Blueberry intelligent disease detection method and system based on multimodal unsupervised learning

Abstract

The present invention is applicable to the field of intelligent agricultural technology and provides a method and system for intelligent blueberry disease detection based on multimodal unsupervised learning. The method comprises the following steps: based on an adaptive contrast deep clustering framework, feature extraction and clustering are performed on preprocessed multimodal data to obtain multimodal preliminary features and preliminary clustering results; based on a multimodal complementary feature fusion mechanism, the multimodal preliminary features are adaptively weighted and fused according to the multimodal preliminary features and preliminary clustering results to obtain multimodal fusion features; unsupervised clustering optimization and disease type identification are performed on the multimodal fusion features to obtain an unsupervised learning model; and the unsupervised learning model is deployed and incrementally learned to detect blueberry diseases. The present invention achieves early and accurate detection of blueberry diseases through an unsupervised learning algorithm, providing a new paradigm for intelligent and precise blueberry management, effectively improving disease prevention and control efficiency and the economic benefits of the industry.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural intelligent technology, and in particular relates to a blueberry intelligent disease detection method and system based on multimodal unsupervised learning. Background Art

[0002] Blueberry, a high-value cash crop, has seen rapid global expansion in recent years. However, its leaves are highly susceptible to fungal and bacterial leaf spot diseases, severely hindering the industry's development. Traditional disease detection methods rely heavily on manual judgment, which is inefficient, costly, and has limited accuracy.

[0003] With the development of deep learning technology, although there have been some breakthroughs in disease detection methods based on image recognition, most of them rely on large amounts of labeled data for training and are not adaptable enough when faced with new varieties or atypical diseases. Summary of the Invention

[0004] The purpose of the present invention is to provide a blueberry intelligent disease detection method based on multimodal unsupervised learning, aiming to solve the above technical problems.

[0005] The present invention is implemented as follows: a blueberry intelligent disease detection method based on multimodal unsupervised learning, comprising the following steps:

[0006] Collecting multimodal data of blueberry samples; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data;

[0007] Preprocessing the multimodal data to obtain preprocessed multimodal data;

[0008] Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results;

[0009] Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features;

[0010] Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model;

[0011] The unsupervised learning model is deployed and incrementally learned to detect blueberry diseases; the blueberry disease is leaf spot.

[0012] Preferably, the step of preprocessing the multimodal data to obtain preprocessed multimodal data specifically includes:

[0013] Multimodal data registration algorithm based on phase consistency is used to spatially align multimodal data;

[0014] After alignment, each modality data is preprocessed separately;

[0015] After preprocessing, the data of each modality were standardized to ensure scale consistency;

[0016] The RGB data preprocessing method is as follows: firstly, the illumination non-uniformity is eliminated by improved homomorphic filtering, and then the adaptive wavelet transform is applied to suppress the noise;

[0017] The preprocessing method for hyperspectral data is as follows: first, dimensionality reduction is performed using minimum noise fraction transform, followed by stepwise optimal band selection;

[0018] The preprocessing method of thermal infrared data is as follows: first, noise suppression is performed through improved bilateral filtering, and then adaptive background correction is applied to eliminate the influence of temperature drift.

[0019] Preferably, based on the adaptive contrast deep clustering framework, the steps of extracting features and clustering the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results specifically include:

[0020] A self-supervised semantic encoder is constructed to extract initial feature representations. The encoder is based on the Vision Transformer architecture and introduces a local-global attention mechanism to enhance feature extraction capabilities.

[0021] Design four types of image enhancement transformations to generate multi-view samples, and construct a local-global contrast loss function and a prototype-perceptual contrast learning module; the four types of image enhancement transformations include random cropping and horizontal flipping, color jittering, affine transformation, Gaussian blurring, and random erasing;

[0022] The disease complexity score is calculated by intra-class distance and inter-class distance. At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect. The temperature parameters are adaptively adjusted according to the training progress and feature distribution to dynamically adjust the number of clusters and prototype distribution.

[0023] Preferably, based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and the preliminary clustering results, the step of adaptively weighting and fusing the multimodal preliminary features to obtain the multimodal fusion features specifically includes:

[0024] Perform intra-modal feature enhancement on the multimodal preliminary features to obtain multi-scale feature maps of each modality; wherein the multimodal preliminary features include RGB modality, hyperspectral modality, and thermal infrared modality; the RGB modality uses an improved DenseNet to extract multi-scale feature maps; the hyperspectral modality uses a 3D-CNN and Transformer architecture to extract multi-scale feature maps; the thermal infrared modality uses a residual attention network to extract multi-scale feature maps;

[0025] A channel recalibration operation is applied to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels; then cross-modal feature complementation is performed to build an inter-modal self-attention mechanism;

[0026] Cross-scale feature integration is achieved through adaptive feature reorganization. Specifically, the feature uncertainty map is first calculated. Then, multi-scale features are adaptively fused based on the uncertainty map. Finally, multimodal fusion features are calculated through attention-weighted summation.

[0027] Preferably, the step of performing unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model specifically includes:

[0028] Use the spectral clustering algorithm to cluster the multimodal fusion features, integrate and optimize the clustering results with the preliminary clustering results, align the two clusters using the Hungarian algorithm, and calculate the consistency score;

[0029] The final cluster labels are obtained by optimizing the matching by maximizing the consistency score. The quality is further improved by using the cluster re-evaluation strategy, which calculates the Mahalanobis distance from the sample to the cluster center to identify possible abnormal assignments.

[0030] After clustering is completed, significant features are extracted from each final cluster label for use in constructing a disease feature library. Specifically, the importance scores of the significant features are calculated. Based on the importance scores, the top several most significant features are extracted to construct a disease feature library. At the same time, the correspondence between these features and common disease symptoms is analyzed, and an expert knowledge base is constructed to obtain a symptom-feature mapping. Based on the similarity between the symptom-feature mapping and the clustering features, a disease type description is automatically generated for each cluster. At the same time, a density-based anomaly detection algorithm is designed to identify potential unknown diseases and calculate the local outlier factor.

