemporal coverage, poor environmental adaptability and low prediction accuracy of dissolved oxygen monitoring in fish ponds through technology integration and signal processing innovation. Specifically, the traditional method relies on on-site sampling at specific sites, with limited spatiotemporal resolution and high cost. The present invention creatively combines empirical mode decomposition (EMD) with machine learning algorithms. First, empirical mode decomposition is used to adaptively decompose nonlinear and non-stationary water body reflectance spectral signals, extract multi-scale intrinsic mode functions (IMFs), effectively separate noise interference and reconstruct key spectral features that are highly correlated with dissolved oxygen concentration; on this basis, the decomposed features are dynamically modeled and trained through machine learning models such as neural networks and support vector machines to construct a high-precision prediction model suitable for small-area fish ponds in the Greater Bay Area, strong artificial interference and complex environments with multiple factors coupling. The core innovation of this technology is reflected in the following aspects: on the one hand, the empirical mode decomposition signal decomposition technology breaks through the bottleneck of feature extraction of traditional spectral analysis in dynamic aquaculture water bodies, and significantly improves the model's ability to analyze the implicit dissolved oxygen information in spectral signals; on the other hand, through the deep adaptation and optimization of machine learning algorithms and regional environmental parameters (such as fish pond types, water body disturbance characteristics), the model has both the scalability of large-scale monitoring and regional specificity, realizing the leap from single-point sampling to regional-scale continuous monitoring. From the perspective of protection scope, the core intellectual property rights of the present invention cover the entire process technology system from empirical mode decomposition of spectral signals, feature screening and reconstruction to dynamic model training, especially the first application of empirical mode decomposition technology to the field of aquaculture water quality monitoring, and the development of a special model parameter optimization method for the unique environment of fish ponds in the Greater Bay Area. The final effect is not only to significantly improve the prediction accuracy of dissolved oxygen concentration, but also to provide real-time and reliable data support for scientific decision-making in aquaculture water quality management through low-cost, high-efficiency wide-area monitoring capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is the location map of the sampling points of the model of the present invention and the location map of the fish ponds actually measured;
[0026] Figure 2 This is the main flow chart of the present invention using EMD and RF models to predict dissolved oxygen concentration;
[0027] Figure 3 The present invention uses the validation dataset to evaluate the performance of the three models, RF, XGBoost, and SVR, in estimating dissolved oxygen concentration. The evaluation shows that the RF model achieves the highest performance graph;
[0028] Figure 4 This is a diagram showing the verification of the model using measured data from fish ponds;
[0029] Figure 5This is the annual change chart of dissolved oxygen concentration in fish ponds in the Greater Bay Area from 2013 to 2023;
[0030] Figure 6 This is the spatial distribution map of dissolved oxygen concentration in fish ponds in the Greater Bay Area in different seasons of the present invention;
[0031] Figure 7 This is a seasonal variation diagram of dissolved oxygen concentration in fish ponds in the Greater Bay Area of the present invention;
[0032] Figure 8 It is a spatial distribution diagram of the dissolved oxygen concentration trend test results of the fish farming pond of the present invention. DETAILED DESCRIPTION
[0033] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.
[0034] A method for estimating dissolved oxygen concentration in aquaculture water by remote sensing based on measured data and machine learning, comprising the following steps:
[0035] Step 1. The present invention is modeled based on the water quality monitoring data of the coastal waters of Guangdong Province from 2020 to 2022, which is derived from the official website of the Guangdong Provincial Department of Ecology and Environment. From 1,962 original monitoring records, 64 valid sample points were screened out based on the conditions such as the Landsat image transit time (within one day before and after) and no cloud cover. Taking into account the prediction needs of the model in the high-value range of dissolved oxygen, 10 high-dissolved oxygen concentration samples from Lake Huron were additionally introduced. A training data set containing 74 samples was constructed. In order to test the accuracy of the established model, this study selected five fish ponds each in Foshan, Huizhou and Zhongshan for field water quality monitoring, and one fish pond in the Hong Kong Special Administrative Region, a total of 16 fish ponds, and the collection time was once in spring, summer, autumn and winter. Among them, Foshan, Huizhou and Zhongshan each selected a fish pond for 48 hours of continuous observation, with an observation interval of two hours, and the remaining four fish ponds were observed once each season; the Hong Kong Special Administrative Region conducted observations once a month. On-site sampling and testing of the surface water (0-15 cm) of the aquaculture ponds was performed using a WTWMulti 3320 instrument. Water quality parameters such as temperature, dissolved oxygen, EC, and salinity were obtained. An anemometer was also used to record environmental factors such as wind speed, air temperature, and air pressure (Fig. 4). Chl-a concentrations were measured by filtering samples through Whatman GF / F (25 mm diameter, 0.47 μm pore size) filters, which were stored at low temperatures before being returned to the laboratory. Chl-a concentrations were determined in the laboratory using a Turner Designs fluorimeter. TSM concentrations were measured by filtering samples through cellulose acetate membranes (47 mm diameter, 0.7 μm pore size) weighed in the laboratory, which were then frozen and tested upon return to the laboratory.
