An image recognition-based farmland micro-plastic rapid identification system

Abstract

The application belongs to the technical field of image recognition, and discloses a farmland micro-plastic rapid identification system based on image recognition, a special four-stage pretreatment algorithm framework is constructed, combined with a special database of farmland soil background, complex background depth is completed, in-situ non-preprocessing collection and background disturbance depth integration of farmland are realized, and feature completion is carried out for small particle size micro-plastics, so that the problem of missed detection of small particle size particles is avoided; meanwhile, a four-dimensional feature fusion framework of shape, spectrum, texture and polarization is built, feature calibration and attention mechanism weighted fusion are carried out combined with the actual characteristics of farmland micro-plastics, effective features are strengthened, redundant interference is weakened, and the false detection and missed detection rate caused by soil organic matter, mineral particles, straw clippings and other impurities is reduced; a recognition model suitable for farmland scenes is constructed based on a YOLO light network, a multi-task parallel output branch and a cooperative optimization mechanism are set, and target detection, material classification and aging degree classification are completed synchronously.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically a rapid identification system for microplastics in farmland based on image recognition. Background Technology

[0002] Microplastics (plastic particles with a diameter of less than 5 mm) have become a new type of persistent pollutant and a core challenge in the remediation of farmland soil pollution in my country. Microplastics in farmland soil can disrupt soil aggregate structure, reduce water and fertilizer retention capacity, inhibit crop growth, and can also accumulate in the human body through the food chain, posing a dual threat to both the ecological environment and human health. Currently, mainstream microplastic detection technologies mainly suffer from the following technical problems: While image recognition-based detection technology improves detection efficiency, it has poor anti-interference capabilities and is easily affected by impurities such as organic matter, mineral particles, and straw debris in the soil, resulting in a high rate of false positives and false negatives. Furthermore, the feature extraction dimension is limited to spectral or morphological features, without combining key distinguishing features such as polarization and texture of microplastics in farmland. In addition, it cannot balance real-time performance and portability, and relies heavily on backend servers for data processing, making it impossible to achieve real-time detection in the field.

[0003] Existing technologies lack sufficient functionality, only capable of identifying the presence or absence of microplastics, and cannot simultaneously perform material classification, aging degree determination, and spatial distribution modeling of pollution, making it difficult to support pollution source tracing and targeted remediation. Furthermore, they fail to achieve deep integration of in-situ, pre-processed soil sample collection in farmland with precise decontamination of complex backgrounds. The collection process requires cumbersome pre-processing of soil samples, which is time-consuming and labor-intensive, and the decontamination algorithm is not customized for the complex backgrounds of farmland, failing to effectively separate microplastics from soil impurities. Multimodal feature fusion is limited to the simple superposition of spectrum and image, without building a dedicated feature system for farmland scenarios. The lightweight edge-end multi-task recognition is not customized and optimized for farmland microplastic detection scenarios, and neither can meet the actual needs of fast, accurate and in-situ detection of farmland microplastics. Summary of the Invention

[0004] The purpose of this invention is to provide a rapid identification system for microplastics in farmland based on image recognition, so as to solve one or more problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a rapid identification system for microplastics in farmland based on image recognition, comprising the following modules: Preferably, the multispectral acquisition module integrates an imaging unit, a multi-band ring light source, an adaptive illumination compensation unit, a polarization filtering unit, and a BeiDou positioning unit. The imaging unit covers two characteristic bands: visible light and near-infrared light. It can simultaneously acquire five sets of characteristic spectral images in the range of 400nm to 1700nm. It has the function of adapting to farmland soil texture and can automatically adjust the imaging focal length and band acquisition parameters according to the content of soil sand and clay particles. The adaptive lighting compensation unit has built-in light and humidity sensors. Combined with the light change patterns in farmland, it predicts the type of light interference and automatically adjusts the brightness and incident angle of the light source to eliminate image noise caused by strong light reflection and shadow occlusion. The polarization filter unit is used to filter diffuse stray light from the soil background, highlighting the edge contours and material characteristics of microplastic particles. It has an adaptive polarization angle adjustment function and optimizes the filter parameters in real time based on the average gray value of the soil background. The BeiDou positioning unit synchronously collects the spatial coordinates of the sampling points and associates them with the soil type information of the sampling points to construct the correspondence between the collected images and the soil environment.

[0006] Preferably, the background descrambling module constructs a four-level preprocessing algorithm framework based on the in-situ soil images and related data acquired by the multispectral acquisition module, which includes farmland background layering modeling, target enhancement, mask segmentation, and small particle size feature completion. This achieves deep integration of in-situ unprocessed farmland acquisition and complex background descrambling. First, it calls a pre-constructed dedicated database of farmland soil background and uses multispectral difference operations and background feature matching to remove the broad-spectrum background interference of soil organic matter and water, generating a differential image with enhanced microplastic features. Secondly, adaptive bilateral filtering and the Retinex algorithm are used to eliminate image grayscale deviation caused by uneven illumination, while preserving sub-pixel edge details of microplastic particles. Then, pixel-level foreground mask segmentation is achieved based on the U-Net network to separate suspected microplastic targets from complex backgrounds, generate ROI regions for single-particle targets, and complete the separation of targets from backgrounds. Finally, a small-particle-size microplastic feature completion module is set up to supplement edge contours and texture features for microplastics with a particle size of less than 1 mm.

[0007] Preferably, the feature fusion module integrates a morphological feature extraction unit, a spectral feature extraction unit, a texture feature extraction unit, and a polarization feature extraction unit. Based on the microplastic feature enhancement differential image and ROI region output by the background descrambling module, a four-dimensional feature fusion extraction framework of morphology, spectrum, texture, and polarization is constructed. A special feature calibration mechanism for farmland microplastics is set up. The morphological feature extraction unit extracts geometric features such as particle size, aspect ratio, roundness, and edge roughness of the target. Combined with the morphological characteristics of farmland microplastics, it has an adaptive adjustment function for morphological feature weights to enhance the distinction from soil impurities. The spectral feature extraction unit extracts spectral features such as reflectance peaks, spectral slopes, and absorption valley positions of targets in multi-band images, and constructs a special spectral fingerprint database for common agricultural microplastics in farmland to distinguish them from non-plastic impurities. The contrast, correlation, entropy and other texture features of the target are extracted by gray-level co-occurrence matrix, and combined with the texture change pattern in the aging process of microplastics, a processing unit with texture feature time-series calibration function is set up to improve the discrimination accuracy of microplastics with different aging levels. The polarization feature extraction unit extracts polarization features such as polarization degree and polarization angle of the target, strengthens the distinguishing weight of polarization features, and finally performs adaptive weighted fusion of four-dimensional features through attention mechanism to strengthen the weight of high distinguishing features, weaken the interference of redundant features, and generate a standardized high distinguishing microplastic feature vector.

