Marine sediment layered collection dynamic monitoring method and system based on Internet of Things

Abstract

The invention provides a method and a system for dynamically monitoring layered collection of marine sediments based on Internet of Things. The method comprises the following steps: acquiring original marine environment parameters collected in real time; calculating to obtain a self-adaptive forgetting factor according to saliency indicating interface transition between sediment layers in the original marine environment parameters, and performing Kalman filtering on the non-image original marine environment parameters by adopting the self-adaptive forgetting factor to obtain filtered environment parameters; selecting a feature subset from a feature set having preset spatial correlation with the current image area, determining division features from the feature subset, and extracting sediment morphological features by adopting a random uniform forest; fusing the filtered environmental parameters and the sediment morphological features to obtain a comprehensive state feature vector; and based on the comprehensive state feature vector, calculating to obtain the adjusted collection depth and collection speed of the marine sediment collection equipment by using a preset layered collection model, and sending the adjusted collection depth and collection speed to a control system of the collection equipment.

Description

Technical Field

[0001] This application belongs to the field of marine sediment monitoring, and particularly relates to a method and system for dynamic monitoring of layered collection of marine sediments based on the Internet of Things. Background Technique

[0002] Marine sediments record key information such as paleoclimate, paleoenvironment, paleo-ocean, and tectonic evolution during the geological history period. Accurately obtaining marine sediment samples with a clear layered structure is crucial for deeply understanding the Earth's evolution process, evaluating the potential of marine resources, and predicting future environmental changes. Traditional marine sediment collection methods, such as gravity coring and piston coring, often rely on empirical judgment during the collection process and are difficult to perceive and adapt to the subtle changes in sediment bedding in real time. Especially in cases where the transition of the sediment layer interface is not obvious, the strata are complex and variable, or there are thin interbeds, traditional methods are prone to loss of layered information, disorder of sample sequences, or confusion of samples from different horizons, seriously affecting the accuracy and scientific research value of subsequent analysis results. Although some modern collection devices integrate some basic sensors for monitoring device status, this information is often only used as a rough reference and lacks the perception of the real-time characteristics of sediments during the collection process.

[0003] With the rapid development of Internet of Things technology, sensor technology, and artificial intelligence algorithms, it provides a new direction for the intelligent upgrade of marine exploration and sampling equipment. Integrating multiple sensors at the front end of the collection device and transmitting real-time collected data to the control center through low-power wide-area Internet of Things (such as LoRa, NB-IoT) or underwater acoustic / optical communication technology lays a data foundation for realizing dynamic monitoring. However, the marine environment is complex and variable, and sensor data often contains a large amount of noise and uncertainty. On the other hand, the visual morphological characteristics of sediments contain rich geological information and are an important basis for judging stratification. Machine learning algorithms such as random forest show good performance in image feature extraction and classification, but traditional random forest algorithms are insensitive to key morphological differences, affecting the accuracy of stratification recognition. Summary of the Invention

[0004] To solve the problems pointed out in the above background technique, this application proposes a method for dynamic monitoring of layered collection of marine sediments based on the Internet of Things, including: Step 1, obtaining the original marine environmental parameters collected in real time by multiple sensor nodes deployed in the marine sediment collection area and transmitted through the Internet of Things communication module; Step 2, calculating an adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters, and using the adaptive forgetting factor to perform Kalman filtering on the non-image original marine environmental parameters to obtain filtered environmental parameters; Step 3: Select a feature subset from the feature set that has a preset spatial correlation with the current image region, determine the partitioning features from the feature subset, and extract sediment morphology features using a random uniform forest. When constructing a decision tree, the random uniform forest determines the selected nodes based on the partitioning features; fuse the filtered environmental parameters and the sediment morphology features to obtain a comprehensive state feature vector; Step 4: Based on the comprehensive state feature vector, calculate the adjusted collection depth and collection speed of the marine sediment collection device using a preset hierarchical collection model, and send the adjusted collection depth and collection speed to the control system of the collection device through the Internet of Things communication module.

[0005] Optionally, calculating the adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters includes: Receive the original marine environmental parameters, and extract multiple indicators related to the transition of the sediment layer interface. The multiple indicators include the pressure change rate, depth change rate, image texture mutation degree, and temperature gradient; Set the weight coefficient of each of the multiple indicators, and perform weighted fusion on the multiple indicators according to the weight coefficient to obtain an initial transition significance score; Obtain historical sediment stratification data and current collection target layer information, and dynamically adjust the threshold of the initial transition significance score according to the historical sediment stratification data and current collection target layer information; When the initial transition significance score exceeds the dynamically adjusted threshold, set the adjustment amplitude of the adaptive forgetting factor to a first preset adjustment amplitude value; otherwise, set the adjustment amplitude of the adaptive forgetting factor to a second preset adjustment amplitude value. The first preset adjustment amplitude value is greater than the second preset adjustment amplitude value; Calculate the final adaptive forgetting factor according to the adjustment amplitude of the adaptive forgetting factor and a preset reference forgetting factor.

