Cadmium and arsenic combined pollution plant response dynamic monitoring system based on time sequence image analysis

Abstract

The application discloses a cadmium-arsenic compound pollution plant response dynamic monitoring system based on time sequence image analysis, and relates to the technical fields of environmental monitoring and image analysis.The system comprises a multi-time-phase image acquisition module, a self-adaptive registration engine, a response feature decoupling extraction module, an illumination environment normalization module and a pollution early warning decision module, can eliminate visual angle deviation through time sequence image accurate registration, can separate pollution response and natural growth through response feature decoupling, can exclude non-pollution interference through illumination environment normalization, and can realize early warning through pollution early warning decision.The system can identify pollution signals when the soil cadmium-arsenic concentration is less than 50% of the national standard limit value, the monitoring accuracy is more than 95%, and the pollution risk can be found 3-5 weeks in advance compared with traditional methods, so that the system provides an effective technical means for soil heavy metal pollution prevention and control.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and image analysis technology, and in particular to a dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis. It falls under the category of non-invasive monitoring technology for soil heavy metal pollution and is specifically applied to the early identification and dynamic assessment of cadmium and arsenic combined pollution through temporal changes in plant phenotypic characteristics. Background Technology

[0002] Heavy metal pollution has become a global environmental problem, especially the combined pollution of cadmium and arsenic, which poses a serious threat to agricultural production and human health. According to research data published in Science in 2024, 14%–17% of farmland worldwide is affected by heavy metal pollution, including more than 2 million hectares of cadmium-contaminated farmland and more than 1.5 million hectares of arsenic-contaminated farmland.

[0003] Plants, as sensitive indicator organisms of the soil environment, can reflect the level of heavy metal pollution in soil through their growth status, morphological characteristics, and physiological changes. Traditional methods for detecting heavy metals in soil, such as atomic absorption spectrometry and inductively coupled plasma mass spectrometry, while offering high measurement accuracy, suffer from drawbacks such as highly destructive sampling, long detection cycles, and high costs, making it difficult to achieve large-scale, real-time dynamic monitoring. Indirect monitoring methods based on plant phenotypes, due to their non-invasiveness, continuity, and cost-effectiveness, are gradually becoming a research hotspot.

[0004] The prior art CN111862194A discloses a method and system for analyzing plant growth models based on computer vision and deep learning. This technology collects plant growth environment data and plant images at different growth cycles, uses deep learning algorithms to calculate the leaf area of ​​plants, and establishes a mapping relationship model between plant growth status and environmental factors. However, this technology has the following shortcomings: (1) It lacks a targeted heavy metal pollution monitoring mechanism, only focuses on general plant growth status monitoring, and cannot effectively distinguish between pollution stress response and natural growth changes; (2) It does not consider environmental interference factors such as light changes and viewing angle shifts during the acquisition of time-series images, resulting in a lack of reliability in the comparison of plant features between images at different time phases; (3) It adopts a simple leaf area statistical method, ignoring the multidimensional phenotypic response characteristics of plants under heavy metal stress, such as leaf color changes, morphological distortions, abnormal growth rates, and other comprehensive manifestations; (4) It lacks early warning capabilities, and can only be identified after plants show obvious symptoms of damage, at which time soil pollution has often reached a high level, missing the best time for prevention and control.

[0005] Therefore, there is an urgent need to develop an intelligent system that can accurately capture the early response characteristics of plants to cadmium and arsenic combined pollution, effectively eliminate environmental interference factors, and realize dynamic monitoring and early warning of pollution. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a dynamic monitoring system for plant responses to cadmium and arsenic compound pollution based on time-series image analysis. It aims to solve the technical problems of existing plant phenotypic monitoring technologies in heavy metal pollution identification, such as insufficient targeting, significant environmental interference, and weak early warning capabilities, and to achieve highly sensitive, dynamic, and non-invasive monitoring of cadmium and arsenic compound pollution in soil.

[0007] This invention achieves continuous temporal image acquisition of target plants through a multi-temporal image acquisition module, ensuring the temporal continuity and integrity of monitoring data. It achieves precise spatial alignment of images from different temporal phases through an adaptive registration engine, eliminating the influence of perspective changes and plant displacement, ensuring the accuracy of temporal comparisons. It effectively separates pollution response features from natural growth features through a response feature decoupling extraction module, highlighting phenotypic abnormalities caused by pollution stress. It dynamically corrects for light fluctuations and environmental factors through a light environment normalization module, eliminating interference from non-pollution factors. Finally, it enables quantitative assessment and early warning of pollution levels through a pollution early warning decision module, identifying pollution when concentrations are below 50% of national standard limits, significantly improving the timeliness and sensitivity of pollution monitoring. A tight data flow and feedback adjustment mechanism is formed among the modules. The registration accuracy of the adaptive registration engine directly affects the feature extraction quality of the response feature decoupling extraction module. The correction effect of the light environment normalization module guides the multi-temporal image acquisition module to optimize its acquisition strategy. The early warning results of the pollution early warning decision module promote the improvement of the feature decomposition algorithm in the response feature decoupling extraction module, achieving continuous optimization of system performance and continuous improvement of monitoring accuracy.

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] (1) Highly targeted pollution control. This invention is specifically designed for monitoring processes of cadmium and arsenic combined pollution. By using a response feature decoupling extraction module, the temporal changes in plant phenotypic characteristics are decomposed into pollution response components and natural growth components, effectively distinguishing between pollution stress and normal growth. This avoids misjudging non-pollution factors such as natural aging and nutrient deficiency as pollution responses, and the monitoring accuracy rate reaches over 95%.

