Crop development period recognition method and device integrating meteorological and image sequences

Abstract

The present invention provides a crop development stage identification method and device that integrates meteorological and image sequences. The method involves obtaining crop image sequence data and associated meteorological element sequence data of the crop to be identified; invoking a pre-tuned crop development stage identification network, loading the crop image sequence data and associated meteorological element sequence data into the pre-tuned crop development stage identification network, mining to obtain a crop image representation array corresponding to the crop image sequence data and a meteorological element representation array corresponding to the associated meteorological element sequence data; performing attention fusion on the crop image representation array and the meteorological element representation array to obtain a fused representation array; and performing development stage classification prediction based on the fused representation array to obtain a crop development stage classification identification result. The present invention enables more accurate and efficient crop development stage identification, providing technical support for refined agricultural management.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a crop development period recognition method and device integrating meteorological and image sequences. Background Art

[0002] In agriculture, accurate identification of crop development stages is crucial for agricultural production management, pest and disease control, and yield prediction. Traditional crop development stage identification relies primarily on manual observation, a time-consuming and labor-intensive method subject to significant subjective influences, resulting in limited accuracy. With the advancement of computer technology and artificial intelligence, the use of image processing and machine learning techniques for automatic crop development stage identification has become a research hotspot. However, existing crop development stage identification methods mostly rely solely on crop image sequence data, overlooking the crucial influence of meteorological factors on crop growth and development. Meteorological factors such as temperature, humidity, and light are critical factors in crop growth and development, directly or indirectly influencing the crop's growth cycle, growth rate, and the occurrence of pests and diseases. Therefore, relying solely on crop image sequence data for development stage identification may not fully and accurately reflect the actual growth and development status of crops. Therefore, to overcome this limitation, a crop development stage identification method that integrates meteorological and image sequence data is urgently needed to improve identification accuracy and efficiency, thereby providing more precise management and decision-making support for agricultural production. Summary of the Invention

[0003] In view of this, the present invention provides a crop growth stage identification method and device that integrates meteorological and image sequences. The technical solution of the present invention is implemented as follows:

[0004] On the one hand, an embodiment of the present invention provides a crop development period identification method that integrates meteorological and image sequences, the method comprising: obtaining crop image sequence data and associated meteorological element sequence data of a crop to be identified; calling a pre-debugged crop development period identification network, loading the crop image sequence data and associated meteorological element sequence data into the pre-debugged crop development period identification network, mining to obtain a crop image representation array corresponding to the crop image sequence data and a meteorological element representation array corresponding to the associated meteorological element sequence data; performing attention fusion on the crop image representation array and the meteorological element representation array to obtain a fused representation array; performing development period classification prediction based on the fused representation array to obtain a crop development period classification identification result.

[0005] On the other hand, the present invention provides an identification device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and the processor implements the steps in the above method when executing the program.

[0006] The present invention generates reference guidance data based on the associated meteorological element data template of the crop image data template, generates meteorological element enhancement data according to the reference guidance data through the meteorological element data enhancement network, generates a first debugging learning template and a second debugging learning template based on the crop image data template, the associated meteorological element data template and the meteorological element enhancement data to merge and obtain multi-type data, and then mines the first characterization vector of the first debugging learning template and the second characterization vector of the second debugging learning template based on the crop development period recognition network, and then determines the metric learning debugging error based on the first characterization vector and the second characterization vector. Because the data type corresponding to the first characterization vector and the data type corresponding to the second characterization vector are different, the crop development period recognition network can be used in the communication. The process of debugging the excessive learning debugging error aligns different types of representation information. Since the meteorological element enhancement data is generated through the meteorological element data enhancement network, the crop development period recognition network can integrate the feature information of the meteorological element data enhancement network through the meteorological element enhancement data when learning the associated meteorological element data template and the crop image data template. That is, the crop development period recognition network can obtain the feature information of the meteorological element data enhancement network, so that the crop development period recognition network can have the ability to perceive generalized information, increase the recognition universality of the crop development period recognition network, and greatly increase the information representation quality of the crop development period recognition network, helping to increase the recognition accuracy of the crop development period recognition network in subsequent development period recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation flow of a crop development stage recognition method integrating meteorological and image sequences provided in an embodiment of the present invention.

[0008] Figure 2 A schematic diagram of a hardware entity of an identification device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] An embodiment of the present invention provides a method for identifying crop growth stages by integrating meteorological and image sequences. The method can be executed by a processor of an identification device, which can be a device with data processing capabilities, such as a server, laptop, tablet, or desktop computer.

[0010] Figure 1 A schematic diagram of the implementation flow of a crop growth stage recognition method integrating meteorological and image sequences provided in an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes the following steps:

[0011] Step 100: Acquire crop image sequence data and associated meteorological element sequence data of the crop to be identified.

[0012] For example, the recognition device can utilize a high-definition camera system installed in the farmland to capture images of crops at regular intervals, such as at specific times each day. These cameras are precisely positioned to ensure a comprehensive and clear capture of the overall crop growth status. The captured image data is recorded in an orderly manner, forming a time-series image collection, known as crop image sequence data. Simultaneously, the recognition device also acquires associated meteorological element sequence data. Associated meteorological element sequence data refers to a data sequence consisting of various meteorological factors closely related to crop growth. These meteorological factors have direct or indirect impacts on crop growth and development. Therefore, combined with the crop image sequence data, it can provide more comprehensive information for accurate identification of crop development stages. Possible associated meteorological elements include temperature, humidity, light intensity, precipitation, wind speed, and so on. The recognition device can obtain this data from meteorological monitoring stations installed near the farmland. These stations are equipped with various specialized meteorological sensors for real-time monitoring and recording of various meteorological elements.

[0013] These meteorological element data are also organized and recorded in chronological order, forming linked meteorological element sequence data. For example, starting at 6:00 AM, temperature data is recorded every hour, in the order of 18°C, 20°C, 22°C, 24°C, 26°C, 28°C, 30°C, 32°C, 30°C, 28°C, 26°C, and 24°C. This constitutes a simple temperature sequence data. Humidity data, light intensity data, and other data are also recorded and organized in a similar manner.

[0014] Step 200: Calling a pre-debugged crop development period recognition network, loading the crop image sequence data and the associated meteorological element sequence data into the pre-debugged crop development period recognition network, and mining to obtain a crop image representation array corresponding to the crop image sequence data and a meteorological element representation array corresponding to the associated meteorological element sequence data.

[0015] The recognition device first invokes a pre-tuned crop development stage recognition network. This network was built through a series of debugging processes, designed to accurately process crop-related data and identify key characteristics. For example, during the debugging process, the network is trained and optimized using a large amount of sample data to ensure it can adapt to different crop varieties, growing environments, and data characteristics. This debugging process involves complex algorithms and parameter adjustments to achieve optimal performance in the crop development stage recognition task. Next, the recognition device loads crop image sequence data into the crop development stage recognition network. Crop image sequence data is an ordered collection of images that contains information about the crop's appearance at different times. The recognition device feeds these images into the network frame by frame or in specific batches, according to the network's input requirements.

[0016] The network uses a series of image processing techniques and algorithms to extract features from crop images. For example, a convolutional neural network (CNN) consists of multiple convolutional layers, pooling layers, and fully connected layers. In the convolutional layers, a kernel is convolved with the image to extract local features. Assuming a 3×3 kernel, for a 224×224 crop image, the kernel slides across the image, calculating the sum of products between the kernel and the corresponding image region with each slide, thus generating a feature map. The pooling layer downsamples the feature map generated by the convolutional layer to reduce the data size while retaining key features. This reduces the data dimension without losing too much important information, improving the network's computational efficiency. After multiple convolutional and pooling layers, the image features are gradually extracted and compressed. Finally, a fully connected layer integrates and maps these features to produce a representation vector corresponding to each crop image. The representation vectors for all crop images are sequentially combined to form a crop image representation array. Simultaneously, the recognition device loads the associated meteorological element sequence data into the crop development stage recognition network. Related meteorological element sequence data consists of multiple different meteorological element data in chronological order. The network also processes this data using methods suitable for processing sequential data, such as recurrent neural networks (RNNs) and their variants, such as long short-term memory networks (LSTMs) or gated recurrent units (GRUs).

[0017] After processing by an LSTM or other similar network structure, the meteorological element data at each time step is converted into a representation vector. The representation vectors of all time steps are sequentially combined to obtain the meteorological element representation array.

[0018] Step 300: Perform attention fusion on the crop image representation array and the meteorological element representation array to obtain a fused representation array.

[0019] The recognition device first obtains a crop image representation array and a meteorological element representation array, which have been obtained through feature extraction. The crop image representation array is formed by sequentially combining the representation vectors after feature extraction of each image in the crop image sequence data. The meteorological element representation array is formed by combining the corresponding representation vectors after feature extraction of each meteorological element data in the associated meteorological element sequence data. Attention fusion enables the recognition device to dynamically assign weights, highlighting information that is more critical for identifying crop development stages. The recognition device uses an attention mechanism to fuse the crop image representation array and the meteorological element representation array. The attention mechanism calculates the degree of correlation between different representation vectors and assigns a weight to each vector, thereby determining the importance of each vector in the fusion process.

[0020] When achieving attention fusion, the recognition device can adopt various methods, one of which is to calculate the attention score. First, the recognition device processes the crop image representation array and the meteorological element representation array. Assume that the crop image representation array is I and the meteorological element representation array is M. The recognition device can map the two arrays to a higher-dimensional feature space through a linear transformation, such as using a weight matrix Transform I and M respectively to obtain Then, the recognition device calculates the attention score. The attention score is used to measure the degree of association between the crop image representation and the meteorological element representation. A simple calculation method is to use the dot product operation. For each vector i' in I' and each vector m' in M', calculate their dot product , get an attention score matrix S, where , I' i is the i-th vector in I', M' j is the jth vector in M'.

