Crop state evaluation method and system based on multi-source data

Abstract

This invention provides a method and system for crop status assessment based on multi-source data, relating to the field of real-time crop physiological status assessment technology. The method includes: acquiring crop images uploaded by users along with their capture timestamps and geographic coordinates; verifying the integrity and numerical range of the captured metadata, rejecting the assessment if it fails; inputting the crop images in parallel into three detection models, outputting the identification results for growth stage, pests and diseases, and fruit maturity respectively; acquiring the real-time temperature at the geographic location, calculating the estimated remaining days for maturity based on the effective accumulated temperature requirement for the growth stage; calculating a comprehensive health score based on fruit maturity distribution, pest and disease categories and confidence levels, combined with growth stage score mapping and confidence level adjustment; and generating a structured assessment result containing physiological phenotypic status levels, environmental stress attribution, and agricultural intervention recommendations. This addresses the problems of existing models having a single dimension for real-time status assessment, lagging stress response representation, and a disconnect between status levels and agricultural interventions.

Description

Technical Field

[0001] This invention relates to the field of real-time assessment technology of crop physiological status, and more specifically, to a method and system for assessing crop status based on multi-source data. Background Technology

[0002] There have been relevant studies on crop growth prediction technology based on multi-source data fusion analysis and multimodal environmental physiological monitoring. For example, patent document CN119398284A discloses a crop growth prediction method based on multi-source data fusion analysis. This method constructs a multi-scale feature tensor and performs anomaly propagation analysis by standardizing and evaluating the reliability of remote sensing, ground sensor, and historical agricultural data. Then, it uses a multi-level prediction model with dynamic weight optimization to output the growth trend prediction results. Patent document CN121212449A discloses a crop growth prediction method and system based on multimodal data. This method extracts dynamic compensation coefficients through correlation analysis between environmental evolution modes and physiological evolution modes, fuses spatiotemporal correlation features using temporal memory and attention mechanisms, and decomposes the fused features into long-term trend components and seasonal cycle components to implement predictions with physical consistency correction.

[0003] However, the aforementioned existing technologies focus on forward-looking predictions of future crop growth indicators or yield trends. Their model architecture and optimization objectives mainly serve time-series extrapolation and trend estimation. They lack systematic means for the refined assessment of real-time health status, stress response level, and comprehensive physiological phenotypic status of crops at specific time points or within short periods. Furthermore, when integrating high-dimensional environmental stress characteristics with crop phenotypic response parameters, existing methods often fail to fully establish an immediate assessment system that reflects the causal relationship of the state. This makes it difficult to achieve multi-dimensional quantitative rating and abnormal attribution of the current growth status of crops when facing sudden environmental fluctuations or early infection by pests and diseases. Based on this, in order to solve the problems of existing predictive models having a single dimension for real-time crop status assessment, lagging stress response representation, and a disconnect between status level and agricultural intervention, there is an urgent need to provide a status assessment method that is different from growth trend prediction and focuses on the coupling relationship between the current physiological phenotype of crops and environmental stress. This method can achieve accurate quantification of crop health, nutrient stress level, and growth and development stage through real-time fusion of multi-source heterogeneous data and reconstruction of status characteristics, providing real-time and interpretable decision-making basis for precise regulation of facility agriculture. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for crop status assessment based on multi-source data, aiming to solve the problems of existing crop growth prediction models having a single dimension for real-time status assessment, lagging stress response representation, and a disconnect between status levels and agricultural interventions.

[0005] This invention is achieved through the following technical solution: A method for assessing crop status based on multi-source data includes the following steps: Obtain crop images uploaded by users and shooting metadata bound to the crop images, the shooting metadata including shooting timestamps and shooting geographic coordinates; The integrity and numerical range of the captured metadata are checked. If the check fails, the subsequent evaluation is rejected and an error response is returned. The crop images that have passed the verification are input in parallel into the pre-built first detection model, second detection model and third detection model, which respectively output the crop growth stage identification results, disease and pest identification results and fruit maturity identification results. The system acquires real-time ambient temperature data corresponding to the captured geographic location coordinates, determines the effective accumulated temperature requirement for the current growth stage of the crop based on the crop growth stage identification results, and calculates the estimated remaining days for crop maturity by combining the real-time ambient temperature data. Based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage scoring mapping relationship and confidence level adjustment item, the comprehensive crop health score is calculated. Based on the crop's comprehensive health score, the estimated remaining days until maturity, and the pest and disease identification results, a structured assessment result is generated, including the crop's current physiological phenotypic status level, environmental stress response attribution, and agricultural intervention recommendations.

[0006] Optionally, the specific construction process of the first detection model is as follows: We collected crop canopy image samples with annotations of different crop types and different growth stages to construct the first training dataset; Data augmentation operations are performed on crop canopy image samples in the first training dataset to generate a first augmented dataset; wherein, the data augmentation operations include at least one of rotation, flipping, brightness perturbation, and Gaussian blur; The first augmented dataset is input into a feature extraction backbone network based on a convolutional neural network, and the output is a set of multi-scale feature maps of crop growth stages. Based on the multi-scale feature map set of crop growth stages, a feature pyramid structure is constructed, and shallow spatial location features and deep semantic features are integrated to generate a growth stage classification feature vector. The growth stage classification feature vector is input into a fully connected classification layer, and the output is the confidence distribution corresponding to each label in the preset crop growth stage label set. The cross-entropy loss function is used to calculate the classification error between the confidence distribution and the pre-labeled crop growth stage labels in the first training dataset. The model weights are then iteratively updated through backpropagation until the classification error converges to a preset first threshold, thus obtaining the first detection model that has been trained.

[0007] Optionally, the specific construction process of the second detection model is as follows: Collect crop organ image samples with labels for different disease categories and different insect pest categories to construct a second training dataset; Random cropping, color gamut dithering, and noise injection operations are sequentially performed on crop organ image samples in the second training dataset to generate the second augmented dataset. The second augmented dataset is input into a deep residual network based on an attention mechanism, which outputs a set of multi-level feature maps of pests and diseases. After the highest semantic level of the multi-level feature map set of diseases and pests, a disease classification branch and a pest detection branch are connected respectively; wherein, the disease classification branch outputs the first confidence distribution of each disease category through global average pooling and the first fully connected layer; the pest detection branch generates pest candidate bounding boxes through a region proposal network and outputs the second confidence distribution of the pest category within each candidate bounding box. The classification loss between the first confidence distribution and the disease category labels in the second training dataset is calculated using the first loss function; the position offset loss of the pest candidate bounding box and the detection loss between the second confidence distribution and the pest category labels in the second training dataset are calculated using the second loss function; the classification loss, the position offset loss and the detection loss are weighted and summed to obtain the joint loss. Based on the joint loss, the weight parameters of the deep residual network, the disease classification branch, and the pest detection branch are iteratively adjusted through gradient backpropagation until the joint loss converges to a preset second threshold, thus obtaining the trained second detection model.