[0010] (2) Good environmental adaptability. This invention dynamically corrects the effects of light intensity and ambient temperature and humidity through a light environment normalization module, and adopts multi-scale light decomposition technology based on Retinex theory. It can maintain the consistency of monitoring results under different time periods and weather conditions, and eliminate the 30%-40% error fluctuation caused by environmental changes in traditional methods.

[0011] (3) Sensitive early warning. The mapping relationship model constructed in this invention can identify pollution signals when there are slight changes in plant phenotypes, and achieve effective early warning when the soil cadmium and arsenic concentration is lower than the national standard limit by 50%. It can detect pollution risks 3-5 weeks earlier than traditional methods, and buy valuable time for pollution prevention and control.

[0012] (4) Accurate temporal analysis. This invention achieves sub-pixel-level accurate alignment of images at different time phases through an adaptive registration engine, with the registration error controlled within 0.5 pixels, ensuring the reliability of temporal comparison; at the same time, it uses empirical mode decomposition and wavelet denoising technology to extract the essential features of the contamination response, which has strong anti-interference ability.

[0013] (5) High system integration. The modules of this invention form a deeply coupled collaborative working mechanism. The output of the adaptive registration engine is directly used as the input of the response feature decoupling extraction module. The correction parameters of the illumination environment normalization module are fed back to the pollution early warning decision module in real time, forming a closed-loop optimization system. The overall performance is better than the simple superposition of a single module. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall structure of the dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis according to the present invention.

[0015] Figure 2 This is a schematic diagram of the processing flow of the adaptive registration engine of the present invention;

[0016] Figure 3 This is a functional structure diagram of the response feature decoupling extraction module of the present invention;

[0017] Figure 4 This is a schematic diagram illustrating the working principle of the illumination environment normalization module of the present invention;

[0018] Figure 5 This is a schematic diagram of the decision logic of the pollution early warning decision module of the present invention. Detailed Implementation

[0019] Please refer to the attached document. Figures 1-5 To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis proposed in accordance with the present invention.

[0020] Reference Figure 1 The dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis of the present invention includes a multi-temporal image acquisition module 1, an adaptive registration engine 2, a response feature decoupling extraction module 3, a light environment normalization module 4, and a pollution early warning decision module 5.

[0021] The multi-temporal image acquisition module 1 is used to continuously acquire RGB images of the target plant within a preset time series, forming a time-series image sequence. Specifically, an industrial-grade CCD camera or a high-resolution digital camera is fixedly installed above the plant growth area, with the camera lens pointing vertically downwards and the shooting distance maintained within the range of 0.5-1.5m to ensure that the plant canopy is fully visible. The image acquisition frequency is set according to the plant growth rate and monitoring needs. For slow-growing plants such as wheat and rice, the acquisition interval is 24 hours, that is, once a day at the same time (e.g., 10:00 AM, when the light is relatively stable). For faster-growing plants such as rapeseed and sunflower, the acquisition interval can be shortened to 12 hours. During each acquisition, the system automatically records environmental parameters such as the current timestamp, light intensity (measured by a light intensity sensor, in lux), ambient temperature (measured by a temperature sensor, in °C), and ambient humidity (measured by a humidity sensor, in %RH), and embeds these parameters as metadata tags into the EXIF ​​information of the image file. Within a complete monitoring cycle (usually 30-90 days), a time-series image sequence containing dozens to hundreds of images is formed.

[0022] In a preferred embodiment, the multi-temporal image acquisition module 1 further includes an image quality assessment unit. This unit performs sharpness detection on each acquired image frame. When the image sharpness is lower than a preset threshold (the variance of the image gradient magnitude is calculated using the Laplacian operator, and the threshold is set to 100), it is determined to be a blurry image, and the system automatically re-acquires the image. This avoids blurry images caused by camera focusing errors, lens contamination, or violent plant swaying from entering the subsequent processing flow, ensuring data quality.

[0023] The adaptive registration engine 2 is connected to the multi-temporal image acquisition module 1, receives the time-series image sequence, performs spatial alignment processing on each frame of the sequence, eliminates spatial position deviations between images caused by factors such as plant growth displacement, wind swaying, and slight camera vibration, and obtains the registered time-series image sequence.

[0024] Reference Figure 2 The registration process of Adaptive Registration Engine 2 is as follows:

[0025] First, the first frame of the time-series image is selected as the reference image, and subsequent frames are used as images to be registered. SIFT (Scale Invariant Feature Transform) feature points are extracted from both the reference and the images to be registered. SIFT feature points have good scale and rotation invariance, allowing for stable detection even when the image is scaled or rotated. After feature point extraction, the nearest neighbor distance ratio method is used for feature point matching: for each SIFT feature point in the image to be registered, the two closest Euclidean distances are found in the SIFT feature point set of the reference image, and the ratio of the nearest distance to the second nearest distance is calculated. If this ratio is less than 0.7, the feature point is considered a successful match; otherwise, it is considered a mismatch and is removed. After matching and removing mismatches, an initial set of feature point pairs is obtained.

[0026] Secondly, based on the initial set of feature point pairs, the RANSAC (Random Sample Consensus) algorithm is used to estimate the initial rigid transformation matrix. The basic idea of ​​the RANSAC algorithm is: randomly select 3 pairs of feature points from the initial set (planar rigid transformation requires at least 3 pairs of points to be determined), calculate a candidate rigid transformation matrix (including rotation and translation parameters) based on these 3 pairs, and then count how many feature point pairs have a transformation error less than a preset threshold (set to 2 pixels) under the action of this candidate matrix. These feature point pairs are considered inliers. This process is repeated 1000 times, and the candidate matrix with the most inliers is selected as the initial rigid transformation matrix. This matrix can describe the overall translation and rotation relationship between the reference image and the image to be registered, but it cannot handle the local non-rigid deformation caused by plant growth.