[0031] Preferably, the steps of deploying and incrementally learning the unsupervised learning model to detect blueberry diseases include:

[0032] Adopt edge-cloud collaborative architecture to deploy unsupervised learning models to actual application scenarios;

[0033] An incremental learning mechanism is designed for newly collected samples, and the model is updated and optimized online through feature distillation and category balance loss.

[0034] Another object of the present invention is to provide a blueberry intelligent disease detection system based on multimodal unsupervised learning, which is used to implement the above-mentioned blueberry intelligent disease detection method based on multimodal unsupervised learning, and includes:

[0035] A multimodal data acquisition module for collecting multimodal data of blueberry samples; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data;

[0036] Preprocessing the multimodal data to obtain preprocessed multimodal data;

[0037] Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results;

[0038] Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features;

[0039] Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model;

[0040] Deploy unsupervised learning models and perform incremental learning to detect blueberry diseases.

[0041] The present invention provides an intelligent blueberry disease detection method based on multimodal unsupervised learning, which innovatively integrates multimodal data such as RGB, hyperspectral and thermal infrared, and realizes early and accurate detection of blueberry diseases through unsupervised learning algorithms. It can adapt to various environmental conditions without a large amount of labeled data, providing a new paradigm for intelligent and precise management of blueberries, effectively improving the efficiency of disease prevention and control and the economic benefits of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the overall process of the blueberry intelligent disease detection method based on multimodal unsupervised learning provided by an embodiment of the present invention.

[0043] Figure 2 A schematic diagram of the multimodal data preprocessing process provided by an embodiment of the present invention.

[0044] Figure 3 A schematic diagram of a process for feature learning using an adaptive contrastive deep clustering framework provided in an embodiment of the present invention.

[0045] Figure 4 A schematic diagram of a process for performing feature fusion using a multimodal complementary feature fusion mechanism provided in an embodiment of the present invention.

[0046] Figure 5 A schematic flow chart of the method for unsupervised clustering optimization and disease type identification provided by an embodiment of the present invention.

[0047] Figure 6 A schematic diagram of the process of deploying and incrementally learning an unsupervised learning model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] It should be noted that, if no explanation is given for the various parameters in the formulas involved in the embodiments of the present invention, they are assumed to have existing publicly known meanings and will not be elaborated on here.

[0050] like Figure 1 As shown, in one embodiment of the present invention, a blueberry intelligent disease detection method based on multimodal unsupervised learning is provided, which includes the following steps:

[0051] S1. Collecting multimodal data of a blueberry sample; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data;

[0052] S2. preprocessing the multimodal data to obtain preprocessed multimodal data;

[0053] S3. Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results;

[0054] S4. Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features;

[0055] S5. Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model;

[0056] S6. Deploy and perform incremental learning on the unsupervised learning model to detect blueberry diseases; the blueberry disease is leaf spot.

[0057] In practical applications, blueberry leaves were selected as samples, and three imaging devices were used to simultaneously collect multimodal data from the blueberry samples: a 4K high-resolution RGB camera (resolution 3840×2160), a hyperspectral camera covering the 400-1000nm band (spectral resolution 2.3nm), and a thermal infrared camera (temperature resolution 0.05°C). The data acquisition system utilizes an octagonal array structure, with a set of sensors installed at each angle to ensure 360° coverage. An integrating sphere light source provides uniform diffuse illumination, eliminating shadows and bright spots. During the acquisition process, the blueberry samples were placed on an automated rotating platform, rotating at 15° / second. The multi-sensor synchronous acquisition was triggered every 2 seconds, resulting in a total of 24 sets of multimodal data for each sample.

[0058] like Figure 2As shown, in a preferred embodiment of the present invention, the step of preprocessing the multimodal data to obtain preprocessed multimodal data specifically includes:

[0059] S21, a multimodal data registration algorithm based on phase consistency to spatially align multimodal data;

[0060] S22, after alignment, preprocessing each modality data separately;

[0061] S23. After the preprocessing is completed, the data of each modality are standardized to ensure the scale consistency;

[0062] The RGB data preprocessing method is as follows: firstly, the illumination non-uniformity is eliminated by improved homomorphic filtering, and then the adaptive wavelet transform is applied to suppress the noise;

[0063] The preprocessing method for hyperspectral data is as follows: first, dimensionality reduction is performed using minimum noise fraction transform, followed by stepwise optimal band selection;

[0064] The preprocessing method of thermal infrared data is as follows: first, noise suppression is performed through improved bilateral filtering, and then adaptive background correction is applied to eliminate the influence of temperature drift.

[0065] In practical applications, after obtaining the original multimodal data (image), the first step is to apply a phase-consistency-based multimodal data registration algorithm to achieve spatial alignment. This algorithm combines SIFT feature point extraction and an improved RANSAC method to calculate the f transformation matrix T:

[0066] ;

[0067] The alignment error is quantified by the root mean square displacement (RMSD): ;

[0068] in, and are the corresponding points in the source image and the target image respectively; the registration optimization adopts the gradient descent method, takes RMSD as the objective function, and iteratively optimizes the transformation matrix to achieve pixel-level accurate matching (RMSD < 0.5 pixels).

[0069] After alignment, each modality data is preprocessed as follows:

[0070] The RGB data is first filtered to eliminate illumination unevenness:

[0071] ;

[0072] in, is the original image, is a Gaussian high-pass filter, is the contrast enhancement parameter (value ranges from 1.5 to 2.2). Adaptive wavelet transform is then applied to suppress noise, Daubechies wavelet (db4) is selected for 4-level decomposition, and the improved Stein unbiased risk estimation (SURE) threshold method is used:

[0073] ;

[0074] Where d is the wavelet coefficient, is the adaptive threshold, is the regularization term (value 10 -6 ).