[0036] Step 2. Since the response of dissolved oxygen concentration changes in the spectrum is small, conventional feature processing methods are difficult to extract features that are strongly correlated with dissolved oxygen concentration changes from the spectrum. Therefore, this study uses a method based on empirical mode decomposition to extract and analyze multispectral reflectance data. First, the reflectance data of 7 bands (443nm, 482nm, 561nm, 655nm, 865nm, 1609nm and 2201nm) of Landsat OLI satellite data are preprocessed (atmospheric correction, water body extraction and cloud mask establishment). Secondly, the reflectance of each pixel is decomposed by empirical mode decomposition. By identifying the local extreme points of the signal sequence, the upper and lower envelopes are constructed using linear interpolation, the mean of the envelope is calculated and the intrinsic mode function components are extracted from the original signal, and this process is repeated until the termination condition is met. Next, the energy contribution of each intrinsic mode function component was calculated to assess its importance. A 3×3 window was established to extract the local statistical features of the intrinsic mode function. A feature validity check mechanism was established (the number of valid non-zero pixels in the window was greater than 5), and the residual term characteristics were analyzed. Finally, the features of all bands were combined to construct a unified feature matrix, and the decomposed features were standardized for subsequent model construction. This method, through adaptive decomposition and multidimensional feature extraction, achieves effective characterization of multispectral reflectance data.
[0037] Step 3: A total of 74 matching data were randomly divided into a training set (n = 51) and a validation set (n = 23). The RF model used features derived from empirical mode decomposition as input variables, including four intrinsic mode functions (IMFs 1-4), one residual, and the corresponding local mean and local standard deviation. The RF model primarily adjusts hyperparameters such as the number of decision trees, the maximum tree depth, and the number of features considered at each split to achieve model robustness and prevent overfitting. We used the GridSearch method in Python, combined with 5-fold cross-validation, to ultimately determine a set of 200 trees, a maximum tree depth of 10, and other optimized parameter combinations. Before model training, to remove some aerosol signals from the Rrc data, the first six bands of the Landsat OLI were subtracted from Rrc (2201). Furthermore, all input features were standardized by removing the mean and scaling to unit variance to improve model fitting stability. Through this series of steps, the RF model can effectively invert the dissolved oxygen concentration in fish ponds in the Greater Bay Area, providing reliable technical support for subsequent water quality monitoring and environmental management.
[0038] In addition, we also constructed two other dissolved oxygen concentration estimation models based on SVR dissolved oxygen concentration estimation model and XGBoost to compare the performance of RF model.
[0039] Step 5: Dissolved oxygen concentrations in fish ponds in the Greater Bay Area are concentrated in the range of 6.80-8.45 mg / L, with an average of 7.44 mg / L. Areas with higher dissolved oxygen concentrations are primarily located in the central part of the study area, including Zhaoqing, western Foshan, northern Jiangmen, and Huizhou. Lower concentrations are primarily located in the outer reaches of the Greater Bay Area, in southern Jiangmen, southern Guangzhou, Zhongshan, and eastern Foshan. Overall, the trend decreases from the periphery to the center. The average annual dissolved oxygen concentration in China's waters is naturally between 7-9 mg / L, but aquaculture activities and management measures have resulted in slightly lower dissolved oxygen concentrations in fish ponds within the Greater Bay Area. Analysis of the inversion results from 2013 to 2023 indicates that, despite fluctuations, the average annual dissolved oxygen concentration has generally remained relatively stable, increasing by 0.8% over the past decade. Specifically, from 2013 to 2014, dissolved oxygen concentrations showed a slight upward trend, increasing by 1.6%. In the following period of 2014-2016, the dissolved oxygen concentration experienced a continuous decline, with a cumulative decrease of 2.2%. Between 2016 and 2018, the dissolved oxygen concentration changed significantly, first experiencing a 2.7% increase, and then a decrease of 1.3%, showing a slight fluctuation feature. During the period of 2019-2023, the dissolved oxygen concentration entered a relatively stable stage, with only slight fluctuations. At the same time, this study found that the dissolved oxygen concentration in aquaculture ponds showed a unique seasonal variation pattern: the results showed that the dissolved oxygen concentration reached its highest value in summer (7.5 mg / L) and dropped to its lowest value in winter (7.16 mg / L). In terms of spatial distribution, in spring and autumn, the dissolved oxygen concentration in the western part of the Greater Bay Area (Zhaoqing, Jiangmen and western Foshan) was significantly higher than that in the eastern region. It is worth noting that aquaculture ponds in Zhuhai and Zhongshan showed lower dissolved oxygen concentrations in all four seasons of the year. In contrast, the seasonal characteristics of the dissolved oxygen concentration in natural water bodies are more obvious. Due to the strong influence of artificial aquaculture activities, dissolved oxygen concentrations in fish ponds experience minimal seasonal fluctuations, with concentrations typically lower in winter and higher in summer. Numerous studies have confirmed that temperature is a key factor influencing seasonal variations in dissolved oxygen concentrations in natural waters. A significant statistical correlation exists between dissolved oxygen concentration and water temperature, indicating a stable negative correlation over long timescales. However, the significant impact of aquaculture activities on the seasonal variation of dissolved oxygen concentrations in water bodies deviates from the normal variation patterns in natural water bodies.