[0008] The quantitative criteria for high discriminative power of microplastic feature vectors are as follows: Based on the recognition model of feature vectors after four-dimensional feature fusion, the inter-class discrimination accuracy of microplastics and non-plastic impurities such as soil organic matter, mineral particles, and straw debris is ≥95%, the inter-class discrimination accuracy of microplastics of different materials and different aging degrees is ≥90%, and the intra-class similarity of feature vectors is ≤10%. If these quantitative indicators are met, the feature vectors are judged to be high discriminative power.

[0009] The quantitative screening criteria for high-discrimination features are as follows: features that contribute ≥80% to the three core tasks of microplastic target identification, material classification, and aging grading, and whose task recognition accuracy decreases by ≥5% after removing the feature, are judged as high-discrimination features. The weight of such features is increased by 30%-50% to ensure the dominant role of effective features.

[0010] Preferably, the edge recognition module receives the standardized high-discrimination microplastic feature vector output by the feature fusion module, constructs an intelligent identification and classification model for farmland microplastics based on the YOLO lightweight network and adapts it to the edge embedded device, sets up a multi-task collaborative optimization mechanism adapted to the special scenario of farmland microplastic detection, sets up a multi-task parallel output branch for the identification and classification model, and simultaneously realizes the three core functions of target detection, material classification and aging degree classification, and handles feature conflicts between multiple tasks through a task collaborative optimization algorithm; Target detection is used to identify microplastic particles within the ROI area and count the number and size distribution of particles. A dedicated detection branch for small-diameter microplastics is set up. Material classification is based on fused feature vectors to classify common agricultural microplastics. The classifier parameters are optimized in combination with the material distribution characteristics of microplastics in farmland to improve classification accuracy. Aging degree classification is based on texture and spectral features to classify microplastics into three levels: unaged, slightly aged, and heavily aged. An aging feature calibration branch is set up and the influence of farmland environment on microplastic aging is combined to improve classification accuracy. Based on a dedicated dataset for farmland microplastics, the model is pre-trained, structured pruning, and INT8 quantization is performed. At the same time, knowledge distillation technology is used to transfer the recognition experience of complex models to lightweight models. The model can also be adaptively fine-tuned according to changes in farmland scenes to meet the needs of farmland microplastic detection in different regions.

[0011] The trigger condition for adaptive fine-tuning of the model is that after continuously detecting 50 sampling points, the average accuracy of target detection of these 50 sampling points is <90%, the average accuracy of material classification is <85%, or the average accuracy of aging grading is <80%. During fine-tuning, only the parameters of the head output layer and the neck feature fusion layer of the model are locally adjusted. The microplastic feature data of farmland in this area stored locally is used as the fine-tuning sample. There is no need to retrain the entire model. The recognition accuracy is verified immediately after fine-tuning. Once the target is met, the adjustment is stopped and the parameters are saved.

[0012] Structured pruning is based on the necessity of extracting microplastic features from farmland. Redundant channels and kernels with absolute values ​​of convolutional layer weights less than 0.01 are removed (the threshold can be fine-tuned according to the farmland scene sample set), while retaining the network structure responsible for extracting core features such as microplastic edges, spectra, and polarization. During the pruning process, the core structures of the model's feature fusion layer and small particle size detection branch are not pruned. After pruning, the compression ratio of model parameters is controlled at 40%-50%, and the core accuracy of target detection is not significantly reduced.

[0013] Preferably, the spatial modeling module, based on the spatial coordinates of sampling points obtained by the multispectral acquisition module and the microplastic pollution data output by the edge recognition module, combines pattern recognition and modeling techniques to construct a gridded spatial model of microplastic pollution in farmland. It also sets up a pollution source tracing accuracy optimization function to improve the accuracy of modeling and source tracing. First, the target farmland is divided into a standard grid of 10m×10m. Based on the Kriging interpolation method, combined with the soil type and microplastic distribution characteristics of the sampling points, spatial interpolation is performed on the pollution data of discrete sampling points to generate a heat map of microplastic pollution distribution in the entire farmland. A risk level calibration module is set up to optimize the pollution risk classification standard by combining farmland planting type and agricultural film usage history. Then, based on pollution concentration, plastic type, and particle size distribution data, and combined with pattern recognition algorithms, each grid is classified into three levels: low risk, medium risk, and high risk. At the same time, combined with related data such as agricultural film usage history, irrigation path, and pesticide packaging placement points, source tracing analysis is carried out on high-pollution areas to locate the core pollution source. Finally, a standardized spatial distribution report of microplastic pollution in farmland is generated.

[0014] The risk level calibration module adjusts the classification thresholds based on two core dimensions of the history of agricultural film use in farmland: for farmland areas that have used agricultural film continuously for 5 years or more, the particle density threshold is lowered by 20% on the basis of the basic classification threshold for microplastic pollution in farmland; for farmland areas where the annual usage of agricultural film exceeds 50 kg / mu, the basic classification threshold for the proportion of small-diameter microplastics is lowered by 15%; for farmland areas with no history of agricultural film use, the basic national limit threshold is adopted to adapt to the needs of farmland pollution risk assessment under different agricultural film use backgrounds.

[0015] Preferably, the result early warning module receives microplastic detection statistics output by the edge recognition module and pollution distribution heat map, pollution source tracing analysis results and risk level data generated by the spatial modeling module. It integrates a touch screen, a voice broadcast unit and a local data storage unit for real-time display and output of multiple core data on site. The core data includes full-dimensional statistical data on the number, particle size distribution, material type and aging degree of microplastic particles. The result early warning module is equipped with a recognition result verification function to determine the rationality of the data. The core data also includes the pollution concentration and risk level of the sampling points, as well as the heat map of farmland pollution distribution and the results of pollution source tracing analysis. The result early warning module has built-in threshold standards corresponding to the control of microplastic pollution in farmland soil. When the pollution concentration at the detection point exceeds the national standard limit, it automatically triggers sound, light and voice warnings to prompt the operator to conduct key verification. The result early warning module supports offline detection and data backtracking in environments without network access. All detection data can be exported on-site as standardized detection reports and encrypted and stored in the on-site detection terminal equipped with a local data storage unit. It supports offline detection and data backtracking in environments without network access and has a data backtracking verification function that can verify the accuracy of detection results through image comparison.