[0008] Optionally, the specific construction process of the third detection model is as follows: Collect local image samples of fruits labeled with different fruit types and maturity stages to construct a third training dataset; Random cropping, contrast transformation, and color space perturbation operations are sequentially performed on the local fruit image samples in the third training dataset to generate the third augmented dataset. The third enhanced dataset is input into a feature extraction network based on a lightweight convolutional neural network, which outputs a set of multi-scale feature maps of fruit ripeness. A maturity classification branch is connected after the highest semantic level of the fruit maturity multi-scale feature map set; wherein, the maturity classification branch outputs the third confidence distribution corresponding to each maturity label through global average pooling and a second fully connected layer; The focus loss function is used to calculate the classification error between the third confidence distribution and the pre-labeled fruit maturity stage labels in the third training dataset. The weight parameters of the feature extraction network and the maturity classification branch are iteratively updated through backpropagation until the classification error converges to the preset third threshold, thus obtaining the trained third detection model.

[0009] Optionally, the specific process of determining the effective accumulated temperature requirement for the current growth stage of the crop based on the crop growth stage identification result, and calculating the estimated remaining days for crop maturity in combination with the real-time ambient temperature data, is as follows: Based on the crop growth stage identification results, the target total effective accumulated temperature required from the current growth stage to the harvest maturity period is retrieved from the preset crop growth stage and effective accumulated temperature requirement mapping table. Using the shooting timestamp as the cutoff time, obtain the daily environmental temperature sequence of the area corresponding to the shooting geographical coordinates since the start of crop growth; The daily average temperature values ​​in the daily ambient temperature sequence are truncated by the lower limit temperature threshold and the upper limit temperature threshold to obtain the daily effective temperature values. The daily effective temperature values ​​are then accumulated and summed to generate the accumulated effective temperature. The remaining effective accumulated temperature requirement is obtained by calculating the difference between the accumulated effective accumulated temperature and the target effective accumulated temperature. Obtain the average daily ambient temperature of the area corresponding to the shooting geographical coordinates for the same period of previous years, and calculate the daily effective accumulated temperature increment benchmark value after the average daily ambient temperature of the same period of previous years is truncated by the lower limit temperature threshold and the upper limit temperature threshold. The remaining effective accumulated temperature requirement is divided by the daily effective accumulated temperature increment benchmark value, and the resulting quotient is used as the number of days remaining until the crop is expected to mature.

[0010] Optionally, the specific process for calculating the comprehensive crop health score is as follows: The fruit maturity recognition results are analyzed, the number of individual fruits corresponding to each maturity label is counted, and the ideal proportion range corresponding to each maturity label of the current growth stage is obtained from the preset growth stage-maturity expected distribution mapping table based on the crop growth stage recognition results. For each maturity label, calculate the actual percentage of the corresponding number of individual fruits to the total number of fruits, and compare the actual percentage with the ideal percentage range. If the actual percentage deviates from the ideal percentage range, generate a maturity deviation deduction item based on the direction and magnitude of the deviation. The results of the disease and pest identification are analyzed, the disease and pest category labels and corresponding confidence scores are extracted, and the baseline deduction weights for each disease category and each pest category are determined according to the preset disease category-harm weight mapping table and pest category-harm weight mapping table. The baseline deduction weight for each disease category is multiplied by the corresponding confidence level to obtain the individual disease damage score, and the baseline deduction weight for each pest category is multiplied by the corresponding confidence level to obtain the individual pest damage score. The scores for all individual diseases and all individual pests are summed to generate a comprehensive deduction item for the overall harm caused by diseases and pests. Based on the crop growth stage identification results, the basic health score corresponding to the current growth stage is retrieved from the preset growth stage-basic health score mapping table; The difference between the basic health score and the sum of the maturity deviation deduction and the comprehensive pest and disease damage deduction is used as the comprehensive crop health score. When the difference is lower than the preset minimum score threshold, the comprehensive crop health score is set to the preset minimum score threshold.

[0011] Optionally, the generation of structured assessment results, including the crop's current physiological phenotypic state level, environmental stress response attribution, and agricultural intervention recommendations, specifically includes: The crop's overall health score is matched with a preset multi-level health status threshold range to determine the crop's current physiological phenotypic status level. Based on the disease and pest categories whose confidence levels exceed the preset warning threshold in the disease and pest identification results, and combined with the recent ambient temperature fluctuation range of the area corresponding to the captured geographic location coordinates, the environmental stress response attribution is determined; wherein, the environmental stress response attribution includes at least one of disease stress, pest stress, and temperature stress; Based on the current physiological phenotypic status of crops and the attribution of environmental stress responses, corresponding intervention recommendations are retrieved from a pre-set agricultural intervention strategy mapping table; Using the current physiological phenotypic status level of the crop as the root node, and the crop comprehensive health score, environmental stress response attribution, estimated remaining days to maturity of the crop, and intervention recommendations as child nodes, a tree-structured assessment result is generated and output.

[0012] Based on the same inventive concept, this invention also provides a crop status assessment system based on multi-source data, used to implement the aforementioned crop status assessment method based on multi-source data, including: The data receiving module is used to acquire crop images uploaded by users and shooting metadata bound to the crop images, the shooting metadata including shooting timestamps and shooting geographic coordinates; The data verification module, connected to the data receiving module, is used to perform integrity verification and numerical range verification on the captured metadata. If the verification fails, subsequent evaluation is rejected and an error response is returned. The parallel state detection module is connected to the data verification module. It has a built-in pre-trained first detection model, second detection model and third detection model. It is used to receive crop images that have passed the verification and output crop growth stage identification results, pest and disease identification results and fruit maturity identification results in parallel. An ambient temperature acquisition and accumulated temperature calculation module is connected to the data receiving module and the parallel state detection module. It is used to acquire real-time ambient temperature data corresponding to the shooting geographical coordinates, determine the effective accumulated temperature requirement required for the current growth stage of the crop based on the crop growth stage identification result, and calculate the estimated number of days remaining for crop maturity in combination with the real-time ambient temperature data. The comprehensive health score calculation module is connected to the parallel state detection module and is used to calculate the comprehensive health score of crops based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage score mapping relationship and confidence level adjustment item. The structured assessment generation module is connected to the comprehensive health score calculation module, the ambient temperature acquisition and accumulated temperature calculation module, and the parallel state detection module, respectively. It is used to generate structured assessment results, including the current physiological phenotype status level of the crop, the attribution of environmental stress response, and agricultural intervention suggestions, based on the crop comprehensive health score, the estimated remaining days to maturity of the crop, and the pest and disease identification results. A preset mapping storage module is connected to the ambient temperature acquisition and accumulated temperature calculation module, the comprehensive health score calculation module, and the structured assessment generation module. It is used to store a mapping table of crop growth stage and effective accumulated temperature requirement, a mapping table of growth stage-maturity expected distribution, a mapping table of disease category-harm weight, a mapping table of pest category-harm weight, a mapping table of growth stage-basic health score, a multi-level health status threshold range, and a mapping table of agricultural intervention strategies.

