Hyperspectral imaging and label learning based mud clam harvesting and sorting system and method

Abstract

The application discloses a mud clam harvesting and sorting system and method based on hyperspectral imaging and label learning. The system comprises: a pre-harvest quality evaluation module for in-situ hyperspectral sampling of mud clams, extracting spectral features, and determining harvesting maturity; a harvesting, seedling protection, and mid-harvest sorting module for separating mud, real-time identifying mud clam seedlings, and pneumatically protecting them, while determining dead or alive state and removing dead clams; a post-harvest sorting module for obtaining three-dimensional topography and reflectance spectrum data of post-harvest mud clams, and based on a pre-trained label learning model, completing dead or alive and freshness discrimination, fatness grade classification, impurity identification, and physical size classification; a central control unit; and a data feedback module for feeding output results and corresponding spectral data to the pre-harvest quality evaluation module or the label learning model, and realizing model iterative optimization. The application solves the problems of mud clam industry, such as difficult determination of harvesting time, low harvesting efficiency, difficult quality sorting, and waste of seedling resources.

Description

Technical Field

[0001] This invention belongs to the field of intelligent processing and photoelectric detection technology of aquatic products, specifically involving an integrated system and method for pre-harvest quality assessment, in-harvest seedling identification and protection, and post-harvest quality sorting and graded distribution of mud clams for dual scenarios of tidal flat tidal and shrimp pond polyculture. Background Technology

[0002] mud clam ( Tegillarca granosa Mud clams are an important economic shellfish in southeastern coastal my country, prized for their delicious meat and rich nutritional value, giving them high market value. However, the mud clam industry currently faces three major challenges: 1. Difficulty in determining the harvesting time: Traditionally, the judgment of the plumpness and maturity of mud clams relies on manual experience, which is highly subjective and easily leads to problems such as harvesting too early (substandard quality) or harvesting too late (increased mortality).

[0003] 2. Low harvesting efficiency: In the mixed farming mode of mudflats and shrimp ponds, mud clam larvae and adults are mixed together, making it difficult to harvest the large ones and leave the small ones, resulting in serious waste of larval resources.

[0004] 3. Difficulty in quality sorting: After harvesting, mud clams need to be sorted according to whether they are alive or dead, freshness, plumpness, and impurities. Traditional manual sorting is inefficient, inconsistent, and cannot effectively identify dead clams and other mixed shellfish.

[0005] In existing technologies, ordinary machine vision can only identify superficial features such as shape and color, and cannot distinguish the internal quality of mud clams (freshness, plumpness, and whether they are alive or dead), and it is difficult to handle other shellfish species mixed in. Summary of the Invention

[0006] This invention provides a harvesting and sorting system and method for mud clams based on hyperspectral imaging and tag learning, in order to solve the technical problems of difficulty in determining the harvesting time, low harvesting efficiency, and difficulty in quality sorting in the prior art.

[0007] A mud clam harvesting and sorting system based on hyperspectral imaging and label learning includes: The pre-harvest quality assessment module is used to perform in-situ hyperspectral sampling of live mud clams in the aquaculture area, extract maturity-related spectral features, and determine the harvest maturity. The harvesting, seedling protection, and sorting modules integrate a mud separation device, a hyperspectral imaging unit, a pneumatic actuator, and a robotic arm. These modules are used to separate mud during harvesting, identify mud clam seedlings in real time, and pneumatically protect them. They also utilize a combination of hyperspectral and visual models to determine the liveness / death status and remove dead clam species. The post-harvest sorting module uses hyperspectral imaging to acquire the three-dimensional morphology and reflectance spectral data of the harvested mud clams. Combined with a mechanical screening device, and based on a pre-trained label learning model, the following tasks are completed: life and freshness discrimination, plumpness grading, impurity identification, and physical size grading. The central control unit connects the various modules, receives data, executes model algorithms, and sends control signals. The impurities include at least: gravel, empty shells, aquaculture ropes, and other shellfish species; The mechanical screening device includes a coaxial drum screen, which is used to achieve three-stage physical separation of large, medium and small sizes. The data feedback module is used to feed back the output results of the post-harvest sorting module and the corresponding spectral data to the pre-harvest quality assessment module or the label learning model to achieve iterative optimization of the model. Furthermore, the system primarily applies to mud clams. In determining whether a clam is alive or dead, it detects the presence or absence of the 540nm / 576nm hemoglobin bimodal peaks in the hinge region or micro-gap at the shell edge to determine whether the clam is alive or dead in the closed shell state. In seedling identification, it utilizes the near-infrared band to obtain information on shell thickness, moisture content (1450nm water absorption peak), and internal soft tissue state, enabling in-depth quality assessment that cannot be obtained by RGB.