[0027] Next, the dense optical flow field between the reference image and the image to be registered is calculated. The optical flow field describes the motion vector of each pixel in the image and can capture non-rigid deformation information between images. This invention uses the Farneback dense optical flow algorithm, which is based on polynomial expansion theory. By calculating the optical flow at each layer of the image pyramid and passing it layer by layer, a dense optical flow field can be obtained efficiently. After the optical flow field is calculated, the non-rigid deformation regions are extracted: the optical flow field is spatially filtered, and the optical flow vector is decomposed into rigid motion components and non-rigid deformation components. When the amplitude of the non-rigid deformation component is greater than 1 pixel, the region is determined to be a non-rigid deformation region. These regions usually correspond to growth deformations such as the extension and bending of plant leaves.

[0028] Finally, based on the initial rigid transformation matrix and the non-rigid deformation regions, the final temporal registration transformation matrix is ​​generated using the thin-plate spline interpolation method. Thin-plate spline interpolation is an elastic registration method that can smoothly interpolate local non-rigid deformations while ensuring overall rigid transformation. Specifically, initial feature point pairs are used as control points. For control points within the non-rigid deformation regions, their target positions are determined by the optical flow field; for control points within the rigid regions, their target positions are determined by the initial rigid transformation matrix. Based on these control points and their target positions, a thin-plate spline energy minimization problem is constructed:

[0029]

[0030] in, For the reference image, the first Coordinates of control points The target location is the corresponding control point in the image to be registered. The total number of control points. For thin plate spline transformation function, Here, is the smoothing coefficient (preferably 0.1). The first term represents the control point matching error, and the second term represents the smoothness constraint of the transformation function. By solving the above energy minimization problem, the optimal thin-plate spline transformation function is obtained. This function is the final temporal registration transformation matrix. This matrix is ​​applied to transform the images to be registered, completing the spatial alignment. The above process is repeated for each frame of the temporal image sequence to be registered, ultimately obtaining a registered temporal image sequence where each frame is precisely aligned in spatial position, with the registration error controlled within 0.5 pixels.

[0031] In a preferred embodiment, the adaptive registration engine 2 further includes a registration quality detection unit. This unit calculates the cross-correlation coefficient between adjacent frames after registration. When the cross-correlation coefficient is lower than 0.85, registration is deemed to have failed, and the system prompts the user to re-acquire images or adjust the camera position. This ensures registration quality and prevents registration errors from being passed on to subsequent processing stages.

[0032] The response feature decoupling extraction module 3 is connected to the adaptive registration engine 2, receives the registered time-series image sequence, extracts the time-series change features of plant phenotypes, and decomposes them into pollution response components and natural growth components to achieve accurate identification of pollution stress response.

[0033] Reference Figure 3 The processing flow of response feature decoupling extraction module 3 is as follows:

[0034] First, a plant target mask is obtained through plant region segmentation. For each frame of the registered temporal image sequence, a threshold segmentation method based on color space is used to extract the plant region. Specifically, the RGB image is converted to the HSV color space. Plant leaves exhibit obvious green characteristics in the HSV space, with hue (H) values ​​concentrated in the range of 60-120, saturation (S) values ​​greater than 0.3, and lightness (V) values ​​in the range of 0.2-0.9. Based on these color features, a threshold is set, and the image is binarized to obtain the initial mask for the plant region. To remove small noise areas in the mask, morphological opening operations (erosion followed by dilation) are used. The structuring elements for the erosion and dilation operations are 5×5 circular structuring elements. The opening operation can remove small connected regions while retaining the main plant canopy regions. Connectivity analysis is performed on the mask after the opening operation, and the connected region with the largest area is selected as the final plant target mask.

[0035] Secondly, stem and leaf morphological parameters and color feature parameters are extracted based on the plant target mask. Stem and leaf morphological parameters include: leaf area (the number of pixels within the mask area multiplied by the actual area corresponding to a single pixel, in cm²). 2 The parameters include: leaf perimeter (calculated using the chain code method to determine the length of the mask boundary, in cm), leaf aspect ratio (calculated by fitting an ellipse to the mask area and calculating the ratio of the major axis to the minor axis of the fitted ellipse), leaf curvature (calculated using the curvature scale space method to determine the average curvature of the mask boundary), stem height (calculated using a skeleton extraction algorithm to extract the main stem skeleton of the plant and calculate the maximum longitudinal length of the skeleton, in cm), stem diameter (calculated at multiple locations on the skeleton and the average value is taken, in mm), and leaf tilt angle (determined by calculating the angle between the leaf's principal axis and the horizontal direction, in degrees). Color feature parameters include: average leaf hue H (calculated by statistically analyzing the H values ​​of all pixels within the mask area and calculating the average), average leaf saturation S (calculated by statistically analyzing the S values ​​of all pixels within the mask area and calculating the average), average leaf brightness V (calculated by statistically analyzing the V values ​​of all pixels within the mask area and calculating the average), leaf color variance (calculated by calculating the sum of the variances of the H, S, and V channel values ​​within the mask area), and chlorophyll index (calculated based on RGB values). Where G is the mean of the green channel and R is the mean of the red channel), and the Normalized Difference Vegetation Index (NDVI) is estimated based on RGB values. (Approximate calculation).