[0013] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described crop status assessment method based on multi-source data.

[0014] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described crop status assessment method based on multi-source data.

[0015] The technical solution of the present invention has at least the following advantages and beneficial effects: By simultaneously acquiring multi-dimensional phenotypic features such as crop growth stage, pests and diseases, and fruit maturity through multiple parallel detection models, and integrating them with real-time ambient temperature, a multi-dimensional quantitative rating of crop health, stress level, and development process is achieved within the same time profile. This overcomes the problem of traditional trend prediction methods having a single dimension for real-time status assessment.

[0016] By utilizing real-time air temperature corresponding to the geographic coordinates captured by the camera and effective accumulated temperature requirements identified based on growth stages, the estimated remaining days for crop maturity can be dynamically calculated. This tightly couples environmental thermal fluctuations with crop development processes, establishing an instant assessment system that reflects the causal relationship of stress. This system can effectively capture the impact of sudden environmental fluctuations on the maturation process and mitigate the lag in stress response characterization.

[0017] Based on a pre-defined growth stage scoring map and confidence adjustment term, the distribution of fruit maturity labels is integrated with pest and disease categories and confidence levels to calculate a comprehensive crop health score. This can transform multi-source, qualitative identification results into continuous quantitative health indicators, sensitively reflecting the subtle impact of early pest and disease infection or abnormal fruit ripening on the overall plant condition.

[0018] The generated structured assessment results simultaneously include the crop's current physiological phenotypic status level, environmental stress response attribution, and agricultural intervention recommendations, directly linking status assessment with precision regulation. This provides facility agriculture managers with immediate and interpretable operational guidance, effectively bridging the gap between status levels and agricultural interventions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the crop status assessment method based on multi-source data according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a crop status assessment system based on multi-source data according to an embodiment of the present invention. Detailed Implementation

[0020] The following is a detailed description of the embodiments, in conjunction with the accompanying drawings.

[0021] Reference Figure 1 A method for assessing crop status based on multi-source data includes the following steps: Step 1: Obtain the crop image uploaded by the user and the shooting metadata associated with the crop image. The shooting metadata includes the shooting timestamp and the shooting geographic coordinates.

[0022] In some embodiments, users can activate the crop status assessment function via a smart mobile terminal, tablet computer, or agricultural information terminal connected to a camera. The client application calls the device system camera interface or local image library selection interface to acquire an image of the crop to be assessed. At the moment of image acquisition or selection, the client simultaneously executes the following two system-level calls: The shooting timestamp is obtained by reading the current time in Coordinated Universal Time or local time through the clock service provided by the device's operating system, with an accuracy of at least the second, and formatting it as a standard timestamp string; The device obtains the geographic location coordinates using its built-in Global Navigation Satellite System (GNSS) positioning module, acquiring the current longitude and latitude in the WGS-84 coordinate system. If the coordinates cannot be obtained immediately due to signal issues, the user will be prompted to manually select a geographic location or try again later until valid coordinate values ​​are obtained.

[0023] The acquired crop images are logically bound to the automatically obtained timestamps and geographic coordinates into a single evaluation request unit. A structured data packet can be constructed, treating the image binary data as a file field and the capture timestamp, longitude, and latitude as independent string or floating-point number fields, placed within the same network request body as the image fields. All fields in this data packet share the same request session identifier, thus achieving the binding. Alternatively, the capture timestamp and geographic coordinates can be written into the Exchangeable Image File Format (EXIF) attribute of the crop image file according to the Exchangeable Image File Format standard (e.g., modifying tags such as GPSInfo and DateTimeOriginal), and then the image file containing complete EXIF ​​information can be uploaded as a single object. In this case, the metadata and image data are physically bound within the file.

[0024] The bound request data is sent to the backend status assessment service interface via the secure hypertext transfer protocol. The interface address is preset in the client. The request method can be POST, and the Content-Type header can be declared as multipart / form-data to ensure reliable transmission of image binary data and text fields.

[0025] After receiving the request, the status assessment backend service performs the following parsing operations to complete the "retrieval" process: When extracting image files, the binary data stream of the image file is read from the requested multipart / form-data body according to the agreed image field names, and temporarily stored in memory or a temporary file system as input for subsequent detection models; When extracting the shooting timestamp, the timestamp string is read from the corresponding text field and the format validity is initially judged. If the EXIF ​​binding method is used, the value of the label is parsed from the EXIF ​​attribute of the image as the timestamp. When extracting the geographic coordinates of the captured location, the longitude and latitude strings are read from the corresponding numerical fields and converted into floating-point numbers; if EXIF ​​binding is used, the longitude and latitude values ​​are parsed from the EXIF ​​attributes of the image. In the server session or database, a unique evaluation task identifier is generated for this request. The extracted crop image, shooting timestamp, longitude, and latitude are attached as a set of associated attributes to this task identifier to complete the logical binding, which is used by step two and subsequent steps.

[0026] Step 2: Perform integrity and numerical range checks on the captured metadata. If the checks fail, refuse to proceed with further evaluation and return an error response.

[0027] In some embodiments, the specific process of performing integrity and numerical range checks on the captured metadata, and rejecting further evaluation and returning an error response if the checks fail, is as follows: First, an integrity check is performed, sequentially verifying the existence and non-empty values ​​(null values ​​or empty strings) of the shooting timestamp field, longitude field, and latitude field obtained in step one. If any field is missing or empty, the integrity check fails, a response object with error code METADATA_INCOMPLETE is immediately generated, specifying the name of the missing field in the response message, and the subsequent verification and evaluation process is terminated.