[0021] In order to convert these attention scores into weights, the recognition device uses the softmax function for normalization. The formula of the softmax function is , where z is the input vector and K is the dimension of the vector. For each row (or column) of the attention score matrix S, apply the softmax function to obtain the normalized attention weight matrix A. For example, for the i-th row S i , obtained by the softmax function , A i Each element A in ij Represents the importance weight of the j-th vector in M' to the i-th vector in I'.

[0022] Next, the recognition device performs weighted summation on the meteorological element representation array M' according to the obtained attention weight matrix A. For each vector I' in the crop image representation array I' i , calculate the corresponding weighted meteorological element representation vector , where n is the number of vectors in the meteorological element representation array M'. In this way, the set of weighted meteorological element representation vectors corresponding to each vector in the crop image representation array I' is obtained. .

[0023] Finally, the recognition device combines the crop image representation array I' and the weighted meteorological element representation vector set To fuse. A simple splicing method can be used to combine each vector in I' and the corresponding Splice together to get the fused vector . All the fused vectors Combined together, we get the fused representation array F.

[0024] Step 400: Perform growth stage classification prediction based on the fused representation array to obtain crop growth stage classification recognition results.

[0025] The recognition device obtains the fused representation array obtained after attention fusion. This array integrates the key information of the crop image representation array and the meteorological element representation array, and is a dataset containing rich features. For example, the fused representation array may be a dimension of where n represents the number of data samples. Assume that n = 100 here, which represents 100 samples obtained from a series of observation data. d is the feature dimension of each sample. Assume that d = 2048, which means that each sample has 2048 features, which comprehensively reflect the comprehensive information of crop images and related meteorological elements.

[0026] To predict crop developmental stages, the recognition device uses an appropriate classification model. Possible classification models include support vector machines (SVMs), decision trees, random forests, and neural networks. Taking a neural network as an example, the recognition device constructs a multi-layer neural network consisting of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is equal to the feature dimension d of the fused representation array, i.e., 2048 neurons. These neurons receive feature data from the fused representation array. In the hidden layers, neurons transform the input data and extract features using nonlinear activation functions. The number of neurons in the output layer is equal to the number of crop developmental stage categories. Assuming the crop developmental stage is divided into six stages: sowing, germination, seedling, flowering, fruiting, and maturity, the output layer would have six neurons. Each neuron outputs a value representing the probability that the input sample belongs to the corresponding developmental stage. The recognition device performs a forward propagation calculation on each sample in the fused representation array, starting from the input layer, proceeding through the hidden layers, and finally reaching the output layer, to obtain the probability distribution of each sample belonging to each developmental stage.

[0027] As an embodiment, the crop development stage recognition network is debugged by the following steps:

[0028] Step 10: Obtain a meteorological element data template associated with the crop image data template, generate reference guidance data based on the associated meteorological element data template, load the reference guidance data into a meteorological element data enhancement network for meteorological element data enhancement processing, and construct meteorological element enhanced data that matches the crop image data template, wherein the reference guidance data is used to guide the meteorological element data enhancement network to construct the meteorological element enhanced data.

[0029] The recognition device first obtains the associated meteorological element data template of the crop image data template. The crop image data template is a representative set of crop images used to debug the network, while the associated meteorological element data template is a set of meteorological element data that is closely related to these crop images in time, space or other logical relationships. Then, the recognition device generates reference guidance data based on the associated meteorological element data template. There are many ways to generate reference guidance data, which depends on the composition of the associated meteorological element data template. If the associated meteorological element data template only includes meteorological attribute data, the recognition device can directly generate reference guidance data based on these meteorological attribute data. For example, the meteorological attribute data contains basic information such as the geographical location of the meteorological monitoring station, the monitoring time range, and the type of meteorological elements. The recognition device can organize and encode this information to form reference guidance data for guiding the meteorological element data enhancement network.

[0030] If the associated meteorological element data template includes meteorological attribute data and meteorological record data, the recognition device generates reference guidance data according to a specific process. First, the meteorological record data is disassembled to obtain multiple meteorological record data sequences. For example, the meteorological record data contains temperature, humidity, and light intensity records for each day of the week. The recognition device disassembles them into temperature sequences, humidity sequences, and light intensity sequences respectively. Then, the crop impact status of these meteorological record data sequences is identified, and the target data sequence is determined in each meteorological record data sequence based on the crop impact status identification results. For example, through analysis, it is found that within a certain temperature range, the growth rate of crops accelerates, so the part of the temperature data sequence within this temperature range may be determined as the target data sequence. Finally, the meteorological attribute data and the target data sequence are combined to generate reference guidance data.

[0031] When the associated meteorological element data template includes meteorological attribute data and meteorological annotation data, the recognition device combines these two to generate reference guidance data. Meteorological annotation data is label-based information summarizing the corresponding meteorological conditions, such as "high temperature" and "suitable humidity." The recognition device combines meteorological attribute data with this annotation information, for example, combining attribute data such as the location of the meteorological monitoring station and the monitoring time range with the "high temperature" annotation to generate reference guidance data.

[0032] If the associated meteorological element data template includes meteorological attribute data, meteorological record data, and meteorological annotation data, the recognition device combines these three to generate reference guidance data. By integrating these different types of data, it can guide the meteorological element data enhancement network more comprehensively and accurately.

[0033] After generating the reference guidance data, the recognition device loads it into the meteorological element data enhancement network for meteorological element data enhancement processing. Taking data expansion as an example, the network can generate more data that is similar to but not exactly the same as the original meteorological element data through a specific algorithm based on the reference guidance data. For example, for temperature data, the meteorological element data enhancement network can generate a new temperature data sequence based on the temperature change trend, range and other information contained in the reference guidance data using methods such as random noise addition and linear transformation. Assuming that the original temperature data sequence is [20, 22, 24], the network may generate a new temperature data sequence by adding a random noise that conforms to the normal distribution to each data point. , generate a new data sequence , thereby achieving data expansion.

[0034] In addition to data augmentation, the meteorological element data augmentation network can also perform other enhancement methods. For example, data transformation can be performed by performing operations such as logarithmic transformation and normalization on the original meteorological element data to change the data distribution and increase data diversity.

[0035] Step 20: Obtain the crop image data template, generate a first debugging learning template based on two of the crop image data template, the associated meteorological element data template, and the meteorological element enhancement data, and generate a second debugging learning template based on the remaining one of the crop image data template, the associated meteorological element data template, and the meteorological element enhancement data after generating the first debugging learning template.

[0036] The recognition device first acquires a crop image data template. This template is a representative collection of crop images, encompassing typical characteristics of the crop at different growth stages and under different environmental conditions. For example, for a particular rice variety, the crop image data template might include images of sparse, tender green plants during the seedling stage after sowing, images of gradually lush plants during the tillering stage, and images of different stages such as heading, flowering, grain filling, and maturity. These images are captured to meet certain quality standards, such as a certain resolution (e.g., 1920×1080 pixels) and appropriate lighting conditions, to ensure that characteristic information such as crop morphology, color, and texture are clearly captured.

[0037] Next, the recognition device generates a first debugging learning template based on two of the crop image data template, the associated meteorological element data template, and the meteorological element enhancement data. This generation method can be combined in various ways. If both the crop image data template and the associated meteorological element data template are selected to generate the first debugging learning template, the recognition device integrates these two data types. For example, crop images of the same growth stage are correlated and matched with the corresponding meteorological element data. For example, during the tillering stage of rice, several images of that period are selected from the crop image data template, while meteorological element data such as temperature, humidity, and light intensity for the same time period are extracted from the associated meteorological element data template. The recognition device can then use data concatenation to concatenate the image data (e.g., pixel value data represented in matrix form) with the meteorological element data (represented in vector form) along specific dimensions to form a sample of the first debugging learning template. In this way, multiple samples are combined to form the first debugging learning template generated based on the crop image data template and the associated meteorological element data template.

[0038] If a crop image data template and meteorological element enhancement data are selected to generate the first debugging learning template, the recognition device will perform similar data integration operations. For example, the meteorological element enhancement data generated by the meteorological element data enhancement network will be matched with the corresponding images in the crop image data template based on temporal, spatial, or other logical associations. Assuming that the meteorological element enhancement data contains data such as temperature and precipitation under a series of simulated climate conditions, the recognition device will find images in the crop image data template taken under similar climate conditions. Then, according to certain rules (such as weighted summation of the image feature vector and the meteorological element data vector), the two are combined into samples to construct the first debugging learning template.

[0039] If the associated meteorological element data template and meteorological element enhancement data are selected to generate the first debugging learning template, the recognition device analyzes the inherent connection between the two types of data. For example, the associated meteorological element data template records the actual observed meteorological data, while the meteorological element enhancement data is a diverse data generated based on it. The recognition device can pair these two types of data with similar characteristics (such as the same meteorological trend or a similar numerical range) and then organize them into a data structure (such as a list, where each element contains the associated meteorological element data and the corresponding meteorological element enhancement data) to form a sample of the first debugging learning template, ultimately constructing the entire first debugging learning template.

[0040] After generating the first debugging learning template, the recognition device generates a second debugging learning template based on the remaining data template. For example, if the first debugging learning template is generated from a crop image data template and an associated meteorological element data template, the recognition device uses the meteorological element enhanced data as the second debugging learning template. This is because the meteorological element enhanced data has undergone specific enhancement processing and has different characteristics and distribution from the original associated meteorological element data. Using it as a separate debugging learning template allows the crop growth stage recognition network to learn the new information and features introduced by the meteorological element data enhancement network.