[0036] Next, a time-series curve of plant phenotypic changes was constructed. For each frame in the registered time-series image sequence, the aforementioned stem and leaf morphological parameters and color feature parameters were extracted and arranged in chronological order to form a curve showing the change of each feature parameter over time. Taking leaf area as an example, the horizontal axis represents time (days), and the vertical axis represents leaf area (cm²). 2Connecting the leaf area values ​​at each time point yields a time-series curve of leaf area variation. Similarly, time-series curves can be obtained for parameters such as leaf perimeter, leaf aspect ratio, leaf curvature, stem height, stem diameter, leaf inclination angle, average leaf hue, average leaf saturation, average leaf brightness, leaf color variance, chlorophyll index, and normalized difference vegetation index. These curves comprehensively reflect the dynamic evolution of plant phenotypes over time.

[0037] Finally, the plant phenotypic time-series change curve is decomposed into a pollution response component and a natural growth component. Plant phenotypic changes are influenced by multiple factors, including normal growth and development (natural growth component) and abnormal changes caused by environmental stresses such as heavy metal pollution (pollution response component). This invention employs Empirical Mode Decomposition (EMD) for signal decomposition. EMD is an adaptive time-frequency analysis method that can decompose complex non-stationary signals into several intrinsic mode function (IMF) components, each IMF component representing an oscillating component within a specific frequency range of the signal. The plant phenotypic time-series change curve is then analyzed. EMD decomposition was performed to obtain multiple IMF components. and a residual component ,Right now:

[0038]

[0039] in, For time, The number of IMF components, usually Between 5 and 10. Classified according to the frequency characteristics of each IMF component: High-frequency IMF components (the first 3 IMF components) reflect rapid fluctuations in plant phenotype, mainly caused by environmental stresses such as heavy metal pollution, and are classified as pollution response components. The low-frequency IMF component (subsequent IMF component) and residual component reflect the slow trend changes in plant phenotype, mainly caused by natural growth and development, and are classified as natural growth components. After decomposition, the pollution response components... Wavelet denoising was performed to remove high-frequency noise interference. A three-level wavelet decomposition using the Daubechies 4 wavelet was employed, and soft thresholding was used to denoise the high-frequency coefficients. An adaptive threshold selection method was used, based on the noise standard deviation. (Estimated using the median absolute deviation method) The threshold is determined. ,in The signal length is denoised. The contamination response components are then reconstructed to obtain the final contamination response characteristic curve.

[0040] Through the above processing, the response feature decoupling extraction module 3 accurately decomposes the complex temporal changes of plant phenotypes into two parts: pollution response and natural growth, providing high-quality feature data for subsequent pollution identification and early warning.

[0041] In a preferred embodiment, the response feature decoupling extraction module 3 further includes a multi-dimensional feature fusion unit. This unit performs weighted fusion of pollution response components of multiple phenotypic parameters (leaf area, chlorophyll index, leaf hue, etc.), determining the weights based on the sensitivity of each parameter to pollution. The fused comprehensive pollution response features can more comprehensively reflect the pollution stress state of plants and improve identification accuracy. Sensitivity is determined through historical data analysis: in experimental environments with known pollution concentration gradients, the variation range of each phenotypic parameter is statistically analyzed; parameters with larger variation ranges have higher sensitivity and correspondingly larger weights.

[0042] The light environment normalization module 4 is connected to the multi-temporal image acquisition module 1 and the response feature decoupling extraction module 3. It receives the environmental parameters (light intensity, ambient temperature and humidity) of the current time phase, normalizes and corrects the plant phenotypic characteristics, eliminates phenotypic changes caused by non-pollution factors, and ensures the reliability of the monitoring results.

[0043] Reference Figure 4 The workflow of the lighting environment normalization module 4 is as follows:

[0044] First, a plant phenotypic baseline database under reference light conditions is established. In a laboratory or greenhouse environment, standard light conditions are set (e.g., light intensity of 30,000 lux, color temperature of 5500 K, ambient temperature of 25℃, and ambient humidity of 60% RH). Healthy, uncontaminated plants are continuously monitored, and data on the changes in their phenotypic parameters (leaf area, chlorophyll index, leaf hue, etc.) under these standard conditions over time are collected to form the reference baseline database. This database describes the normal growth curves of plants under ideal environmental conditions and serves as a reference standard for subsequent normalization correction.

[0045] Secondly, the illumination intensity and ambient temperature and humidity values ​​for the current time phase are acquired. These environmental parameters have already been acquired in real time in the multi-temporal image acquisition module 1 and recorded in the image metadata, which can be directly read by the illumination environment normalization module 4.

[0046] Secondly, based on the illumination intensity value, the image is decomposed into reflection and illumination components using Retinex theory, and illumination correction coefficients are calculated according to reference illumination conditions. Retinex theory posits that an image can be decomposed into the product of a reflection component (the intrinsic color of an object, independent of illumination) and an illumination component (the influence of ambient light). This invention employs the Multi-Scale Retinex (MSR) algorithm for decomposition. For the current phase image... First, logarithmic transformations are performed on the three RGB channels respectively. Then, the illumination components are extracted using Gaussian filter banks of different scales. Reflection component Determined by the logarithmic domain difference between the image and the illumination component:

[0047]

[0048] in, For pixel coordinates, For the scale quantity (preferred) (corresponding to small scale, medium scale, and large scale). Weights for each scale (preferred) ), For the first Gaussian filters of various scales This represents the convolution operation. A Gaussian filter is defined as:

[0049]

[0050] in, For the first Gaussian standard deviation at each scale (preferably) Pixels Pixels (pixels). Through MSR decomposition, the reflectance component reflecting the true color of the plant is obtained. and the light component that reflects the influence of light. Compare the average illumination components at the current time phase. Average light component under reference illumination conditions (Extracted from the benchmark database), calculate the illumination correction factor:

[0051]

[0052] This coefficient reflects the degree of deviation between the current illumination and the reference illumination.