[0028] After the integrity check passes, the numerical range check is performed, which includes the validity check of the timestamp and the validity check of the geographic location coordinates.

[0029] For timestamp validity verification, the captured timestamp string is parsed into a time object according to ISO8601 or a preset time format template. If parsing fails and cannot be recognized as a valid time value, the error code INVALID_TIMESTAMP is returned with a time format error message. If parsing is successful, the current server system time is used as a reference time, and the captured timestamp is checked to see if it is later than the current server time plus a preset maximum clock deviation tolerance value. This maximum clock deviation tolerance value can be set to 5 to 10 minutes depending on network latency and device clock asynchrony. If the time represented by the captured timestamp exceeds this tolerance range, it indicates possible clock instability or time spoofing, and the error code FUTURE_TIMESTAMP is returned with a timestamp expiration message. Optionally, the captured timestamp can be checked to see if it is earlier than the system-set valid evaluation start date; for example, if only images acquired after a certain historical time point are accepted, and the image is earlier than that start point, the error code EXPIRED_TIMESTAMP is returned with a message indicating that the image has expired, to prevent the use of outdated data for status evaluation.

[0030] For validating geographic coordinates, the longitude and latitude strings extracted from the request or EXIF ​​are converted to double-precision floating-point numbers. If the conversion fails, the error code INVALID_COORDINATE_FORMAT is returned. If the conversion is successful, the longitude value is checked to ensure it falls within the range [-180.0, 180.0] and the latitude value within the range [-90.0, 90.0]. If either value exceeds these limits, the error code COORDINATE_OUT_OF_RANGE is returned. Furthermore, default invalid values ​​from the Global Navigation Satellite System are excluded. For example, (longitude, latitude) values ​​are checked to ensure they are not equal to (0.0, 0.0) or fall within a very small range near (0.0, 0.0) to avoid invalid coordinates being passed due to positioning failure. If such invalid fixed values ​​are detected, the error code INVALID_GPS_FIX is returned. It can also determine whether the coordinates are located in a known crop-growing area or at least in a land area based on global agricultural zoning data or land masking; if the coordinates fall into an uncultivable area such as the ocean or a barren polar region, it returns the error code COORDINATE_NOT_IN_FARMLAND and indicates that the geographical location is not within the scope of valid agricultural assessment.

[0031] If all checks pass, the process is allowed to proceed to step three. If any of the above checks fail, the subsequent evaluation is immediately terminated, a unified structured error response body is constructed, which includes a status code (e.g., 400), an error code, readable error description information, and the associated evaluation task identifier, and is returned to the client via HTTPS response. Subsequent steps such as image detection and growth calculation are no longer executed.

[0032] Step 3: Input the crop images that have passed the verification into the pre-constructed first detection model, second detection model and third detection model in parallel, and output the crop growth stage identification results, pest and disease identification results and fruit maturity identification results respectively.

[0033] In some embodiments, the specific construction process of the first detection model is as follows: A first training dataset is constructed by collecting crop canopy image samples labeled with different crop species and different growth stages. Data augmentation operations are performed on the crop canopy image samples in the first training dataset to generate a first augmented dataset. The data augmentation operations include at least one of rotation, flipping, brightness perturbation, and Gaussian blur. The first augmented dataset is input into a feature extraction backbone network based on a convolutional neural network, outputting a set of multi-scale feature maps of crop growth stages. Based on the set of multi-scale feature maps of crop growth stages, a feature pyramid structure is constructed, fusing shallow spatial location features and deep semantic features to generate a growth stage classification feature vector. The growth stage classification feature vector is input into a fully connected classification layer, outputting a confidence distribution corresponding to each label in a preset set of crop growth stage labels. The cross-entropy loss function is used to calculate the classification error between the confidence distribution and the pre-labeled crop growth stage labels in the first training dataset, and the model weights are iteratively updated through backpropagation until the classification error converges to a preset first threshold, resulting in a first detection model that has been trained.

[0034] The first training dataset can be constructed from crop canopy image samples labeled with different crop species and growth stages. :

[0035] in, Indicates the first Zhang crop canopy images; For the corresponding number The growth stage labels corresponding to crop canopy images. To preset the total number of growth stage categories, This represents the total number of samples in the first training dataset.

[0036] right Each image Apply random data augmentation transformations to generate enhanced images. This constitutes the first augmented dataset. , .

[0037] The randomized data augmentation transformation randomly selects and combines operations from a set of operations such as rotation, flipping, brightness perturbation, and Gaussian blur. The augmented dataset is used to improve the model's robustness to changes in viewpoint and lighting conditions.

[0038] Enhance the image The input is a backbone network based on a convolutional neural network, and the output is a set of multi-scale feature maps. Let the feature maps output at different stages of the backbone network be... These correspond to feature representations that progressively decrease in spatial resolution and progressively increase in semantic information from shallow to deep layers. The elements in the multi-scale feature map set are defined by the following formula:

[0039] in, This represents the feature extraction function corresponding to the backbone network based on a convolutional neural network; The four levels represent the layer indexes of the feature extraction stage in the backbone network. These four levels correspond to different stages of the backbone network from shallow to deep. The smaller the level, the closer the feature map is to the input (shallow layer), the higher the spatial resolution, and the lower the semantic information. Indicates the first Zhang Zengqiang Image In the backbone network Feature maps output by the stage , For the first Spatial height of the stage feature map For the first Spatial width of the stage feature map For the first The number of channels in the stage feature map.

[0040] A Feature Pyramid Network (FPN) is constructed to fuse shallow spatial location features with deep semantic features. First, a 1×1 convolution is performed on the highest-level feature map to obtain the top-level feature map in the Feature Pyramid Network. Then, the layers are fused layer by layer from top to bottom. After upsampling and adding the feature maps of the corresponding layers, the layers are smoothed by 3×3 convolution, as shown in the following formula:

[0041] in, The feature pyramid network represents the first time. The final feature map output by the layer; This indicates that the backbone convolutional neural network is in the first... Feature maps output by each stage; This indicates that the kernel size is Two-dimensional convolution operation; This is an upsampling operation used to upsample the lower-resolution, more semantically rich feature map from the previous layer. The size is enlarged to perfectly match the spatial dimensions (height and width) of the current layer feature map so that element-wise addition and fusion can be performed. This can be achieved using nearest neighbor interpolation or bilinear interpolation. Finally, the most refined feature map after fusion is selected. Global average pooling (GAP) is then applied to obtain a fixed-dimensional growth stage classification feature vector. As shown in the following formula:

[0042] in, This is a global average pooling operation; For feature map The height (number of rows in the spatial dimension); For feature map Width (number of columns in the spatial dimension); For feature map Spatial location index in the height direction, with a value range of: arrive ; For feature map Spatial position index in the width direction, with a value range of: arrive ; Indicates the first Feature map of each sample In spatial coordinates The eigenvector at that location.