[0041] If the first debugging learning template is generated from the crop image data template and the meteorological element enhanced data, the recognition device will associate the meteorological element data template as the second debugging learning template. This allows the network to compare and learn the differences and connections between the original observed meteorological element data and the enhanced data.

[0042] If the first debugging learning template is generated by associating a meteorological element data template with meteorological element enhancement data, the recognition device uses the crop image data template as the second debugging learning template. This allows the network to focus on learning the characteristics of the crop image itself and the relationship between these characteristics and the meteorological element data features in the first debugging learning template.

[0043] Step 30: Load the first debugging learning template and the second debugging learning template into a crop development stage recognition network, and mine a first representation vector of the first debugging learning template and a second representation vector of the second debugging learning template.

[0044] The recognition device loads the first and second debugging learning templates into the crop development stage recognition network. These two templates contain information generated through different combinations of crop image data templates, associated meteorological element data templates, and meteorological element enhancement data, and serve as important inputs for network learning and analysis. For example, suppose the first debugging learning template is generated from the crop image data template and the associated meteorological element data template. It may contain 100 data sets, each combining images of crops at a certain growth stage with meteorological element data from the same period. The second debugging learning template is the meteorological element enhancement data, containing 80 enhanced meteorological element samples.

[0045] For the first debugging learning template, the recognition device mines its first characterization vector using relevant components in the network. Take the network structure including the first meteorological element characterization vector mining component, the second meteorological element characterization vector mining component, the crop image characterization vector mining component and the feature integration component as an example (when the second debugging learning template is generated based on meteorological element enhancement data or an associated meteorological element data template). Based on the first meteorological element characterization vector mining component, the recognition device extracts features from the data in the associated meteorological element data template. For example, the temperature data in the associated meteorological element data template can be extracted by the component through a series of convolution operations (such as convolution operation of the two-dimensional convolution kernel K with the temperature data matrix T, the formula is , where C is the convolution result matrix, M and N are the sizes of the convolution kernel), extract the local features of the temperature data, such as the trend and fluctuation of temperature change, and convert it into a feature vector as the first feature vector of the associated meteorological element data template.

[0046] Simultaneously, the recognition device extracts features from the image within the crop image data template, based on the crop image representation vector mining component. For example, for a crop image, this component can use a deep convolutional neural network (CNN) through multiple convolutional and pooling layers. In the convolutional layers, different convolution kernels extract various image features, such as edges and textures. The pooling layer downsamples the convolutional layer output, reducing the data volume while retaining key features. After these multiple layers of processing, the image is converted into a feature vector, which serves as the third feature vector of the crop image data template.

[0047] If the second debugging learning template is generated based on the meteorological element enhanced data, the recognition device, based on the feature integration component, interacts (or fuses) the first and third feature vectors to obtain a first representation vector for the first debugging learning template. For example, the feature integration component may concatenate the first and third feature vectors in a dimensionally concatenated manner to form a new vector, which serves as the first representation vector for the first debugging learning template. The second representation vector for the second debugging learning template is directly derived from the second feature vector mined from the meteorological element enhanced data by the second meteorological element representation vector mining component.

[0048] If the second debugging learning template is generated based on the associated meteorological element data template, based on the feature integration component, the recognition device interacts the second feature vector and the third feature vector to obtain the first characterization vector of the first debugging learning template, and uses the first feature vector as the second characterization vector of the second debugging learning template.

[0049] If the second debugging learning template is generated based on the crop image data template, the crop development period recognition network includes a first meteorological element representation vector mining component and a crop image representation vector mining component. Based on the first meteorological element representation vector mining component, the recognition device sequentially mines the representation vectors of the associated meteorological element data template and the meteorological element enhancement data to obtain the first representation vector of the first debugging learning template. For example, the associated meteorological element data template is first processed through a series of linear transformations (such as y = Wx + b, where x is the input meteorological element data vector, W is the weight matrix, b is the bias vector, and y is the output feature vector) and nonlinear activation functions (such as ReLU function ), extracting its features to obtain a vector; similar processing is then performed on the meteorological element enhancement data to obtain another vector, and these two vectors are then combined (e.g., added or concatenated) to obtain a first representation vector for the first debugging learning template. Alternatively, the recognition device may combine the associated meteorological element data template and the meteorological element enhancement data and then perform representation vector mining based on the first meteorological element representation vector mining component to obtain the first representation vector for the first debugging learning template. Simultaneously, the crop image representation vector mining component performs representation vector mining on the crop image data template to obtain a second representation vector for the second debugging learning template.

[0050] Step 40: Determine a metric learning debugging error based on the first characterization vector and the second characterization vector, and debug the crop development stage recognition network based on the metric learning debugging error.

[0051] The recognition device first determines the metric learning debugging error based on the first and second representation vectors. The first and second representation vectors are mined from the first and second debugging learning templates, respectively, and they carry data features from various aspects, such as crop images and meteorological factors. The goal of metric learning is to enable the network to learn the similarities and differences between different data types, thereby better identifying crop development stages.

[0052] To determine the metric learning debugging error, the recognition device analyzes the relationship between the first representation vector and the second representation vector from multiple perspectives. Assuming there are multiple first representation vectors, each of which corresponds one-to-one to a plurality of second representation vectors. The recognition device determines, for each first representation vector, a first correlation coefficient between it and the corresponding second representation vector. This coefficient reflects the degree of commonality between the two vectors. For example, the first correlation coefficient is determined by calculating the cosine similarity between the two vectors.

[0053] Next, the recognition device performs a logarithmic operation on the ratio between the first correlation coefficient and the first cumulative correlation coefficient to determine the first component debugging error. Assume that the first correlation coefficient is , the first cumulative correlation coefficient is , then the first component is the debugging error .

[0054] Similarly, for each second characterization vector, the recognition device determines a third correlation coefficient between the second characterization vector and the corresponding first characterization vector, as well as a fourth correlation coefficient between each of the other non-corresponding first characterization vectors. A second cumulative correlation coefficient is determined based on the sum of the third correlation coefficient and each of the fourth correlation coefficients. A logarithmic operation is then performed on the ratio between the third correlation coefficient and the second cumulative correlation coefficient to determine a second component debugging error.

[0055] Finally, the recognition device determines the metric learning debugging error based on the first component debugging error and the second component debugging error, for example, by adding them together.

[0056] After determining the metric learning debugging error, the recognition device uses this error to debug the crop development stage recognition network. This process involves the coordinated debugging of multiple components within the network, such as the first meteorological element representation vector mining component, the second meteorological element representation vector mining component, the crop image representation vector mining component, and the feature integration component. The recognition device can use optimization algorithms such as gradient descent to adjust network parameters. By continuously iteratively updating these parameters, the recognition device gradually reduces the metric learning debugging error.

[0057] During actual debugging, the recognition device repeats the above error determination and parameter adjustment process multiple times until the metric learning debugging error reaches a satisfactory range. In this way, the crop growth stage recognition network can better align different types of representation information, improving the network's understanding and processing capabilities of crop images and meteorological element data, thereby enhancing the network's accuracy and generalization ability in crop growth stage recognition tasks.

[0058] As an embodiment, if the second debugging learning template is generated based on the meteorological element enhancement data or the associated meteorological element data template, the crop development period recognition network includes a first meteorological element representation vector mining component, a second meteorological element representation vector mining component, a crop image representation vector mining component, and a feature integration component, step 30 of mining the first representation vector of the first debugging learning template and the second representation vector of the second debugging learning template includes:

[0059] Step S31: mining a first feature vector of the associated meteorological element data template based on the first meteorological element representation vector mining component, mining a second feature vector of the meteorological element enhanced data based on the second meteorological element representation vector mining component, and mining a third feature vector of the crop image data template based on the crop image representation vector mining component;

[0060] Step S32: If the second debugging learning template is generated based on the meteorological element enhancement data, interacting the first feature vector and the third feature vector based on the feature integration component to obtain a first representation vector of the first debugging learning template, and using the second feature vector as a second representation vector of the second debugging learning template;

[0061] Step S33: If the second debugging learning template is generated based on the associated meteorological element data template, the second feature vector and the third feature vector are interacted based on the feature integration component to obtain the first characterization vector of the first debugging learning template, and the first feature vector is used as the second characterization vector of the second debugging learning template.

[0062] In step S31, the recognition device uses the first meteorological element representation vector mining component to extract features from the associated meteorological element data template. This component, for example, includes a series of algorithms and model structures designed to extract feature information from the associated meteorological element data template that is important for identifying crop growth stages. Taking temperature data as an example, assuming that the temperature data in the associated meteorological element data template is presented in the form of a time series, such as temperature values ​​recorded every hour over a period of time, The first meteorological element representation vector mining component may first pre-process the data, such as normalization operation.

[0063] Next, the component can employ a convolutional neural network (CNN) architecture to extract features. For example, a CNN model with multiple convolutional and pooling layers can be used. Using different convolutional kernels, various local features in the temperature data, such as temperature trends and fluctuations, can be extracted. The pooling layer downsamples the output of the convolutional layer, ultimately generating a vector representing the temperature characteristics of the associated meteorological element data template—the first eigenvector.

[0064] For meteorological element enhanced data, the recognition device uses the second meteorological element representation vector mining component to process it. Since meteorological element enhanced data is generated through specific enhancement operations, its feature extraction method may be slightly different from that of the associated meteorological element data template, but the principles are similar.