[0006] Optionally, performing Kalman filtering on the non-image original marine environmental parameters using the adaptive forgetting factor to obtain filtered environmental parameters includes: Initialize the state vector and covariance matrix of the Kalman filter. The state vector includes non-image marine environmental parameters to be estimated; Based on the system dynamics model, perform a one-step prediction on the state vector and covariance matrix at the current moment to obtain a predicted state vector and a predicted covariance matrix; Obtain the non-image original marine environmental parameters at the current moment as the observation value; Calculate the Kalman gain by using the predicted state vector, the predicted covariance matrix, the observed value, and the adaptive forgetting factor; Update the state vector by combining the predicted state vector, the observed value, and the Kalman gain to obtain the filtered environmental parameters at the current moment; Update the covariance matrix for filtering at the next moment.

[0007] Optionally, the selecting a feature subset from a set of features having a preset spatial correlation with the current image region includes: Preprocess the image data in the original marine environmental parameters, extract the underlying visual features, and preliminarily group the underlying visual features according to physical meaning and spatial scale to form an initial feature set; Obtain the macroscopic features of the current image region to be processed, where the macroscopic features include the average depth information and color information of the region; Judge the type of sediment morphological features that need to be distinguished currently according to the macroscopic features; Based on the judged type of sediment morphological features, adjust the prior probability that the feature group related to the judged type of sediment morphological features and having a strong spatial correlation with the current image region in the initial feature set is selected to obtain an adjusted feature set; Randomly extract a specified number of features from the adjusted feature set according to the adjusted prior probability to form the feature subset.

[0008] Optionally, when using a random uniform forest to extract sediment morphological features, the random uniform forest determines the selected nodes according to the partitioning features when constructing a decision tree, including: Step 3.1: Initialize the random uniform forest, and set the number of decision trees and the maximum depth of each tree; Step 3.2: For each decision tree to be constructed, perform bootstrap sampling with replacement on the training samples to form a training subset of the current decision tree; Step 3.3: Recursively construct a decision tree starting from the root node. For the current node to be split, obtain the feature subset for the samples of the image region represented by the current node; randomly select a feature from the feature subset as the partitioning feature, and randomly generate a split point; partition the samples of the current node into child nodes according to the partitioning feature and the split point; if the stop splitting condition is met, mark the current node as a leaf node and assign a class label; otherwise, repeat Step 3.3 for the child nodes; Step 3.4: Combine all the constructed decision trees into a random uniform forest model; Step 3.5: Input the image data of the morphological features to be extracted into the random uniform forest model, and obtain the final sediment morphological features through majority voting or averaging.

[0009] This application also proposes an IoT-based dynamic monitoring system for layered collection of marine sediments, including: A data acquisition unit, configured to obtain the original marine environmental parameters that are transmitted through the IoT communication module and are collected in real time by a plurality of sensor nodes deployed in the marine sediment collection area; A filtering unit, configured to calculate an adaptive forgetting factor of the Kalman filtering algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters, and perform Kalman filtering on the non-image original marine environmental parameters by using the adaptive forgetting factor to obtain filtered environmental parameters; A feature extraction unit, configured to select a feature subset from a feature set including features having a preset spatial correlation with the current image area, determine a partitioning feature from the feature subset, extract sediment morphological features by using a random uniform forest, and the random uniform forest determines the selected nodes according to the partitioning feature when constructing a decision tree; fuse the filtered environmental parameters and the sediment morphological features to obtain a comprehensive state feature vector; A monitoring unit, configured to calculate the adjusted collection depth and collection speed of the marine sediment collection device based on the comprehensive state feature vector and by using a preset layered collection model, and send the adjusted collection depth and collection speed to the control system of the collection device through the IoT communication module.

[0010] Optionally, calculating the adaptive forgetting factor of the Kalman filtering algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters includes: Receiving the original marine environmental parameters, and extracting a plurality of indicators related to the transition of the sediment layer interface, where the plurality of indicators include a pressure change rate, a depth change rate, an image texture mutation degree, and a temperature gradient; Setting weight coefficients for the plurality of indicators respectively, and performing weighted fusion on the plurality of indicators according to the weight coefficients to obtain an initial transition significance score; Obtaining historical sediment layering data and current collection target layer information, and dynamically adjusting the threshold of the initial transition significance score according to the historical sediment layering data and the current collection target layer information; When the initial transition significance score exceeds the dynamically adjusted threshold, setting the adjustment amplitude of the adaptive forgetting factor to a first preset adjustment amplitude value, otherwise, setting the adjustment amplitude of the adaptive forgetting factor to a second preset adjustment amplitude value, where the first preset adjustment amplitude value is greater than the second preset adjustment amplitude value; Adjust the amplitude according to the adaptive forgetting factor and the preset reference forgetting factor, and calculate the final adaptive forgetting factor.

[0011] Optionally, using the adaptive forgetting factor to perform Kalman filtering on the original marine environmental parameters of non-images to obtain filtered environmental parameters, including: Initialize the state vector and covariance matrix of the Kalman filter, where the state vector contains the non-image marine environmental parameters to be estimated; Based on the system dynamics model, perform a one-step prediction on the state vector and covariance matrix at the current moment to obtain a predicted state vector and a predicted covariance matrix; Obtain the non-image original marine environmental parameters at the current moment as the observed value; Use the predicted state vector, predicted covariance matrix, the observed value, and the adaptive forgetting factor to calculate the Kalman gain; Combine the predicted state vector, the observed value, and the Kalman gain to update the state vector to obtain the filtered environmental parameters at the current moment; Update the covariance matrix for filtering at the next moment.