[0043] The growth stage classification feature vector simultaneously encodes shallow details and deep semantics, achieving a comprehensive representation of the spatial phenotype of crop canopy.

[0044] Growth stage classification feature vectors Input a fully connected layer and obtain the confidence distribution of each preset growth stage using the softmax function. As shown in the following formula:

[0045] in, Represents the normalization function; This is the weight matrix. The dimension of the feature vector; This is the bias vector. Output satisfy , Indicates the first The sample is the first Predicted probabilities of seed growth stages.

[0046] The classification error between the predicted confidence distribution and the true label is calculated using the cross-entropy loss function, and all learnable parameters are updated via backpropagation. The cross-entropy loss function is shown in the following equation:

[0047] in, Classification loss values ​​for growth stages; This is the training batch size for the first detection model; Indicates the first Does the _ sample belong to the _ ... The growth stage is labeled; a value of 1 indicates a stage, and a value of 0 indicates a stage not belonging to the growth stage. Through iterative optimization... Converging to a preset first threshold The first detection model, after training, is obtained.

[0048] In some embodiments, the specific construction process of the second detection model is as follows: Crop organ image samples containing labels for different disease categories and different pest categories are collected to construct a second training dataset. Random cropping, color gamut dithering, and noise injection operations are sequentially performed on the crop organ image samples in the second training dataset to generate a second augmented dataset. The second augmented dataset is input into a deep residual network based on an attention mechanism, which outputs a multi-level feature map set of diseases and pests. After the highest semantic level of the multi-level feature map set, a disease classification branch and a pest detection branch are connected respectively. The disease classification branch outputs the first confidence distribution of each disease category through global average pooling and a first fully connected layer. The pest detection branch generates candidate bounding boxes for pests through a region proposal network. The system outputs a second confidence distribution for each pest category within each candidate bounding box; it calculates the classification loss between the first confidence distribution and the labeled pest categories in the second training dataset using a first loss function; it calculates the positional offset loss of the pest candidate bounding boxes and the detection loss between the second confidence distribution and the labeled pest categories in the second training dataset using a second loss function; it performs a weighted summation of the classification loss, the positional offset loss, and the detection loss to obtain a joint loss; based on the joint loss, it iteratively adjusts the weight parameters of the deep residual network, the pest classification branch, and the pest detection branch through gradient backpropagation until the joint loss converges to a preset second threshold, thus obtaining the trained second detection model.

[0049] We can set the second training dataset as follows: ,in, For the second training dataset, the first Images of crop organs from a sample; For the second training dataset, the first Disease category label corresponding to each sample , Number of disease categories; For the second training dataset, the first The set of pest labeling information corresponding to each sample , For the first In the sample, compared with the first sample... The target is to regress the coordinates of the true pest bounding boxes matched with the candidate boxes. For the first Pest category labels for each pest instance. , Number of pest categories; This represents the total number of samples in the second training dataset.

[0050] The confidence distribution of the disease classification branch output is shown in the following formula:

[0051] in, For the first The disease classification confidence distribution vector of crop organ image samples, with dimension . Each element This indicates that the sample was predicted to be the first. The probability of a disease, and the sum of the probabilities of all categories is 1; The weight matrix of the fully connected layer for the disease classification branch; For the first The disease classification feature vector is obtained by extracting the samples through a deep residual network and global average pooling. Here is the bias vector of the fully connected layer for the disease classification branch; the corresponding loss is represented by cross-entropy, as shown in the following equation:

[0052] in, This represents the cross-entropy loss value of the disease classification branch; This is the training batch size for the first detection model; For the first In the one-hot encoded form of the true disease label of the sample, corresponding to the first... Values ​​for different types of diseases.

[0053] The pest detection branch outputs the predicted offset of the candidate bounding boxes and the distribution of pest categories. The pest category distribution is shown in the following formula:

[0054] in, Indicates the first In the crop organ image sample, the first The probability distribution vector of the predicted pest category for each pest candidate bounding box; The weight matrix of the fully connected layer for the insect pest classification branch; For the first In the nth sample The feature vector obtained after region pooling and subsequent feature extraction of the candidate insect pest boxes; This is the bias vector of the fully connected layer in the insect pest classification branch; The position offset loss is defined as the smooth-L1 loss, as shown in the following equation:

[0055] in, This represents the bounding box regression position offset loss value; This represents the set of positive candidate bounding box indices. The total number of positive samples; The model predicts the first The sample, the first The bounding box coordinate offset of each candidate box; This is the smoothed L1 loss function.

[0056] The loss for pest category detection is the cross-entropy, as shown in the following formula:

[0057] in, This represents the cross-entropy loss value for pest category detection; The total number of candidate boxes participating in the classification; For the first In the nth sample The one-hot encoding of the true pest category label of each pest candidate box in the 1st... The value of the class, if the true category of the candidate box is ,but Otherwise ; For the first In the nth sample The insect pest candidate box was predicted by the model to be the [number]th [type of] [pest]. The probability value of the pest category, which comes from the category distribution vector output by the pest detection branch. The Each component satisfies .

[0058] The joint loss is the weighted sum of the three losses, as shown in the following formula:

[0059] in, This is the joint loss value; , and These are the weight coefficients corresponding to the loss. By minimizing... Until convergence to the preset second threshold The second detection model, after training, is obtained.

[0060] In some embodiments, the specific construction process of the third detection model is as follows: A third training dataset is constructed by collecting local image samples of fruits labeled with different fruit types and maturity stages. Random cropping, contrast transformation, and color space perturbation operations are sequentially performed on the local image samples in the third training dataset to generate a third augmented dataset. This third augmented dataset is then input into a feature extraction network based on a lightweight convolutional neural network, which outputs a set of multi-scale feature maps of fruit maturity. A maturity classification branch is connected after the highest semantic level of this set. This maturity classification branch outputs a third confidence distribution corresponding to each maturity label through global average pooling and a second fully connected layer. The classification error between the third confidence distribution and the pre-labeled fruit maturity stage labels in the third training dataset is calculated using a focus loss function. The weight parameters of the feature extraction network and the maturity classification branch are iteratively updated through backpropagation until the classification error converges to a preset third threshold, resulting in a trained third detection model.