[0065] For crop image data templates, the recognition device uses a crop image representation vector mining component to extract features. This component is based on the convolutional neural network (CNN) architecture used in deep learning. For example, a crop image might be represented as a matrix with pixel values ​​of I. The crop image representation vector mining component first processes the image through a series of convolutional layers. For example, the first convolutional layer convolves the image with multiple different convolution kernels, each extracting a local feature in the image, such as edges or texture. The convolution operation generates multiple feature maps, which are then downsampled through a pooling layer to reduce the data volume. After alternating convolutional and pooling layers, the image features are gradually extracted and compressed. Finally, a fully connected layer integrates and maps these features to produce a vector representing the crop image features, the third feature vector.

[0066] After completing step S31, the recognition device performs further processing according to steps S32 and S33. The feature integration component may use various methods to fuse the first and third feature vectors. For example, a splicing method may be used to splice the first and third feature vectors in terms of dimension. The second representation vector of the second debugging learning template directly uses the second feature vector obtained by mining the meteorological element enhancement data using the second meteorological element representation vector mining component. This is because, in this case, the second debugging learning template primarily focuses on the features of the meteorological element enhancement data. This second feature vector can fully represent the key information of this data and is used for subsequent comparison with the first representation vector and network debugging. If the second debugging learning template is generated based on the associated meteorological element data template, the recognition device, based on the feature integration component, interacts with the second and third feature vectors to obtain the first representation vector of the first debugging learning template, and uses the first feature vector as the second representation vector of the second debugging learning template.

[0067] The feature integration component may fuse the second feature vector and the third feature vector by concatenation or a more complex weighted fusion method.

[0068] In this case, the first eigenvector serves as the second representation vector of the second debugging learning template. This is because, in this case, the second debugging learning template is centered around the associated meteorological element data template, and the first eigenvector is extracted from this template. It accurately represents the characteristics of the associated meteorological element data template and is used for subsequent network debugging and comparative analysis.

[0069] Through the above steps, the recognition device can effectively extract key features from different data templates and, based on the basis for generating the second debug learning template, rationally integrate these features to generate a first representation vector for the first debug learning template and a second representation vector for the second debug learning template. These representation vectors not only contain the core information of the crop image, associated meteorological elements, and meteorological element-enhanced data, but also, through a rational fusion method, better reflect the relationships and characteristics between different data types.

[0070] As an embodiment, in step 40, debugging the crop development stage recognition network based on the metric learning debugging error includes:

[0071] Step 41: Based on the metric learning debugging error, the first meteorological element representation vector mining component, the second meteorological element representation vector mining component, the crop image representation vector mining component and the feature integration component are collaboratively debugged.

[0072] The first step is to debug the first meteorological element representation vector mining component. The main function of this component is to mine valuable feature information from the associated meteorological element data template and convert it into the corresponding first feature vector. The recognition device analyzes the accuracy and effectiveness of the component in the feature extraction process based on the metric learning debugging error. If the metric learning debugging error indicates that the features extracted by the component are not accurate or complete, the recognition device may adjust the internal parameter settings of the component. For example, the convolution kernel size and step size in the convolutional neural network (CNN) or the number of hidden layer nodes and weight coefficients in the recurrent neural network (RNN) can be changed to improve its feature extraction ability for meteorological element data, so that the mined first feature vector can more accurately reflect the inherent characteristics of the associated meteorological element data template.

[0073] Next, the second meteorological element representation vector mining component is debugged. This component is responsible for mining the second feature vector from the meteorological element enhanced data. The recognition device also evaluates the performance of this component based on the metric learning debugging error. Assume that the meteorological element enhanced data is obtained by performing data augmentation on the original meteorological data and contains more meteorological variation patterns. If the metric learning debugging error of the second meteorological element representation vector mining component is large when processing this data, it indicates that it may not fully mine the effective information in the enhanced data. The recognition device then takes appropriate measures, such as adjusting the parameters of the component's attention mechanism.

[0074] The debugging of the crop image representation vector mining component is also a key link. This component needs to mine the third eigenvector from the crop image data template. The recognition device judges its extraction effect on the crop image features based on the metric learning debugging error. For example, the crop image data template contains crop images at different growth stages, and the component needs to extract features such as the morphology, color, and texture of the crop. If the metric learning debugging error indicates that the features extracted by the component cannot well reflect the actual development of the crop, the recognition device can adjust the image processing algorithm in the component. For example, for a convolutional neural network model based on deep learning, the number of convolutional layers can be increased or decreased, and the method of pooling operation can be changed to optimize the feature extraction of the crop image, so that the generated third eigenvector can more accurately represent the features of the crop image data template.

[0075] Finally, the feature integration component is debugged. The function of this component is to interact and integrate feature vectors from different sources to obtain a more comprehensive and accurate representation vector. The recognition device analyzes the effect of this component in the feature fusion process based on the metric learning debugging error. For example, when the first debugging learning template is generated based on the associated meteorological element data template and the crop image data template, the feature integration component needs to effectively interact the first feature vector and the third feature vector. If the metric learning debugging error shows that the integrated feature vector cannot well reflect the comprehensive information of the two data, the recognition device may adjust the feature integration method. For example, the feature vectors are integrated using a weighted summation method, and weights are assigned according to the importance of different feature vectors. By adjusting the weight coefficient, the integrated representation vector can better integrate feature information from different sources.

[0076] Throughout the collaborative debugging process, the recognition device continuously adjusts and optimizes these four components based on the metric learning debugging error. By gradually reducing the error, the components can work better together, thereby improving the crop development stage recognition network's ability to process crop images and meteorological data, ultimately enhancing the accuracy and reliability of crop development stage recognition.

[0077] In one embodiment, the number of the first characterization vectors is multiple, and the multiple first characterization vectors correspond one-to-one to the multiple second characterization vectors. In step 40, determining the metric learning debugging error based on the first characterization vector and the second characterization vector includes:

[0078] Step 42: For each first characterization vector, determine a first correlation coefficient between the first characterization vector and the corresponding second characterization vector, determine a second correlation coefficient between the first characterization vector and each other non-corresponding second characterization vector, determine a first cumulative correlation coefficient based on the sum of the first correlation coefficient and each second correlation coefficient, perform a logarithmic operation on the ratio between the first correlation coefficient and the first cumulative correlation coefficient, and determine a first component debugging error;

[0079] Step 43: For each second characterization vector, determine a third correlation coefficient between the second characterization vector and the corresponding first characterization vector, determine a fourth correlation coefficient between the second characterization vector and each other non-corresponding first characterization vector, determine a second cumulative correlation coefficient based on the sum of the third correlation coefficient and each of the fourth correlation coefficients, perform a logarithmic operation on the ratio between the third correlation coefficient and the second cumulative correlation coefficient, and determine a second component debugging error;

[0080] Step 44: Determine a metric learning debugging error based on the first component debugging error and the second component debugging error.

[0081] In step 42, the recognition device performs a series of calculations on each first characterization vector. First, the first correlation coefficient between the first characterization vector and the corresponding second characterization vector is determined. The first characterization vector is a set of feature vectors mined from a specific data template, representing the key feature information of the data; the second characterization vector is similar and has a corresponding relationship with the first characterization vector. Taking a specific scenario as an example, assuming that in the crop development period recognition task, the first characterization vector set It is mined from the associated meteorological element data template and reflects the characteristics of meteorological elements. The second characterization vector set It is mined from meteorological element enhancement data or other related data, and One-to-one correspondence. Identification device calculation The first correlation coefficient between them, such as the cosine similarity formula ,in is the vector dot product, and They are vectors and The formula measures the similarity between two vectors in direction, and its value range is [-1,1] The larger the value, the more similar the two vectors are, that is, the larger the first correlation coefficient, the closer the corresponding relationship.

[0082] Next, the recognition device determines the second correlation coefficient between the first characterization vector and each of the other non-corresponding second characterization vectors. , except for the corresponding In addition to calculating the first correlation coefficient, we also need to calculate and 、 and The second correlation coefficient also uses the cosine similarity formula.

[0083] The first cumulative correlation coefficient is determined based on the sum of the first correlation coefficient and each second correlation coefficient. Then, the recognition device performs a logarithmic operation on the ratio between the first correlation coefficient and the first cumulative correlation coefficient to determine the first component debugging error. Let the first component debugging error be , the calculation formula is Through logarithmic operations, the proportional relationship is converted into an error value that is easier to analyze and process. Indicates that the correlation between the first representation vector and the corresponding second representation vector accounts for a larger proportion in the overall correlation, that is, the corresponding relationship is more accurate; conversely, a larger This indicates that there is a problem, perhaps the correspondence is inaccurate or there is a deviation in feature extraction.

[0084] The recognition device repeats the above calculation process for each first characterization vector to obtain a series of first component debugging errors. , calculate its The first correlation coefficient ,and 、 The second correlation coefficient , the first cumulative correlation coefficient , and the first component debugging error ;for Likewise, get .

[0085] In step 43, the recognition device performs similar calculations for each second characterization vector. A third correlation coefficient between the second characterization vector and the corresponding first characterization vector is determined. For example, calculate its The third correlation coefficient still uses the cosine similarity formula .

[0086] Then determine the fourth correlation coefficient between the second characterization vector and each of the other non-corresponding first characterization vectors. and 、 and The fourth correlation coefficients are and .

[0087] The second cumulative correlation coefficient is determined based on the sum of the third correlation coefficient and each fourth correlation coefficient. , its second cumulative correlation coefficient .