[0016] Preferably, the cloud iteration module receives anonymized detection data and manually reviewed and labeled data uploaded by the aforementioned modules, and constructs a closed-loop iteration system of data collection, model optimization, and terminal update. First, the cloud iteration module classifies and stores the received data, constructs a dynamically updated national farmland microplastic image database, and classifies and organizes the data according to soil type, regional climate, and pollution type. Secondly, based on the incremental learning algorithm, the recognition and classification model is continuously fine-tuned and optimized. At the same time, the cloud platform can manage the model's full lifecycle version and push the optimized lightweight model to the edge terminal.

[0017] The beneficial effects of this invention are as follows: 1. This invention constructs a dedicated four-level preprocessing algorithm framework, which, combined with a dedicated database of farmland soil background, completes in-depth de-scratching of complex backgrounds, achieving deep integration of in-situ preprocessing-free data acquisition and background de-scratching of farmland. It also performs feature completion for small-diameter microplastics, avoiding the problem of missed detection of small-diameter particles. At the same time, it builds a four-dimensional feature fusion framework of morphology, spectrum, texture, and polarization, and performs feature calibration and attention mechanism weighted fusion based on the actual characteristics of farmland microplastics, strengthening effective features and weakening redundant interference, thereby reducing the false detection and false detection rates caused by impurities such as soil organic matter, mineral particles, and straw debris.

[0018] 2. This invention constructs a recognition model adapted to farmland scenarios based on the YOLO lightweight network, sets up a multi-task parallel output branch and collaborative optimization mechanism, and simultaneously completes target detection, material classification, and aging degree grading; the model is lightweighted through structured pruning, knowledge distillation and other methods, and can run smoothly on edge embedded devices without relying on a backend server. It can also adaptively fine-tune according to changes in farmland scenarios, realizing in-situ unprocessed data collection in farmland and real-time detection in the field.

[0019] 3. This invention visualizes pollution distribution through rasterized spatial modeling, classifies pollution risks by combining farmland planting types and agricultural film usage history, and accurately traces the source of high-pollution areas based on related data such as irrigation paths and application points. The result early warning module can display full-dimensional detection data on-site, automatically triggering early warnings when thresholds are exceeded, and also supports offline storage, data backtracking, and result verification. The cloud iteration module builds a dynamically updated farmland microplastic database, continuously optimizes the model through incremental learning, and achieves remote updates, enabling the system's recognition accuracy and scene adaptability to continuously improve. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall workflow of the system of the present invention; Figure 2 This is a flowchart of the image preprocessing and feature fusion sub-process of the present invention; Figure 3 This is a flowchart of the intelligent recognition and cloud-based iterative process of the present invention. Detailed Implementation

[0021] 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.

[0022] like Figures 1 to 3As shown, this embodiment of the invention provides a rapid identification system for microplastics in farmland based on image recognition, comprising the following modules: The multispectral acquisition module integrates an imaging unit, a multi-band ring light source, an adaptive illumination compensation unit, a polarization filtering unit, and a Beidou positioning unit. The imaging unit covers two characteristic bands: visible light and near-infrared light. It can simultaneously acquire five sets of characteristic spectral images in the range of 400nm to 1700nm. It has the function of adapting to farmland soil texture and can automatically adjust the imaging focal length and band acquisition parameters according to the content of soil sand and clay. Soil texture is categorized and matched with parameters based on sand and clay content: For sandy soils with sand content >70% and clay content <10%, the imaging focal length is adjusted to 10mm, focusing on acquiring characteristic images in the 400-600nm visible light band and the 1400-1700nm near-infrared band; for clayey soils with clay content >30%, the imaging focal length is adjusted to 15mm, focusing on acquiring characteristic images in the 600-1000nm visible light-near-infrared transition band; for loamy soils, a general parameter of 12mm focal length is used, and 5 sets of full-band characteristic images are acquired simultaneously to adapt to the imaging needs of different soil textures.

[0023] The adaptive illumination compensation unit has built-in light and humidity sensors. Combining the light variation patterns in farmland with the real-time light intensity and light angle data collected by the light sensor, it predicts the type of light interference and automatically adjusts the light source brightness and incident angle to eliminate image noise caused by strong light reflection and shadow occlusion. The adaptive illumination compensation unit collects ambient light intensity (unit: lux) and illumination angle in real time through a light sensor. When the light intensity is >80,000 lux, it is determined to be strong midday light interference, and the brightness of the light source is reduced by 30%-50%, and the incident angle is adjusted to 45° oblique illumination. When the light intensity is <5,000 lux and there is backlight angle, it is determined to be backlight interference in the morning or evening, and the brightness of the light source is increased by 40%-60%, and the side-mounted supplementary lighting unit of the multi-band ring light source is turned on. When shadow occlusion is detected, the local brightness of the light source is adjusted point by point according to the shadow area ratio (>20%) to achieve illumination uniformity compensation.

[0024] The polarization filter unit is used to filter diffuse stray light from the soil background, highlighting the edge contours and material characteristics of microplastic particles. It has an adaptive polarization angle adjustment function and optimizes the filter parameters in real time based on the average gray value of the soil background. The polarization filtering unit first extracts the average gray value of the soil background. When the gray value is in the range of 0-80 (dark soil), the polarization angle is adjusted to 0°-30° and the filter transmittance is increased to 80%. When the gray value is in the range of 81-180 (medium brightness soil), the polarization angle is adjusted to 30°-60° and the filter transmittance is set to 60%. When the gray value is in the range of 181-255 (bright soil), the polarization angle is adjusted to 60°-90° and the filter transmittance is reduced to 40%. Through the linkage adjustment of gray value and polarization parameter, stray light is accurately filtered.

[0025] The BeiDou positioning unit synchronously collects the spatial coordinates of the sampling points and associates them with the soil type information of the sampling points to construct the correspondence between the collected images and the soil environment.

[0026] The background descrambling module is based on the in-situ soil images and related data acquired by the multispectral acquisition module. It constructs a four-level preprocessing algorithm framework of farmland background layer modeling, target enhancement, mask segmentation, and small particle size feature completion. This achieves deep integration of in-situ unprocessed farmland acquisition and complex background descrambling. First, it calls a pre-constructed dedicated database of farmland soil background, which contains image features of different soil types, organic matter content, straw debris, and mineral particles. Through multispectral difference operation and background feature matching, it removes the broad-spectrum background interference of soil organic matter and moisture, and generates a differential image with enhanced microplastic features. Farmland background stratification modeling is divided into four layers based on the characteristic differences of soil background: organic matter layer, mineral particle layer, straw debris layer, and water layer. Based on five sets of feature spectral images acquired by multispectral acquisition, the spectral reflectance characteristics and spatial distribution characteristics of each layer are extracted to construct background feature models for each layer. Each layer model is associated with the feature threshold of the corresponding soil type, providing a stratification matching basis for subsequent multispectral difference operations.