[0053] Next, based on the environmental temperature and humidity values, the plant phenotypic benchmark database is queried to calculate the environmental correction coefficient. Plant growth rate and phenotypic characteristics are significantly affected by environmental temperature and humidity. The current environmental temperature and humidity are queried from the benchmark database. The corresponding plant phenotypic parameter values And reference ambient temperature and humidity ( ℃, Plant phenotypic parameter values ​​corresponding to %RH Calculate the environmental correction factor:

[0054]

[0055] This coefficient reflects the degree to which changes in environmental temperature and humidity affect plant phenotype. If the current environment deviates little from the reference environment, then... If the deviation is close to 1, a larger correction is required.

[0056] Finally, the illumination correction factor and environmental correction factor Weighted fusion is performed to obtain the normalized correction coefficients:

[0057]

[0058] in, Weighting coefficients (preferably selected) (This indicates that the effect of light is relatively more significant). Normalized correction coefficients were applied to correct the stem and leaf morphological parameters and color feature parameters extracted by the response feature decoupling extraction module 3:

[0059]

[0060] in, The original phenotypic parameter values ​​obtained by measurement, These are the normalized phenotypic parameter values. Normalized parameters eliminate the influence of light and ambient temperature and humidity, more accurately reflecting the plant's true response to pollution.

[0061] Through the above normalization process, the light environment normalization module 4 effectively eliminates 30%-40% of environmental factor interference, ensuring that plant phenotypic data obtained under different time phases and environmental conditions are comparable, and significantly improving the reliability and accuracy of pollution monitoring.

[0062] In a preferred embodiment, the light environment normalization module 4 further includes a dynamic threshold adjustment unit. This unit dynamically adjusts various threshold parameters (such as the standard deviation of the Gaussian filter, weighting coefficients, etc.) in the normalization process based on statistical analysis of long-term monitoring data, adapting to the characteristics of different plant species and growth stages, and further improving the normalization effect.

[0063] The pollution early warning decision module 5 is connected to the response feature decoupling extraction module 3 and the light environment normalization module 4. It receives the normalized pollution response components, establishes a mapping relationship model between pollution response features and heavy metal concentration, calculates the current pollution level index, and generates a pollution early warning signal when the pollution level index exceeds the preset threshold, so as to realize the early identification and dynamic early warning of cadmium and arsenic compound pollution.

[0064] Reference Figure 5 The decision-making process of pollution early warning decision module 5 is as follows:

[0065] First, a mapping model between pollution response characteristics and heavy metal concentrations is established. In laboratory or field experimental environments, concentration gradients for cadmium and arsenic are set (e.g., cadmium concentrations of 0, 0.1, 0.3, 0.5, 1.0, and 2.0 mg / kg, and arsenic concentrations of 0, 5, 15, 30, 60, and 100 mg / kg). Target plants are planted at each concentration level, and their phenotypic response characteristics are continuously monitored. Simultaneously, soil samples are collected periodically for heavy metal concentration determination, forming a training sample set containing plant phenotypic data and corresponding soil heavy metal concentration data. For each training sample, a multidimensional pollution response feature vector is extracted. ,in Representing the Individual pollution response characteristics (such as leaf area change rate, chlorophyll index decrease, leaf hue shift, etc.). The feature dimension (preferred) (This corresponds to the pollution response component features of the 12 phenotypic parameters extracted from the response feature decoupling extraction module 3). The corresponding soil heavy metal concentration vector is... ,in Soil cadmium concentration, This represents the soil arsenic concentration.

[0066] A multidimensional pollution response feature vector is established using the Support Vector Regression (SVR) algorithm. With heavy metal concentration The nonlinear mapping relationship. The basic idea of ​​SVR is to find an optimal regression function. This minimizes the prediction error of the training samples while also minimizing the complexity of the function (achieved by maximizing the margin). Separate SVR models are built for cadmium and arsenic concentrations:

[0067]

[0068]

[0069] in, The number of training samples. These are Lagrange multipliers (determined by solving a quadratic programming problem). The kernel function is preferably a radial basis function (RBF) kernel. ,in For kernel function parameters, the preferred values ​​are... ), and This is the bias term. The two functions mentioned above... and This is the required mapping model, which can be based on the pollution response feature vector of plants. Predict cadmium and arsenic concentrations in the soil.

[0070] To optimize the parameters of the mapping relationship model (such as the RBF kernel function parameters) SVR penalty coefficient The model employs a 5-fold cross-validation method: the training sample set is randomly divided into 5 subsets. Each time, 4 subsets are selected as the training set, and the remaining subset is used as the validation set. The SVR model is trained, and its prediction performance is evaluated on the validation set (using root mean square error (RMSE) as the evaluation metric). This process is repeated 5 times, and the parameter combination with the smallest average RMSE from the 5 validations is taken as the optimal parameter. After cross-validation optimization, the model's prediction accuracy is significantly improved, with the RMSE for cadmium concentration prediction being less than 0.05 mg / kg and the RMSE for arsenic concentration prediction being less than 3 mg / kg.

[0071] Secondly, calculate the current pollution level index. This is done using the pollution response feature vector normalized for the current time phase. The data is then input into the trained SVR mapping model to predict the current cadmium concentration in the soil. and arsenic concentration :

[0072]

[0073]

[0074] Calculate the single-phase pollution index based on the predicted concentration. The single-phase pollution index is calculated using the Pollution Load Index (PLI) method:

[0075]

[0076] in, For the number of pollutant types (for cadmium and arsenic combined pollution), ), For the first The pollution coefficient of a pollutant is defined as follows: , For the first Predicted concentrations of various pollutants The limits for this pollutant are the national standard limits (according to the "Soil Environmental Quality Standard for Agricultural Land Soil Pollution Risk Control (GB 15618-2018)", the screening value for cadmium is 0.3 mg / kg, and the screening value for arsenic is 30 mg / kg). For cadmium and arsenic combined pollution, the single-phase pollution index is:

[0077]

[0078] when When, it indicates that the pollution level is lower than the national standard; when When the pollution level reaches or exceeds the national standard, it indicates that the pollution level has reached or exceeded the national standard.

