Nondestructive intelligent staging method and system for oocytes of blue crabs

Abstract

The invention discloses a lossless intelligent staging method and system for oocytes of scylla paramamosain. The lossless intelligent staging method comprises the following steps: obtaining morphological parameters of living scylla paramamosain at least including the first length and the body weight; the morphological parameters are input into an XGBoost-based feature enhancement regression prediction model which is subjected to Pipeline packaging training in advance and has optimal parameters, and an oocyte diameter prediction value is obtained; accurately matching a preset staging standard according to the predicted value, and determining an ovarian development staging; and when the prediction confidence coefficient is low, triggering identification and carrying out incremental learning on the model by utilizing confirmation data so as to adaptively optimize the model precision. According to the method, machine learning is used for replacing anatomy to realize nondestructive measurement of the oocyte diameter of the blue crab, and a diameter direct connection ovary development stage table automatic matching development stage is predicted. The simplest configuration of the method only needs the measurement of the nail length and the body weight as effective parameters, the lightweight model can be applied to a quick online lossless ovarian development staging system of the blue crabs, and the intelligent staging can effectively guide the actual breeding decision.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, specifically to a non-destructive intelligent staging method and system for mud crab oocytes based on machine learning technology, and in particular, a method and system for automatically detecting and evaluating the degree of ovarian development using morphological parameters. Background Technology

[0002] The mud crab (Scylla paramamosain) is an important marine aquaculture species in coastal areas of my country. The degree of ovarian development directly affects the market price of mud crabs with roe. Red-roe female crabs are plump with roe and meat, and their gonads have high nutritional value, making them a prized seafood ingredient. Furthermore, gonad development is fundamental to crab reproduction; a good level of gonad development can improve the reproductive efficiency of breeding crabs and ensure a large quantity of high-quality seedlings.

[0003] The mud crab has a thick carapace with deep pigment deposition, making it difficult to accurately obtain ovarian development characteristics using conventional optical methods or high-cost infrared imaging. Traditional methods for detecting ovarian development in mud crabs rely on manual experience or dissection and microscopy, which are invasive—dissection and sampling lead to the death of the crab and make repeated monitoring impossible; subjective—visual observation is easily affected by experience differences and has a high misjudgment rate; and inefficient—single-animal testing is time-consuming and difficult to scale up. Judging ovarian development solely based on the direct relationship between plumpness or weight and stage has significant errors. Even quantifying the total gonadal content of the mud crab cannot directly determine the size of the oocyte and the accurate ovarian development stage based on the total gonadal content.

[0004] Crabs typically develop rapidly after reproductive molting. Based on changes in egg cell composition and morphology, the first ovarian development in female crabs is divided into six stages (stages I-VI). Ovarian development begins in stage II after reproductive molting, yolk sac formation mainly occurs in stage III when the ovary matures, and development is essentially complete by stage V, culminating in spawning. However, in actual production, the number of development days or food intake does not directly correspond to the actual ovarian development. Malnutrition, disease, and environmental fluctuations can all affect the gonadal development progress of mud crabs, leading to consequences such as precocious or delayed sexual maturation, thus hindering industry development. Therefore, there is an urgent need to develop a non-destructive, low-cost, high-precision, rapid, and efficient method for predicting the developmental stages of mud crab oocytes. Summary of the Invention

[0005] Oocyte diameter defines cell maturity and oocyte stage. Currently, oocyte diameter is the most scientific and accurate biological benchmark for researchers or aquaculture technicians to classify oocyte stages (i.e., ovarian development stages). This invention overcomes the limitations of subjective staged stage based on experience by using a feature-enhanced regression model based on XGBoost to replace traditional anatomy. Through machine learning, it accurately captures the complex nonlinear mapping between growth traits such as shell length, body weight, and ovarian development, constructing a non-destructive prediction model for the ovarian development of mud crabs. XGBoost uses a gradient descent framework to optimize a specific objective function. The direction of new tree construction is along the negative gradient of the loss function relative to the current model's prediction, reducing bias and gradually improving model performance. This model achieves non-destructive detection of live mud crabs in 1 second per crab, with a survival rate of 100%. The correlation coefficient between the actual and predicted oocyte diameter values ​​in the training set is 0.99, and the correlation coefficient between the actual and predicted oocyte diameter values ​​in the test set is 0.83. Small-scale ovarian stage prediction is 100% accurate. This invention achieves non-destructive, rapid, and accurate ovarian staging in live mud crabs through precise prediction of oocyte diameter, and possesses continuous self-optimization capabilities. It can guide aquaculture decisions at different ovarian development stages, avoiding losses from early / late harvesting and obtaining high-quality commercial mud crabs / breeding crabs. This has significant theoretical and practical value for the preservation, breeding, and promotion of new mud crab varieties, thus addressing the problems mentioned in the background art.