[0027] Multispectral differential operation selects the near-infrared band (1400-1700nm) and the visible light band (400-600nm) where the microplastic characteristics of farmland are most significant for pixel-level differential operation. The difference gray value is obtained by subtracting the gray value of the visible light band from the gray value of the near-infrared band pixel. Then, combined with the differential threshold range of the dedicated database of farmland soil background, the differential gray values ​​are filtered to retain the abnormal differential gray values ​​corresponding to microplastics and remove the normal differential gray values ​​of organic matter and moisture. Finally, a differential image with enhanced microplastic characteristics is generated.

[0028] Secondly, adaptive bilateral filtering and Retinex algorithm are used to eliminate image grayscale deviation caused by uneven illumination, while preserving subpixel-level edge details of microplastic particles and avoiding feature loss of small-diameter particles. Adaptive bilateral filtering dynamically adjusts the filtering window and smoothing factor for farmland soil scenes. For sandy soil (strong granularity), a 3×3 filtering window is set with a smoothing factor of 0.1; for clay soil (fine texture), a 5×5 filtering window is set with a smoothing factor of 0.3; and for loamy soil, a 4×4 filtering window is set with a smoothing factor of 0.2. During the filtering process, pixels in the edge area of ​​microplastics are protected, preserving the abrupt change characteristics of pixel grayscale values, and smoothing and denoising are performed only on the soil background area.

[0029] To address the uneven illumination characteristics of farmland soil images, a multi-scale Retinex algorithm was used for processing. Three scale factors adapted to the farmland scene were selected to complete the image grayscale correction. During the correction process, grayscale values ​​were suppressed in the highlight areas corresponding to microplastic particles and grayscale values ​​were enhanced in the dark areas. The algorithm only processed the unevenly illuminated areas of the soil background, preserving the grayscale characteristics and edge details of the microplastic particles themselves, and avoiding the distortion of microplastic features caused by the algorithm processing.

[0030] Then, based on the U-Net network (equipped with a dedicated training set for farmland microplastic samples), pixel-level foreground mask segmentation is achieved to separate suspected microplastic targets from complex backgrounds, generating ROI regions for single-particle targets and completing the separation of targets from the background. Finally, a small-particle-size microplastic feature completion module is set up to supplement edge contours and texture features for microplastics with a particle size of less than 1 mm, effectively improving the accuracy of small-particle-size microplastic recognition.

[0031] The U-Net network was adapted to farmland soil scenarios with lightweight modifications. The core structure of 4 downsampling layers at the encoder and 4 upsampling layers at the decoder was retained, while the last convolutional layer at the encoder was removed. Upsampling was performed using transposed convolution and feature fusion. The network input was a 256×256 pixel multispectral difference image, and the output was a binary mask image of the same size, with 0 representing the background and 1 representing suspected microplastic targets. The training batch size was set to 16, the initial learning rate was 0.001, and the training was iterated for 200 epochs. The training set consisted of microplastic labeled samples of farmland soil, including 6 types of soil backgrounds such as sandy and clayey, with a sample size of ≥10,000 images.

[0032] The small-particle-size microplastic feature completion module first extracts the missing edge points of small-particle-size microplastics using an edge detection operator, and then completes the continuous edge contour based on the gray values ​​of neighboring pixels. Next, for areas with missing texture features, it refers to the texture feature library of microplastics of the same material and particle size class, and uses a neighborhood texture mapping method to supplement texture details. In the completion process, the uniqueness of the original features of microplastics is preserved to avoid confusion with soil impurity features.

[0033] The feature fusion module integrates a morphological feature extraction unit, a spectral feature extraction unit, a texture feature extraction unit, and a polarization feature extraction unit. Based on the microplastic feature enhancement differential image and ROI region output by the background descrambling module, it constructs a four-dimensional feature fusion extraction framework of morphology, spectrum, texture, and polarization. It sets up a special feature calibration mechanism for farmland microplastics to enhance feature discrimination. The morphological feature extraction unit extracts geometric features such as particle size, aspect ratio, roundness, and edge roughness of the target. Combined with the morphological characteristics of farmland microplastics, such as the fact that broken agricultural film microplastics are mostly flakes and pesticide packaging microplastics are mostly granular, it has an adaptive adjustment function for morphological feature weights to enhance the discrimination against soil impurities (irregular mineral particles and straw debris). The morphological feature weight adjustment is based on soil type and microplastic morphological characteristics. For farmland with a high proportion of microplastics in sheet agricultural film, the feature weights of aspect ratio and edge roughness are increased (each by 20%). For farmland with a high proportion of microplastics in granular pesticide packaging, the feature weights of roundness and particle size are increased (each by 25%). For areas with a lot of irregular mineral particles in the soil, the weight of edge roughness is further increased by 15%. Through weight adjustment, the morphological differentiation between microplastics and soil impurities is strengthened.

[0034] The spectral feature extraction unit extracts spectral features such as reflectance peaks, spectral slopes, and absorption valley positions of targets in multi-band images, and constructs a dedicated spectral fingerprint database for common agricultural microplastics such as PE, PP, PVC, and PET, enabling them to be distinguished from non-plastic impurities such as organic matter and insect eggs. The spectral fingerprint database stores the core characteristic values ​​and tolerance ranges of agricultural microplastics such as PE, PP, PVC, and PET, including the peak reflectance and absorption valley positions. During matching, the extracted target spectral features are compared one by one with the data in the fingerprint database. If the feature matching degree is ≥85%, it is determined to be the corresponding type of microplastic. If the matching degree is <60%, it is directly determined to be a non-plastic impurity. In the 60%-85% range, polarization characteristics are combined for further verification and differentiation.

[0035] The texture feature extraction unit extracts texture features such as contrast, correlation, and entropy of the target through the gray-level co-occurrence matrix (GLCM). Combined with the texture change pattern during the aging process of microplastics, it has a texture feature temporal calibration function to improve the differentiation accuracy of microplastics with different aging levels. Texture feature temporal calibration is based on the aging time sequence of microplastics in farmland. First, a texture feature benchmark library of unaged, lightly aged, and heavily aged microplastics is established. After extracting texture features such as contrast, correlation, and entropy value of the target to be detected, a temporal comparison is performed with the texture features of microplastics of the same region and material in the benchmark library. Texture feature deviations are corrected according to the temperature and humidity of the farmland environment. Under high temperature and high humidity conditions, the entropy value features are positively calibrated by 10%-15%. After calibration, they are incorporated into the feature fusion system.