[0075] Hyperspectral data is reduced in dimension using the minimum noise fraction transform (MNF). First, the covariance matrix is ​​estimated:

[0076] ;

[0077] in, represents the neighborhood of pixel i;

[0078] Then, solve the generalized eigenvalue problem:

[0079] ;

[0080] Preserve eigenvalues The eigenvector of

[0081] Next, the stepwise optimal band selection (SOBS) algorithm is applied to calculate the mutual information between bands:

[0082] ;

[0083] By maximizing the amount of information and minimizing redundancy, 20-30 most discriminative bands are selected. The band selection process is optimized by an entropy-based objective function:

[0084] ;

[0085] Among them, b i represents the band, C represents the categorical variable, and is a trade-off parameter (the values ​​are 0.7 and 0.3 respectively).

[0086] Thermal infrared data is noise-reduced using improved bilateral filtering:

[0087] ;

[0088] Among them, f r and g s are the Gaussian kernel functions in the range and space, respectively, W p is the normalization coefficient;

[0089] Then, adaptive background correction is applied to remove the effects of temperature drift:

[0090] ;

[0091] in, is the local background temperature distribution, estimated by large-scale Gaussian filtering, is the average background temperature.

[0092] After preprocessing, the modal data are normalized to ensure scale consistency:

[0093] ;

[0094] in, and are the mean and standard deviation of each modality data respectively; to speed up subsequent processing, the standardized data is stored in HDF5 format with indexes to facilitate random access and batch processing.

[0095] like Figure 3 As shown, in a preferred embodiment of the present invention, based on the adaptive contrastive deep clustering framework (ACDCF), the steps of extracting features and clustering the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results specifically include:

[0096] S31. Build a self-supervised semantic encoder to extract initial feature representations. The encoder is based on the Vision Transformer architecture and introduces a local-global attention mechanism to enhance feature extraction capabilities.

[0097] S32. Design four types of image enhancement transformations to generate multi-view samples, and construct a local-global contrast loss function and a prototype perception contrast learning module; the four types of image enhancement transformations include random cropping and horizontal flipping, color jittering, affine transformation, Gaussian blurring, and random erasing;

[0098] S33. The disease complexity score is calculated by intra-class distance and inter-class distance. At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect. The temperature parameters are adaptively adjusted according to the training progress and feature distribution to dynamically adjust the number of clusters and prototype distribution.

[0099] In practical applications, based on the data preprocessed in step S2, the ACDCF framework first constructs a self-supervised semantic encoder to extract initial feature representations. The encoder is based on the Vision Transformer architecture and introduces a local-global attention mechanism to enhance feature extraction capabilities. Specifically,

[0100] The input multimodal data (image) Split into non-overlapping patches, where P is the patch size (here 16); each patch is linearly projected and then added to the position code. The specific calculation formula is as follows:

[0101] ;

[0102] in, is the patch embedding matrix; Encode for position;

[0103] The local-global attention mechanism consists of two parallel branches, namely the local branch and the global branch; the local branch uses window attention:

[0104] ;

[0105] Among them, B is the relative position offset, and the window size M=7;

[0106] The global branch uses sparse attention to reduce computational complexity through layered random sampling:

[0107] ;

[0108] Among them, S is the sampling matrix, and the sampling rate is 25%;

[0109] The outputs of the two branches are adaptively fused to obtain multimodal preliminary features:

[0110] ;

[0111] in, and is a learnable parameter, LN is layer normalization, and MLP is a two-layer feedforward network; the encoder contains 12 such local-global Transformer blocks, and the final output feature vector , where D=768;

[0112] In a preferred embodiment of the present invention, the steps of designing four types of image enhancement transformations to generate multi-view samples and constructing a local-global contrast loss function and a prototype-perceptual contrast learning module specifically include:

[0113] Design four types of image enhancement transforms Generate multi-view samples, where: t1 is random cropping and horizontal flipping; t2 is color jitter (brightness, contrast, saturation, hue); t3 is affine transformation (rotation, scaling, translation); t4 is Gaussian blur and random erasure; each original sample x generates 4 enhanced views ;

[0114] Constructing a local-global contrast loss function:

[0115] ;

[0116] in, is the temperature parameter (the initial value is set to 0.07), It represents the feature representation of the i-th sample after the j-th enhancement;

[0117] The prototype-aware contrastive learning module defines K disease category prototype vectors , the association between samples and prototypes is established through the soft assignment matrix Q:

[0118] ;

[0119] in, is the eigenvector Z i With prototype C k The distance between (cosine distance can be used), is the scaling factor (value 0.1); the prototype vector is updated by moving average:

[0120] ;

[0121] in, is the momentum parameter (set to 0.99);

[0122] On this basis, the prototype contrast loss is optimized to ensure the clustering consistency of similar samples:

[0123] ;

[0124] To avoid trivial solutions, cluster consistency loss and entropy regularization term are introduced:

[0125] ;

[0126] Where P is the target distribution, which is calculated by promoting high confidence distribution and balancing the category distribution:

[0127] ;

[0128] The entropy regularization term is:

[0129] ;

[0130] The overall optimization goal is:

[0131] ;

[0132] in, are weight coefficients, and their initial values ​​are set to 1.0, 1.0, 0.5, and 0.1 respectively.