[0006] To achieve the above-mentioned objectives, one aspect of the present invention provides a non-destructive intelligent staging method for mud crab oocytes, comprising the following steps:

[0007] Step S1: Collect the first ovarian development period of the mud crab modeling sample and obtain the morphological parameters of the live mud crab. The morphological parameters include at least carapace length and body weight.

[0008] Step S2: Input the morphological parameters into the pre-trained oocyte diameter prediction model to obtain the predicted oocyte diameter value; wherein, the oocyte diameter prediction model is based on a machine learning algorithm, with the carapace length and weight of the sample crab as independent variables and the anatomical true value of the oocyte diameter as the dependent variable for training;

[0009] Step S3: Match the predicted oocyte diameter value with the preset ovarian development stage standard to non-destructively determine the current ovarian development stage of the live mud crab.

[0010] Step S4: When the confidence level of the prediction result is lower than the threshold, a manual verification request is triggered, and the data confirmed by the manual verification is used as a new sample to perform incremental learning on the prediction model in order to adaptively optimize the model accuracy.

[0011] Furthermore, in step S1, the mud crab sample is a live mud crab in the ovarian development stage. The weight of the mud crab sample is measured in grams (g), with a measurement accuracy of 0.1g. The carapace length of the mud crab sample is the distance from the central concave tooth at the front of the shell to the central tooth at the end of the shell, measured in millimeters (mm), with a measurement accuracy of 0.01mm.

[0012] Furthermore, in step S2, the true value of the oocyte diameter of the mud crab modeling sample is the average value of the oocyte diameter measured 6 times in the stained sections of ovarian tissue after dissecting the mud crab, in micrometers (μm), with a measurement accuracy of 0.01 μm.

[0013] Furthermore, step S2 specifies that the machine learning algorithm is the XGBoost algorithm, specifically a feature-enhanced regression model based on XGBoost:

[0014] ,

[0015] Where: y^log is the logarithmic transformation value of the oocyte diameter predicted by the model; K is the number of trees; It is the k-th decision tree, and each tree maps the feature vector x to the score of a leaf node; x is the feature vector after feature engineering (only some important features are retained after feature selection).

[0016] Furthermore, step S2 includes feature engineering, which generates 14 new features based on the input morphological parameters (arm length and weight) using a custom converter. These new features are automatically filtered based on the importance of XGBoost using SelectFromModel, ultimately retaining the most effective subset of features for modeling. The 14 new features include:

[0017] 1. A1 (length of A)_squared: The squared value of the characteristic length of A.

[0018] The formula is: A1 2 .

[0019] 2. A2 (weight)_squared: The squared value of the characteristic weight.

[0020] The formula is: A2 2 .

[0021] 3. A1 (Ag length) _ A2 (Weight) _ interaction: The interaction feature (product) between ag length and weight.

[0022] The formula is: A1×A2.

[0023] 4. A1 (length of A)_log: Logarithmic transformation of length A.

[0024] The formula is: log(1+A1).

[0025] 5. A2 (weight)_over_A1 (arm length)_cubed: weight divided by the cube of arm length.

[0026] The formula is: .

[0027] 6. A1 (length of A) _normalized: Min-Max normalization based on the length of A in the training set.

[0028] The formula is: .

[0029] 7. A2 (body weight)_zscore: Z-Score standardization based on the weight of the training group (μ is the mean, σ is the standard deviation).