[0036] To address the textural characteristics of microplastics in farmland, a 3×3 gray-level co-occurrence matrix calculation window was set up. Texture features were calculated in four directions: 0°, 45°, 90°, and 135°. The average of the calculation results in the four directions was taken as the final texture feature values ​​such as contrast, correlation, and entropy. The step size was set to 1, and the gray-level quantization was set to 64 levels, which is suitable for the feature extraction requirements of microplastics in farmland, which are mostly small-diameter particles with few texture details.

[0037] The polarization feature extraction unit extracts polarization features such as polarization degree and polarization angle of the target. To address the difficulty of distinguishing transparent microplastics from water droplets and glass fragments, the distinguishing weight of polarization features is strengthened. Finally, an attention mechanism (equipped with a farmland microplastic feature priority judgment module) is used to adaptively weight and fuse the four-dimensional features, strengthen the weight of high-discrimination features, weaken the interference of redundant features, and generate a standardized high-discrimination microplastic feature vector.

[0038] The CBAM attention mechanism is adopted and the weights are adjusted according to the microplastic features of farmland. The weights are calculated sequentially by channel attention and spatial attention. The channel attention assigns basic weights of morphology, spectrum, texture and polarization to the four-dimensional features as 0.2, 0.35, 0.25 and 0.2 respectively. The weights are then dynamically adjusted according to the background type of farmland soil. For sandy soil, the weight of polarization features is increased to 0.3 and for clay soil, the weight of spectral features is increased to 0.4. Spatial attention focuses on the edge region of microplastics and the convolution kernel is set to 3×3. The input of the attention mechanism is a 256-dimensional feature matrix after the four-dimensional features are concatenated, and the output is a standardized 128-dimensional microplastic feature vector, which is normalized to the interval [0, 1].

[0039] First, morphological features (4 dimensions such as particle size and aspect ratio), spectral features (128 dimensions such as reflectance peak), texture features (64 dimensions such as contrast and entropy), and polarization features (8 dimensions such as degree of polarization and polarization angle) are sequentially spliced ​​in dimensional order to form a 204-dimensional original feature vector. Then, the feature values ​​of each dimension are mapped to the [0, 1] interval using the maximum-minimum normalization method. Outlier detection is used to remove abnormal feature values. Finally, a standardized 128-dimensional microplastic feature vector is generated by compression, which ensures the uniformity and comparability of the feature vector in the edge recognition module.

[0040] The edge recognition module receives the standardized high-discrimination microplastic feature vector output by the feature fusion module. Based on the YOLO lightweight network, it constructs an intelligent identification and classification model for farmland microplastics that is adapted to edge-embedded devices. It sets up a multi-task collaborative optimization mechanism adapted to the special scenario of farmland microplastic detection. The identification and classification model sets up multi-task parallel output branches to simultaneously realize the three core functions of target detection, material classification and aging degree classification. The feature conflict between multiple tasks is handled through the task collaborative optimization algorithm. The multi-task collaborative optimization mechanism first decouples and extracts features from the three tasks of target detection, material classification, and aging degree classification, separating the specific features of each task from the shared features. When feature conflicts occur, the specific features of each task are retained, and the shared features are weighted according to the importance of the task. The weighted fusion is then input into the output branches of each task, with target detection receiving 0.4, material classification receiving 0.3, and aging degree classification receiving 0.3. This weighted fusion is then input into the output branches of each task to avoid the decrease in recognition accuracy caused by feature conflicts.

[0041] The feature division criteria are defined according to task requirements: shared features are basic features such as edge contour, particle size, and core spectral peak of microplastics, which are jointly relied upon by the three tasks of target detection, material classification, and aging degree classification; target detection-specific features are particle pixel coordinates and contour area; material classification-specific features are spectral fingerprint and polarization features; and aging degree classification-specific features are texture entropy value and spectral reflectance attenuation rate. Feature decoupling is completed according to this standard to ensure the relevance of feature extraction for each task.

[0042] Target detection is used to identify microplastic particles within the ROI area and count the number and size distribution of particles. A dedicated detection branch for small-diameter microplastics is set up to improve detection accuracy. Material classification is based on fused feature vectors to classify common agricultural microplastics such as PE, PP, PVC, and PET. The classifier parameters are optimized in combination with the material distribution characteristics of microplastics in farmland to improve classification accuracy. A dedicated detection branch accesses the shallow features (80×80 pixels) of the YOLOv8n neck feature fusion layer to specifically extract edge features of microplastics with a particle size <1mm. The branch input is a 32×32 pixel ROI sub-region feature map, and the output is the microplastic pixel coordinates, particle size pixel value (1 pixel corresponds to 0.02mm in reality), and detection confidence. The confidence threshold is set to 0.5. Targets below the threshold will trigger a secondary verification by the small particle size microplastic feature completion module.

[0043] Material classification was conducted for the five most common agricultural microplastics in farmland: PE, PP, PVC, PET, and PS. The classifier used was the Softmax classifier. The classification weights were adjusted according to the actual pollution situation in the farmland, with PE accounting for 40%, PP for 30%, and PVC / PET / PS each accounting for 10%. At the same time, a confidence level was set for the classification results. Results with a confidence level <0.7 were combined with polarization characteristics for secondary judgment to further improve the classification accuracy.

[0044] The aging degree classification is based on texture and spectral characteristics, dividing microplastics into three levels: unaged, slightly aged, and heavily aged. An aging characteristic calibration branch is set up and the influence of farmland environment (temperature and humidity) on microplastic aging is combined to improve the classification accuracy. The aging feature calibration branch uses texture features (entropy value, contrast) and spectral features (reflectance peak offset) as the core calibration basis. Combined with the annual average temperature and annual humidity data of farmland sampling points, it corrects the aging features of microplastics under different environments. For example, in farmland areas with an annual average temperature >25℃ and annual humidity >60%, the aging judgment threshold of reflectance peak offset is reduced by 10%, and the judgment threshold of texture entropy value is increased by 15% to avoid misjudgment of aging degree due to environmental differences. At the same time, this branch will remove feature anomalies caused by soil stains and retain the aging features of microplastics themselves.