[0012] Optionally, the selecting a feature subset from a feature set including features having a preset spatial correlation with the current image region includes: Preprocess the image data in the original marine environmental parameters, extract the underlying visual features, and preliminarily group the underlying visual features according to physical meaning and spatial scale to form an initial feature set; Obtain the macroscopic features of the current image region to be processed, where the macroscopic features include the average depth information and color information of the region; Judge the type of sediment morphology features to be distinguished currently according to the macroscopic features; Based on the judged type of sediment morphology features, adjust the prior probability of selection of the feature group related to the judged type of sediment morphology features and having a strong spatial correlation with the current image region in the initial feature set to obtain an adjusted feature set; Randomly extract a specified number of features from the adjusted feature set according to the adjusted prior probability to form the feature subset.

[0013] Optionally, using a random uniform forest to extract sediment morphology features, where the random uniform forest determines the selected nodes according to the partitioning features when constructing a decision tree, including: Step 3.1: Initialize the random uniform forest, and set the number of decision trees and the maximum depth of each tree; Step 3.2: For each decision tree to be constructed, perform bootstrap sampling with replacement on the training samples to form the training subset for the current decision tree; Step 3.3: Recursively construct the decision tree starting from the root node. For the current node to be split, obtain the feature subset for the samples in the image region represented by the current node; randomly select a feature from the feature subset as the splitting feature, and randomly generate a splitting point; divide the samples of the current node into child nodes according to the splitting feature and the splitting point; if the stopping splitting condition is met, mark the current node as a leaf node and assign a class label; otherwise, repeat Step 3.3 for the child nodes; Step 3.4: Combine all the constructed decision trees into a random uniform forest model; Step 3.5: Input the image data for which the morphological features are to be extracted into the random uniform forest model, and obtain the final sediment morphological features through majority voting or averaging.

[0014] Compared with the prior art, the present application has the following beneficial effects: The adaptive forgetting factor Kalman filter sensitive to interlayer transitions can capture the interlayer changes of sediments, quickly adapt to the characteristics of new layers, and accurately locate the stratification interface. At the same time, the spatially context-guided random uniform forest can more intelligently extract key morphological features from images, providing a reliable visual basis for stratification judgment. It realizes more accurate and adaptable dynamic monitoring and control of the marine sediment stratification sampling process, thereby obtaining higher-quality sediment samples and providing more valuable data for marine scientific research. Description of the Drawings

[0015] Figure 1 It is a flowchart of Embodiment 1; Figure 2 It is a calculation process of the comprehensive state feature vector; Figure 3 It is a flowchart of the random uniform forest for extracting sediment morphological features; Figure 4 It is a schematic diagram of a single decision tree. Detailed Embodiments

[0016] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0017] In the description of the present application, the terms "first", "second" and the corresponding term numbers in the specification, claims and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that such terms can be interchanged under appropriate circumstances, which is only a way of distinguishing objects with the same attributes when describing the embodiments of the present application. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device comprising a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these processes, methods, products or devices.

[0018] In addition, in the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more. The term "and / or" or the character " / " in the present application is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B, or A / B, can represent: A exists alone, A and B exist simultaneously, and B exists alone.

[0019] Specific embodiment, a dynamic monitoring method for layered collection of marine sediments based on the Internet of Things, as Figure 1 shown, includes: Step 1, obtaining the original marine environmental parameters collected in real time by a plurality of sensor nodes deployed in the marine sediment collection area through the Internet of Things communication module; Receive the original marine environmental parameters collected and sent in real time by a plurality of sensor nodes deployed in the marine sediment collection area through a preset Internet of Things communication module. The Internet of Things communication module uses a variety of wireless or wired communication technologies, such as low-power wide-area network technologies, such as LoRa, NB-IoT, Sigfox, etc. The sensor nodes are integrated on the marine sediment collection device, and their types include but are not limited to depth sensors, pressure sensors, temperature sensors, and image sensors. These sensors work together to sense the key environmental parameters and sediment characteristics during the collection process in different dimensions.

[0020] Step 2, calculating an adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters, and using the adaptive forgetting factor to perform Kalman filtering on the non-image original marine environmental parameters to obtain filtered environmental parameters; Analyze the received original marine environmental parameters, especially the characteristic information that can indicate the upcoming or ongoing transition of the sediment layer interface. The determination of significance is based on the change trend, change rate, appearance of specific patterns, etc. of single or multiple parameters. For example, rapid changes in pressure, stability or sudden changes in depth, abnormal temperature gradients, or macroscopic texture changes initially extracted from image data, etc.

[0021] Based on the evaluation results of the transition significance, dynamically calculate or adjust the adaptive forgetting factor used in the Kalman filtering algorithm. The role of the adaptive forgetting factor is to control the weight allocation of the filter to historical data and current observation data during state estimation. A properly adjusted forgetting factor can enable the filter to converge to a new true state faster during mutations such as layer interface transitions. There are various calculation methods. For example, when it is determined that the transition significance increases, a certain preset strategy is adopted to increase the forgetting effect, that is, to assign greater weight to the current observation and forget the historical state faster, and vice versa to reduce the forgetting effect.

[0022] Use the adaptively calculated or adjusted forgetting factor to perform Kalman filtering on the non-image data in the original marine environmental parameters, such as time series data from depth sensors, pressure sensors, temperature sensors, etc. Kalman filtering can estimate the state of a dynamic system from a series of incomplete and noisy measurements. Through filtering, the random noise in the sensor data is effectively removed, and the accuracy and stability of the parameters are improved, and finally the filtered environmental parameters are obtained.