[0030] The formula is: .

[0031] 8.A1 (length of A)_capped: Truncates outliers using the 1% and 99% quantiles of the length of A in the training set.

[0032] The formula is: clip(A1, Q) 1% (A1 train ), Q 99% (A1 train )).

[0033] 9.A1 (length of A)_inverse: the reciprocal of length A.

[0034] The formula is: .

[0035] 10. A1 (Ag length) _ A2 (Weight) _ ratio: The ratio of ag length to weight.

[0036] The formula is: .

[0037] 11.A1 (length of A) _cubed: The cube value of length A.

[0038] The formula is: A1 3 .

[0039] 12.A2(weight)_sqrt: the square root of the weight.

[0040] The formula is: .

[0041] 13. A1 (length of A) _sin: Sine transformation of length A (capturing periodicity).

[0042] The formula is: sin(A1).

[0043] 14. A1 (length of A) cos: cosine transform of length A (capturing periodicity).

[0044] The formula is: cos(A1).

[0045] The statistical results are as follows:

[0046] category Number of features Example Polynomial characteristics 3 A1², A2², A1³ Interactive features 2 A1×A2, A1 / A2 Statistical Transformation 3 Normalization, standardization, truncation Nonlinear transformation 6 log, sin, cos, sqrt

[0047] Furthermore, the feature engineering performs a logarithmic transformation on the target variable, oocyte diameter.

[0048] Furthermore, in step S2, a feature enhancement regression model based on XGBoost is constructed using a Pipeline. This Pipeline includes feature engineering, standardization, feature selection, and XGBoost regression. The Pipeline combines multiple data processing steps (such as feature extraction, preprocessing, and model training) into a single machine learning pipeline, preventing data leakage and enabling simultaneous tuning of the parameters of the preprocessing steps and the model, thus facilitating process management and reuse.

[0049] Furthermore, in step S2, the synchronous tuning of model parameters employs a randomized search (Randomized SearchCV) with 3-fold cross-validation, performing 30 parameter combination evaluations. The optimal configuration is automatically selected from a simplified parameter grid, with negative mean absolute error (MAE) as the optimization objective. The tuning parameters include:

[0050] a) Model structure: n_estimators (number of trees), max_depth (tree depth);

[0051] b) Learning control: learning_rate (learning rate), subsample (sample sampling rate);

[0052] c) Feature sampling: colsample_bytree (feature sampling rate);

[0053] d) Regularization: reg_alpha(L1), reg_lambda(L2), gamma (split threshold);

[0054] e) Subtree control: min_child_weight (child node weight);

[0055] f) Feature selection: threshold (importance threshold).

[0056] Furthermore, in step S2, the predictive performance of the optimal model is determined by comparing the coefficients of determination (R²) among different models. 2The mean absolute error (MAE), the Pearson correlation coefficient (r) between the predicted and actual values, and the mean absolute percentage error (MAPE) are used for evaluation.

[0057] The coefficient of determination (R²) measures the proportion of the variance of the target variable that the model explains. Its formula is:

[0058] R² = 1 - (SS_res / SS_tot),

[0059] in:

[0060] (Sum of squared residuals)

[0061] (Total sum of squares)

[0062] Mean Absolute Error (MAE): This is the average of the absolute errors between the predicted and actual values. Its formula is:

[0063] ,

[0064] Where n is the number of samples.

[0065] Pearson Correlation Coefficient (r): The Pearson correlation coefficient measures the degree of linear correlation between two variables. Its formula is as follows:

[0066] ,

[0067] Calculate the predicted value The correlation coefficient between the actual value (y) and the true value (y), therefore x_i in the formula is... , The mean of the predicted values y_i is the actual value. It is the mean of the true values.

[0068] Therefore, specifically regarding the correlation coefficient between predicted and actual values:

[0069] .

[0070] Mean Absolute Percentage Error (MAPE): A metric for evaluating prediction accuracy, calculated using the following formula:

[0071] ,

[0072] Where y_i is the actual value. It is a predicted value. It is the average of the true values.