[0088] Then, the ratio between the third correlation coefficient and the second cumulative correlation coefficient is logarithmically calculated to determine the second component debugging error. Let the second component debugging error be , the calculation formula is Likewise, smaller Indicates that the correlation between the second representation vector and the corresponding first representation vector accounts for a larger proportion in the overall correlation. This indicates that there may be an issue with inaccurate association.

[0089] The recognition device performs this process for each second characterization vector, and obtains a series of second component debugging errors. ,calculate ;for ,calculate .

[0090] In step 44, the recognition device determines the metric learning debugging error based on the first component debugging error and the second component debugging error. A feasible calculation method is to accumulate all the first component debugging errors and the second component debugging errors or use a weighted average method. Assume that the first component debugging error set is , the second component debugging error set is , simple cumulative metric learning debugging error If the weighted average method is used, the weight of the first component debugging error is , the weight of the second component debugging error is ,and , then the metric learning debugging error The weight setting can be adjusted according to the actual situation to highlight the impact of errors in different parts on the overall debugging.

[0091] As another embodiment, if the second debugging learning template is generated based on the crop image data template, the crop development period recognition network includes a first meteorological element representation vector mining component and a crop image representation vector mining component, the step 30 of mining the first representation vector of the first debugging learning template and the second representation vector of the second debugging learning template includes:

[0092] Step 301: performing representation vector mining on the associated meteorological element data template and the meteorological element enhancement data in sequence based on the first meteorological element representation vector mining component to obtain a first representation vector of the first debugging learning template, or combining the associated meteorological element data template and the meteorological element enhancement data and performing representation vector mining based on the first meteorological element representation vector mining component to obtain a first representation vector of the first debugging learning template;

[0093] Step 302: Performing representation vector mining on the crop image data template based on the crop image representation vector mining component to obtain a second representation vector of the second debugging learning template.

[0094] In step 301, assuming a specific crop growth stage identification task scenario, the associated meteorological element data template contains meteorological attribute data and meteorological record data for a period of time. For example, daily meteorological record data such as temperature, humidity, and sunshine duration, as well as meteorological attribute data such as weather type (sunny, cloudy, rainy, etc.) and season. The meteorological element enhancement data is obtained by performing data augmentation operations on the associated meteorological element data template, such as adding data samples simulating extreme weather conditions.

[0095] The first meteorological element representation vector mining component is implemented, for example, using a deep learning model such as a convolutional neural network (CNN) or a recurrent neural network (RNN) and its variants (such as LSTM, GRU).

[0096] When mining the combined data templates of associated meteorological elements and enhanced meteorological elements, the recognition device first combines the two data types according to certain rules. For example, the associated meteorological element data templates and enhanced meteorological element data are aligned in the time dimension, and then the data at the same time point or time period are merged to form a new combined data matrix.

[0097] This combined data matrix is ​​then input into the first meteorological element representation vector mining component. This component also extracts features from the combined data through convolutional neural network operations such as convolution and pooling. During this process, the convolution kernel simultaneously captures and fuses the features of the associated meteorological element data and the meteorological element enhanced data. For example, when scanning the data, the convolution kernel considers both the normal variation trends in the original meteorological record data and the abnormal characteristics introduced by the meteorological element enhanced data, thereby extracting more comprehensive and representative features. After a series of processing, the first representation vector of the first debugging learning template is directly obtained.

[0098] In step 302, assume that the crop image data template contains crop images at different growth stages. These images contain rich visual information, such as crop morphology, color, and texture. The crop image representation vector mining component utilizes, for example, a deep learning model specifically designed for image data processing, such as a convolutional neural network (CNN) architecture, such as VGG and ResNet models.

[0099] After processing through multiple convolutional layers, pooling layers, and residual blocks, the image data is gradually converted into a representative feature vector. Ultimately, the feature vector output by the crop image representation vector mining component serves as the second representation vector of the second debugging learning template. This second representation vector incorporates key feature information from the crop image data template, including crop morphology, color, texture, and other aspects, accurately representing the feature representation of the crop image under the current network model.

[0100] Through step 301 and step 302 , the recognition device mines a first representation vector of the first debugging learning template and a second representation vector of the second debugging learning template based on the first meteorological element representation vector mining component and the crop image representation vector mining component respectively.

[0101] As an embodiment, in step 40, determining the metric learning debugging error based on the first characterization vector and the second characterization vector includes:

[0102] Step 402A: If the first representation vector is obtained by sequentially performing representation vector mining on the associated meteorological element data template and the meteorological element enhancement data, a first sub-metric learning debugging error is determined based on the first representation vector and the second representation vector corresponding to the associated meteorological element data template, a second sub-metric learning debugging error is determined based on the first representation vector and the second representation vector corresponding to the meteorological element enhancement data, and a metric learning debugging error is determined based on the first sub-metric learning debugging error and the second sub-metric learning debugging error.

[0103] Alternatively, step 402B: if the first characterization vector is obtained by combining the associated meteorological element data template and the meteorological element enhancement data and performing characterization vector mining based on the first meteorological element characterization vector mining component, then the metric learning debugging error is determined based on the first characterization vector and the second characterization vector.

[0104] In step 402A, for a specific crop development stage identification project, the associated meteorological element data template records daily meteorological element information such as temperature, humidity, and wind speed for the past month. The meteorological element enhanced data is obtained by performing enhancement operations such as interpolation and noise addition on the associated meteorological element data. The identification device uses a specific deep learning model, such as a recurrent neural network (RNN)-based architecture, to sequentially mine representation vectors for the associated meteorological element data template and the meteorological element enhanced data.

[0105] For the associated meteorological element data template, the recognition device inputs it into the first meteorological element representation vector mining component based on RNN. RNN can process data with sequence characteristics. Its core idea is to remember past information through hidden states. At each time step t, the hidden state The update formula is: ;in is the input data at time step t, is the weight matrix from hidden state to hidden state, is the weight matrix input to the hidden state, is the bias term, is an activation function, such as tanh or sigmoid function. After a series of time steps, the first representation vector corresponding to the associated meteorological element data template is finally obtained. .

[0106] Similarly, the recognition device processes the meteorological element enhancement data and obtains the first representation vector corresponding to the meteorological element enhancement data through the same first meteorological element representation vector mining component. .

[0107] Next, the recognition device needs to determine the first sub-metric learning debugging error. In order to measure the first representation vector corresponding to the associated meteorological element data template and the second characterization vector The recognition device uses cosine similarity to calculate the similarity between them. The calculation formula of cosine similarity is: ;in is the dot product of two vectors, are the norms of the two vectors. If the directions of the two vectors are more similar, the cosine similarity is closer to 1; conversely, if the direction difference is greater, the cosine similarity is closer to -1. The recognition device hopes and There is a high similarity between them because they should represent relevant feature information. Assuming that ideally, the expected similarity is , then the first sub-metric learning debugging error It can be calculated by the following formula: This formula uses mean square error to measure the difference between actual similarity and ideal similarity. The square operation makes the error more obvious, which is convenient for subsequent analysis and adjustment.

[0108] The first characterization vector corresponding to the meteorological element enhancement data and the second characterization vector The recognition device also calculates the second sub-metric learning and debugging error according to the above method .

[0109] Finally, the recognition device learns and debugs the error based on the first sub-metric and the second sub-metric learning debugging error Determine the metric learning debugging error E. One way is to directly add the two, that is, E = + .

[0110] Through such calculation, the recognition device can accurately determine the difference between the characterization vector obtained based on the associated meteorological element data template and the meteorological element enhanced data and the second characterization vector, thereby providing a specific direction for adjusting the crop development period recognition network.

[0111] In step 402B, continuing with the above project as an example, the recognition device first combines the associated meteorological element data template and the meteorological element enhancement data. Assuming a simple splicing method is used, the matrix of the associated meteorological element data template is combined. and matrix of meteorological element enhancement data Concatenate columns into a new matrix M=[ M 1 , M 2 ] .

[0112] The recognition device then inputs this combined matrix M into the first meteorological element representation vector mining component. This first meteorological element representation vector mining component can be a model based on a convolutional neural network (CNN). CNNs excel at processing image and matrix data, extracting data features through operations such as convolutional and pooling layers.

[0113] Next, the recognition device uses the first characterization vector And the second representation vector Determine the metric learning debugging error. The cosine similarity is also used to measure the similarity between them, and the calculation method is the same as the method of calculating the similarity of two vectors in step 402A. Assume that the expected similarity is , then the metric learning debugging error E can be calculated by the following formula: This formula also uses mean squared error to measure the difference between actual and ideal similarity. Using this error value, the recognition device can determine the degree of match between the features mined from combining the associated meteorological element data template and the enhanced meteorological element data and the features represented by the second representation vector.

[0114] As an embodiment, in step 30, before loading the first debugging learning template and the second debugging learning template into the crop growth stage recognition network, the crop growth stage recognition method integrating meteorological and image sequences further includes:

[0115] Step 30a: shielding the target data item in the associated meteorological element data template or the meteorological element enhanced data to obtain associated meteorological element shielded data, and shielding the target image area in the crop image data template to obtain crop image shielded data;

[0116] Step 30b: loading the associated meteorological element shielded data and the crop image shielded data into the crop growth stage recognition network, performing reasoning on the shielded target data item to obtain a first reasoning result, and performing reasoning on the shielded target image area to obtain a second reasoning result;

[0117] Step 30c: determining a first restoration error based on the first inference result and the target data item, and determining a second restoration error based on the second inference result and the target image area;

[0118] Step 30d: Debugging the crop development stage recognition network based on the first restoration error and the second restoration error.

[0119] In step 30a, the recognition device masks the target data item in the associated meteorological element data template or the meteorological element enhanced data to obtain associated meteorological element masked data, and masks the target image area in the crop image data template to obtain crop image masked data.