[0023] Step 3, select a feature subset from the feature set that has a preset spatial correlation with the current image region, and determine the partitioning feature from the feature subset. Use a random uniform forest to extract sediment morphological features. The random uniform forest determines the selected nodes according to the partitioning feature when constructing decision trees; fuse the filtered environmental parameters and the sediment morphological features to obtain a comprehensive state feature vector; When extracting the morphological features of sediments through images, when performing feature selection, the spatial correlation between features within the image region will be considered. Instead of simply looking at each feature in isolation, attention will also be paid to feature combinations that are spatially adjacent or jointly represent certain geological structures such as bedding, nodules, bioturbation traces, etc. In one embodiment, a set of spatially correlated feature sets related to specific sediment morphology is predefined or obtained through learning.

[0024] Then, one or more feature subsets are selected from this feature set with preset spatial correlation, and partitioning features for subsequent model construction are determined from these subsets. The selection methods include but are not limited to importance evaluation of features, mutual information between features, etc. The random uniform forest algorithm is used to extract the morphological features of sediments. During the construction of each decision tree, when a node needs to be selected for splitting, the previously determined partitioning features are used to guide the selection and splitting of the node. The growth of the decision tree is not completely random but is guided by the feature information related to the sediment morphology and spatial correlation. Through the ensemble learning of the entire forest, morphological features that can characterize information such as sediment color, texture, particle size, bedding structure, etc. are finally output. The filtered (non-image) environmental parameters obtained are fused with the extracted sediment morphological features at the data layer or feature layer. Information from different sensors describing different aspects is integrated to form a comprehensive state feature vector that can more comprehensively and accurately reflect the current state of marine sediment collection. The fusion methods include but are not limited to simple feature splicing, weighted averaging, or fusion using neural networks or Bayesian networks, etc., Figure 2 shows the calculation process of the comprehensive state feature vector.

[0025] Step 4, based on the comprehensive state feature vector, and using a preset hierarchical collection model, calculate the adjusted collection depth and collection speed of the marine sediment collection device, and send the adjusted collection depth and collection speed to the control system of the collection device through the Internet of Things communication module.

[0026] Take the obtained comprehensive state feature vector as input, and combine it with a preset hierarchical collection model to calculate the collection depth and collection speed that the marine sediment collection device needs to adjust. Preferably, the hierarchical collection model is a rule-based expert system or a machine learning model trained from historical data, such as a regression model, a reinforcement learning model, etc. The hierarchical collection model judges whether the collection strategy needs to be adjusted currently according to the sediment state characterized by the comprehensive state feature vector in real time, and how to adjust to optimize the collection of samples at the target layer while avoiding interlayer confusion. The calculated adjusted collection depth and collection speed parameters are sent back to the control system of the marine sediment collection device in real time through the Internet of Things communication module. After receiving these parameters, the control system of the collection device will correspondingly adjust the operation of actuators such as the propulsion system, thereby realizing the closed-loop dynamic control of the sediment hierarchical collection process. For example, when it is judged that a new sediment layer is about to be reached, the collection speed will be slowed down and the collection depth will be finely adjusted to ensure accurate sampling.

[0027] In an optional embodiment, calculating an adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environment parameters includes: Receiving the original marine environment parameters, and extracting a plurality of indicators related to the transition of the sediment layer interface, where the plurality of indicators include a pressure change rate, a depth change rate, an image texture mutation degree, and a temperature gradient; Setting weight coefficients for each of the plurality of indicators, and performing weighted fusion on the plurality of indicators according to the weight coefficients to obtain an initial transition significance score; Obtaining historical sediment stratification data and current acquisition target layer information, and dynamically adjusting a threshold of the initial transition significance score according to the historical sediment stratification data and the current acquisition target layer information; When the initial transition significance score exceeds the dynamically adjusted threshold, setting the adjustment amplitude of the adaptive forgetting factor to a first preset adjustment amplitude value; otherwise, setting the adjustment amplitude of the adaptive forgetting factor to a second preset adjustment amplitude value, where the first preset adjustment amplitude value is greater than the second preset adjustment amplitude value; Calculating the final adaptive forgetting factor according to the adjustment amplitude of the adaptive forgetting factor and a preset reference forgetting factor.

[0028] Specifically, preprocessing the image data in the obtained original marine environment parameters includes image correction, region of interest (ROI) extraction, image enhancement, etc. Extracting underlying visual features from the preprocessed image, where the underlying visual features include, but are not limited to, Gabor features, local binary patterns, gray-level co-occurrence matrix features, color histograms, etc. In one embodiment, it also includes shape features, such as the shape after extracting the particle contour.

[0029] Grouping the extracted underlying visual features preliminarily according to their geological or morphological meanings, such as a group of Gabor features related to the bedding direction, a group of LBP or GLCM features related to particle roughness, a group of color features related to color changes, and spatial scales, to form a structured initial feature total set.

[0030] During the construction process of each decision tree in the random uniform forest, when selecting an internal node , which represents a part of the training samples, that is, a part of the image area or pixels, for splitting: Calculate the macroscopic context features of the sample represented by the node. The macroscopic features include the average depth value associated with the depth sensor data synchronized with the image, the average color obtained by clustering the pixel colors in the region or calculating the average RGB value, the operating state of the current acquisition device, etc. If the average color of the sample area of the node is dark and the average depth value is large, it indicates that the current situation is in dark muddy sediments. Based on the macroscopic context features and combined with a preset geological knowledge rule base or an auxiliary classifier, dynamically determine what type of sediment morphological feature is most likely to need further differentiation and identification in the sample of the current node. For example, based on the above example context, it is determined that what needs to be focused on differentiating currently is whether there are fine laminations or whether there are signs of bioturbation, rather than macroscopic color differences.