[0073] Furthermore, in step S3, the oocyte diameter of the mud crab test sample calculated by the optimal model is restored to its original size through inverse transformation. Then, the ovarian stage is divided according to the correspondence table between the ovarian development stage and the oocyte diameter, as shown in the table below:

[0074] Correspondence table of ovarian development stages and oocyte diameter in blue crabs

[0075] (Aquaculture 2019 Volume 515, 734560)

[0076] Developmental stage Average diameter (μm) Dominant cell type Phase I 5～30 Oogonia (Og), a small number of stage I oocytes (Oc1) Phase II 30～70 Stage II oocytes (Oc2) Phase III 70～120 Stage III oocytes (Oc3) Phase IV 120～200 Stage IV oocytes (Oc4) Phase V 200～260 Stage IV oocytes (mature state)

[0077] Another aspect of the present invention provides a non-destructive intelligent staging system for mud crab oocytes, used to assess the developmental stages of live mud crab oocytes, comprising: an ovarian development data collection module, a lightweight model construction module, an intelligent staging decision module, and a model performance adaptive optimization module, wherein:

[0078] Ovarian development data collection module: responsible for collecting standardized multimodal data required for model training and validation.

[0079] Implementation method:

[0080] We collected mud crabs at different stages of their first ovarian development as modeling samples.

[0081] Determine the morphological parameters of the sample: measure and record the nail length, and simultaneously record the weight.

[0082] Determining the diameter of oocytes in the samples: The above-mentioned mud crab samples were dissected, ovarian histological sections were prepared, and the diameter of oocytes was accurately measured by microscopic imaging. This data was used as the ground truth for model training.

[0083] The module will automatically generate a standardized data record for each mud crab sample, containing [shell length, weight, oocyte diameter], and upload it to establish an ovarian development database.

[0084] Lightweight model building module: Utilizes the data from the collection module to build and select the regression model with the highest prediction accuracy.

[0085] Implementation method:

[0086] Data cleaning and preprocessing were performed using the length and weight of the oocyte obtained from the collection module as independent variables and the diameter of the oocyte as the dependent variable.

[0087] This module constructs a Pipeline-encapsulated feature-enhanced regression machine learning model based on XGBoost. The model's predicted oocyte diameter is compared with the actual oocyte diameter to validate its performance and obtain optimal model parameters for deployment of the optimized lightweight model. The module outputs a trained model file (e.g., .pkl format) and a corresponding model performance evaluation report.

[0088] Intelligent staging decision module: Uses the optimal model to perform non-destructive testing on live mud crabs and outputs the ovarian development staging results.

[0089] Implementation method:

[0090] Users measure the carapace length and weight of the live mud crabs to be tested.

[0091] The module calls the optimal model that has been deployed (in the cloud or locally), inputs morphological parameters, and calculates the predicted oocyte diameter in real time.

[0092] The system has a pre-built knowledge base of phased rules, which can be customized and modified by experts according to different crab species.

[0093] The module automatically matches the predicted diameter value with the knowledge base rules and immediately outputs the ovarian development stage and prediction confidence level.

[0094] Model performance adaptive optimization module: Enables continuous iteration and performance improvement of the model in practical applications, forming a closed-loop learning system.

[0095] Implementation method: a complete closed loop of discovering uncertainty -> manual annotation -> incremental learning.

[0096] The module monitors the output confidence level of the intelligent staging decision module in real time. When the confidence level is lower than a preset threshold, the system automatically marks the prediction as a "low-confidence prediction." For low-confidence samples, the system will proactively prompt the user to submit the crab sample to an expert for manual identification. The identified oocyte diameter and developmental stage, along with the corresponding carapace length and weight data, constitute a new high-quality training sample pair. This module periodically or quantitatively initiates an incremental training process, integrating the new sample set into the original training set, fine-tuning and updating the optimal basic model, and generating an upgraded model.

[0097] After verification, the upgraded model can seamlessly replace the old model, achieving silent model upgrades, continuously evolving the system's intelligence and accuracy, and adapting to the differences in mud crabs under different regions and farming conditions.