[0008] The system, in which the pre-harvest quality assessment module uses the hemocyanin activity characteristic peak of 550-580nm and the water absorption peak of 960-980nm as key spectral indicators for determining the maturity of mud clams, and the selection of the 550-580nm band is based on the absorption characteristics of respiratory pigments in the hemolymph of mud clams.

[0009] The system described above uses a post-harvest sorting module where the plumpness grade is based on individual density measurements. An offline buoyancy method is used to establish a density-spectral mapping model, and density is retrieved online using hyperspectral data. The grading standard is: extra-grade density <1.85 g / cm³. 3 Primary density: 1.85-2.00 g / cm³ 3 Secondary density: 2.00-2.15 g / cm³ 3 Tertiary density > 2.15 g / cm³ 3 .

[0010] The system described above uses Ground Truth to determine the liveness and freshness of the organism. This is achieved by calibrating the shell-closing reaction intensity: a closing time of <0.5 seconds after touch indicates a healthy organism, 0.5-2 seconds indicates a stressed / dying organism, and >2 seconds or no reaction indicates a dead organism. This is verified using hemolymph pH: healthy individuals have a pH of 7.5-8.0, dying individuals have a pH of 7.0-7.5, and dead individuals have a pH <7.0. The model training utilizes the physiological characteristic of mud clams being rich in hemoglobin, using absorption peaks at 540nm and 576nm as activity criteria (the double peaks are significant in live organisms, and disappear synchronously after death due to hemoglobin degradation). Three-dimensional morphology data is also introduced to enhance the accuracy of the determination. During online detection, the hyperspectral probe focuses on the hinge area and shell edge of the mud clam, as these are the thinnest parts of the shell. After death, the ligaments loosen and the shell edge slightly opens, allowing the spectrum to effectively penetrate or enter the shell through micro-gap to capture the hemoglobin signal.

[0011] In the aforementioned system, for other shellfish species, the system only identifies them as "non-clam impurities" and removes them, without outputting specific species names to simplify the calculation model.

[0012] The system described herein uses a label learning model that is a multimodal fusion network, comprising a visual branch and a spectral branch. The visual branch employs a convolutional neural network based on the YOLO architecture for spatial localization, seedling size measurement, and initial screening of impurity morphology. The spectral branch uses a one-dimensional convolutional neural network (1D-CNN) to extract hyperspectral features, while simultaneously optimizing sub-tasks such as life and death discrimination, freshness discrimination, and plumpness grading. The weights of the loss function are dynamically adjusted according to accuracy requirements.

[0013] In the system described above, the harvesting, seedling protection, and sorting modules employ a screw-type device for separating mudflats. Seedling identification calculates the shell diameter of mud clams based on acquired spatial information. When the shell diameter is less than 15mm, it is determined to be a seedling and pneumatic protection is triggered, with an air pressure range of 0.2–0.5MPa.

[0014] An integrated control method for harvesting and quality sorting of mud clams based on the aforementioned system includes the following steps: Pre-harvest hyperspectral sampling was used to determine the maturity of mud clams by utilizing hemocyanin activity and water absorption characteristics. During the harvesting process, mud and silt are separated, seedlings are identified in real time and pneumatically protected, while dead shellfish are removed using a robotic arm; after harvesting, hyperspectral sampling and mechanical sieving are performed, and label learning models are used to simultaneously determine whether the shellfish are alive or dead, freshness, plumpness, impurities, and physical size. Streams are sorted by grade, and impurities are removed; Post-harvest hyperspectral sampling, through a label learning model, simultaneously completes the distinction between live and dead animals and freshness, the grading of plumpness, and the identification of impurities; Streams are sorted by grade, and impurities are removed; Post-harvest data is fed back to the pre-harvest model to achieve iterative optimization.

[0015] A non-transitory readable storage medium storing a computer program that, when executed by a processor, implements all or part of the steps of the method.