[0031] Based on the determined type of sediment morphological feature that needs to be differentiated most in the current node, such as fine laminations, adjust the prior probabilities of each feature or feature group in the initial total feature set being selected into the candidate feature subset of the current node. For example, if it is determined that fine laminations need to be differentiated, then in the initial total feature set, the Gabor feature group describing horizontal or nearly horizontal textures, or the LBP feature group that can reflect subtle gray-scale changes, the basic probabilities of being selected will be temporarily increased. For example, increase the sampling weights of these related features from to , where c>1 is an amplification factor, and at the same time, the weights of other features are correspondingly reduced to ensure that the total probability sum is 1. According to the prior probability distribution dynamically adjusted by the context information, randomly select m features with weights from the total image features. The greater the weight, the greater the probability of being selected, to form a candidate feature subset for splitting the current node , where m is much smaller than , preferably . From the constructed candidate feature subset with context-guided bias , determine the splitting feature and the splitting point threshold for splitting the current node in the standard manner of the random uniform forest algorithm . In one embodiment, randomly select a feature from , and then randomly generate a splitting point within its value range for , that is, and . According to the selected splitting feature and splitting point, divide the samples of the current node into its left and right child nodes.

[0032] Repeat the context analysis, guided feature subset construction, and node splitting for the newly generated child nodes until the conditions for the decision tree to stop growing are met. For example, the node reaches the preset maximum depth, the number of samples in the node is too small to continue splitting, etc.

[0033] Repeat the above steps multiple times. Each tree is independently constructed based on a different training subset obtained by bootstrap sampling with replacement from the original training samples of the sediment image regions with labels and their corresponding morphological categories. All these independently constructed decision trees together form a random uniform forest model.

[0034] When new, unlabeled sediment image data is input, the image data or its extracted underlying features are input into each decision tree in the forest. Each tree independently gives a classification result for the input data. For example, it determines whether the image region belongs to predefined categories such as homogeneous mud, parallel - bedded sand, or cross - bedded silt, or a regression value. By integrating the output results of all decision trees in the forest, for example, for classification tasks, the majority voting principle is adopted; for regression tasks, the average of the output values of all trees is adopted, to obtain the final sediment morphological feature description or classification result for the input image data.

[0035] In an optional embodiment, the obtaining the filtered environmental parameters by performing Kalman filtering on the non - image original marine environmental parameters using the adaptive forgetting factor includes: Initialize the state vector and covariance matrix of the Kalman filter. The state vector contains the non - image marine environmental parameters to be estimated. Based on the system dynamics model, perform a one - step prediction on the state vector and covariance matrix at the current moment to obtain the predicted state vector and predicted covariance matrix. Obtain the non - image original marine environmental parameters at the current moment as the observation value. Use the predicted state vector, predicted covariance matrix, the observation value, and the adaptive forgetting factor to calculate the Kalman gain. Combine the predicted state vector, the observation value, and the Kalman gain to update the state vector to obtain the filtered environmental parameters at the current moment. Update the covariance matrix for filtering at the next moment.

[0036] Specifically, initialize the state vector of the system. The state vector includes non-image ocean environment parameters to be estimated. For example, the state vector includes the current depth, speed, and / or salinity of the acquisition device, etc. And initialize the error covariance matrix of the state vector. The covariance matrix characterizes the degree of uncertainty of the initial state estimate. Based on the system dynamics model, such as uniform motion, predict the state and error covariance at the current moment according to the posterior optimal estimate state and posterior error covariance at the previous moment. Obtain the non-image raw ocean environment parameters transmitted through the IoT communication module at the current moment and collected in real time by the sensor nodes as the observation values at the current moment. For example, a vector composed of the actual speed, water temperature, and / or salinity data measured by the sensor at this moment.

[0037] Adjust the parameters in the Kalman filter equation according to the calculated adaptive forgetting factor. For example, by affecting the effective values of the process noise covariance and the observation noise covariance, or directly adjusting the update weight of the prior error covariance. When a significant transition of the sediment layer interface is detected, the forgetting factor is adjusted so that the filter focuses more on the current new observation data, thus quickly responding to parameter changes; conversely, when the parameters are stable, it focuses more on historical data to smooth the noise. Use the calculated Kalman gain, combined with the prior state estimate and the current observation value, to correct the state vector to obtain the optimal state estimate at the current moment, that is, the filtered environmental parameters. Finally, update the error covariance matrix to reflect the degree of uncertainty of the posterior state estimate. This updated covariance matrix will be used in the prediction stage of the next moment.

[0038] Among them, when calculating the Kalman gain using the predicted state vector, the predicted covariance matrix, the observation value, and the adaptive forgetting factor, in one embodiment, the predicted covariance matrix is weighted by the adaptive forgetting factor, and then the Kalman gain is calculated using the weighted predicted covariance matrix, the predicted state vector, the observation value, etc.