[0098] Compared with existing technologies, this system and method have the following advantages:

[0099] 1. Methodological innovation: For the first time, this method has achieved accurate prediction of the diameter of internal oocytes by using only external non-destructive measurement parameters such as shell length and body weight, replacing anatomy. It also provides a quantitative and objective standard for judging the stage of ovarian development (predicted diameter value), replacing the traditional subjective and qualitative visual observation that relies on experience, thus improving accuracy and consistency and overcoming the technical challenge of assessing the ovarian development status of live mud crabs.

[0100] 2. Model Innovation: A dedicated lightweight prediction model based on feature engineering and XGBoost ensemble learning, encapsulated in Pipeline, was constructed. By exploring the deep nonlinear relationship between morphological parameters and biological state, the high accuracy and reliability of the prediction were ensured.

[0101] 3. System Innovation: A self-evolution mechanism based on active learning and incremental learning is introduced, enabling the system to continuously optimize using expert feedback. This achieves self-evolution and improvement of performance in practical applications, laying a solid foundation for industrial applications. Attached Figure Description

[0102] Figure 1 This is a flowchart of an intelligent staging method for mud crab oocytes based on morphological parameters.

[0103] Figure 2 A scatter plot showing the linear correlation between predicted and actual values.

[0104] Figure 3 This is a graph showing the operation of the prediction model.

[0105] Figure 4 Architecture diagram of a self-learning, non-destructive online assessment system for ovarian development in blue crabs Detailed Implementation

[0106] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0107] Example 1: such as Figure 1 The diagram shown is a flowchart of the method of the present invention. An embodiment of the present invention provides a non-destructive prediction algorithm for the diameter and developmental stage of mud crab oocytes. The specific steps are as follows:

[0108] Step S1, data acquisition.

[0109] 424 mud crabs in their first ovarian development stage that were complete and without deformities were collected. Their body weight was measured to an accuracy of 0.1g, the carapace length was measured to an accuracy of 0.01mm, and the diameter of the actual oocytes of the dissected crabs was measured to an accuracy of 0.01μm through stained sections. The crabs were divided into a training set (296 crabs) and a test set (128 crabs) to ensure consistent data distribution.

[0110] Step S2: Construct a feature-enhanced regression model based on XGBoost using the training set.

[0111] First, a custom transformer generates 14 new features (squared terms, interaction terms, statistical features, etc.) based on the input nail length and weight. These features are then automatically filtered using SelectFromModel based on XGBoost importance, ultimately retaining the most effective subset for modeling. The filtered feature set consists of 8 features: A1 (nails length), A2 (weight), A1_A2_interact, A2_over_A1_cubed, A1_capped, A1_A2_ratio, A1_sin, and A1_cos.

[0112] Then, a logarithmic transformation was performed on the target variable, oocyte diameter.

[0113] Further establish a pipeline that includes feature engineering, standardization, feature selection, and XGBoost regression to prevent data leakage, enable simultaneous tuning of preprocessing steps and model parameters, and facilitate process management and reuse.

[0114] Then, a randomized search (Randomized SearchCV) with 3-fold cross-validation is used to perform 30 parameter combination evaluations, with negative mean absolute error (MAE) as the optimization objective, automatically selecting the optimal configuration from the simplified parameter grid. The tuning parameters include:

[0115] a) Model structure: n_estimators (number of trees), max_depth (tree depth)

[0116] b) Learning control: learning_rate (learning rate), subsample (sample sampling rate)

[0117] c) Feature sampling: colsample_bytree (feature sampling rate)

[0118] d) Regularization: reg_alpha (L1), reg_lambda (L2), gamma (split threshold)

[0119] e) Subtree control: min_child_weight (child node weight)

[0120] f) Feature selection: threshold (importance threshold)