[0079] In addition to the single-phase pollution index, the pollution trend index also needs to be calculated. It reflects the trend of pollution levels over time. The pollution trend index is determined based on the rate of change of pollution response components over multiple consecutive time phases (preferably the most recent 5 time phases):

[0080]

[0081] in, The number of time phases to be considered (take) ), For the first The amplitude of the pollution response components in each time phase, For the first The time of each phase. When When the pollution response is on the rise, it indicates that the pollution may be worsening; when When the pollution response is declining, it indicates that the pollution may be decreasing; when When the pollution level is relatively stable, it indicates that the pollution level is relatively stable.

[0082] Comprehensive single-phase pollution index and pollution trend index Calculate the comprehensive pollution level :

[0083]

[0084] in, and Weighting coefficients (preferred) (This indicates that the impact of the current pollution level is more significant). According to The values ​​are used to classify pollution levels: Cleanliness level, The pollution level is light. The pollution level is moderate. It is classified as a severely polluted area.

[0085] Finally, determine whether to generate a pollution warning signal. (Preset threshold) The threshold was set to 0.5, corresponding to 50% of the cadmium and arsenic limits in the national soil environmental quality standards, representing the plant response characteristic threshold (determined statistically from the aforementioned training sample set; when the soil cadmium concentration was 0.15 mg / kg and the arsenic concentration was 15 mg / kg, detectable minute changes in plant phenotypes were observed). (This value is taken as the warning threshold). When a pollution risk is determined to exist, the pollution early warning decision module 5 generates a pollution early warning signal. The early warning signal includes the following: (1) Pollution type: based on and (1) The relative size of the pollutants determines whether the pollution is primarily cadmium, primarily arsenic, or a combination of cadmium and arsenic; (2) The degree of pollution: based on the relative size of the pollutants, the pollution is determined to be either primarily cadmium, primarily arsenic, or a combination of cadmium and arsenic; (2) Provide the numerical value and give the specific pollution level (clean, lightly polluted, moderately polluted, heavily polluted); (3) Predict the pollution concentration range: give the numerical value and give the specific pollution level (clean, lightly polluted, moderately polluted, heavily polluted); and The predicted values ​​and their confidence intervals are determined based on the prediction variance of the SVR model, with a confidence level of 95%. The early warning signal is presented to the user through a visualization output module, facilitating timely implementation of pollution control measures.

[0086] Through the above processing, the pollution early warning decision module 5 achieves quantitative assessment and early warning of cadmium and arsenic compound pollution. Compared with traditional methods that require plants to show obvious symptoms of damage (such as yellowing leaves and stunted growth) before pollution can be identified (at which point the soil cadmium and arsenic concentrations are often several times higher than the national standard), this invention can identify pollution signals when the pollution concentration is 50% lower than the national standard limit, discovering pollution risks 3-5 weeks in advance, thus gaining valuable time for pollution prevention and control, and has important practical value and social significance.

[0087] In a preferred embodiment, the system further includes a data storage module and a visualization output module. The data storage module is connected to each functional module and stores data such as time-series image sequences, registered time-series image sequences, plant phenotypic time-series change curves, normalized feature parameters, mapping relationship models, pollution level indicators, and pollution early warning signals, forming a complete monitoring data archive. The visualization output module is connected to the pollution early warning decision module and the data storage module and presents pollution level indicators, pollution early warning signals, and historical monitoring data in chart form, including pollution level change curves over time, comparison charts of predicted and actual concentrations, and pollution spatial distribution heat maps, allowing users to intuitively understand the pollution status and trends.

[0088] Example 1

[0089] In a farmland area contaminated with cadmium and arsenic, rice was selected as the monitoring target, and the dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis of this invention was deployed. The multi-temporal image acquisition module 1 uses a 5-megapixel industrial CCD camera, installed at a height of 1.2m, acquiring RGB images once daily at 10:00 AM for 60 consecutive days. The light intensity recorded by the light intensity sensor fluctuated within the range of 15,000-45,000 lux, the ambient temperature varied within the range of 18-32℃, and the ambient humidity varied within the range of 40%-80%RH.

[0090] Adaptive registration engine 2 performed registration processing on 60 frames of acquired time-series images. During the feature point matching stage, an average of approximately 1500 SIFT feature points were detected per pair of adjacent images. After screening using the nearest neighbor distance ratio method, approximately 800 reliable matching points were retained. The initial rigid transformation matrix estimated by the RANSAC algorithm showed a translation of 2-5 pixels and a rotation of 0.5-2 degrees between adjacent images. Optical flow field calculations indicated that the area of ​​non-rigid deformation regions caused by growth extension in rice leaves accounted for 15%-25% of the total plant area. After thin-plate spline interpolation registration, the cross-correlation coefficient between adjacent images reached over 0.92, and the registration error was controlled within 0.4 pixels, meeting the accuracy requirements for subsequent time-series comparative analysis.