[0039] In an optional embodiment, the selecting a feature subset from the feature set having a preset spatial correlation with the current image region includes: Preprocess the image data in the original ocean environment parameters, extract the underlying visual features, and preliminarily group the underlying visual features according to physical meaning and spatial scale to form an initial feature set; Obtain the macroscopic features of the current image region to be processed. The macroscopic features include the average depth information and color information of this region; Judge the type of sediment morphological features that need to be distinguished currently according to the macroscopic features; Based on the determined sediment morphological feature type, adjust the prior probability of selection of the feature group in the initial feature set that is related to the determined sediment morphological feature type and has a strong spatial correlation with the current image region, to obtain an adjusted feature set; From the adjusted feature set, randomly extract a specified number of features according to the adjusted prior probability to form the feature subset.

[0040] Specifically, after image preprocessing, extract features from the image that can describe its basic visual content, including but not limited to features such as texture and shape. Classify and organize the extracted multiple low-level visual features according to their physical meanings, such as color-related features, texture-related features, shape-related features, and spatial scales, such as features describing large-scale changes and features describing detailed textures. For example, all color-related features are grouped into the color feature group, and all features describing local texture are grouped into the texture feature group. These feature groups together constitute the initial feature set.

[0041] Obtain other information from other sensors such as depth sensors. The other information includes: the absolute water depth at the current image acquisition location. For example, the average depth of the current area is 1500 meters; average color information, etc. The macroscopic features provide context information for subsequent determination of sediment types and adjustment of feature selection strategies. For example, in marine environments with different depths, the sediment types and optical properties are significantly different.

[0042] First, make a preliminary judgment to predict the most likely existing or the most important sediment morphological type to be distinguished in the current image region. Preferably, use a prior knowledge base or a pre-trained model. The knowledge base contains common sediment types and their distinguishing points in different depths and different color regions. For example, if the average depth is relatively shallow and the average color is light yellow, it is judged that the main types to be distinguished currently are fine sand, coarse sand, or shell debris. If the average depth is very deep and the average color is dark brown or grayish black, it is judged that the initial forms of different types of deep-sea soft mud, clay, or manganese nodules need to be concerned.

[0043] Associate the determined sediment morphological feature type that is most likely to need to be distinguished with the specific feature groups in the initial feature set. If it is determined that fine sand and coarse sand need to be distinguished, then the texture feature group and part of the edge feature group in the initial feature set will be considered highly relevant. At the same time, consider the matching degree of the feature with the content of the current image region. For example, if the current image region visually presents a very smooth texture, then although the feature used to describe the rough texture belongs to the texture feature group, its strong spatial correlation with the current region may be low. For those features that are considered highly relevant to the target sediment type and have a high matching degree with the content of the current image region, the prior probability or the selection weight for being selected into the final feature subset will be increased. Conversely, the prior probability of irrelevant or redundant features will be decreased. Continuing with the above example, if it is determined that fine sand and coarse sand need to be distinguished, then the prior probability of the feature describing the granularity information in the texture feature group will be significantly increased, while the prior probability of the global color feature in the color feature group will be appropriately decreased. The adjusted feature set attaches an updated selection probability to each feature in the initial feature set. The set itself does not reduce the number of elements, but the importance of each element, that is, the chance of being selected, has changed. A weighted random sampling method is used to form the feature subset.

[0044] In an alternative embodiment, as Figure 3 shown, the sediment morphological features are extracted using a random uniform forest. When constructing a decision tree, the random uniform forest determines the selected nodes according to the division features, including: Step 3.1: Initialize the random uniform forest and set the number of decision trees and the maximum depth of each tree. Step 3.2: For each decision tree to be constructed, perform bootstrap sampling with replacement on the training samples to form the training subset of the current decision tree. Step 3.3: Recursively construct a decision tree starting from the root node. For the current node to be split, obtain the feature subset for the sample of the image region represented by the current node; randomly select a feature from the feature subset as the division feature and randomly generate a split point; divide the samples of the current node into child nodes according to the division feature and the split point; if the stop splitting condition is met, mark the current node as a leaf node and assign a class label; otherwise, repeat Step 3.3 for the child nodes. Step 3.4: Combine all the constructed decision trees into a random uniform forest model. Step 3.5: Input the image data of the sediment morphological features to be extracted into the random uniform forest model, and obtain the final sediment morphological features through majority voting or averaging.

[0045] Specifically, the basic parameters of the forest are set, including the total number of decision trees constituting the forest, such as 100 decision trees, and the maximum depth allowed for each decision tree, such as 20 layers. For each decision tree to be constructed in the forest, a bootstrap sampling method with replacement is adopted. That is, from the total training samples, for example, samples containing multiple sediment image areas and their known morphological features, samples are randomly selected to form a training subset exclusive to the decision tree.

[0046] Each decision tree starts from its root node and is constructed by recursive splitting. For the node to be split, the pre-selected feature subset corresponding to the image area sample represented by the node is first obtained, and a feature is randomly selected from the feature subset as the partition feature of the current split. For example, if the feature subset contains {average particle size, porosity}, the average particle size may be randomly selected as the partition feature. Then, a split point is randomly generated for this selected partition feature. For example, if the partition feature is the average particle size, and its value range in the node sample is 0.01mm to 2mm, a split point such as 0.5mm may be randomly generated. Based on the partition feature and the split point, the samples in the current node are assigned to two child nodes. For example, samples with an average particle size less than 0.5mm enter the left child node, while samples with an average particle size greater than or equal to 0.5mm enter the right child node. The splitting process will continue until the preset stop condition is met.