[0121] The optimized parameter grid is: param_grid = {

[0122] 'model__n_estimators': [100, 200],

[0123] 'model__max_depth': [5, 10],

[0124] 'model__learning_rate': [0.05, 0.1],

[0125] 'model__subsample': [0.8, 1.0],

[0126] 'model__colsample_bytree': [0.8, 1.0],

[0127] 'model__reg_alpha': [0, 0.5],

[0128] 'model__reg_lambda': [0, 0.5],

[0129] 'model__gamma': [0, 0.1],

[0130] 'model__min_child_weight': [1, 3],

[0131] 'feature_selection__threshold': [0.01, 0.05, 'mean', 'median'],

[0132] }

[0133] The optimal parameter configuration is as follows:

[0134] parameter value model__subsample 0.8 model__reg_lambda 0 model__reg_alpha 0 model__n_estimators 200 model__min_child_weight 3 model__max_depth 10 model__learning_rate 0.05 model__gamma 0 model__colsample_bytree 0.8 feature_selection__threshold 0.01

[0135] Step S3, Model Prediction. The carapace length and weight of samples from the training set (296 crabs) and test set (128 crabs) of 424 crabs were input into the model with the above optimal parameters. An inverse transformation was performed to obtain the predicted oocyte diameter for each sample, which was then compared with the actual oocyte diameter. The comparison results are shown in Table 1 and... Figure 2As shown. Using macroscopic phenotypes to predict cell parameters, the correlation coefficient between the true and predicted values ​​in the model training set is as high as 0.9996, and the correlation coefficient between the true and predicted values ​​in the model test set is greater than 0.8, reflecting a linear relationship between the prediction and the true values, indicating high model accuracy. The R² evaluation of the overall model goodness of fit is greater than 0.7, indicating a strong ability to explain variance. MAE represents the absolute magnitude of the prediction error; the actual oocyte prediction error is 10 μm, which is within an acceptable range. The MAPE value represents the average percentage difference between the predicted and actual values; the average deviation of the prediction from the actual value is 8.2%, and less than 10% indicates good accuracy.

[0136] Table 1 Performance Indicators of the Blue Crab Oocyte Diameter Prediction Model

[0137] index training set test set Correlation coefficient r 0.9996 0.8256 MAE 0.7501 10.0581 R² Score 0.9992 0.7664 MAPE 0.60 8.20

[0138] Step S4, Model Application. Collect 10 live mud crab samples, measure the carapace length and weight of the crabs, and input...

[0139] The model with the above optimal parameters was used to obtain the predicted oocyte diameter of live mud crabs through inverse transformation, and matched to the ovarian stage. Dissection was then performed to verify the prediction accuracy, and the actual oocyte diameter was measured by stained sections for comparison. The final measurement results are shown in Table 2 and... Figure 3 As shown.

[0140] Table 2. Comparison of machine learning predictions and anatomical true values ​​for the diameter of mud crab oocytes.

[0141] Serial Number Length of nail (mm) Body weight (g) Actual oocyte diameter (μm) Predicting oocyte diameter (μm) Relative error (%) Actual ovarian staging Predicting ovarian staging 1 95.34 421.0 100.00 100.20 0.20 Ⅲ Ⅲ 2 89.08 369.0 60.00 59.91 0.15 Ⅱ Ⅱ 3 91.96 418.0 90.00 90.75 0.83 Ⅲ Ⅲ 4 88.22 408.0 66.67 66.79 0.18 Ⅱ Ⅱ 5 91.20 391.1 99.66 107.73 8.10 Ⅲ Ⅲ 6 93.23 310.0 100.00 101.35 1.35 Ⅲ Ⅲ 7 85.20 327.0 136.67 138.09 0.10 Ⅳ Ⅳ 8 88.50 400.0 85.51 76.58 10.44 Ⅲ Ⅲ 9 87.80 369.0 160.00 153.59 4.00 Ⅳ Ⅳ 10 80.34 277.58 203.48 201.01 1.21 Ⅴ Ⅴ

[0142] The results of measurements on 10 randomly selected mud crabs showed that the above model had high accuracy in predicting the diameter of mud crab oocytes, with an average relative error of 2.66% and a minimum of 0.1%, and 100% accuracy in predicting ovarian stage. This indicates that the model is stable and performs well. The entire process, from input parameters to output, takes only 1 second per crab, making it easy to operate. The entire process only requires measurements of carapace length and weight, and the predicted diameter automatically matches the developmental stage, achieving in vivo non-destructive assessment and intelligent staging of mud crab ovarian development. Compared with manual staging, it significantly reduces costs and increases efficiency, making it worthy of widespread application.