[0120] For the associated meteorological element data template and meteorological element enhancement data, the recognition device randomly or according to specific rules selects some target data items to mask. For example, suppose the associated meteorological element data template contains daily data such as temperature, humidity, wind speed, and sunshine duration. The recognition device may select and mask the temperature data of certain days as target data items. For example, this masking operation can be achieved by setting the data at the corresponding position in the data matrix to a specific placeholder (such as 0 or a special mark).

[0121] For crop image data templates, the recognition device similarly selects specific target image areas for masking. For example, if the crop image data template consists of a series of images of crops at different growth stages, the recognition device might choose to mask a specific portion of the crop, such as a portion of a leaf. This can be achieved through image processing techniques, using the image's coordinate system to determine the area to be masked, and then setting the pixel values ​​in that area to a specific value (such as complete black or a fixed grayscale value).

[0122] In step 30b, after the masked data for the associated meteorological elements is loaded into the crop development stage recognition network, the network attempts to infer the value of the masked target data item based on the unmasked data. For example, using the previously masked temperature data, the network calculates and infers the value using the connection weights and activation functions between neurons in the network, based on other meteorological elements such as humidity, wind speed, and sunshine duration, as well as its own learned patterns and regularities.

[0123] For masked crop image data, the crop development stage recognition network uses information from unmasked areas of the image, such as the crop's overall morphology and color distribution, to infer the content of the masked target image area. For example, in an image recognition model based on a convolutional neural network (CNN), the convolutional layer extracts local features of the image, and the pooling layer downsamples these features to reduce the data volume. After processing through multiple convolutional and pooling layers, the fully connected layer integrates these features and outputs a prediction for the masked area, which is the second inference result.

[0124] In step 30c, when calculating the first restoration error, the recognition device compares the first inference result with the actual target data item. Taking temperature data as an example, if the actual masked temperature value is , and the first inference result obtained by network inference is , then the mean square error (MSE) can be used to calculate the first restoration error , , where N is the number of masked temperature data, and are the kth actual temperature value and the inferred temperature value, respectively. The mean square error measures the average square of the error between the predicted value and the actual value. A larger error value indicates a lower accuracy of the network's inference on the target data item.

[0125] For the second restoration error, the recognition device compares the second inference result with the actual target image area. In the image field, the structural similarity index (SSIM) can be used to measure the similarity between images and then calculate the second restoration error. Assume that the original target image area is , the image region obtained by inference is , the calculation formula of SSIM is, for example, ,in and They are images and The mean of and They are images and The standard deviation of yes and The covariance of and is a constant used to stabilize the calculation. Can be achieved through calculate, The larger the value, the greater the difference between the inferred image area and the actual target image area.

[0126] In step 30d, the recognition device adjusts the parameters of the crop development period recognition network based on the magnitude of the first restoration error and the second restoration error. If the first restoration error is large, it indicates that the network has problems processing the associated meteorological element data. The recognition device may adjust the parameters of the layers in the network related to meteorological element data processing, such as adjusting the connection weights and bias terms of the neurons in the first meteorological element representation vector mining component. For example, the stochastic gradient descent (SGD) algorithm is used to update the weight w, and the update formula is: ,in is the learning rate, It is an error The gradient of the weight w. By continuously adjusting the weight, the network can make the inference of meteorological element data more accurate, thereby reducing the first reduction error.

[0127] If the second restoration error is large, the recognition device adjusts the network components related to crop image data processing, such as adjusting the parameters of the crop image representation vector mining component. For example, in a convolutional neural network, parameters such as the size, number, or step size of the convolution kernel can be adjusted. Assuming the current convolution kernel size is k×k, if the second restoration error is large, the recognition device may try increasing or decreasing the convolution kernel size, retraining the network, and observing the changes in the second restoration error until it finds a suitable parameter setting that improves the network's reasoning ability for crop image data and reduces the second restoration error.

[0128] During the actual debugging process, the recognition device comprehensively considers the first restoration error and the second restoration error. For example, different weights are assigned to the two. and , and then calculate a comprehensive error The recognition device comprehensively adjusts and optimizes the parameters of the crop development period recognition network, aiming to minimize the overall error E. Through multiple iterative adjustments, the overall error is continuously reduced, enabling the network to achieve better performance when processing both correlated meteorological element data and crop image data, thereby improving the network's accuracy in identifying crop development periods.

[0129] Then, the associated meteorological element shielded data and crop image shielded data are loaded into the crop growth stage recognition network for inference. After the network calculation, the first inference result for the shielded wind speed data and the second inference result for the shielded leaf area are obtained. Then, the recognition device determines the first restoration error and the second restoration error respectively according to the above calculation method. Assuming the first restoration error = 0.3, second restoration error = 0.2. The recognition device assigns weights to the first restoration error and the second restoration error respectively. = 0.6 and = 0.4, and the calculated comprehensive error is E = 0.6 × 0.3 + 0.4 × 0.2 = 0.26.

[0130] Based on this comprehensive error, the recognition device began to debug the crop development period recognition network. First, in the case of large first restoration error, the connection weights of some neurons in the first meteorological element representation vector mining component were adjusted. After several iterative adjustments, the first restoration error was reduced to = 0.2. At the same time, in order to address the second restoration error, the convolution kernel size of the convolution layer in the crop image representation vector mining component was adjusted. After retraining and adjustment, the second restoration error was reduced to = 0.15. The combined error E is recalculated as 0.6 × 0.2 + 0.4 × 0.15 = 0.18. By continuously performing this cycle of masking, inference, error calculation, and network parameter adjustment, the recognition device can gradually optimize the crop growth stage recognition network, making it more accurate and reliable when processing meteorological and image data, ultimately improving the accuracy and efficiency of crop growth stage recognition.

[0131] As an implementation manner, the step 10 of generating reference guidance data based on the associated meteorological element data template includes any one of the following four methods:

[0132] Method 1: if the associated meteorological element data template includes meteorological attribute data, generating reference guidance data based on the meteorological attribute data;

[0133] Method 2: If the associated meteorological element data template includes meteorological attribute data and meteorological record data, the meteorological attribute data and the meteorological record data are combined to generate reference guidance data;

[0134] Method 3: If the associated meteorological element data template includes meteorological attribute data and meteorological annotation data, the meteorological attribute data and the meteorological annotation data are combined to generate reference guidance data;

[0135] Method 4: If the associated meteorological element data template includes meteorological attribute data, meteorological record data and meteorological annotation data, the meteorological attribute data, the meteorological record data and the meteorological annotation data are combined to generate reference guidance data.

[0136] In Method 1, if the associated meteorological element data template includes meteorological attribute data, the recognition device generates reference guidance data based on this meteorological attribute data. Meteorological attribute data is a qualitative or quantitative description of meteorological characteristics, such as weather type (sunny, cloudy, rainy, snowy, etc.), seasonal attributes (spring, summer, autumn, winter), and climate zone attributes (tropical, subtropical, temperate, frigid, etc.). The recognition device directly uses this meteorological attribute data to generate reference guidance data.

[0137] In the second method, meteorological record data is the actual observation record of meteorological elements, such as temperature, humidity, wind speed, precipitation and other specific numerical records.

[0138] Assume that the associated meteorological element data template contains meteorological attribute information (such as weather type, season, etc.) for each day of a week and detailed meteorological record data (such as daily maximum temperature, minimum temperature, average humidity, wind speed, etc.). The recognition device will first organize and analyze the meteorological record data. For different meteorological record parameters, normalization processing may be performed to unify the scale of the data to facilitate subsequent combination with meteorological attribute data. For example, for temperature data T, the normalization formula is used ,in and are the minimum and maximum values ​​of the temperature during the time period, respectively. This way, the temperature data can be mapped to the [0, 1] interval.

[0139] As previously mentioned, meteorological attribute data is processed using methods such as one-hot encoding. The recognition device then combines the processed meteorological attribute data with the meteorological record data. One feasible combination method is to concatenate the encoded meteorological attribute vector with the normalized meteorological record data vector. For example, assuming that the meteorological attribute data is one-hot encoded to obtain an m-dimensional vector A, and the meteorological record data is normalized to obtain a k-dimensional vector B, the combined reference guidance data is an m + k-dimensional vector C = [A; B] (where [;] represents a vector concatenation operation).

[0140] The reference guidance data generated in this way incorporates both meteorological attribute characteristics and specific meteorological observation numerical information. For the meteorological element data augmentation network, this rich information can guide the network to generate more realistic and diverse meteorological element augmentation data. For example, when generating new meteorological data samples, the network can generate various possible meteorological data combinations that match the scenario based on the meteorological attributes in the reference guidance data (such as summer and sunny days) and the temperature and humidity ranges reflected in the meteorological record data.

[0141] In method three, meteorological annotation data is a mark or description of the relationship between meteorological data and crop growth, such as marking whether a certain meteorological condition is favorable, unfavorable or has no obvious effect on crop growth, or marking what growth stage the crop is in under the current meteorological conditions.

[0142] For example, the associated meteorological element data template records meteorological attribute information (such as climate type, season, etc.) over a period of time and the corresponding meteorological annotation data, with the annotation content being "Current meteorological conditions are favorable for wheat growth, and wheat is in the jointing stage." When processing, the recognition device first encodes the meteorological attribute data, such as using one-hot encoding to convert attributes such as climate type and season into vector form. For meteorological annotation data, the recognition device classifies and encodes them according to the specific content of the annotation. If the annotation is about the impact on crop growth, "favorable," "unfavorable," and "no obvious impact" may be encoded as different numerical values ​​or vectors respectively; if the annotation is about the crop growth stage, the different growth stages will also be encoded.