[0091] The response feature decoupling extraction module 3 extracts plant phenotypic features from the registered images. The rice leaf area increased from an initial 85 cm². 2 It gradually increased to 320cm at the end of the monitoring period. 2 The chlorophyll index decreased from an initial 0.42 to 0.31 at the end of the monitoring period, and the average leaf hue shifted from an initial 110 degrees to 95 degrees at the end of the monitoring period. Empirical mode decomposition (EMD) decomposed the temporal variation curves of each phenotypic parameter into four IMF components and one residual component. The oscillation period of the first two high-frequency IMF components was 3-7 days, reflecting the rapid response of plants to pollution stress; the variation period of the subsequent low-frequency IMF component and residual component was 30-60 days, reflecting the natural growth trend of plants. After wavelet denoising, the signal-to-noise ratio of the extracted pollution response components was improved by approximately 12 dB.

[0092] The illumination environment normalization module 4 normalizes and corrects phenotypic features. Multi-scale Retinex decomposition results show that the average illumination component values ​​vary from 0.45 to 0.78 across different time phases, exhibiting a significant deviation compared to the reference illumination conditions (average illumination component value of 0.60). Illumination correction coefficients. The range of variation is 0.77-1.33. Environmental temperature and humidity correction coefficient. The variation range was 0.92–1.08. After normalization correction, the environmental factor interference of phenotypic parameters in different time phases was reduced by approximately 35%, improving the accuracy of pollution identification.

[0093] The pollution early warning decision module 5 predicts soil cadmium and arsenic concentrations based on normalized pollution response characteristics and a pre-trained SVR mapping model. Prediction results show that initially, soil cadmium concentrations were approximately 0.12 mg / kg and arsenic concentrations were approximately 12 mg / kg, both below 50% of the national standard limits. During the monitoring period, the predicted concentrations showed an upward trend, reaching 0.28 mg / kg for cadmium and 27 mg / kg for arsenic by the end of the period, approaching the national standard limits. The calculated single-phase pollution index gradually increased from an initial 0.4 to 0.9 by the end, with a pollution trend index of +0.08, indicating a worsening pollution trend. (Comprehensive pollution level...) On the 15th day of monitoring, the concentration exceeded the preset threshold of 0.5 for the first time, and the system generated a pollution warning signal, indicating that there was a risk of cadmium and arsenic combined pollution. The pollution level was light pollution, with a predicted cadmium concentration range of 0.13-0.17 mg / kg (95% confidence) and a predicted arsenic concentration range of 13-18 mg / kg (95% confidence).

[0094] To verify the accuracy of the system's early warning, soil samples were collected on days 15, 30, 45, and 60 of monitoring and sent to the laboratory for atomic absorption spectrometry (AAS) analysis. The measured cadmium concentrations were 0.14, 0.19, 0.23, and 0.27 mg / kg, respectively, and the measured arsenic concentrations were 14, 19, 23, and 28 mg / kg, respectively. The relative errors between the measured concentrations and the system's predicted concentrations were all less than 15%. The early warning time (day 15 of monitoring) was 23 days earlier than the time when obvious damage symptoms appeared in the plants (day 38 of monitoring), verifying the system's early warning capability and predictive accuracy.

[0095] Example 2

[0096] In a soil-contaminated site near an industrial zone, sunflowers were selected as the monitoring target, and the system of this invention was deployed to monitor cadmium and arsenic compound pollution. Unlike Example 1, sunflowers grow faster, allowing the multi-temporal image acquisition module 1 to shorten the acquisition interval to 12 hours (acquiring images twice daily at 10:00 AM and 4:00 PM), with continuous monitoring for 30 days, resulting in 60 images. The site experienced significant variations in light and environmental conditions, with a 2-3 fold difference in light intensity between morning and afternoon, and a 8-12°C difference in ambient temperature.

[0097] Through precise registration by the adaptive registration engine 2 and normalization correction by the light environment normalization module 4, the system effectively eliminated the influence of time-varying differences and environmental fluctuations. The pollution response components extracted by the response feature decoupling extraction module 3 showed that sunflowers were more sensitive to cadmium and arsenic pollution, with a chlorophyll index decrease rate approximately 40% faster than rice and a greater shift in leaf hue. The pollution early warning decision module 5 generated an early warning signal on the 6th day of monitoring (monitoring for only 6 days), predicting a soil cadmium concentration of 0.18 mg / kg and an arsenic concentration of 20 mg / kg. Laboratory measurements showed that the actual cadmium concentration was 0.16 mg / kg and the arsenic concentration was 22 mg / kg, with a system prediction error of less than 12%, verifying the system's applicability and high sensitivity to different plant species.

[0098] The above embodiments fully demonstrate the technical solution and beneficial effects of the present invention. The dynamic monitoring system for plant response to cadmium and arsenic combined pollution based on time-series image analysis of the present invention achieves highly sensitive, accurate, dynamic, and non-invasive monitoring and early warning of cadmium and arsenic combined pollution in soil through the close collaboration of five modules: multi-temporal image acquisition, adaptive registration, response feature decoupling extraction, light environment normalization, and pollution early warning decision-making. This provides an effective technical means for the prevention and control of heavy metal pollution in soil and has significant theoretical value and practical application prospects.