[0040] To further analyze the changing trends in dissolved oxygen concentrations in fish ponds in the Greater Bay Area, this study conducted a MK trend test on fish ponds in the Greater Bay Area. The study divided the Greater Bay Area into ten regions based on administrative divisions: Dongguan, Foshan, Guangzhou, the Hong Kong Special Administrative Region, Huizhou, Jiangmen, Shenzhen, Zhaoqing, Zhongshan, and Zhuhai. (Macao, due to its limited number of fish ponds, was not included in the study.) The average Z-score for all ten regions was less than 1.69, indicating that changes in dissolved oxygen concentrations in fish ponds were not significant in any region.
[0041] However, from the trend test results, it can be seen that the dissolved oxygen concentration in some fish ponds shows a clear upward trend, such as in the eastern part of Jiangmen and Zhaoqing; while the fish ponds with a clear downward trend in dissolved oxygen concentration are scattered in various regions without aggregation. Although there is no significant trend of change in the dissolved oxygen concentration in the fish ponds in the Greater Bay Area in the MK test, the Sen's slope test results show that the dissolved oxygen concentration still shows an overall slow upward trend from 2013 to 2023 (β = 0.08). Among the ten regions, only Zhaoqing has an average β value less than 0, and the average β values of the remaining regions are greater than 0, with the maximum value occurring in Shenzhen (β = 0.239). The dissolved oxygen concentration in most areas is slowly rising, and only in a small number of areas is the dissolved oxygen concentration slowly falling. In general, the dissolved oxygen concentration in the Greater Bay Area is slowly rising as a whole, but the trend is not significant.
[0042] Compared with traditional dissolved oxygen monitoring technology, the core advantage of the present invention is that it systematically solves the problems of insufficient spatiotemporal coverage, poor environmental adaptability and low prediction accuracy of dissolved oxygen monitoring in fish ponds through technology integration and signal processing innovation. Specifically, the traditional method relies on on-site sampling at specific sites, with limited spatiotemporal resolution and high cost. The present invention creatively combines empirical mode decomposition (EMD) with machine learning algorithms. First, empirical mode decomposition is used to adaptively decompose nonlinear and non-stationary water reflectance spectral signals, extract multi-scale intrinsic mode functions (IMFs), effectively separate noise interference and reconstruct key spectral features that are highly correlated with dissolved oxygen concentration; on this basis, the decomposed features are dynamically modeled and trained through machine learning models such as neural networks and support vector machines to construct a high-precision prediction model suitable for small-area fish ponds in the Greater Bay Area, strong artificial interference and complex environments with multiple coupling factors. The core innovation of this technology is reflected in the following aspects: on the one hand, the empirical mode decomposition signal decomposition technology breaks through the bottleneck of feature extraction of traditional spectral analysis in dynamic aquaculture water bodies, and significantly improves the model's ability to analyze the implicit dissolved oxygen information in spectral signals; on the other hand, through the deep adaptation and optimization of machine learning algorithms and regional environmental parameters (such as fish pond types, water body disturbance characteristics), the model has both the scalability of large-scale monitoring and regional specificity, realizing the leap from single-point sampling to regional-scale continuous monitoring. From the perspective of protection scope, the core intellectual property rights of the present invention cover the entire process technology system from empirical mode decomposition of spectral signals, feature screening and reconstruction to dynamic model training, especially the first application of empirical mode decomposition technology to the field of aquaculture water quality monitoring, and the development of a special model parameter optimization method for the unique environment of fish ponds in the Greater Bay Area. The final effect is not only to significantly improve the prediction accuracy of dissolved oxygen concentration, but also to provide real-time and reliable data support for scientific decision-making in aquaculture water quality management through low-cost, high-efficiency wide-area monitoring capabilities.
[0043] Matters not covered by the present invention are known technologies.