[0133] A key innovation of ACDCF is its adaptive cluster optimization mechanism, which dynamically adjusts the number of clusters and prototype distribution. Specifically, in a preferred embodiment of the present invention, the disease complexity score is calculated using intra-class and inter-class distances. A dynamic temperature adjustment mechanism is designed to optimize comparative learning effects. The temperature parameter is adaptively adjusted based on training progress and feature distribution to dynamically adjust the number of clusters and prototype distribution. The steps specifically include:

[0134] Design disease complexity score S c , calculated by intra-class distance and inter-class distance:

[0135] ;

[0136] When S c Exceeding the preset threshold When , the number of clusters is increased; when it is lower than the preset threshold When , similar clusters are merged; cluster merging is based on the prototype similarity matrix M, where , automatically determine the merge objects through spectral clustering;

[0137] At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect, and the temperature parameters are adaptively adjusted according to the training progress and feature distribution:

[0138] ;

[0139] Among them, t is the current iteration number, T is the total iteration number, and 0.05 and 0.2 respectively.

[0140] The ACDCF model ultimately outputs high-quality feature representations (preliminary multimodal features) and preliminary clustering results. The feature representation dimension is 768, and the number of clusters is adaptively determined based on data complexity, typically ranging from 5 to 12, corresponding to different disease types and healthy blueberries. The ACDCF model is trained using the Adam optimizer with an initial learning rate of 3e-4 and a cosine annealing strategy for 200 epochs.

[0141] like Figure 4 As shown, in a preferred embodiment of the present invention, based on the multimodal complementary feature fusion mechanism (MCFF), according to the multimodal preliminary features and the preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features, specifically including:

[0142] S41. Perform intra-modal feature enhancement on the multimodal preliminary features to obtain multi-scale feature maps of each modality; wherein the multimodal preliminary features include RGB modality, hyperspectral modality, and thermal infrared modality; the RGB modality uses an improved DenseNet to extract multi-scale feature maps; the hyperspectral modality uses a 3D-CNN and Transformer architecture to extract multi-scale feature maps; the thermal infrared modality uses a residual attention network to extract multi-scale feature maps;

[0143] S42. Apply channel recalibration to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels; then perform cross-modal feature complementation and build an inter-modal self-attention mechanism;

[0144] S43. Cross-scale feature integration is achieved through adaptive feature recombination. Specifically, the feature uncertainty map is first calculated. Then, multi-scale features are adaptively fused based on the uncertainty map. Finally, multimodal fusion features are calculated through attention weighted summation.

[0145] In practical applications, MCFF begins to construct a multimodal fusion representation using the multimodal preliminary features and preliminary clustering results generated by ACDCF in step S3 above.

[0146] In a preferred embodiment of the present invention, the step of performing intra-modal feature enhancement on the multimodal preliminary features to obtain a multi-scale feature map of each modality specifically includes:

[0147] The RGB modality uses an improved DenseNet to extract features. The original DenseNet structure is modified, the SENet channel attention mechanism is introduced, and the SE module is added after each dense block:

[0148] ;

[0149] Among them, GAP is global average pooling, and is the weight of the fully connected layer, r is the dimensionality reduction ratio (value is 16); at the same time, depthwise separable convolution is introduced to reduce the computational complexity, and the standard convolution is replaced by:

[0150] ;

[0151] The hyperspectral modality uses the 3D-CNN and Transformer architecture. The 3D-CNN part contains four convolutional blocks, and the structure of each convolutional block is:

[0152] ;

[0153] The number of convolution kernels increases layer by layer (32-64-128-256) to capture the joint spatial-spectral features; the Transformer part contains 4 encoder blocks, each of which contains a self-attention layer and a feedforward network:

[0154] ;

[0155] Among them, MSA is multi-head self-attention, the number of heads is 8, and the hidden dimension is 512;

[0156] The thermal infrared modality uses a residual attention network. The basic structure is ResNet50, but a hybrid attention mechanism is introduced in each residual block, considering both channel and spatial dimensions:

[0157] ;

[0158] Among them, GMP is the global maximum pooling, A C and A S are channel and spatial attention maps respectively;

[0159] Each modality network outputs a multi-scale feature map , where i represents the modality and j represents the scale level; the feature dimensions are: RGB modality (64, 128, 256, 512), hyperspectral modality (64, 128, 256, 512), and thermal infrared modality (64, 128, 256, 512).

[0160] In a preferred embodiment of the present invention, a channel recalibration operation is applied to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels; then, cross-modal feature complementation is performed to construct an inter-modal self-attention mechanism, specifically including:

[0161] Apply channel recalibration to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels:

[0162] ;

[0163] in, and is the weight of the fully connected layer, reducing the channel dimension from C to C / r and then restoring it to C, with r taking the value of 16;

[0164] Then, we perform cross-modal feature complementation and build an inter-modal self-attention mechanism. Specifically, we first project the features of different modalities into a common feature space:

[0165] ;

[0166] in, is the projection matrix; then calculate the inter-modal attention matrix:

[0167] ;

[0168] Among them, i and Represent different modalities; update features through the attention matrix:

[0169] ;

[0170] in, is the inter-modal complementary coefficient for adaptive learning, initialized to a uniform distribution , optimized through back-propagation; The update adopts a gating mechanism:

[0171] ;

[0172] in, is the learnable weight matrix, Represents feature splicing.

[0173] In a preferred embodiment of the present invention, the steps of implementing cross-scale feature integration through adaptive feature recombination specifically include:

[0174] Compute the feature uncertainty map:

[0175] ;

[0176] in, is the mean of the feature map, N is the number of feature maps;

[0177] Adaptive fusion of multi-scale features based on uncertainty graph:

[0178] ;

[0179] in, is the upsampling operation, is the temperature parameter (value 10);

[0180] Calculate the multimodal fusion features through attention weighted summation:

[0181] ;

[0182] in, To adapt the modality importance weight, we use a learnable attention network to calculate:

[0183] ;

[0184] To enhance feature representation learning, an auxiliary contrast learning objective is introduced:

[0185] ;

[0186] in, is the final multimodal fusion feature, is an enhanced version of the same sample; the overall optimization goal is:

[0187] ;

[0188] in, As the main task loss, the clustering loss in the deep clustering framework of step S2 is used for adaptive comparison. and 0.1 and 0.01 respectively.