[0099] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic monitoring system for plant response to cadmium-arsenic combined pollution based on time-series image analysis, characterized in that, The method comprises the steps of: A multi-time image acquisition module is used to continuously acquire RGB images of target plants within a preset time sequence to form a time sequence image sequence, each frame of image in the time sequence image sequence carrying a timestamp and an environmental parameter label; An adaptive registration engine is connected with the multi-time image acquisition module, used to receive the time sequence image sequence, generate a time phase registration transformation matrix based on feature point matching results and optical flow field estimation results of adjacent time phase images, and perform spatial alignment processing on each frame of image in the time sequence image sequence according to the time phase registration transformation matrix to obtain a registered time sequence image sequence; A response feature decoupling extraction module is connected with the adaptive registration engine, used to receive the registered time sequence image sequence, obtain a plant target mask through plant region segmentation, extract stem and leaf morphological parameters and color feature parameters based on the plant target mask, construct a plant phenotype time sequence change curve, decompose the plant phenotype time sequence change curve by using an empirical mode decomposition method to obtain a plurality of intrinsic mode function components and a residual component, and classify high-frequency intrinsic mode function components as pollution response components and low-frequency intrinsic mode function components and the residual component as natural growth components according to frequency characteristics of each intrinsic mode function component; An illumination environment normalization module is connected with the multi-time image acquisition module and the response feature decoupling extraction module, used to receive environmental parameters of a current time phase, generate a normalization correction coefficient based on illumination intensity and environmental temperature and humidity, and correct the stem and leaf morphological parameters and the color feature parameters according to the normalization correction coefficient to eliminate phenotype changes caused by non-pollution factors; A pollution early warning decision module is connected with the response feature decoupling extraction module and the illumination environment normalization module, used to receive the normalized pollution response component, establish a mapping relationship model of pollution response features and heavy metal concentrations, calculate a current pollution level index according to the mapping relationship model, and generate a pollution early warning signal when the pollution level index exceeds a preset threshold.

2. The system according to claim 1, wherein, The adaptive registration engine generates the time phase registration transformation matrix based on the feature point matching results and the optical flow field estimation results of the adjacent time phase images, and the process comprises the steps of: SIFT feature points of adjacent time phase images are extracted, feature point matching is performed by using a nearest neighbor distance ratio method, and an initial feature point pair set is obtained after eliminating mis-matching points; An initial rigid transformation matrix is estimated by using a RANSAC algorithm based on the initial feature point pair set; A dense optical flow field of adjacent time phase images is calculated, and a non-rigid deformation region in the optical flow field is extracted; The final time phase registration transformation matrix is generated by using a thin plate spline interpolation method according to the initial rigid transformation matrix and the non-rigid deformation region.

3. The system according to claim 1, wherein the system is characterized by, When the response feature decoupling extraction module classifies the high-frequency intrinsic mode function components as the pollution response components, the first three intrinsic mode function components are classified as the pollution response components; and when the low-frequency intrinsic mode function components and the residual component are classified as the natural growth components, the fourth and subsequent intrinsic mode function components and the residual component are classified as the natural growth components. The response feature decoupling extraction module is further configured to perform wavelet denoising processing on the pollution response component to remove noise interference.

4. The system according to claim 1, wherein the system is characterized by, The stem and leaf morphological parameters extracted by the response feature decoupling extraction module include: blade area, blade circumference, blade aspect ratio, blade curvature, stem height, stem diameter, blade inclination angle; The color feature parameters include: blade average hue, blade average saturation, blade average lightness, blade color variance, chlorophyll index, normalized vegetation index.

5. The system according to claim 1, wherein the system is characterized by, The process of generating the normalization correction coefficient by the illumination environment normalization module includes: establishing a plant phenotype benchmark database under a reference illumination condition; obtaining an illumination intensity value and an environmental temperature and humidity value at a current phase; based on the illumination intensity value, decomposing an image into a reflection component and an illumination component using Retinex theory, and calculating an illumination correction coefficient according to the reference illumination condition; based on the environmental temperature and humidity value, querying the plant phenotype benchmark database to calculate an environmental correction coefficient; weighting and fusing the illumination correction coefficient and the environmental correction coefficient to obtain the normalization correction coefficient.

6. The system according to claim 5, wherein the system is characterized by, When the illumination environment normalization module decomposes the image into the reflection component and the illumination component based on the Retinex theory, a multi-scale Retinex algorithm is used to extract the illumination component through a different scale of Gaussian filter set, and the reflection component is determined by the difference between the logarithmic domain of the image and the illumination component.

7. The system according to claim 1, wherein the system is characterized by, The process of establishing the mapping relationship model by the pollution early warning decision module includes: collecting plant phenotype data under a known heavy metal concentration gradient as training samples; extracting a multi-dimensional pollution response feature vector in the training samples; establishing a non-linear mapping relationship between the multi-dimensional pollution response feature vector and the heavy metal concentration using a support vector regression algorithm; optimizing the parameters of the mapping relationship model through cross-validation.

8. The system according to claim 1, wherein the system is characterized by, The pollution level index calculated by the pollution early warning decision module includes: a single-phase pollution index, determined based on the amplitude of the pollution response component at the current phase; a pollution trend index, determined based on the change rate of the pollution response component at consecutive multiple phases; a comprehensive pollution level, obtained by weighting calculation of the single-phase pollution index and the pollution trend index.

9. The system according to claim 1, wherein the system is characterized by, The preset threshold is a plant response feature threshold corresponding to 50% of the national soil environmental quality standard limit value of cadmium and arsenic. When the pollution level index exceeds the preset threshold, the pollution early warning signal generated by the pollution early warning decision module includes the pollution type, the pollution degree, and the predicted pollution concentration range.

10. The system according to claim 1, wherein the system is characterized by, Further comprising: a data storage module connected with the multi-phase image acquisition module, the adaptive registration engine, the response feature decoupling extraction module, the illumination environment normalization module, and the pollution early warning decision module, configured to store the time sequence image sequence, the registered time sequence image sequence, the plant phenotype time sequence change curve, the normalized feature parameters, the mapping relationship model, and the pollution early warning signal. A visual output module, connected with the pollution early warning decision module and the data storage module, is configured to present the pollution level index, the pollution early warning signal and historical monitoring data in a chart form.

Images