[0016] The beneficial effects of this invention are: 1. By using hyperspectral imaging and tag learning, the maturity of mud clams can be quantitatively assessed, avoiding the blind harvesting based on experience; 2. Identify seedlings during harvesting and provide pneumatic protection to achieve "harvesting large seedlings and leaving small ones, enabling sustainable harvesting"; 3. Post-harvest, the system can complete the determination of dead and alive status and freshness, the grading of plumpness, and the identification of impurities in one go, improving overall efficiency by more than 10 times; 4. Impurity identification covers other shellfish species, eliminating the need for complex multi-species classification models; the system is simple and reliable. 5. The data feedback loop enables continuous model optimization, making it more accurate the more it is used. Attached Figure Description

[0017] Figure 1 : Schematic diagram of the overall structure of the system of the present invention.

[0018] Figure 2 : Schematic diagram of the pre-harvest quality assessment module workflow.

[0019] Figure 3 Schematic diagram of the harvesting and seedling protection module.

[0020] Figure 4 : Schematic diagram of the post-harvest sorting module workflow.

[0021] Figure 5 : Schematic diagram of the tag learning model structure.

[0022] Figure 6 Flowchart of data feedback and model iteration.

[0023] Figure 7: Hyperspectral imaging feature map of mud clam.

[0024] Figure 8 A modeling diagram of the system of this invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Example 1: Overall System Structure like Figure 1 , Figure 8As shown, the system of this invention includes: a pre-harvest quality assessment module, a harvesting and seedling protection and in-harvest sorting module, a post-harvest sorting module, and a central control unit. The pre-harvest quality assessment module is deployed in the aquaculture area and uses a portable hyperspectral imager to perform non-contact spectral acquisition of mud clams underwater or on the surface of tidal flats. This hyperspectral imaging system has a spectral resolution better than 3 nm and a three-dimensional morphology measurement accuracy better than 0.02 mm, ensuring the precision of feature extraction. This module also includes a mud separation device (such as...). Figure 3 (As shown), used for preliminary purification.

[0027] The harvesting and seedling protection module is integrated at the front end of the harvesting equipment, specifically including a screw-type mud separator, a hyperspectral imaging unit, a pneumatic nozzle array, and a control robotic arm. For example... Figure 3 As shown, the collected mud clams are separated from the mud by a screw and water flow. Then, they enter the hyperspectral imaging evaluation stage. If seedlings (shell diameter <15mm) are identified, the control system triggers a pneumatic nozzle to blow them back; if dead clams are identified (through the "dead clams sorting during collection" process), the control system controls a robotic arm to adsorb and remove them, achieving "collecting large ones and keeping small ones, removing dead ones and keeping live ones".

[0028] Post-harvest sorting module: This is an independent sorting production line. It includes a conveyor belt, a hyperspectral imaging darkroom, and a coaxial drum screen (such as...). Figure 4 (As shown).

[0029] The hyperspectral imaging darkroom is responsible for acquiring three-dimensional morphology and reflectance spectral data for internal quality analysis (life / death, plumpness). The mud clams then enter the "mud clam grading and screening" unit, where they are physically separated into large, medium, and small sizes using a coaxial drum screen with three different aperture sizes. The screens feature a quick-release design for easy maintenance.

[0030] Central control unit: such as Figure 1 As shown on the right, it is responsible for receiving data from each module, executing model algorithms, monitoring equipment status, sending control commands (such as robotic arm movements and pneumatic valve switching), and coordinating linkage control and storage management.

[0031] Data Feedback Module: Used to feed back the output results of the post-harvest sorting module and the corresponding spectral data to the pre-harvest quality assessment module or label learning model, enabling iterative optimization of the model (e.g., Figure 6 (As shown).

[0032] Example 2: Pre-harvest quality assessment Combination Figure 2 The process shown is as follows: The pre-harvest quality assessment module uses the characteristic peaks of hemocyanin activity (550-580nm) and water absorption peaks (960-980nm) of mud clam as key spectral indicators for maturity determination.

[0033] A hyperspectral camera is used to identify the status of mud clams and seedlings in real time. The process is as follows: the shell diameter is calculated. If the shell diameter is determined to be ≥15mm (adult), the clams are sent for post-harvest sorting; if the shell diameter is determined to be <15mm (seedling), the clams are sent to the seedling identification module.