[0045] Based on a dedicated dataset for farmland microplastics, including different soil types, pollution levels, microplastic materials, and aging grades, the model is pre-trained, structured pruning, and INT8 quantization is performed. Knowledge distillation technology is also employed to transfer the recognition experience of complex models to lightweight models. This achieves a lightweight model while maintaining high recognition accuracy, enabling smooth operation on low-power embedded edge devices. The single-point image recognition processing speed is ≤5s, allowing for real-time detection in farmland. Furthermore, the model can adaptively fine-tune according to changes in the farmland scene, adapting to the microplastic detection needs of different regions.

[0046] YOLOv8n was selected as the basic lightweight network. The neck feature fusion structure was modified to adapt to the farmland scene. One convolutional layer after removing the SPPF module was added, and a small-particle-size feature fusion branch was added. The model pre-training learning rate adopted the cosine annealing strategy with an initial learning rate of 0.0005, a batch size of 32, and 150 epochs. INT8 quantization was performed using farmland microplastic samples as the calibration set, and the number of model parameters after quantization was ≤5M. Knowledge distillation was performed using YOLOv8x as the teacher model. After distillation, the core accuracy of the model's target detection decreased by ≤3%. The distillation temperature was set to 4, and the loss weights were assigned as classification loss 0.3, detection loss 0.5, and aging grading loss 0.2.

[0047] The pre-training uses a dedicated dataset of microplastics in farmland as the training set, with an initial learning rate of 0.0005 and a batch size of 32. Structured pruning removes redundant convolutional and fully connected layers from the network. INT8 quantization uses farmland soil samples as the calibration set. Knowledge distillation uses a YOLO model with a large number of parameters as the teacher model, with a distillation temperature of 4, enabling the lightweight model to perform fast inference on embedded devices at the edge, with a single-point image recognition processing speed of ≤5s.

[0048] The 5-second time consumption for single-point image recognition is allocated as follows: feature vector input and preprocessing 0.5s, identification of small-particle-size microplastics special detection branch 1s, target detection main branch statistics of particle number and particle size distribution 1s, material classification branch operation 1s, aging degree grading branch operation 0.5s, and multi-task result fusion and output 1s. Each step adopts parallel computing mode and is processed synchronously on the independent computing core of the edge embedded device to ensure that the total consumption time is controlled within 5s. If any feature abnormality occurs in any step, a secondary verification of 0.5s will be automatically triggered without occupying the overall basic consumption time.

[0049] The spatial modeling module, based on the spatial coordinates of sampling points obtained by the multispectral acquisition module and the microplastic pollution data output by the edge recognition module, combines pattern recognition and modeling techniques to construct a gridded spatial model of microplastic pollution in farmland. It also sets up a pollution source tracing accuracy optimization function to improve the accuracy of modeling and source tracing. First, the target farmland is divided into a standard grid of 10m×10m. Based on the Kriging interpolation method, combined with the soil type and microplastic distribution characteristics of the sampling points, spatial interpolation is performed on the pollution data of discrete sampling points to generate a heat map of microplastic pollution distribution in the entire farmland, thereby improving the interpolation accuracy. To address the spatial characteristics of microplastic pollution distribution in farmland, Kriging interpolation uses a spherical variogram as the fitting function. Interpolation sub-regions are divided according to soil type, with different variogram parameters set for sandy soil and clay soil. The interpolation search neighborhood is set as a fixed neighborhood with ≥10 sampling points within the neighborhood. Discrete abnormal pollution data that exceed a reasonable range are preprocessed and removed to ensure the authenticity and accuracy of the pollution heat map after interpolation.

[0050] A risk level calibration module is set up to optimize the pollution risk classification standard by combining the type of farmland planting and the history of agricultural film use. Then, based on pollution concentration, plastic type and particle size distribution data, combined with pattern recognition algorithm, each grid is classified into three levels: low risk, medium risk and high risk. Random forest was selected as the pattern recognition algorithm for pollution risk classification. The algorithm inputs 5-dimensional feature data for each 10m×10m grid, including microplastic particle density, small particle size ratio, PE / PP agricultural plastic ratio, aged microplastic ratio, and soil organic matter content. The output is a grid risk level label, 0=low risk, 1=medium risk, and 2=high risk. The algorithm training sample size is ≥5000 groups, the number of decision trees is set to 100, the maximum depth is set to 8, and the model classification accuracy is ≥90% after cross-validation.

[0051] Based on the national limit standard for microplastic pollution in farmland soil, quantitative classification thresholds are set by combining the density of microplastic particles and the proportion of small-diameter particles (<1mm) within the grid: low risk is defined as microplastic particle density <50 particles / kg and small-diameter particle proportion <30%; medium risk is defined as microplastic particle density 50-200 particles / kg or small-diameter particle proportion 30%-60%; high risk is defined as microplastic particle density >200 particles / kg or small-diameter particle proportion >60%. For different types of farmland such as fruits and vegetables and grains, the thresholds are slightly adjusted by 20% to adapt to the pollution risk assessment needs of different planting scenarios.

[0052] Simultaneously, by combining relevant data such as the history of agricultural film use, irrigation routes, and pesticide packaging placement points in farmland, source tracing analysis is conducted on highly polluted areas to pinpoint the core sources of pollution, improve the accuracy of source tracing, and ultimately generate a standardized spatial distribution report of microplastic pollution in farmland.

[0053] The pollution source tracing will be carried out in three steps: The first step is to extract the core microplastic characteristics (material type, particle size distribution, and aging degree) of the highly polluted grid and match them with the plastic types and years of use in the history of agricultural film use to initially identify suspected sources of pollution. The second step is to analyze the pollution concentration gradient between the high-pollution grid and the upstream and downstream sampling points based on the water flow direction of the irrigation path to determine the direction of pollution diffusion. The third step involves overlaying the spatial coordinates of pesticide packaging placement points and agricultural film recycling points, calculating the distance from each point to the high-pollution grid and the pollution similarity. Points with a pollution similarity ≥90% and the closest distance are identified as the core source of pollution. If multiple suspected sources exist, the time-series characteristics of microplastic aging are combined for further precise identification.