[0047] After all the preset number of decision trees, for example 100, are independently constructed according to step 3.3, these decision trees are organically combined to form the final random uniform forest model. Figure 4 A schematic diagram of a single decision tree is shown. When it is necessary to extract the morphological features of new, unknown sediment image data, the image data is input into the trained random uniform forest model. Each decision tree in the model will independently analyze the input data and output a prediction result (such as a category label). Finally, by taking a majority vote on the output results of all decision trees, for example, if 70 out of 100 trees judge a certain area as sandy sediment, the final morphological features of the area are sandy sediments or the average value is calculated to obtain the sediment morphological features corresponding to the input image data.

[0048] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some feature data can be ignored or not executed. Another point is that the displayed or discussed coupling, direct coupling, or communication connection to each other can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be in electrical, mechanical, or other forms.

[0049] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0050] In addition, each functional unit in various embodiments of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0051] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

Claims

1. An IoT-based dynamic monitoring method for layered sampling of marine sediments, characterized in that, Including: Step 1: Obtain the original marine environmental parameters that are transmitted through the Internet of Things communication module and are collected in real time by multiple sensor nodes deployed in the marine sediment collection area; Step 2: Calculate the adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters, and use the adaptive forgetting factor to perform Kalman filtering on the non-image original marine environmental parameters to obtain the filtered environmental parameters; Step 3: Select a feature subset from the feature set including features having a preset spatial correlation with the current image area, and determine the partitioning features from the feature subset, and extract the sediment morphology features using a random uniform forest. The random uniform forest determines the selected nodes according to the partitioning features when constructing the decision tree; fuse the filtered environmental parameters and the sediment morphology features to obtain a comprehensive state feature vector; Step 4: Based on the comprehensive state feature vector, and use a preset hierarchical acquisition model to calculate the adjusted acquisition depth and acquisition speed of the marine sediment collection device, and send the adjusted acquisition depth and acquisition speed to the control system of the collection device through the Internet of Things communication module.

2. The method according to claim 1, wherein The calculating the adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters includes: Receiving the original marine environmental parameters, and extracting multiple indicators related to the transition of the sediment layer interface. The multiple indicators include the pressure change rate, depth change rate, image texture mutation degree, and temperature gradient; Setting the weight coefficients of the multiple indicators respectively, and performing weighted fusion on the multiple indicators according to the weight coefficients to obtain an initial transition significance score; Obtaining historical sediment stratification data and current acquisition target layer information, and dynamically adjusting the threshold of the initial transition significance score according to the historical sediment stratification data and current acquisition target layer information; When the initial transition significance score exceeds the dynamically adjusted threshold, setting the adjustment amplitude of the adaptive forgetting factor to a first preset adjustment amplitude value, otherwise, setting the adjustment amplitude of the adaptive forgetting factor to a second preset adjustment amplitude value, where the first preset adjustment amplitude value is greater than the second preset adjustment amplitude value; Calculating the final adaptive forgetting factor according to the adjustment amplitude of the adaptive forgetting factor and a preset reference forgetting factor.

3. The method according to claim 1, wherein The performing Kalman filtering on the non-image original marine environmental parameters using the adaptive forgetting factor to obtain the filtered environmental parameters includes: Initializing the state vector and covariance matrix of the Kalman filter, where the state vector includes the non-image marine environmental parameters to be estimated; Based on the system dynamics model, performing a one-step prediction on the state vector and covariance matrix at the current moment to obtain a predicted state vector and a predicted covariance matrix; Obtaining the non-image original marine environmental parameters at the current moment as the observation value; Calculating the Kalman gain using the predicted state vector, predicted covariance matrix, the observation value, and the adaptive forgetting factor; Update the state vector by combining the predicted state vector, the observation value, and the Kalman gain to obtain the filtered environmental parameters at the current moment; Update the covariance matrix for filtering at the next moment.

4. The method according to claim 1, characterized in that The selecting a feature subset from a set of features having a preset spatial correlation with the current image region includes: Preprocess the image data in the original marine environmental parameters, extract underlying visual features, and preliminarily group the underlying visual features according to physical meaning and spatial scale to form an initial set of features; Obtain the macroscopic features of the current image region to be processed, where the macroscopic features include the average depth information and color information of the region; Judge the type of sediment morphological features to be distinguished currently according to the macroscopic features; Based on the judged type of sediment morphological features, adjust the prior probability that a feature group related to the judged type of sediment morphological features and having a strong spatial correlation with the current image region in the initial feature set is selected to obtain an adjusted feature set; Randomly extract a specified number of features from the adjusted feature set according to the adjusted prior probability to form the feature subset.

5. The method according to claim 1, characterized in that, The adopting a random uniform forest to extract sediment morphological features, where when constructing a decision tree, the random uniform forest determines the selected nodes according to the partitioning features, includes: Step 3.1: Initialize the random uniform forest and set the number of decision trees and the maximum depth of each tree; Step 3.2: For each decision tree to be constructed, perform bootstrap sampling with replacement on the training samples to form a training subset of the current decision tree; Step 3.3: Recursively construct a decision tree starting from the root node. For the current node to be split, obtain the feature subset for the samples of the image region represented by the current node; randomly select a feature from the feature subset as the partitioning feature and randomly generate a split point; partition the samples of the current node into child nodes according to the partitioning feature and the split point; if the stop splitting condition is met, mark the current node as a leaf node and assign a class label; otherwise, repeat Step 3.3 for the child nodes; Step 3.4: Combine all the constructed decision trees into a random uniform forest model; Step 3.5: Input the image data for which the morphological features are to be extracted into the random uniform forest model, and obtain the final sediment morphological features through majority voting or averaging.