[0034] Example 3: Harvesting and Seedling Care like Figure 3 and Figure 2 As shown: After the screw-type mud separator outputs clean mud clams, a hyperspectral camera collects data in real time. The system calculates the projected area or three-dimensional volume of the mud clams using image processing algorithms, and then estimates the shell diameter. When a shell diameter < 15mm is detected, it is identified as a juvenile; otherwise (shell diameter ≥ 15mm), it is identified as an adult. For juveniles, the control system triggers a pneumatic nozzle to blow the juveniles back to the aquaculture environment with an air pressure of 0.2-0.5MPa.

[0035] Furthermore, this invention leverages the unique advantages of hyperspectral imaging in seedling identification. Unlike RGB vision, which can only observe shell size, hyperspectral imaging can acquire crucial information that RGB cannot obtain: Based on the above advantages, this invention realizes a layered seedling protection strategy of "RGB rapid positioning + hyperspectral depth verification": RGB vision (optional auxiliary) is responsible for rapid size screening, and hyperspectral is responsible for health and activity verification. Only when the target simultaneously meets the requirements of "size is seedling" and "spectral detection is live and healthy" will the pneumatic return operation be performed to avoid putting dead or diseased seedlings back into the mudflats.

[0036] Example 4: Post-harvest sorting and label learning 4.1 Explanation of the Optical-Physical Principles of Mud Clam Quality Grading In this invention, the quality grade of mud clams is not based on traditional size or weight grading, but rather on a comprehensive judgment of plumpness and freshness. The core physical indicator of plumpness is individual density (the ratio of mass to volume), which is negatively correlated with the fullness of the mud clam meat. The specific principle is as follows: Mud clams consist of two parts: a high-density shell (mainly composed of calcium carbonate, with a density of approximately 2.70–2.95 g / cm³) and low-density soft tissue (clam meat, with a density of approximately 1.05–1.10 g / cm³). In a living state, the geometric volume of the shell is basically fixed, while the filling material inside the shell determines the overall density.

[0037] When the mud clam meat is plump (high plumpness), the cavity inside the shell is mostly occupied by low-density clam meat, and the overall density is close to the weighted average of the meat and the shell, which is relatively low (typical value 1.82–1.98 g / cm³).

[0038] When the mud clam meat shrinks (low plumpness) or dies, the cavity inside the shell is partially or completely filled with seawater (density 1.00 g / cm³) or mud and sand (density 1.60–2.65 g / cm³). Since the density of the filling material is greater than or equal to the density of the clam meat, and the high-density shell remains unchanged, the overall density actually increases (typical value 2.15–2.30 g / cm³).

[0039] Therefore, lower density indicates fuller shellfish meat and better quality; higher density indicates a higher probability of empty shells, thin shellfish, or dead shellfish. This rule contradicts the traditional intuitive feeling that "the greater the weight, the better the quality," but it has been confirmed by experimental data.

[0040] Based on the above principles, this invention uses the offline buoyancy method to determine the true density of the sample as the Ground Truth, and uses online hyperspectral inversion of density as the key basis for the grading of fatness level.

[0041] 4.2 Label System Construction (Ground Truth Labeling Method) In this invention, the Ground Truth of the quality grade of mud clams is obtained by the following method, which combines two dimensions: plumpness and freshness / vitality.

[0042] (1) Fullness calibration (buoyancy method) (used for training set construction) The density of individual mud clams was determined non-destructively using the buoyancy method (displacement volume method). Weigh the mud clams on a precision electronic balance and record the mass M (g). The mud clams were completely immersed in pure water (water temperature 20±1℃), and the buoyancy F (g) was measured using the suspension method. According to Archimedes' principle, the volume of the mud clam V = F / (ρwater × g), where ρwater ≈ 1.00 g / cm³; Calculate the individual density ρ = M / V (g / cm³).

[0043] The fatness level is classified as follows: (2) Freshness / Viability Assessment (Closing Reaction Intensity + Hemolymph pH Verification) Live mud clam samples were collected at different vitality states (healthy, stressed, near death, and dead), and the true state was determined using the following methods: (3) Online detection feature extraction (based on the data in Figure 7) In the actual production line, the model uses hyperspectral data to simulate the above conditions. As shown in Figure 7, the blood of the mud clam... The HB2 uptake curve of hemoglobin in lymph showed significant uptake near 482nm, 540nm, and 576nm. Characteristic peaks are observed. As the activity of the mud clam decreases, the position and intensity of these absorption peaks shift, affecting the system. The liveness and health status of mud clams can be determined online by monitoring reflectance changes in the 540nm and 576nm wavelengths. Simultaneously, analyzing the three-dimensional morphological characteristics of mud clams (such as shell closure degree and surface texture) can help determine their vitality.