[0054] The result early warning module receives microplastic detection statistics output by the edge recognition module and pollution distribution heat map, pollution source analysis results and risk level data generated by the spatial modeling module. It integrates a high-definition touch screen, a voice broadcast unit and a local data storage unit for real-time display and output of multiple core data. The core data includes full-dimensional statistical data on the number, particle size distribution, material type and aging degree of microplastic particles. The result early warning module is equipped with a recognition result verification function to judge the rationality of the data and avoid misjudgment. The identification results are verified using a dual verification method. The first step is data logic verification, which verifies the logical rationality of the numerical relationships between the number of microplastic particles, particle size distribution, and pollution concentration, and removes abnormal data that clearly does not conform to the distribution pattern of microplastic pollution in farmland. The second step is feature comparison verification, which extracts the core features of microplastics in the detection results and compares them with the basic feature database of farmland microplastics in the local area. If the feature deviation exceeds a reasonable range, a re-detection prompt is triggered to ensure the accuracy of the detection data.

[0055] The core data also includes the pollution concentration and risk level of the sampling points, as well as the heat map of farmland pollution distribution and the results of pollution source tracing analysis. The result early warning module has built-in threshold standards corresponding to the control of microplastic pollution in farmland soil. When the pollution concentration at the detection point exceeds the national standard limit, it automatically triggers sound, light and voice warnings to prompt the operator to conduct key verification. All test data can be exported on-site as standardized test reports, and simultaneously encrypted and stored in the on-site test terminal equipped with a local data storage unit. It supports offline testing and data backtracking in environments without network access, is suitable for farmland and field conditions without network access, and has a data backtracking and verification function that can verify the accuracy of test results through image comparison.

[0056] During offline data backtracking, operators can retrieve historical test data and original collected images in the local data storage unit by sampling time, sampling point coordinates, and soil type through the touch screen. The system supports exporting and viewing single / batch data. The image comparison and verification compares the microplastic feature image corresponding to the test result to be backtracked with the standard feature image of the same type of microplastic in the same area in the past at the pixel level. If the feature matching degree is ≥80%, the test result is deemed valid. If the matching degree is <80%, it is marked as a suspicious result. At the same time, the original collected image is retrieved for manual review by the operator.

[0057] Local data is stored using a dual encryption method of hardware encryption and data anonymization. At the hardware level, the data is encrypted at the underlying level through the embedded encryption chip of the detection terminal, and can only be unlocked with the terminal's exclusive verification key. At the data anonymization level, the precise coordinates of the sampling points are blurred (retained down to the village level), and the original detection images are encrypted with watermarks. The detection data is stored in a dedicated encrypted format, and a data verification code is generated to ensure that the data is not tampered with or leaked during offline storage.

[0058] The cloud-based iteration module receives anonymized detection data and manually reviewed and labeled data uploaded by the aforementioned multispectral acquisition module, edge recognition module, and result warning module. It constructs a closed-loop iteration system for data acquisition, model optimization, and terminal updates, and sets up a dedicated incremental learning mechanism for farmland scenarios to improve the model's cross-scenario adaptability and iteration efficiency. First, the cloud-based iteration module classifies and stores the received data, constructs a dynamically updated national farmland microplastic image database, and classifies and organizes the data according to soil type, regional climate, and pollution type to provide targeted data support for model optimization. The specific execution process of the closed-loop iterative system is as follows: 1. The edge device uploads anonymized detection data and corresponding original collected images daily. The uploaded data must include the soil type of the sampling point and the microplastic detection results. The data format is standardized JSON format. 2. The cloud platform performs initial screening of uploaded data, removing invalid data before pushing it to the manual review stage. Manual review requires marking any questionable points in the test results and small-diameter microplastic identification results, and then sending the marked data back to the cloud. Invalid data is rejected based on a triple judgment standard: 1) Detection data with missing sampling point coordinates or spatial offsets >5m and the corresponding original acquired image; 2) Original data with blurred images or noise levels >30%, making it impossible to extract microplastic features; 3) Abnormal data where microplastic detection results exceed the reasonable pollution range for farmland (microplastic particle density >1000 particles / kg). The cloud-based algorithm automatically verifies data integrity, image clarity, and numerical reasonableness to quickly filter invalid data. 3. The cloud adds the verified and labeled valid data to the incremental training set, and completes model optimization after the incremental training conditions are met; 4. The optimized model is tested and verified in the cloud. After the recognition accuracy is improved, a lightweight update package is generated and pushed to the edge terminal in the corresponding area to complete the terminal model update. The model OTA remote update is pushed precisely by region, pushing the adapted model update package only to edge terminals corresponding to the soil type and climate region. The update package adopts an incremental push method (only transmitting the optimized parameters of the model) and supports breakpoint resume. After receiving the update package, the edge terminal automatically calls the local test sample set to perform a self-check of the model accuracy. If the comprehensive accuracy of target detection, material classification, and aging grading is improved by ≥1% compared with the original model, the model is automatically replaced and the original model is saved as a backup. If the accuracy does not meet the standard, the update package is automatically discarded and the original model is kept running, while the update anomaly information is reported to the cloud.

[0059] Secondly, based on the incremental learning algorithm (equipped with priority weights for farmland scene features), the identification and classification model is continuously fine-tuned and optimized. Without retraining the entire model, it can quickly adapt to farmland scenes with different soil types, climate regions, and pollution types. A fine-tuning incremental learning algorithm is adopted, using anonymized labeled data uploaded from the edge terminal as the incremental training set. Model fine-tuning is triggered when the number of incremental samples (including labeled images and corresponding detection data) in a single batch is ≥500. Fine-tuning only trains the head output layer of the YOLO lightweight network, with a learning rate of 0.0001, a batch size of 16, and 20 epochs. The model update threshold is when the recognition accuracy of the new sample set is improved by ≥2%. After the threshold is reached, a lightweight model update package is generated in the cloud and pushed to the edge terminal via OTA.

[0060] Meanwhile, the cloud platform can manage the entire lifecycle of the model, push the optimized lightweight model to the edge terminal, realize remote OTA update of the terminal model, and continuously improve the system's recognition accuracy and scene adaptability.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] 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 rapid identification system for microplastics in farmland based on image recognition, characterized in that, Includes the following modules: The multispectral acquisition module is used to acquire multi-band spectral images of in-situ soil in farmland, simultaneously obtain the spatial coordinates of sampling points and soil type information, complete the pre-processing linkage between acquisition and preprocessing, and reduce environmental interference. The background descrambling module, based on the raw data from the multispectral acquisition module, removes various background impurities from farmland soil through a multi-step preprocessing algorithm, while preserving the edge and detail features of microplastic particles, thus achieving separation of microplastic targets from the background. The feature fusion module extracts multi-dimensional features of microplastics, including geometry, spectrum, texture, and polarization, based on the processed data after background descrambling. It strengthens effective features and weakens redundant interference through an adaptive weighted fusion mechanism to generate standardized, highly discriminative microplastic feature vectors. The edge recognition module receives the feature vector output by the feature fusion module, constructs a lightweight edge intelligent recognition model adapted to farmland scenarios, and simultaneously realizes multi-task processing of microplastic target detection, material classification and aging degree classification, and has the ability to adapt and fine-tune according to the scenario. The spatial modeling module, based on sampling point coordinate data and microplastic pollution data, constructs a rasterized model of farmland microplastic pollution, completes pollution risk classification and source tracing analysis, and generates a standardized distribution report; The results early warning module integrates data display, storage and verification functions, has built-in pollution threshold standards, and realizes on-site data output, abnormal early warning and offline backtracking; The cloud-based iteration module constructs a closed-loop iteration system, optimizing the lightweight edge-end intelligent recognition model through incremental learning, and realizing remote model updates and multi-terminal data aggregation and analysis.