[0189] like Figure 5 As shown, in a preferred embodiment of the present invention, the steps of performing unsupervised clustering optimization and disease type identification on multimodal fusion features to obtain an unsupervised learning model specifically include:

[0190] S51. Use the spectral clustering algorithm to cluster the multimodal fusion features, integrate and optimize the clustering results with the preliminary clustering results, align the two clusters using the Hungarian algorithm, and calculate the consistency score;

[0191] S52. Optimize the matching by maximizing the consistency score to obtain the final cluster label; use the cluster re-evaluation strategy to further improve the quality and identify possible abnormal assignments by calculating the Mahalanobis distance from the sample to the cluster center;

[0192] S53. After clustering is completed, significant features are extracted from each final cluster label for use in constructing a disease feature library. Specifically, the importance scores of the significant features are calculated. Based on the importance scores, the top several most significant features are extracted to construct a disease feature library. At the same time, the correspondence between these features and common disease symptoms is analyzed to construct an expert knowledge base and obtain a symptom-feature mapping. Based on the similarity between the symptom-feature mapping and the clustering features, a disease type description is automatically generated for each cluster. At the same time, a density-based anomaly detection algorithm is designed to identify potential unknown diseases and calculate the local outlier factor.

[0193] In practical applications, based on the multimodal fusion features output by MCFF in step S4 above, the embodiment of the present invention further optimizes the clustering results and realizes automatic identification of disease types.

[0194] In a preferred embodiment of the present invention, a spectral clustering algorithm is used to cluster the multimodal fusion features, the clustering result is integrated and optimized with the preliminary clustering result, the two clusters are aligned using the Hungarian algorithm, and the steps of calculating the consistency score specifically include:

[0195] Use the spectral clustering algorithm to cluster the multimodal fusion features and construct the sample affinity matrix S:

[0196] ;

[0197] in, is the kernel parameter, which is determined by local scale adaptation:

[0198] ;

[0199] in, represents the k nearest neighbors of sample i, where k is 15;

[0200] Construct the degree matrix D and the normalized Laplace matrix :

[0201] ;

[0202] extract The first K eigenvectors of form the feature matrix After normalizing each row, K-means is applied to complete clustering to obtain the clustering results; the number of clusters K is automatically determined by eigenvalue analysis:

[0203] ;

[0204] in, for The kth eigenvalue of (in ascending order);

[0205] The clustering result is integrated and optimized with the preliminary clustering result obtained in step S3 above. The two clusters are aligned using the Hungarian algorithm and the consistency score is calculated:

[0206] ;

[0207] in, and denote the i-th and j-th clusters obtained by the adaptive contrastive deep clustering framework and the multimodal complementary feature fusion mechanism, respectively;

[0208] The matching is optimized by maximizing the consistency score to obtain the final cluster label. The cluster re-evaluation strategy is used to further improve the quality. The steps of identifying possible abnormal assignments are carried out by calculating the Mahalanobis distance from the sample to the cluster center. Specifically,

[0209] Optimize the matching by maximizing the consistency score and obtain the final cluster label :

[0210] ;

[0211] A cluster re-evaluation strategy is used to further improve the quality by calculating the Mahalanobis distance from the sample to the cluster center to identify possible abnormal assignments:

[0212] ;

[0213] in, is the covariance matrix of the kth cluster; when When , redistribute samples to suboptimal clusters, is the distance threshold (value is 3.0).

[0214] In a preferred embodiment of the present invention, after clustering is completed, the step of extracting significant features from each final cluster label for use in constructing a disease feature library specifically includes:

[0215] For the kth cluster, calculate its feature importance score:

[0216] ;

[0217] in, is the mean of cluster k on feature j, and are the global feature mean and standard deviation respectively;

[0218] Based on the importance score, the top 10 most significant features are extracted to construct a disease feature library:

[0219] ;

[0220] At the same time, we analyze the correspondence between these characteristics and common disease symptoms to build an expert knowledge base:

[0221] ;

[0222] Among them, s i Indicates the description of symptoms, is a set of related features;

[0223] Based on the similarity between the symptom-feature map and the cluster characteristics, a disease type description is automatically generated for each cluster:

[0224] ;

[0225] The symptom with the highest similarity was selected as the cluster description;

[0226] At the same time, a density-based anomaly detection algorithm is designed to identify potential unknown diseases and calculate the local outlier factor (LOF):

[0227] ;

[0228] in, is the local reachability density of point x, are the k nearest neighbors of x; samples with LOF significantly greater than 1 are marked as abnormal and may represent unknown disease types; by applying the DBSCAN algorithm to re-cluster abnormal samples, we try to discover new disease patterns:

[0229] ;

[0230] in, is the neighborhood radius (1.5 times the average nearest neighbor distance), and MinPts is the density threshold (5).

[0231] The above steps ultimately output the disease type represented by each cluster and its visual feature description, including: cluster identifier and disease type inference, key visual features and their importance scores, abnormal sample identification results and potential new disease warnings, and disease severity assessment (based on the distance between the sample and the center of the healthy category).

[0232] like Figure 6 As shown, in a preferred embodiment of the present invention, the steps of deploying an unsupervised learning model and performing incremental learning to detect blueberry diseases specifically include:

[0233] S61. Use edge-cloud collaborative architecture to deploy unsupervised learning models to actual application scenarios;

[0234] S62. Design an incremental learning mechanism for newly collected samples, and implement online model update optimization through feature distillation and category balance loss.

[0235] In practice, the trained models from the previous steps are deployed to real-world scenarios using an edge-cloud collaborative architecture. The edge device (based on the Jetson Xavier NX) performs data acquisition, preprocessing, and preliminary feature extraction. Model quantization uses 16-bit floating-point numbers, and TensorRT optimizes inference performance, achieving a processing speed of 8-10 samples per second. The cloud (based on a GPU server) performs deep feature fusion and clustering analysis, handles batch tasks, and updates the global model.