[0143] The recognition device then combines the encoded meteorological attribute data with the meteorological annotation data. For example, the meteorological attribute encoding vector A is concatenated with the meteorological annotation encoding vector B to form reference guidance data C = [A; B]. This combined reference guidance data provides the meteorological element data enhancement network with key information about the relationship between weather and crop growth. Based on this information, the meteorological element data enhancement network can generate meteorological element enhancement data that has different effects on crop growth under different meteorological attributes or corresponds to different growth stages.

[0144] In Method 4, the associated meteorological element data template contains detailed meteorological information for a month, including meteorological attributes (such as weather type, season, and climate zone), meteorological record data (such as daily specific values ​​of temperature, humidity, wind speed, and precipitation), and meteorological annotation data (such as the labeling of a specific crop growth stage and the evaluation of its impact on crop growth). The recognition device processes these three types of data in sequence.

[0145] Meteorological attribute data is converted into vector form using methods such as one-hot encoding. Meteorological record data is normalized to uniform data scale. Meteorological annotation data is categorized and encoded based on the annotation content. For example, meteorological attribute data is encoded as vector A, meteorological record data is normalized to form vector B, and meteorological annotation data is encoded as vector C.

[0146] Finally, the recognition device combines these three vectors to form reference guidance data. One feasible combination is to directly concatenate them sequentially, resulting in a new vector D = [A; B; C]. This generated reference guidance data contains the richest meteorological information, covering both basic meteorological attributes and specific observational values, as well as annotations on the relationship between weather and crop growth.

[0147] As an implementation manner, if the associated meteorological element data template includes meteorological attribute data and meteorological record data, in the second approach, combining the meteorological attribute data and the meteorological record data to generate reference guidance data includes:

[0148] Step S121: disassembling the meteorological record data to obtain multiple meteorological record data sequences;

[0149] Step S122: performing crop impact status identification on the meteorological record data sequence, and determining a target data sequence in each meteorological record data sequence based on the crop impact status identification result;

[0150] Step S123: Combine the meteorological attribute data and the target data sequence to generate reference guidance data.

[0151] In step S121, the meteorological record data is, for example, a set of observation values ​​of multiple meteorological elements recorded in a certain time sequence, which contains rich information on meteorological changes. The recognition device decomposes the data according to the time dimension or different meteorological element categories.

[0152] Assume that the associated meteorological element data template records the meteorological data of a certain area for one consecutive year. The meteorological record data covers daily temperature, humidity, wind speed, and precipitation. The recognition device first classifies and decomposes these meteorological record data according to different meteorological elements. For temperature data, the temperature values ​​of each day of the year are extracted in sequence to form a temperature data sequence with a length of 365. ,in Represents the temperature value of the i-th day. Similarly, for humidity data, generate humidity data sequence , wind speed data series , precipitation data series .

[0153] In step S122 , crop impact status recognition is a process of determining the impact of meteorological record data on crop growth and development. This requires the recognition device to use pre-built models or rules for analysis.

[0154] Taking the temperature data sequence T as an example, the recognition device may use a threshold-based judgment model. For a certain crop, assuming that the temperature range suitable for growth is arrive The recognition device traverses the temperature data sequence T, and for each temperature value , to determine whether arrive between.

[0155] Based on these judgment results, the recognition device determines the target data sequence in each meteorological data sequence. If the temperature data sequence determines that the temperature of certain time periods is favorable for crop growth, then the subsequence consisting of the temperature values ​​corresponding to these time periods is the temperature target data sequence.

[0156] In step S123, the meteorological attribute data is a qualitative description of meteorological characteristics, such as season, weather type (sunny, rainy, etc.). The recognition device first encodes the meteorological attribute data so that it can be effectively combined with the target data sequence.

[0157] Assume that meteorological attribute data includes season information (spring, summer, autumn, winter) and weather type information (sunny, cloudy, rainy, snowy). The recognition device can encode this using one-hot encoding. For the season attribute, if there are four season categories, the one-hot encoding vector for spring is [1, 0, 0, 0], summer is [0, 1, 0, 0], autumn is [0, 0, 1, 0], and winter is [0, 0, 0, 1]. For the weather type attribute, assuming there are four types, the one-hot encoding vector for sunny is [1, 0, 0, 0], cloudy is [0, 1, 0, 0], rainy is [0, 0, 1, 0], and snowy is [0, 0, 0, 1].

[0158] Then, the recognition device combines the encoded meteorological attribute data with the target data sequence. For example, assuming the temperature target data sequence is , the humidity target data sequence is , the wind speed target data sequence is , the precipitation target data sequence is , and the season coding vector is S, and the weather type coding vector is Wt. The recognition device can combine these data in the following way to generate reference guidance data. A simple combination method is to splice them into a large vector in sequence. Assuming that the reference guidance data vector is R, then R=[S;Wt; T target ; H target ; W target ; P target ] (where [;] represents a vector concatenation operation).

[0159] In this way, the generated reference guidance data not only contains the basic attribute characteristics of the meteorology (season, meteorological type), but also incorporates meteorological record data information that has a specific impact on crop growth.

[0160] As an embodiment, the debugging process of the crop development stage recognition network further includes:

[0161] Step 50: Obtain a fine-tuned priori label of the crop image data template, load the crop image data template and the associated meteorological element data template into the debugged crop development period recognition network, perform characterization vector mining on the associated meteorological element data template to obtain a template meteorological element data characterization vector, perform characterization vector mining on the crop image data template to obtain a crop image data template characterization vector, interact the template meteorological element data characterization vector and the crop image data template characterization vector to obtain a template interaction characterization vector, and perform inference on the template interaction characterization vector to obtain a template inference result;

[0162] In step 50, the recognition device first obtains fine-tuned prior labels for the crop image data template. These fine-tuned prior labels are predetermined based on domain knowledge, expert experience, or a large amount of previous experimental data, and are used to guide the network's understanding and recognition of crop image data. For example, in a project targeting wheat developmental stage recognition, the crop image data template included images of wheat at different growth stages. Based on the annotation and research of agricultural experts, key fine-tuned prior labels were identified, such as growth stage labels such as "wheat three-leaf stage," "wheat jointing stage," and "wheat heading stage," as well as labels related to crop appearance characteristics, such as "dark green leaf color" and "thick stems." These labels provide the recognition device with additional information about the crop image data, helping it to more accurately understand and process the image data.

[0163] Next, the recognition device loads the crop image data template and the associated meteorological element data template into the debugged crop growth stage recognition network. Assume that the associated meteorological element data template records meteorological data such as temperature, humidity, and wind speed over a period of time. The recognition device inputs this data into the network, and the various components within the network begin to operate.

[0164] For the associated meteorological element data template, the recognition device mines its representation vector through the first meteorological element representation vector mining component to obtain the template meteorological element data representation vector. Taking a deep learning-based network as an example, the first meteorological element representation vector mining component may be a multi-layer neural network. When processing the associated meteorological element data template, the input layer receives meteorological data, such as temperature T, humidity H, wind speed V, etc., and represents it in vector form. = [T, H, V]. Then, the data is processed by multiple hidden layers in the network. The neurons in the hidden layer extract features from the data through weighted summation and activation function. Assume that the input of neuron i in the hidden layer is , whose output The formula can be Calculate, where is the connection weight, is the bias term, Is an activation function, such as the ReLU function After processing through multiple hidden layers, the final output is the template meteorological element data representation vector ,This vector contains the key feature information of the associated meteorological element data template.

[0165] Simultaneously, the recognition device uses a crop image representation vector mining component to mine representation vectors for the crop image data template, obtaining a representation vector for the crop image data template. Assuming the crop image data template is a series of color images, the crop image representation vector mining component can be a convolutional neural network (CNN). When processing an image, the convolution kernel in the convolution layer performs a convolution operation on the image to extract local features.

[0166] Then, the recognition device characterizes the template meteorological element data vector and crop image data template representation vector Interact to obtain the template interaction representation vector. This interaction can be achieved in many ways, such as splicing operation, splicing two vectors into a new vector v → Interaction =[ v → meteorological ; v → image ] Alternatively, more complex fusion methods, such as attention-based fusion, can be employed. This method dynamically assigns importance to the fused vector by calculating the attention weights of the two vectors. Finally, the recognition device infers the template interaction representation vector to obtain the template inference result. This inference process is accomplished, for example, by the network's output layer (e.g., a fully connected layer).

[0167] Step 60: Determine a fine-tuning error based on the template reasoning result and the fine-tuning priori mark, and debug the crop development stage recognition network again based on the fine-tuning error.

[0168] In step 60, the recognition device determines the fine-tuning error based on the template inference result and the fine-tuning prior label. In step 50, the recognition device infers the template interaction representation vector to obtain the template inference result, and the fine-tuning prior label is a pre-given, highly accurate labeling information about the crop development period. To determine the fine-tuning error, the recognition device can use an appropriate loss function to measure the difference between the template inference result and the fine-tuning prior label. For example, the loss function is the cross-entropy loss function, whose formula is: , where n is the number of categories (n = 4 in this example, i.e., four developmental categories), is the true probability of the i-th class in the fine-tuned prior label (in the example of “flowering period”, = 0, ), is the predicted probability of the i-th class in the template inference result (i.e., the value in the above probability distribution vector). Substituting the specific value into the cross entropy loss function yields: The calculated result is the fine-tuning error in the current situation, which reflects the degree of inconsistency between the template inference result and the fine-tuning prior mark. The larger the error value, the greater the gap between the network's prediction result and the true mark.