[0061] We can set the third training dataset as... , For the third training dataset, the first Local images of the fruit in each sample; For image The corresponding actual maturity stage label, , This represents the total number of maturity categories. Let be the total number of samples in the third training dataset. The confidence distribution of the maturity classification branch output is shown in the following formula:

[0062] in, Indicates the first The confidence distribution vector of maturity classification for local image samples of a fruit; This is the weight matrix of the second fully connected layer in the maturity classification branch; From the Feature vectors extracted from local image samples of a fruit; This represents the bias vector of the second fully connected layer in the maturity classification branch. Focus loss is used to address class imbalance, as shown in the following equation:

[0063] in, The value of the focus loss function; This is the training batch size for the third detection model; For the first The one-hot vector of the true maturity category label of each sample is in the th... The value that a dimension can take; For the third detection model to the first The sample is predicted to belong to the _th ... The probability value of class maturity, which comes from the confidence distribution vector output by the maturity classification branch. The Each component satisfies ; For category Balance weights; To focus parameters, This is used to adjust the contribution of easy and difficult samples to the loss. The larger the value, the better the model predicts the probability of easily classified samples (i.e., the higher the probability of prediction). The higher the value of the loss, the more it is suppressed, thus making the training focus more on hard-to-classify samples. Iterative optimization makes... Converging to the preset third threshold The third detection model was obtained.

[0064] Step 4: Obtain the real-time ambient temperature data corresponding to the captured geographical coordinates, determine the effective accumulated temperature requirement required for the current growth stage of the crop based on the crop growth stage identification results, and calculate the estimated remaining days for crop maturity by combining the real-time ambient temperature data.

[0065] In some embodiments, the specific process of determining the effective accumulated temperature requirement for the current growth stage of the crop based on the crop growth stage identification result, and calculating the estimated remaining days for crop maturity in combination with the real-time ambient temperature data, is as follows: Based on the crop growth stage identification results, the target total effective accumulated temperature required from the current growth stage to the harvest maturity period is retrieved from the preset crop growth stage and effective accumulated temperature requirement mapping table. Using the shooting timestamp as the cutoff time, obtain the daily environmental temperature sequence of the area corresponding to the shooting geographical coordinates since the start of crop growth; The daily average temperature values ​​in the daily ambient temperature sequence are truncated by the lower limit temperature threshold and the upper limit temperature threshold to obtain the daily effective temperature values. The daily effective temperature values ​​are then accumulated and summed to generate the accumulated effective temperature. The remaining effective accumulated temperature requirement is obtained by calculating the difference between the accumulated effective accumulated temperature and the target effective accumulated temperature. Obtain the average daily ambient temperature of the area corresponding to the shooting geographical coordinates for the same period of previous years, and calculate the daily effective accumulated temperature increment benchmark value after the average daily ambient temperature of the same period of previous years is truncated by the lower limit temperature threshold and the upper limit temperature threshold. The remaining effective accumulated temperature requirement is divided by the daily effective accumulated temperature increment benchmark value, and the resulting quotient is used as the number of days remaining until the crop is expected to mature.

[0066] Step 5: Based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage scoring mapping relationship and confidence level adjustment item, calculate the crop comprehensive health score.

[0067] In some embodiments, the specific process for calculating the comprehensive crop health score is as follows: The fruit maturity recognition results are analyzed, the number of individual fruits corresponding to each maturity label is counted, and the ideal proportion range corresponding to each maturity label of the current growth stage is obtained from the preset growth stage-maturity expected distribution mapping table based on the crop growth stage recognition results. For each maturity label, calculate the actual percentage of the corresponding number of individual fruits to the total number of fruits, and compare the actual percentage with the ideal percentage range. If the actual percentage deviates from the ideal percentage range, generate a maturity deviation deduction item based on the direction and magnitude of the deviation. The results of the disease and pest identification are analyzed, the disease and pest category labels and corresponding confidence scores are extracted, and the baseline deduction weights for each disease category and each pest category are determined according to the preset disease category-harm weight mapping table and pest category-harm weight mapping table. The baseline deduction weight for each disease category is multiplied by the corresponding confidence level to obtain the individual disease damage score, and the baseline deduction weight for each pest category is multiplied by the corresponding confidence level to obtain the individual pest damage score. The scores for all individual diseases and all individual pests are summed to generate a comprehensive deduction item for the overall harm caused by diseases and pests. Based on the crop growth stage identification results, the basic health score corresponding to the current growth stage is retrieved from the preset growth stage-basic health score mapping table; The difference between the basic health score and the sum of the maturity deviation deduction and the comprehensive pest and disease damage deduction is used as the comprehensive crop health score. When the difference is lower than the preset minimum score threshold, the comprehensive crop health score is set to the preset minimum score threshold.

[0068] Step 6: Based on the crop's comprehensive health score, the estimated remaining days until crop maturity, and the pest and disease identification results, generate a structured assessment result that includes the crop's current physiological phenotype status level, environmental stress response attribution, and agricultural intervention recommendations.

[0069] In some embodiments, generating structured assessment results including the crop's current physiological phenotypic state level, environmental stress response attribution, and agricultural intervention recommendations specifically includes: The crop's overall health score is matched with a preset multi-level health status threshold range to determine the crop's current physiological phenotypic status level. Based on the disease and pest categories whose confidence levels exceed the preset warning threshold in the disease and pest identification results, and combined with the recent ambient temperature fluctuation range of the area corresponding to the captured geographic location coordinates, the environmental stress response attribution is determined; wherein, the environmental stress response attribution includes at least one of disease stress, pest stress, and temperature stress; Based on the current physiological phenotypic status of crops and the attribution of environmental stress responses, corresponding intervention recommendations are retrieved from a pre-set agricultural intervention strategy mapping table; Using the current physiological phenotypic status level of the crop as the root node, and the crop comprehensive health score, environmental stress response attribution, estimated remaining days to maturity of the crop, and intervention recommendations as child nodes, a tree-structured assessment result is generated and output.