[0236] Edge-to-cloud communication adopts an asynchronous message queue mechanism based on the MQTT protocol and designs a three-level data transmission strategy: 1. Normal mode: only the extracted feature vector (768 dimensions) is transmitted; 2. Enhanced mode: the feature vector and key image areas are transmitted; 3. Full mode: full-resolution multimodal data is transmitted.

[0237] Design an incremental learning mechanism for newly collected samples. When the accumulation exceeds the threshold T newWhen new samples (set to 500) are added or potential new disease categories are detected, feature space update and cluster fine-tuning are triggered. The importance score of new samples is calculated by sample density and uncertainty:

[0238] ;

[0239] in, is the balance parameter (value is 0.7), Calculated by -NN density estimation, Estimation via Monte Carlo Dropout:

[0240] ;

[0241] Where H is the entropy function, is the predicted probability distribution under the t-th dropout sampling, and T is the number of sampling times (value 20).

[0242] Feature distillation technology is used in incremental learning to retain original knowledge:

[0243] ;

[0244] in, and are the feature extractors of the old model and the new model respectively, is the cosine similarity weight (value is 0.5).

[0245] At the same time, ACDCF and MCFF are combined to update the loss function for fine-tuning optimization:

[0246] ;

[0247] in, and are 0.8 and 0.5 respectively, To balance the loss for the categories, prevent new samples from causing category imbalance:

[0248] ;

[0249] in, is the category weight, n k is the number of samples of category k, is the balance factor (value is 0.9999).

[0250] Design an automatic knowledge base expansion mechanism. When a new disease cluster is found and the confidence exceeds a threshold (set to 0.85), a new disease category is automatically created and its feature distribution is recorded. The confidence is evaluated by the sample similarity and boundary clarity within the cluster:

[0251] ;

[0252] Among them, silhouette is the silhouette coefficient, and density-ratio is the ratio of the density within the cluster to the density at the edge.

[0253] Expert annotations are collected through an online feedback interface, and the system designs a lightweight active learning module to determine the most valuable annotated samples:

[0254] ;

[0255] in, They are 0.4, 0.3, and 0.3 respectively; diversity is evaluated by the maximum distance between the sample and the labeled set; representativeness is evaluated by the distance between the sample and the center of the cluster to which it belongs.

[0256] The solution provided by the embodiment of the present invention also includes an adaptive sampling strategy to adjust the data collection frequency. Based on environmental change monitoring (temperature, humidity, etc.) and historical disease occurrence patterns, the sampling frequency is increased during critical periods to optimize resource utilization:

[0257] ;

[0258] in, is the basic sampling frequency, and are environmental risk scores and temporal risk scores, respectively. and is the weight coefficient (all are 0.5).

[0259] The entire solution provided by the embodiment of the present invention forms a closed-loop optimization system that can adapt to new blueberry varieties and growing environments without requiring a large amount of labeled data.

[0260] In another embodiment of the present invention, a blueberry intelligent disease detection system based on multimodal unsupervised learning is provided, which is used to implement the above-mentioned blueberry intelligent disease detection method, specifically comprising:

[0261] A multimodal data acquisition module for collecting multimodal data of blueberry samples; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data;

[0262] Preprocessing the multimodal data to obtain preprocessed multimodal data;

[0263] Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results;

[0264] Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features;

[0265] Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model;

[0266] Deploy unsupervised learning models and perform incremental learning to detect blueberry diseases.

[0267] It should be noted that the above modules can be implemented in the form of a computer program, which can be run on a computer device. The computer program constituting each module can be stored in the memory of the computer device so that the processor executes each step of the above method.

[0268] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0269] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0270] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A blueberry intelligent disease detection method based on multimodal unsupervised learning, characterized in that: The following steps are involved: Collecting multimodal data of blueberry samples; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data; Preprocessing the multimodal data to obtain preprocessed multimodal data; Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results; Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features; Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model; Deploy an unsupervised learning model and perform incremental learning to detect blueberry diseases; the blueberry disease is leaf spot; Based on the adaptive contrastive deep clustering framework, the steps of extracting features and clustering the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results include: A self-supervised semantic encoder is constructed to extract initial feature representations. The encoder is based on the Vision Transformer architecture and introduces a local-global attention mechanism to enhance feature extraction capabilities. Design four types of image enhancement transformations to generate multi-view samples, and construct a local-global contrast loss function and a prototype-perceptual contrast learning module; the four types of image enhancement transformations include random cropping and horizontal flipping, color jittering, affine transformation, Gaussian blurring, and random erasing; The disease complexity score is calculated by intra-class distance and inter-class distance. At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect. The temperature parameters are adaptively adjusted according to the training progress and feature distribution to dynamically adjust the number of clusters and prototype distribution. Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features. Specifically, the steps include: Performing intra-modal feature enhancement on the multimodal preliminary features to obtain a multi-scale feature map of each modality; wherein the multimodal preliminary features include RGB modality, hyperspectral modality, and thermal infrared modality; A channel recalibration operation is applied to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels; then cross-modal feature complementation is performed to build an inter-modal self-attention mechanism; Cross-scale feature integration is achieved through adaptive feature recombination. Specifically, the feature uncertainty map is first calculated. Then, multi-scale features are adaptively fused based on the uncertainty map. Finally, multimodal fusion features are calculated through attention weighted summation. The steps for performing unsupervised clustering optimization and disease type identification on multimodal fusion features to obtain an unsupervised learning model include: Use the spectral clustering algorithm to cluster the multimodal fusion features, integrate and optimize the clustering results with the preliminary clustering results, align the two clusters using the Hungarian algorithm, and calculate the consistency score; The final cluster labels are obtained by optimizing the matching by maximizing the consistency score. The quality is further improved by using the cluster re-evaluation strategy, and abnormal assignments are identified by calculating the Mahalanobis distance from the sample to the cluster center. After clustering is completed, significant features are extracted from each final cluster label for use in constructing a disease feature library. Specifically, the importance scores of the significant features are calculated. Based on the importance scores, the top several most significant features are extracted to construct a disease feature library. At the same time, the correspondence between these features and common disease symptoms is analyzed, and an expert knowledge base is constructed to obtain a symptom-feature mapping. Based on the similarity between the symptom-feature mapping and the clustering features, a disease type description is automatically generated for each cluster. At the same time, a density-based anomaly detection algorithm is designed to identify potential unknown diseases and calculate the local outlier factor.

2. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 1 is characterized in that: The step of preprocessing the multimodal data to obtain preprocessed multimodal data specifically includes: Multimodal data registration algorithm based on phase consistency is used to spatially align multimodal data; After alignment, each modality data is preprocessed separately; After preprocessing, the data of each modality were standardized to ensure scale consistency; The RGB data preprocessing method is as follows: firstly, the illumination non-uniformity is eliminated by improved homomorphic filtering, and then the adaptive wavelet transform is applied to suppress the noise; The preprocessing method for hyperspectral data is as follows: first, dimensionality reduction is performed using minimum noise fraction transform, followed by stepwise optimal band selection; The preprocessing method of thermal infrared data is as follows: first, noise suppression is performed through improved bilateral filtering, and then adaptive background correction is applied to eliminate the influence of temperature drift.

3. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 1 is characterized in that: The steps to construct a self-supervised semantic encoder to extract initial feature representations include: The input multimodal data Split into non-overlapping patches, where P is the patch size; each patch is linearly projected and then added to the position code. The specific calculation formula is as follows: ； in, is the patch embedding matrix; Encode for position; The local-global attention mechanism consists of two parallel branches, namely the local branch and the global branch; the local branch uses window attention: ； Among them, B is the relative position offset, and the window size M=7; The global branch uses sparse attention and reduces computational complexity through layered random sampling: ； Among them, S is the sampling matrix, and the sampling rate is 25%; The outputs of the two branches are adaptively fused to obtain multimodal preliminary features: ； in, and is a learnable parameter, LN is layer normalization, and MLP is a two-layer feedforward network; the encoder contains 12 such local-global Transformer blocks, and the final output feature vector , where D=768; The steps of designing four types of image enhancement transformations to generate multi-view samples and constructing the local-global contrast loss function and prototype-aware contrast learning module include: Design four types of image enhancement transforms Generate multi-view samples, where: t1 is random cropping and horizontal flipping; t2 is color jittering; t3 is affine transformation; t4 is Gaussian blur and random erasure; each original sample x generates 4 enhanced views ; Constructing a local-global contrast loss function: ； in, is the temperature parameter, It represents the feature representation of the i-th sample after the j-th enhancement; The prototype-aware contrastive learning module defines K disease category prototype vectors , the association between samples and prototypes is established through the soft assignment matrix Q: ； in, is the eigenvector Z i With prototype C k The distance between is the scaling factor; the prototype vector is updated by moving average: ； in, is the momentum parameter; On this basis, the prototype contrast loss is optimized to ensure the clustering consistency of similar samples: ； To avoid trivial solutions, cluster consistency loss and entropy regularization term are introduced: ； Where P is the target distribution, which is calculated by promoting high confidence distribution and balancing the category distribution: ； The entropy regularization term is: ； The overall optimization goal is: ； in, is the weight coefficient; The disease complexity score is calculated by intra-class distance and inter-class distance. At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect. The temperature parameters are adaptively adjusted according to the training progress and feature distribution to dynamically adjust the number of clusters and prototype distribution. The specific steps include: Design disease complexity score S c , calculated by intra-class distance and inter-class distance: ； When S c Exceeding the preset threshold When , the number of clusters is increased; when it is lower than the preset threshold When , similar clusters are merged; cluster merging is based on the prototype similarity matrix M, where , automatically determine the merge objects through spectral clustering; At the same time, a dynamic temperature adjustment mechanism is designed to optimize the comparative learning effect, and the temperature parameters are adaptively adjusted according to the training progress and feature distribution: ； Among them, t is the current iteration number, T is the total iteration number, and 0.05 and 0.2 respectively.

4. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 1 is characterized in that: The RGB modality uses an improved DenseNet to extract multi-scale feature maps; the hyperspectral modality uses a 3D-CNN and Transformer architecture to extract multi-scale feature maps; and the thermal infrared modality uses a residual attention network to extract multi-scale feature maps.

5. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 4 is characterized in that: The steps of performing intra-modal feature enhancement on the multi-modal preliminary features to obtain the multi-scale feature maps of each modality include: The RGB modality uses an improved DenseNet to extract features. The original DenseNet structure is modified, the SENet channel attention mechanism is introduced, and the SE module is added after each dense block: ； Among them, GAP is global average pooling, and is the weight of the fully connected layer, and r is the dimensionality reduction ratio. At the same time, depthwise separable convolution is introduced to reduce computational complexity, replacing the standard convolution with: ； The hyperspectral modality uses the 3D-CNN and Transformer architecture. The 3D-CNN part contains four convolutional blocks, and the structure of each convolutional block is: ； The number of convolution kernels increases layer by layer to capture the joint spatial-spectral features; the Transformer part contains 4 encoder blocks, each of which contains a self-attention layer and a feedforward network: ； Among them, MSA is multi-head self-attention, the number of heads is 8, and the hidden dimension is 512; The thermal infrared modality uses a residual attention network. The basic structure is ResNet50, but a hybrid attention mechanism is introduced in each residual block, considering both channel and spatial dimensions: ； Among them, GMP is the global maximum pooling, A C and A S are channel and spatial attention maps respectively; Each modality network outputs a multi-scale feature map , where i represents the mode and j represents the scale level; The channel recalibration operation is applied to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels. Then, cross-modal feature complementation is performed to build an inter-modal self-attention mechanism. Specifically, the following steps are performed: Apply channel recalibration to the multi-scale feature maps of each modality to strengthen the discriminative feature channels and suppress redundant channels: ； in, and is the weight of the fully connected layer, reducing the channel dimension from C to C / r and then restoring it to C, where r is 16; Then, we perform cross-modal feature complementation and build an inter-modal self-attention mechanism. Specifically, we first project the features of different modalities into a common feature space: ； in, is the projection matrix; then calculate the inter-modal attention matrix: ； Among them, i and Represent different modalities; update features through the attention matrix: ； in, is the inter-modal complementary coefficient for adaptive learning, initialized to a uniform distribution , optimized through back-propagation; The update adopts a gating mechanism: ； in, is the learnable weight matrix, Represents feature splicing; The steps to achieve cross-scale feature integration through adaptive feature recombination include: Compute the feature uncertainty map: ； in, is the mean of the feature map, N is the number of feature maps; Adaptive fusion of multi-scale features based on uncertainty graph: ； in, is the upsampling operation, is the temperature parameter; Calculate the multimodal fusion features through attention weighted summation: ； in, To adapt the modality importance weight, we use a learnable attention network to calculate: ； To enhance feature representation learning, an auxiliary contrast learning objective is introduced: ； in, is the final multimodal fusion feature, is an enhanced version of the same sample; the overall optimization goal is: ； in, As the main task loss, the clustering loss in the adaptive contrast deep clustering framework is adopted. and 0.1 and 0.01 respectively.

6. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 1 is characterized in that: The steps of clustering the multimodal fusion features using the spectral clustering algorithm, integrating and optimizing the clustering results with the preliminary clustering results, aligning the two clusters using the Hungarian algorithm, and calculating the consistency score include: Use the spectral clustering algorithm to cluster the multimodal fusion features and construct the sample affinity matrix S: ； in, is the kernel parameter, which is determined by local scale adaptation: ； in, represents the k nearest neighbors of sample i, where k is 15; Construct the degree matrix D and the normalized Laplace matrix : ； extract The first K eigenvectors of form the feature matrix After normalizing each row, K-means is applied to complete clustering to obtain the clustering results; the number of clusters K is automatically determined by eigenvalue analysis: ； in, for The kth eigenvalue of ; The clustering result is integrated and optimized with the above preliminary clustering result. The two clusters are aligned using the Hungarian algorithm and the consistency score is calculated: ； in, and denote the i-th and j-th clusters obtained by the adaptive contrastive deep clustering framework and the multimodal complementary feature fusion mechanism, respectively; The matching is optimized by maximizing the consistency score to obtain the final cluster label. The cluster re-evaluation strategy is used to further improve the quality. The steps of identifying possible abnormal assignments are carried out by calculating the Mahalanobis distance from the sample to the cluster center. Specifically, Optimize the matching by maximizing the consistency score and obtain the final cluster label : ； A cluster re-evaluation strategy is used to further improve the quality by calculating the Mahalanobis distance from the sample to the cluster center to identify possible abnormal assignments: ； in, is the covariance matrix of the kth cluster; when When , redistribute samples to suboptimal clusters, is the distance threshold; After clustering is completed, the steps of extracting significant features from each final cluster label for building a disease feature library include: For the kth cluster, calculate its feature importance score: ； in, is the mean of cluster k on feature j, and are the global feature mean and standard deviation respectively; Based on the importance score, the top 10 most significant features are extracted to construct a disease feature library: ； At the same time, we analyze the correspondence between these characteristics and common disease symptoms to build an expert knowledge base: ； Among them, s i Indicates the description of symptoms, is a set of related features; Based on the similarity between the symptom-feature map and the cluster characteristics, a disease type description is automatically generated for each cluster: ； The symptom with the highest similarity was selected as the cluster description; At the same time, a density-based anomaly detection algorithm is designed to identify potential unknown diseases and calculate the local outlier factor (LOF): ； in, is the local reachability density of point x, are the k nearest neighbors of x; samples with LOF significantly greater than 1 are marked as abnormal and may represent unknown disease types; by applying the DBSCAN algorithm to re-cluster abnormal samples, we try to discover new disease patterns: ； in, is the neighborhood radius, and MinPts is the density threshold.

7. The blueberry intelligent disease detection method based on multimodal unsupervised learning according to claim 1, characterized in that: The steps for deploying an unsupervised learning model and performing incremental learning to detect blueberry diseases include: Adopt edge-cloud collaborative architecture to deploy unsupervised learning models to actual application scenarios; An incremental learning mechanism is designed for newly collected samples, and the model is updated and optimized online through feature distillation and category balance loss.

8. A blueberry intelligent disease detection system based on multimodal unsupervised learning, used to implement the blueberry intelligent disease detection method based on multimodal unsupervised learning according to any one of claims 1 to 7, characterized in that: include: A multimodal data acquisition module for collecting multimodal data of blueberry samples; the multimodal data includes RGB data, hyperspectral data, and thermal infrared data; Preprocessing the multimodal data to obtain preprocessed multimodal data; Based on the adaptive contrast deep clustering framework, feature extraction and clustering are performed on the preprocessed multimodal data to obtain preliminary multimodal features and preliminary clustering results; Based on the multimodal complementary feature fusion mechanism, according to the multimodal preliminary features and preliminary clustering results, the multimodal preliminary features are adaptively weighted and fused to obtain the multimodal fusion features; Perform unsupervised clustering optimization and disease type identification on the multimodal fusion features to obtain an unsupervised learning model; Deploy unsupervised learning models and perform incremental learning to detect blueberry diseases.

Images