[0169] After determining the fine-tuning error, the recognition device re-tunes the crop development stage recognition network based on this error. This tuning process aims to adjust the network parameters so that the network can more accurately identify crop development stages in subsequent processing. The recognition device can use a variety of optimization algorithms to adjust network parameters, such as stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, and Adam. Taking the stochastic gradient descent algorithm as an example, its core idea is to calculate the gradient of the loss function with respect to network parameters (such as weights w and biases b), and then update the parameters in the opposite direction of the gradient to gradually reduce the value of the loss function.

[0170] Assume that the weight matrix of a layer in the network is W, the bias vector is b, and the gradient of the loss function L with respect to the weight W is , the gradient of the bias b is In the stochastic gradient descent algorithm, the update formulas for weights and biases are and ,in is the learning rate, which controls the step size of each parameter update.

[0171] For the crop image representation vector mining component, if fine-tuning errors indicate that the component has problems extracting crop image features, the recognition device may change the pooling layer strategy, such as switching from maximum pooling to average pooling, or adjust the number of filters in the convolutional layer. For example, if the convolutional layer originally has 32 filters, the recognition device may increase the number of filters to 64 to enhance the component's ability to capture crop image features.

[0172] In terms of the feature integration component, if the fine-tuning error indicates that the component is not effective in fusing meteorological and image features, the recognition device can adjust the way the features are fused.

[0173] By repeatedly determining the fine-tuning error and adjusting network parameters, the recognition device gradually optimizes the crop development period recognition network. As debugging progresses, the fine-tuning error decreases, indicating that the difference between the network's predictions and the fine-tuning priors is narrowing, and the network's recognition accuracy for crop development periods is improving.

[0174] Figure 2 A schematic diagram of a hardware entity of an identification device provided by an embodiment of the present invention, such as Figure 2 As shown, the hardware entity of the identification device 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and when the processor 1001 executes the program, the steps in the method of any of the above embodiments are implemented.

Claims

1. A crop growth stage recognition method integrating meteorological and image sequences, characterized in that: Methods include: Acquire crop image sequence data and associated meteorological element sequence data of the crop to be identified; Calling a pre-debugged crop development period recognition network, loading the crop image sequence data and the associated meteorological element sequence data into the pre-debugged crop development period recognition network, and mining to obtain a crop image representation array corresponding to the crop image sequence data and a meteorological element representation array corresponding to the associated meteorological element sequence data; Perform attention fusion on the crop image representation array and the meteorological element representation array to obtain a fused representation array; Based on the fusion representation array, the crop development period classification prediction is performed to obtain the crop development period classification recognition results; The crop development period recognition network was debugged using the following steps: Obtaining a meteorological element data template associated with the crop image data template, generating reference guidance data based on the associated meteorological element data template, loading the reference guidance data into a meteorological element data enhancement network for meteorological element data enhancement processing, and constructing meteorological element enhanced data that matches the crop image data template, wherein the reference guidance data is used to guide the meteorological element data enhancement network in constructing the meteorological element enhanced data; obtaining a crop image data template, generating a first debugging learning template based on two of the crop image data template, the associated meteorological element data template, and the meteorological element enhancement data, and generating a second debugging learning template based on the remaining one of the crop image data template, the associated meteorological element data template, and the meteorological element enhancement data; Based on the first meteorological element representation vector mining component, a first feature vector of the associated meteorological element data template is mined; based on the second meteorological element representation vector mining component, a second feature vector of the meteorological element enhanced data is mined; based on the crop image representation vector mining component, a third feature vector of the crop image data template is mined; if the second debugging learning template is generated based on the meteorological element enhanced data, the first feature vector and the third feature vector are interacted based on the feature integration component to obtain a first feature vector of the first debugging learning template, and the second feature vector is used as the second feature vector of the second debugging learning template; if the second debugging learning template is generated based on the associated meteorological element data template, the second feature vector and the third feature vector are interacted based on the feature integration component to obtain a first feature vector of the first debugging learning template, and the first feature vector is used as the second feature vector of the second debugging learning template; There are multiple first characterization vectors, and the multiple first characterization vectors correspond one-to-one to the multiple second characterization vectors. For each first characterization vector, a first correlation coefficient between the first characterization vector and the corresponding second characterization vector is determined, and a second correlation coefficient between the first characterization vector and each other non-corresponding second characterization vector is determined. A first cumulative correlation coefficient is determined based on the sum of the first correlation coefficient and each second correlation coefficient, and a logarithmic operation is performed on the ratio between the first correlation coefficient and the first cumulative correlation coefficient to determine a first component debugging error. For each second characterization vector, a third correlation coefficient is determined between the second characterization vector and the corresponding first characterization vector, and a fourth correlation coefficient between the second characterization vector and each other non-corresponding first characterization vector is determined. A second cumulative correlation coefficient is determined based on the sum of the third correlation coefficient and each fourth correlation coefficient, and a logarithmic operation is performed on the ratio between the third correlation coefficient and the second cumulative correlation coefficient to determine a second component debugging error. Based on the first component debugging error and the second component debugging error, a metric learning debugging error is determined. Based on the metric learning debugging error, the first meteorological element characterization vector mining component, the second meteorological element characterization vector mining component, the crop image characterization vector mining component, and the feature integration component are collaboratively debugged.

2. The crop growth stage identification method integrating meteorological data and image sequences according to claim 1, wherein: If the second debugging learning template is generated based on the crop image data template, the crop growth period recognition network includes a first meteorological element representation vector mining component and a crop image representation vector mining component, mining the first representation vector of the first debugging learning template and the second representation vector of the second debugging learning template, including: performing representation vector mining on the associated meteorological element data template and the meteorological element enhancement data in sequence based on the first meteorological element representation vector mining component to obtain a first representation vector of the first debugging learning template, or combining the associated meteorological element data template and the meteorological element enhancement data and then performing representation vector mining based on the first meteorological element representation vector mining component to obtain a first representation vector of the first debugging learning template; Performing representation vector mining on the crop image data template based on the crop image representation vector mining component to obtain a second representation vector of the second debugging learning template; Determining the metric learning debugging error based on the first representation vector and the second representation vector includes: If the first representation vector is obtained by sequentially mining the representation vectors of the associated meteorological element data template and the meteorological element enhancement data, then determining a first sub-metric learning debugging error based on the first representation vector and the second representation vector corresponding to the associated meteorological element data template, determining a second sub-metric learning debugging error based on the first representation vector and the second representation vector corresponding to the meteorological element enhancement data, and determining a metric learning debugging error based on the first sub-metric learning debugging error and the second sub-metric learning debugging error; Alternatively, if the first characterization vector is obtained by combining the associated meteorological element data template and the meteorological element enhancement data and performing characterization vector mining based on the first meteorological element characterization vector mining component, the metric learning debugging error is determined based on the first characterization vector and the second characterization vector.

3. The crop growth stage identification method integrating meteorological data and image sequences according to claim 1 or 2, characterized in that: Before loading the first debugging learning template and the second debugging learning template into the crop growth stage recognition network, the crop growth stage recognition method integrating meteorological and image sequences further includes: Shielding the target data item in the associated meteorological element data template or the meteorological element enhanced data to obtain associated meteorological element shielded data, and shielding the target image area in the crop image data template to obtain crop image shielded data; Loading the associated meteorological element shielded data and the crop image shielded data into the crop development stage recognition network, performing reasoning on the shielded target data item to obtain a first reasoning result, and performing reasoning on the shielded target image area to obtain a second reasoning result; determining a first restoration error based on the first inference result and the target data item, and determining a second restoration error based on the second inference result and the target image area; The crop development period recognition network is debugged based on the first restoration error and the second restoration error.

4. The crop growth stage identification method integrating meteorological data and image sequences according to claim 1, wherein: Generate reference guidance data based on the associated meteorological element data template, including: If the associated meteorological element data template includes meteorological attribute data, reference guidance data is generated based on the meteorological attribute data; or; If the associated meteorological element data template includes meteorological attribute data and meteorological record data, the meteorological attribute data and meteorological record data are combined to generate reference guidance data; or; If the associated meteorological element data template includes meteorological attribute data and meteorological annotation data, the meteorological attribute data and meteorological annotation data are combined to generate reference guidance data; or; If the associated meteorological element data template includes meteorological attribute data, meteorological record data, and meteorological annotation data, the meteorological attribute data, meteorological record data, and meteorological annotation data are combined to generate reference guidance data; If the associated meteorological element data template includes meteorological attribute data and meteorological record data, the meteorological attribute data and meteorological record data are combined to generate reference guidance data, including: Disassemble the meteorological record data to obtain multiple meteorological record data sequences; Identify the crop impact status of meteorological record data sequences, and determine the target data sequence in each meteorological record data sequence based on the crop impact status identification results; The meteorological attribute data and target data sequence are combined to generate reference guidance data.

5. The crop growth stage identification method integrating meteorological data and image sequences according to claim 1, wherein: Crop growth stage recognition methods that integrate meteorological and image sequences also include: Obtain a fine-tuned prior label of the crop image data template, load the crop image data template and the associated meteorological element data template into the debugged crop development period recognition network, perform characterization vector mining on the associated meteorological element data template to obtain the template meteorological element data characterization vector, perform characterization vector mining on the crop image data template to obtain the crop image data template characterization vector, interact the template meteorological element data characterization vector and the crop image data template characterization vector to obtain the template interaction characterization vector, perform inference on the template interaction characterization vector to obtain the template inference result; The fine-tuning error is determined based on the template reasoning results and the fine-tuning prior markers, and the crop development period recognition network is debugged again based on the fine-tuning error.

6. An identification device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 5 are implemented.

Images