[0044] Method 1: Intensity of shell-closing reaction (non-destructive, primary label) Gently touch the water outlet of the mud clam with a probe and record the time it takes for the shell to completely close after the touch. Measure each sample three times and take the average. Method 2: Hemolymph pH verification (destructive, only used for verification of some representative samples) A small amount of hemolymph (approximately 0.1-0.2 mL) was extracted from the adductor muscle of the mud clam using a syringe, and measured using a portable pH meter. Note: The pH method requires piercing the mud clam and is only used for validation of a small number of samples (<5%) in the training set to confirm the accuracy of the closed-shell reaction intensity label. It is not used for online detection on the production line.

[0045] The principle of hyperspectral identification of dead closed shell clams: Live mud clams are rich in hemoglobin, and their oxygenated state exhibits two significant characteristic absorption peaks at 540 nm and 576 nm. After death, hemoglobin is rapidly oxidized and degraded, and the two characteristic peaks weaken or disappear simultaneously.

[0046] To address the challenge of indistinguishable appearances among dead shellfish in their closed state, this system focuses its hyperspectral detection on the hinge region and shell edge—the thinnest parts of the shell. After death, the ligaments loosen and the shell edge slightly opens, allowing the spectrum to effectively penetrate or enter the shell through micro-gap, capturing hemoglobin signals. By detecting the presence or absence of the 540nm / 576nm bimodal peaks, high-precision differentiation between live and dead shells in their closed state is achieved, fundamentally overcoming the generalization bottleneck of RGB vision's inability to distinguish between shells with identical appearances.

[0047] (3) Comprehensive grade classification rules Based on a combination of plumpness and freshness / vitality, mud clams are classified into the following grades: 4.3 Spectral Data Acquisition and Preprocessing

[0048] The reflectance spectra of each mud clam were acquired using a 400-1000 nm hyperspectral camera. The raw spectra were subjected to dark current correction, whiteboard normalization, and ROI average spectrum extraction, and were preprocessed using Savitzky-Golay smoothing, first derivative, and standard normal variable transformation (SNV).

[0049] Pay special attention to the following characteristic bands: 550-580nm: Hemocyanin activity peak, sensitive to hemolymph pH and activity status; 960-980nm: Moisture absorption peak, reflecting the opening and closing of the shell slits and the moisture state.

[0050] Simultaneously, three-dimensional topographic data is collected to enhance the model's ability to perceive the morphology and volume of shells, thereby improving the accuracy of plumpness inversion. 4.4 Multimodal Label Learning Model Structure

[0051] The label learning model used in this invention is a multimodal fusion architecture, combined with, for example... Figure 5 The visual detection structure and hyperspectral analysis module shown: A. Visual Branch (Spatial and Morphological Detection): such as Figure 5 As shown, the input image is processed by Backbone (YOLO11sCSPDarknet) to extract features, enhanced by Neck (FPN+PAN+attention mechanism), and finally output by the decoupled detection head. This branch is mainly used for impurity identification, seedling size measurement, and spatial localization.

[0052] B. Spectral Branch (Intrinsic Quality Detection): For hyperspectral data, a one-dimensional convolutional neural network (1D-CNN) is used as the feature extraction layer. Feature peaks such as 540nm / 576nm are used for live / dead detection, and density is retrieved from the full spectrum data for fatness classification.

[0053] C. Fusion Decision: The outputs of the visual branch (such as impurity category and location) and the outputs of the spectral branch (live / dead state, plumpness level) are jointly optimized. The loss function is L = w1Lfreshness + w2Lplumpness + w3Limpurities. 4.5 Model Training and Update

[0054] like Figure 6 As shown, this invention continuously performs data feedback and iteration to establish the optimal model.

[0055] Initial training set: 3000 mud clam samples were collected, covering different plumpness levels, different freshness / vitality states, and different impurity types, and the ground truth was calibrated according to the method in Section 4.2.