[0070] Based on the same inventive concept, and corresponding to any of the above embodiments, refer to... Figure 2 This invention provides a crop status assessment system based on multi-source data, used to implement the aforementioned crop status assessment method based on multi-source data, comprising: The data receiving module is used to acquire crop images uploaded by users and shooting metadata bound to the crop images, the shooting metadata including shooting timestamps and shooting geographic coordinates; The data verification module, connected to the data receiving module, is used to perform integrity verification and numerical range verification on the captured metadata. If the verification fails, subsequent evaluation is rejected and an error response is returned. The parallel state detection module is connected to the data verification module. It has a built-in pre-trained first detection model, second detection model and third detection model. It is used to receive crop images that have passed the verification and output crop growth stage identification results, pest and disease identification results and fruit maturity identification results in parallel. An ambient temperature acquisition and accumulated temperature calculation module is connected to the data receiving module and the parallel state detection module. It is used to acquire real-time ambient temperature data corresponding to the shooting geographical coordinates, determine the effective accumulated temperature requirement required for the current growth stage of the crop based on the crop growth stage identification result, and calculate the estimated number of days remaining for crop maturity in combination with the real-time ambient temperature data. The comprehensive health score calculation module is connected to the parallel state detection module and is used to calculate the comprehensive health score of crops based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage score mapping relationship and confidence level adjustment item. The structured assessment generation module is connected to the comprehensive health score calculation module, the ambient temperature acquisition and accumulated temperature calculation module, and the parallel state detection module, respectively. It is used to generate structured assessment results, including the current physiological phenotype status level of the crop, the attribution of environmental stress response, and agricultural intervention suggestions, based on the crop comprehensive health score, the estimated remaining days to maturity of the crop, and the pest and disease identification results. A preset mapping storage module is connected to the ambient temperature acquisition and accumulated temperature calculation module, the comprehensive health score calculation module, and the structured assessment generation module. It is used to store a mapping table of crop growth stage and effective accumulated temperature requirement, a mapping table of growth stage-maturity expected distribution, a mapping table of disease category-harm weight, a mapping table of pest category-harm weight, a mapping table of growth stage-basic health score, a multi-level health status threshold range, and a mapping table of agricultural intervention strategies.

[0071] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the crop status assessment method based on multi-source data of the embodiments.

[0072] Alternatively, the aforementioned electronic device may be a server.

[0073] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the crop status assessment method based on multi-source data of the embodiment.

[0074] It is understood that the processor in the embodiments of the present invention may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0075] The method steps in the embodiments of the present invention can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0076] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive (SSD)).

Claims

1. A method for assessing crop status based on multi-source data, characterized in that, Includes the following steps: Obtain crop images uploaded by users and shooting metadata bound to the crop images, the shooting metadata including shooting timestamps and shooting geographic coordinates; The integrity and numerical range of the captured metadata are checked. If the check fails, the subsequent evaluation is rejected and an error response is returned. The crop images that have passed the verification are input in parallel into the pre-built first detection model, second detection model and third detection model, which respectively output the crop growth stage identification results, disease and pest identification results and fruit maturity identification results. The system acquires real-time ambient temperature data corresponding to the captured geographic location coordinates, determines the effective accumulated temperature requirement for the current growth stage of the crop based on the crop growth stage identification results, and calculates the estimated remaining days for crop maturity by combining the real-time ambient temperature data. Based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage scoring mapping relationship and confidence level adjustment item, the comprehensive crop health score is calculated. Based on the crop's comprehensive health score, the estimated remaining days until maturity, and the pest and disease identification results, a structured assessment result is generated, including the crop's current physiological phenotypic status level, environmental stress response attribution, and agricultural intervention recommendations.

2. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The specific construction process of the first detection model is as follows: We collected crop canopy image samples with annotations of different crop types and different growth stages to construct the first training dataset; Data augmentation operations are performed on crop canopy image samples in the first training dataset to generate a first augmented dataset; wherein, the data augmentation operations include at least one of rotation, flipping, brightness perturbation, and Gaussian blur; The first augmented dataset is input into a feature extraction backbone network based on a convolutional neural network, and the output is a set of multi-scale feature maps of crop growth stages. Based on the multi-scale feature map set of crop growth stages, a feature pyramid structure is constructed, and shallow spatial location features and deep semantic features are integrated to generate growth stage classification feature vectors. The growth stage classification feature vector is input into a fully connected classification layer, and the output is the confidence distribution corresponding to each label in the preset crop growth stage label set. The cross-entropy loss function is used to calculate the classification error between the confidence distribution and the pre-labeled crop growth stage labels in the first training dataset. The model weights are then iteratively updated through backpropagation until the classification error converges to a preset first threshold, thus obtaining the first detection model that has been trained.

3. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The specific construction process of the second detection model is as follows: Collect crop organ image samples with labels for different disease categories and different pest categories to construct a second training dataset; Random cropping, color gamut dithering, and noise injection operations are sequentially performed on crop organ image samples in the second training dataset to generate the second augmented dataset. The second augmented dataset is input into a deep residual network based on an attention mechanism, which outputs a set of multi-level feature maps of pests and diseases. After the highest semantic level of the multi-level feature map set of diseases and pests, a disease classification branch and a pest detection branch are connected respectively; wherein, the disease classification branch outputs the first confidence distribution of each disease category through global average pooling and the first fully connected layer; the pest detection branch generates pest candidate bounding boxes through a region proposal network and outputs the second confidence distribution of the pest category within each candidate bounding box. The classification loss between the first confidence distribution and the disease category labels in the second training dataset is calculated using the first loss function; the position offset loss of the pest candidate bounding box and the detection loss between the second confidence distribution and the pest category labels in the second training dataset are calculated using the second loss function; the classification loss, the position offset loss and the detection loss are weighted and summed to obtain the joint loss. Based on the joint loss, the weight parameters of the deep residual network, the disease classification branch, and the pest detection branch are iteratively adjusted through gradient backpropagation until the joint loss converges to a preset second threshold, thus obtaining the trained second detection model.

4. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The specific construction process of the third detection model is as follows: Collect local image samples of fruits labeled with different fruit types and maturity stages to construct a third training dataset; Random cropping, contrast transformation, and color space perturbation operations are sequentially performed on the local fruit image samples in the third training dataset to generate the third augmented dataset. The third enhanced dataset is input into a feature extraction network based on a lightweight convolutional neural network, which outputs a set of multi-scale feature maps of fruit ripeness. A maturity classification branch is connected after the highest semantic level of the fruit maturity multi-scale feature map set; wherein, the maturity classification branch outputs the third confidence distribution corresponding to each maturity label through global average pooling and a second fully connected layer; The focus loss function is used to calculate the classification error between the third confidence distribution and the pre-labeled fruit maturity stage labels in the third training dataset. The weight parameters of the feature extraction network and the maturity classification branch are iteratively updated through backpropagation until the classification error converges to the preset third threshold, thus obtaining the trained third detection model.

5. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The specific process of determining the effective accumulated temperature requirement for the current growth stage of the crop based on the crop growth stage identification results, and calculating the estimated remaining days for crop maturity in combination with the real-time ambient temperature data, is as follows: Based on the crop growth stage identification results, the target total effective accumulated temperature required from the current growth stage to the harvest maturity period is retrieved from the preset crop growth stage and effective accumulated temperature requirement mapping table. Using the shooting timestamp as the cutoff time, obtain the daily environmental temperature sequence of the area corresponding to the shooting geographical coordinates since the start of crop growth; The daily average temperature values ​​in the daily ambient temperature sequence are truncated by the lower limit temperature threshold and the upper limit temperature threshold to obtain the daily effective temperature values. The daily effective temperature values ​​are then accumulated and summed to generate the accumulated effective temperature. The remaining effective accumulated temperature requirement is obtained by calculating the difference between the accumulated effective accumulated temperature and the target effective accumulated temperature. Obtain the average daily ambient temperature of the area corresponding to the shooting geographical coordinates for the same period of previous years, and calculate the daily effective accumulated temperature increment benchmark value after the average daily ambient temperature of the same period of previous years is truncated by the lower limit temperature threshold and the upper limit temperature threshold. The remaining effective accumulated temperature requirement is divided by the daily effective accumulated temperature increment benchmark value, and the resulting quotient is used as the number of days remaining until the crop is expected to mature.

6. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The specific process for calculating the comprehensive crop health score is as follows: The fruit maturity recognition results are analyzed, the number of individual fruits corresponding to each maturity label is counted, and the ideal proportion range corresponding to each maturity label of the current growth stage is obtained from the preset growth stage-maturity expected distribution mapping table based on the crop growth stage recognition results. For each maturity label, calculate the actual percentage of the corresponding number of individual fruits to the total number of fruits, and compare the actual percentage with the ideal percentage range. If the actual percentage deviates from the ideal percentage range, generate a maturity deviation deduction item based on the direction and magnitude of the deviation. The results of the disease and pest identification are analyzed, the disease and pest category labels and corresponding confidence scores are extracted, and the baseline deduction weights for each disease category and each pest category are determined according to the preset disease category-harm weight mapping table and pest category-harm weight mapping table. The baseline deduction weight for each disease category is multiplied by the corresponding confidence level to obtain the individual disease damage score, and the baseline deduction weight for each pest category is multiplied by the corresponding confidence level to obtain the individual pest damage score. The scores for all individual diseases and all individual pests are summed to generate a comprehensive deduction item for the overall harm caused by diseases and pests. Based on the crop growth stage identification results, the basic health score corresponding to the current growth stage is retrieved from the preset growth stage-basic health score mapping table; The difference between the basic health score and the sum of the maturity deviation deduction and the comprehensive pest and disease damage deduction is used as the comprehensive crop health score. When the difference is lower than the preset minimum score threshold, the comprehensive crop health score is set to the preset minimum score threshold.

7. The crop status assessment method based on multi-source data as described in claim 1, characterized in that, The generated structured assessment results include the crop's current physiological phenotypic state level, environmental stress response attribution, and agricultural intervention recommendations, specifically including: The crop's overall health score is matched with a preset multi-level health status threshold range to determine the crop's current physiological phenotypic status level. Based on the disease and pest categories whose confidence levels exceed the preset warning threshold in the disease and pest identification results, and combined with the recent ambient temperature fluctuation range of the area corresponding to the captured geographic location coordinates, the environmental stress response attribution is determined; wherein, the environmental stress response attribution includes at least one of disease stress, pest stress, and temperature stress; Based on the current physiological phenotypic status of crops and the attribution of environmental stress responses, corresponding intervention recommendations are retrieved from a pre-set agricultural intervention strategy mapping table; Using the current physiological phenotypic status level of the crop as the root node, and the crop comprehensive health score, environmental stress response attribution, estimated remaining days to maturity of the crop, and intervention recommendations as child nodes, a tree-structured assessment result is generated and output.

8. A crop status assessment system based on multi-source data, used to implement the crop status assessment method based on multi-source data as described in any one of claims 1-7, characterized in that, include: The data receiving module is used to acquire crop images uploaded by users and shooting metadata bound to the crop images, the shooting metadata including shooting timestamps and shooting geographic coordinates; The data verification module, connected to the data receiving module, is used to perform integrity verification and numerical range verification on the captured metadata. If the verification fails, subsequent evaluation is rejected and an error response is returned. The parallel state detection module is connected to the data verification module. It has a built-in pre-trained first detection model, second detection model and third detection model. It is used to receive crop images that have passed the verification and output crop growth stage identification results, pest and disease identification results and fruit maturity identification results in parallel. An ambient temperature acquisition and accumulated temperature calculation module is connected to the data receiving module and the parallel state detection module. It is used to acquire real-time ambient temperature data corresponding to the shooting geographical coordinates, determine the effective accumulated temperature requirement required for the current growth stage of the crop based on the crop growth stage identification result, and calculate the estimated number of days remaining for crop maturity in combination with the real-time ambient temperature data. The comprehensive health score calculation module is connected to the parallel state detection module and is used to calculate the comprehensive health score of crops based on the number of fruits corresponding to each maturity label in the fruit maturity identification results and the disease category and confidence level in the disease and pest identification results, combined with the preset growth stage score mapping relationship and confidence level adjustment item. The structured assessment generation module is connected to the comprehensive health score calculation module, the ambient temperature acquisition and accumulated temperature calculation module, and the parallel state detection module, respectively. It is used to generate structured assessment results, including the current physiological phenotype status level of the crop, the attribution of environmental stress response, and agricultural intervention suggestions, based on the crop comprehensive health score, the estimated remaining days to maturity of the crop, and the pest and disease identification results. A preset mapping storage module is connected to the ambient temperature acquisition and accumulated temperature calculation module, the comprehensive health score calculation module, and the structured assessment generation module. It is used to store a mapping table of crop growth stage and effective accumulated temperature requirement, a mapping table of growth stage-maturity expected distribution, a mapping table of disease category-harm weight, a mapping table of pest category-harm weight, a mapping table of growth stage-basic health score, a multi-level health status threshold range, and a mapping table of agricultural intervention strategies.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform the crop status assessment method based on multi-source data as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the crop status assessment method based on multi-source data as described in any one of claims 1-7.

Images