[0143] Example 2: Intelligent system for non-destructive online assessment of ovarian development in mud crabs with automatic measurement of morphological parameters

[0144] The system of this invention can integrate mobile phone camera ranging technology and a dedicated app for Bluetooth electronic scales. By defining the measurement process through software, it greatly simplifies hardware requirements and balances portability and rapid measurement.

[0145] Simple Sensor Configuration Implementation Examples ( Figure 4 ):

[0146] I. Length measurement (using mobile phone camera + computer vision):

[0147] Technical implementation: Users simply place the crab next to a simple, standardized reference object (such as a card with a QR code and a known size ruler) and take a picture with a mobile app.

[0148] Workflow:

[0149] The app automatically recognizes the reference objects and the outline of the blue crab.

[0150] Using image recognition algorithms, the carapace length of the blue crab is automatically calculated based on the known dimensions of a reference object.

[0151] This method eliminates the hassle of manual measurement for users, is highly accurate, and is extremely easy to operate, making it ideal for use in fields or near aquaculture ponds.

[0152] II. Weight Measurement (using a Bluetooth electronic scale):

[0153] Technical Implementation: A low-cost Bluetooth smart electronic scale is used. Users place the crabs on the scale, and the weight data is automatically and wirelessly transmitted to a mobile app via Bluetooth.

[0154] Advantages: It avoids the errors and inconveniences of manual data entry and achieves automatic data collection and synchronization.

[0155] III. Data Synchronization and App Integration:

[0156] Process: The user opens the App, places the mud crab on the reference card and takes a picture (automatically obtaining the shell length), then places the mud crab on the Bluetooth scale (the weight is automatically synchronized to the App). The App automatically sends the "shell length + weight" data to the cloud or a locally deployed lightweight model, instantly obtaining the predicted value of the oocyte diameter and the developmental stage results.

[0157] Advantages: The model finally deployed according to this invention is a lightweight model that has been optimized by pruning and quantization and is encapsulated in Pipeline. It can be embedded in a mobile app to run, realize offline prediction, protect user data privacy, and adapt to field environments with poor network conditions.

[0158] IV. Hardware Configuration:

[0159] The entire system requires only a low-cost Bluetooth electronic scale and a standardized reference card as external hardware. The sensor configuration is extremely simple and the cost is very low, making it very suitable for large-scale promotion.

[0160] V. Database Updates and Continuous Model Optimization:

[0161] The system automatically generates a standardized data record for each mud crab sample, containing [carapace length, body weight, oocyte diameter], and uploads it to a cloud database.

[0162] The system monitors the output confidence level of the intelligent staging decision module in real time. When the confidence level falls below a preset threshold (e.g., <85%), the system automatically marks the prediction as a "low-confidence prediction." For low-confidence samples, the app proactively prompts the user to submit the crab sample to an expert for manual identification (morphological or anatomical assessment). The expert identification results (confirmed oocyte diameter and developmental stage) along with the corresponding carapace length and weight data constitute a new high-quality training sample pair. This module periodically (e.g., monthly) or quantitatively (e.g., after adding 100 samples) initiates an incremental training process, integrating the new sample set into the original training set, fine-tuning and updating the optimal base model, and generating an upgraded model. After validation, the upgraded model can seamlessly replace the old model, achieving silent model upgrades. This makes the system increasingly intelligent with use, continuously evolving its accuracy, and adapting to the differences in crabs across different regions and farming conditions.

[0163] This system only requires measurements of carapace length and weight throughout the entire process. Compared with manual staging, it improves the efficiency and accuracy of judging ovarian development, reduces the cost of accidental capture, and obtains high-quality commercial crabs / breeding crabs.