[0056] Training / Validation / Testing = 6:2:2.

[0057] Model update: During system operation, samples with a prediction confidence level below 0.85 are pushed to the review interface. After review by the quality inspector, new labeled samples are generated. Incremental training is triggered every 500 samples accumulated.

[0058] During training, the system utilizes the three-dimensional morphology and spectral data obtained from hyperspectral imaging to construct a more comprehensive feature representation, significantly improving the model's accuracy in judging the quality of mud clams.

[0059] In addition, the data feedback module regularly synchronizes the post-harvest sorting results and spectral / morphological data to the central server to update the pre-harvest evaluation model and achieve closed-loop optimization of the entire chain from pre-harvest to harvesting and post-harvest. 4.6 Impurity Identification and Physical Sorting

[0060] like Figure 4 As shown, post-harvest sorting includes not only algorithmic sorting but also mechanical assistance. For other shellfish species and impurities such as gravel, the system identifies and removes them using a visual + spectral model. Simultaneously, a coaxial drum screen is used for physical grading. The quick-release screen design facilitates cleaning and maintenance, while the anti-collision design ensures the integrity of the mud clams.

[0061] In this embodiment of the invention, mud clams are used as an example for model training and validation. For other mixed shellfish species (such as clams and scallops), the system identifies them as impurities and removes them without requiring the output of specific species names. Through subsequent incremental learning, the model can be gradually expanded into a multi-species classification system.

[0062] Impurities include: gravel, empty shells, aquaculture ropes, foreign objects, and other shellfish species.

[0063] During the impurity identification process, the system simultaneously utilizes three-dimensional morphological data (such as shell shape, size, and texture) and spectral data for comprehensive judgment, significantly improving the identification accuracy. 4.7 Effect Verification

[0064] The RGB visual comparison data is based on the YOLO model independently trained by the inventors in the early stage (accuracy of 0.97 under controlled lighting conditions in the laboratory, which drops to about 0.82 under actual working conditions on the tidal flats). This system improves the generalization ability of life and death judgment by introducing a hyperspectral module.

[0065] Example 5: Data Feedback and Iteration The grade results, spectral data, and verification labels output by the post-harvest sorting module are periodically synchronized to the pre-harvest evaluation model on the central server. The pre-harvest model is retrained every two weeks, and the updated model parameters are sent to the pre-harvest quality evaluation module, forming a closed-loop optimization across the entire "pre-harvest-harvest-post-harvest" chain (e.g., Figure 6 (As shown).

[0066] During the data feedback process, the system synchronizes not only spectral data but also three-dimensional topographic data, making the model iteration more comprehensive and accurate.

[0067] Example 6: Model Training and Initial Training Set Update: 3000 mud clam samples were collected, covering different plumpness levels, different freshness / vitality states, and different impurity types. Ground Truth was calibrated according to the method in Section 4.2. Training / Validation / Testing ratio = 6:2:2.

[0068] Model Update: During system operation, samples with a prediction confidence level below 0.85 are pushed to the review interface. After review by quality inspectors, new labeled samples are generated, and incremental training is triggered every 500 accumulated samples. After two weeks of testing, the model's accuracy in identifying novel impurities has significantly improved.

[0069] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalents.

Claims

1. A mud clam harvesting and sorting system based on hyperspectral imaging and tag learning, characterized in that, include: The pre-harvest quality assessment module is used to perform in-situ hyperspectral sampling of live mud clams in the aquaculture area, extract maturity-related spectral features, and determine the harvest maturity. The harvesting, seedling protection, and sorting module integrates a mud separation device, a hyperspectral imaging unit, a pneumatic actuator, and a robotic arm. It is used to separate mud during the harvesting process, identify mud clam seedlings in real time and pneumatically protect them, and at the same time use hyperspectral and visual models to jointly determine the liveness and death status and remove dead clam. The post-harvest sorting module acquires the three-dimensional morphology and reflectance spectrum data of the harvested mud clams through hyperspectral imaging, and, in conjunction with a mechanical screening device, completes the following tasks based on a pre-trained label learning model: life and death determination and freshness classification, plumpness grade classification, impurity identification, and physical size classification. The central control unit is used to connect the various modules, receive data and execute model algorithms, and send control signals. The impurities include at least: gravel, empty shells, aquaculture ropes, and other shellfish species; The mechanical screening device includes a coaxial drum screen, which is used to achieve three-stage physical separation of large, medium and small sizes. The data feedback module is used to feed back the output results of the post-harvest sorting module and the corresponding spectral data to the pre-harvest quality assessment module or the label learning model to achieve iterative optimization of the model. Furthermore, the system primarily applies to mud clams. In determining whether a clam is alive or dead, it detects the presence or absence of the 540nm / 576nm hemoglobin bimodal peaks in the hinge region or micro-gap at the shell edge to determine whether the clam is alive or dead in the closed shell state. In seedling identification, it utilizes the near-infrared band to obtain information on shell thickness, moisture content, and internal soft tissue status, enabling a depth quality assessment that cannot be obtained by RGB.