6. An ocean sediment layered sampling dynamic monitoring system based on the Internet of Things, characterized in that, including: A data acquisition unit for obtaining the original marine environmental parameters that are transmitted through the Internet of Things communication module and are collected in real time by a plurality of sensor nodes deployed in the marine sediment collection area; A filtering unit for calculating an adaptive forgetting factor of the Kalman filtering algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters, and performing Kalman filtering on the non-image original marine environmental parameters by using the adaptive forgetting factor to obtain the filtered environmental parameters; A feature extraction unit, which is used to select a feature subset from a set of features having a preset spatial correlation with the current image region, determine a partitioning feature from the feature subset, and extract sediment morphology features using a random uniform forest. When constructing a decision tree, the random uniform forest determines the selected nodes according to the partitioning feature; fuse the filtered environmental parameters and the sediment morphology features to obtain a comprehensive state feature vector; A monitoring unit, which is used to calculate the adjusted collection depth and collection speed of the marine sediment collection device based on the comprehensive state feature vector and using a preset hierarchical collection model, and send the adjusted collection depth and collection speed to the control system of the collection device through the Internet of Things communication module.

7. The system according to claim 6, characterized in that, The calculation of the adaptive forgetting factor of the Kalman filter algorithm according to the significance indicating the transition of the sediment layer interface in the original marine environmental parameters includes: Receiving the original marine environmental parameters, and extracting a plurality of indicators related to the transition of the sediment layer interface. The plurality of indicators include a pressure change rate, a depth change rate, an image texture mutation degree, and a temperature gradient; Setting weight coefficients for each of the plurality of indicators, and performing weighted fusion on the plurality of indicators according to the weight coefficients to obtain an initial transition significance score; Obtaining historical sediment stratification data and current collection target layer information, and dynamically adjusting the threshold of the initial transition significance score according to the historical sediment stratification data and the current collection target layer information; When the initial transition significance score exceeds the dynamically adjusted threshold, setting the adjustment amplitude of the adaptive forgetting factor to a first preset adjustment amplitude value; otherwise, setting the adjustment amplitude of the adaptive forgetting factor to a second preset adjustment amplitude value. The first preset adjustment amplitude value is greater than the second preset adjustment amplitude value; Calculating the final adaptive forgetting factor according to the adjustment amplitude of the adaptive forgetting factor and a preset reference forgetting factor.

8. The system according to claim 6, wherein The use of the adaptive forgetting factor to perform Kalman filtering on the non-image original marine environmental parameters to obtain filtered environmental parameters includes: Initializing the state vector and covariance matrix of the Kalman filter. The state vector includes non-image marine environmental parameters to be estimated; Based on the system dynamics model, performing a one-step prediction on the state vector and covariance matrix at the current moment to obtain a predicted state vector and a predicted covariance matrix; Obtaining the non-image original marine environmental parameters at the current moment as the observed value; Calculating the Kalman gain using the predicted state vector, the predicted covariance matrix, the observed value, and the adaptive forgetting factor; Combining the predicted state vector, the observed value, and the Kalman gain to update the state vector to obtain the filtered environmental parameters at the current moment; Updating the covariance matrix for filtering at the next moment.

9. The system according to claim 6, characterized in that, The selection of a feature subset from a set of features having a preset spatial correlation with the current image region includes: Preprocess the image data in the original marine environmental parameters, extract the underlying visual features, and preliminarily group the underlying visual features according to physical meaning and spatial scale to form an initial feature set; Obtain the macroscopic features of the current image area to be processed, where the macroscopic features include the average depth information and color information of the area; Judge the type of sediment morphological features that need to be distinguished currently according to the macroscopic features; Based on the judged type of sediment morphological features, adjust the prior probability of selection of the feature group in the initial feature set that is related to the judged type of sediment morphological features and has a strong spatial correlation with the current image area to obtain an adjusted feature set; Randomly extract a specified number of features from the adjusted feature set according to the adjusted prior probability to form the feature subset.

10. The system according to claim 6, wherein When using a random uniform forest to extract sediment morphological features, the random uniform forest determines the selected nodes according to the partitioning features when constructing a decision tree, including: Step 3.1: Initialize the random uniform forest and set the number of decision trees and the maximum depth of each tree; Step 3.2: For each decision tree to be constructed, perform bootstrap sampling with replacement on the training samples to form the training subset of the current decision tree; Step 3.3: Recursively construct a decision tree starting from the root node. For the current node to be split, obtain the feature subset for the sample of the image area represented by the current node; randomly select a feature from the feature subset as the partitioning feature and randomly generate a split point; partition the samples of the current node into child nodes according to the partitioning feature and the split point; if the stop splitting condition is met, mark the current node as a leaf node and assign a class label; otherwise, repeat Step 3.3 for the child nodes; Step 3.4: Combine all the constructed decision trees into a random uniform forest model; Step 3.5: Input the image data of the sediment morphological features to be extracted into the random uniform forest model, and obtain the final sediment morphological features through majority voting or averaging.

Images