[0044] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A remote sensing method for estimating dissolved oxygen concentration in aquaculture water based on measured data and machine learning, characterized by: The method comprises the following steps: Step 1: Establish a variable dataset for dissolved oxygen modeling, which includes remote sensing reflectance data and empirical mode decomposition data; Step 2: Construct three machine learning models for estimating dissolved oxygen concentration based on RF, SVR, and XGBoost, and train them on the variable dataset and the measured dissolved oxygen concentration data on the day of the satellite pass; Step 3: Compare and evaluate the prediction accuracy of the RF model, the SVR dissolved oxygen concentration estimation machine learning model, and the XGBoost dissolved oxygen concentration estimation machine learning model based on the R2, RMSE, and MAPE indices. Select the model with the highest accuracy as the final concentration estimation model, and make predictions on the measured dissolved oxygen concentration data and variable data sets obtained on the day of the satellite transit.
2. The method for remote sensing estimation of dissolved oxygen concentration in aquaculture water based on measured data and machine learning according to claim 1, characterized in that: In step 1, the remote sensing reflectance data is the remote sensing reflectance data of the relevant bands after the remote sensing data is corrected by Rayleigh scattering. The remote sensing reflectance data includes reflectance data of the bands of 443nm, 482nm, 561nm, 655nm, 865nm, 1609nm and 2201nm. Then the reflectance of each pixel of the remote sensing reflectance data is decomposed by empirical mode decomposition. By identifying the local extreme points of the signal sequence, the upper and lower envelopes are constructed using linear interpolation, the mean of the envelope is calculated, and the intrinsic mode function components are extracted from the original signal. The calculation is repeated until the termination condition is met, and then the energy contribution rate of each intrinsic mode function component is calculated to evaluate its importance. A 3×3 window is established to extract the local statistical features of the intrinsic mode function, and a feature validity check mechanism is established. The number of valid non-zero pixels in the window is greater than 5. At the same time, the residual term features are analyzed. Finally, the features of all bands are combined to construct a unified feature matrix, and the decomposed features are standardized for subsequent model establishment. Through adaptive decomposition and multi-dimensional feature extraction, effective characterization of multispectral reflectance data is achieved.
3. The method for remote sensing estimation of dissolved oxygen concentration in aquaculture water based on measured data and machine learning according to claim 1, characterized in that: In step 1, Landsat OLI level-1 image data was used as remote sensing data. In SeaDAS8.4.0, atmospheric correction was performed on the first-level remote sensing image products to achieve Rayleigh-corrected reflectance. After correction, the image showed excellent results in identifying water bodies. The reflectance data of the first six bands of Landsat OLI were subtracted to 2201 nm, and some aerosol signals in the RRC data were removed. The quality band of the Landsat satellite image was used as a reference to exclude pixels affected by clouds.
4. The method for remote sensing estimation of dissolved oxygen concentration in aquaculture water based on measured data and machine learning according to claim 1, characterized in that: In step 1, the remote sensing data is preprocessed by performing radiometric calibration, atmospheric correction, and image mosaicking on the images covering the study area. Then, the boundary information of the aquaculture ponds in the study area is extracted by combining the SVM object-oriented classification and visual interpretation methods, and the normalized difference water index is used as a threshold filter to extract the area water body.
5. The method for remote sensing estimation of dissolved oxygen concentration in aquaculture water based on measured data and machine learning according to claim 1, characterized in that: In step 1, the remote sensing data are first screened, including:
1. Excluding any sampling points covered by clouds in the RGB image; 2. Adding concentration samples outside the study area that are higher than the set value; 3. Screening out valid sample points based on conditions such as image transit time and no cloud cover; 4. When obtaining Rrc data, extracting the reflectance value in a 3×3 pixel window centered on each sampling point and averaging it to ensure consistency among surrounding pixels.
6. The method for estimating dissolved oxygen concentration in aquaculture water by remote sensing based on measured data and machine learning according to claim 1, characterized in that: The RF dissolved oxygen concentration estimation machine learning model uses the features decomposed by empirical mode decomposition as input variables, including four intrinsic mode functions, one residual, and the corresponding local mean and local standard deviation. The RF model achieves model robustness and prevents overfitting by adjusting hyperparameters such as the number of decision trees, the maximum tree depth, and the number of features considered at each split. The GridSearch method in Python, combined with 5-fold cross-validation, ultimately determined a combination of 200 trees, a maximum tree depth of 10, and tuned parameters.
7. The method for estimating dissolved oxygen concentration in aquaculture water by remote sensing based on measured data and machine learning according to claim 1, characterized in that: In the model training data in step 2, 5-fold cross validation is used to ensure that the training dataset is randomly divided into different parts, and 70% of the data is selected as the model training dataset and 30% as the validation dataset.