2. The rapid identification system for microplastics in farmland based on image recognition according to claim 1, characterized in that, The multispectral acquisition module integrates an imaging unit, a multi-band ring light source, an adaptive illumination compensation unit, a polarization filtering unit, and a BeiDou positioning unit. The imaging unit covers two characteristic bands: visible light and near-infrared, and can acquire five sets of characteristic spectral images in the range of 400nm to 1700nm. It has the function of adapting to farmland soil texture and can automatically adjust imaging parameters according to the content of soil sand and clay particles. The adaptive illumination compensation unit combines changes in outdoor lighting to automatically adjust the light source parameters to eliminate image noise caused by strong light reflection and shadow occlusion. The polarization filtering unit is used to filter diffuse reflection stray light in the soil background and highlight the edge contours and material characteristics of microplastic particles, and has the function of adaptive polarization angle adjustment. The BeiDou positioning unit synchronously acquires the coordinates of sampling points and associates them with soil type information to establish a correspondence between the acquired images and the soil environment.

3. The rapid identification system for microplastics in farmland based on image recognition according to claim 2, characterized in that, The background descrambling module constructs a four-level preprocessing algorithm framework for farmland background layering modeling, target enhancement, mask segmentation, and small-particle-size feature completion, achieving the integration of in-situ, preprocess-free farmland acquisition and complex background descrambling. First, multispectral difference operations and background feature matching are used to remove background interference, including soil organic matter and moisture, generating a microplastic feature-enhanced image. Then, adaptive bilateral filtering and the Retinex algorithm are used to eliminate the influence of uneven illumination and preserve the edge details of microplastics. Subsequently, pixel-level segmentation based on the U-Net network is used to separate suspected microplastic targets from the background, generating single-particle target ROI regions. Finally, a small-particle-size microplastic feature completion module is set up to supplement edge and texture features for microplastics with a particle size of less than 1 mm.

4. The rapid identification system for microplastics in farmland based on image recognition according to claim 3, characterized in that, The feature fusion module integrates a morphological feature extraction unit, a spectral feature extraction unit, a texture feature extraction unit, and a polarization feature extraction unit to construct a four-dimensional feature fusion extraction framework and set up a dedicated feature calibration mechanism; The morphological feature extraction unit extracts geometric features of microplastics, including particle size, aspect ratio, roundness, and edge roughness, and has an adaptive feature weight adjustment function; the spectral feature extraction unit extracts spectral features of microplastics and distinguishes them from non-plastic impurities based on a dedicated spectral fingerprint database of common agricultural microplastics in farmland; the texture feature extraction unit extracts texture features of microplastics and has a time-series calibration function that incorporates the texture change patterns during the aging process of microplastics; the polarization feature extraction unit extracts polarization features of microplastics and strengthens the distinguishing weights; finally, the four-dimensional features are adaptively weighted and fused through an attention mechanism to generate a standardized feature vector.

5. The rapid identification system for microplastics in farmland based on image recognition according to claim 4, characterized in that, The edge recognition module is based on the YOLO lightweight network to build a lightweight edge-end intelligent recognition model for microplastics in farmland, adapted to embedded edge devices. It features a dedicated multi-task collaborative optimization mechanism for farmland scenarios. The model has parallel output branches for multiple tasks, simultaneously achieving target detection, material classification, and aging degree grading. Target detection can identify microplastics within the ROI area and statistically analyze particle number and size distribution parameters, with a dedicated detection branch for small-diameter microplastics. Material classification uses fused feature vectors to classify common agricultural microplastic types. Aging degree grading divides microplastics into three levels: unaged, slightly aged, and heavily aged, with an aging feature calibration branch. The model is optimized through pre-training, structured pruning, INT8 quantization using farmland microplastic samples as a calibration set, and knowledge distillation, enabling it to adaptively fine-tune according to changes in farmland scenarios.

6. A rapid identification system for microplastics in farmland based on image recognition according to claim 5, characterized in that, The spatial modeling module, combined with pattern recognition technology, divides the target farmland into standard 10m×10m grids and generates a heat map of microplastic pollution distribution across the entire farmland using Kriging interpolation. A risk level calibration module is set up to optimize risk classification standards based on farmland planting type and agricultural film usage history, classifying each grid based on pollution-related data. Simultaneously, by combining associated data including agricultural film usage history, irrigation paths, and pesticide packaging placement points, source tracing analysis is conducted on high-pollution areas to pinpoint the core pollution sources, ultimately generating a standardized spatial distribution report of pollution.

7. A rapid identification system for microplastics in farmland based on image recognition according to claim 6, characterized in that, The result warning module integrates a touch screen, a voice broadcast unit, and a local data storage unit. It displays and outputs real-time comprehensive statistical data, including the number of microplastic particles, particle size distribution, material type, and aging degree, as well as pollution information and pollution analysis results at sampling points. It has a result verification function and incorporates the national limit standard for microplastic pollution in farmland soil. When the pollution concentration exceeds the standard, it automatically triggers sound, light, and voice warnings. The detection data can be exported on-site as a standardized report, encrypted and stored on the local terminal. It supports offline detection and data backtracking in environments without network access and has a data backtracking verification function based on image comparison.

8. A rapid identification system for microplastics in farmland based on image recognition according to claim 7, characterized in that, The cloud-based iterative module receives anonymized detection data and manually reviewed and labeled data uploaded by the aforementioned modules, constructs a dynamically updated national farmland microplastic image database, and categorizes and organizes it according to soil type, regional climate, and pollution type. Based on the incremental learning algorithm, it continuously optimizes the identification and classification model, realizes full lifecycle version management of the model through the cloud platform, and pushes the optimized lightweight model to the edge terminal.

Images