[0164] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-destructive intelligent staging method and system for blue crab oocyte, characterized in that, The method comprises the following steps: Step S1, collecting the first modeling sample of Scylla paramamosain in the first ovary development stage, and obtaining morphological parameters of the live crab, wherein the morphological parameters at least include carapace length and body weight; Step S2, inputting the morphological parameters into a pre-trained oocyte diameter prediction model to obtain a predicted oocyte diameter value; wherein the oocyte diameter prediction model is a Pipeline-encapsulated feature enhancement regression model based on XGBoost, and the carapace length and body weight of the sample crab are used as independent variables, and the anatomical true value of the oocyte diameter is used as a dependent variable to obtain the oocyte diameter prediction model through optimization training; Step S3, matching the predicted oocyte diameter value with the preset ovary development staging standard to determine the current ovary development staging of the live Scylla paramamosain without loss; Step S4, when the confidence of the prediction result is lower than a threshold, triggering an artificial identification request, and using the data confirmed by the artificial identification as a new sample to perform incremental learning on the prediction model to adaptively optimize the model precision.

2. The method according to claim 1, wherein, The prediction model in step S2 is a feature enhancement regression model based on XGBoost, and the formula is as follows: ， wherein: y^log is the model predicted log-transformed value of the oocyte diameter; K is the number of trees; is the kth decision tree, each tree maps the feature vector x to a score of a leaf node; x is the feature vector after feature engineering.

3. The method according to claim 2, wherein, The feature enhancement regression model based on XGBoost in step S2 includes feature engineering, which generates 14 new features from the input morphological parameters through a self-defined converter, and the new features are automatically selected based on the importance of XGBoost through SelectFromModel, and finally the most effective feature subset is reserved for modeling.

4. The method according to claim 3, wherein, The feature engineering performs logarithmic transformation on the target variable oocyte diameter.

5. The method of claim 1, wherein the method is characterized by, In step S2, the Pipeline encapsulation establishes a Pipeline overall pipeline including feature engineering, standardization, feature selection and XGBoost regression, prevents data leakage, optimizes the parameters of the preprocessing step and the model at the same time, and facilitates process management and reuse.

6. The method of claim 1, wherein the method is characterized by, In step S2, the model structure, learning control, feature sampling, regularization, subtree control and feature selection parameters are simultaneously optimized, a random search with 3-fold cross-validation is adopted, 30 parameter combination evaluations are performed, the negative mean absolute error is used as the optimization target, and the optimal configuration is automatically selected from the simplified parameter grid.

7. The method of claim 1, wherein the method is characterized by, In step S3, the accurate division of the ovary staging is based on the oocyte diameter, the oocyte diameter of the crab detection sample obtained by inputting the prediction model is reduced to the original scale after inverse transformation, and the accurate division of the ovary staging is performed according to the corresponding table of the ovary development stage and the oocyte diameter of the crab.

8. A non-destructive intelligent staging system for blue crab oocyte, for evaluating the development stage of oocyte in live blue crab, characterized in that, The system comprises an ovary development data collection module, a lightweight model construction module, an intelligent staging decision module and a model performance adaptive optimization module, wherein: The ovary development data collection module collects standardized multi-modal data required for model training and verification; The lightweight model construction module is used to construct an XGBoost-based feature enhancement regression machine learning model packaged by Pipeline, taking the determined modeling sample of the blue crab's carapace length and body weight as independent variables and the oocyte diameter as dependent variable. The model performance is verified by comparing the predicted oocyte diameter with the actual oocyte diameter, and the optimal model parameters are obtained to deploy the optimized lightweight model. The intelligent staging decision module stores the oocyte diameter prediction optimal model obtained through machine learning training. The module receives the morphological parameters and outputs the predicted oocyte diameter value, and matches the ovarian development stage according to the oocyte diameter staging table to finally output the ovarian development stage result without loss. The model performance self-adaptive optimization module introduces an active learning mechanism. When the system finds that the prediction result has low confidence, the user can be prompted to submit the sample to an expert for manual identification. The data confirmed by the expert will automatically form a new training sample and return to the model for incremental training, so that the model can continuously evolve and improve the accuracy during use.

Images