2. The system according to claim 1, characterized in that, The pre-harvest quality assessment module uses the hemocyanin activity characteristic peak at 550-580nm and the water absorption peak at 960-980nm as key spectral indicators for determining the maturity of mud clams, and the selection of the 550-580nm band is based on the absorption characteristics of respiratory pigments in the hemolymph of mud clams.

3. The system according to claim 1, characterized in that, The plumpness grade in the post-harvest sorting module is based on individual density measurement. A density-spectral mapping model is established using the offline buoyancy method, and density is retrieved online using hyperspectral data. The classification standard is: extra-grade density <1.85 g / cm³. 3 Primary density: 1.85-2.00 g / cm³ 3 Secondary density: 2.00-2.15 g / cm³ 3 Tertiary density > 2.15 g / cm³ 3 .

4. The system according to claim 1, characterized in that, The Ground Truth for determining liveness and freshness is achieved by calibrating the shell-closing response intensity: a closing time of <0.5 seconds after touch indicates a healthy live organism, 0.5-2 seconds indicates a live organism under stress / dying, and >2 seconds or no response indicates a dead organism. This is verified using hemolymph pH: healthy individuals have a pH of 7.5-8.0, dying individuals have a pH of 7.0-7.5, and dead individuals have a pH <7.

0. The model training utilizes the hemoglobin-rich physiological characteristics of mud clams, using absorption peaks at 540nm and 576nm as activity criteria, and incorporates three-dimensional morphology data to enhance accuracy. During online detection, the hyperspectral probe focuses on the hinge area and shell edge of the mud clam, as these are the thinnest parts of the shell, and after death, the ligaments loosen and the shell edge slightly opens, allowing the spectrum to effectively penetrate or enter the shell through micro-gap to capture hemoglobin signals.

5. The system according to claim 1, characterized in that, In the impurity identification process, for other shellfish species, the system only determines them as "non-clam impurities" and removes them, without outputting the specific species name to simplify the calculation model.

6. The system according to claim 1, characterized in that, The label learning model is a multimodal fusion network, including a visual branch and a spectral branch. The visual branch uses a convolutional neural network based on the YOLO architecture for spatial localization, seedling size measurement, and initial screening of impurity morphology. The spectral branch uses a one-dimensional convolutional neural network (1D-CNN) to extract hyperspectral features, while optimizing sub-tasks such as life and death discrimination, freshness discrimination, and plumpness grading. The weights of the loss function are dynamically adjusted according to the accuracy requirements.

7. The system according to claim 1, characterized in that, In the harvesting, seedling protection, and sorting modules, the mud separation uses a screw-type device; the seedling identification calculates the shell diameter of the mud clam by acquiring spatial information, and when the shell diameter is less than 15mm, it is determined to be a seedling and triggers pneumatic protection, with an air pressure range of 0.2-0.5MPa.

8. A method for integrated control of mud clam harvesting and quality sorting based on the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: Pre-harvest hyperspectral sampling was used to determine the maturity of mud clams by utilizing hemocyanin activity and water absorption characteristics. During the harvesting process, mud and silt are separated, seedlings are identified in real time and pneumatically protected, while dead shellfish are removed using a robotic arm; after harvesting, hyperspectral sampling and mechanical sieving are performed, and label learning models are used to simultaneously determine whether the shellfish are alive or dead, freshness, plumpness, impurities, and physical size. Streams are sorted by grade, and impurities are removed; Post-harvest data is fed back to the pre-harvest model to achieve iterative optimization.

9. A non-transitory readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements all or part of the steps of the method of claim 8.

Images