Crop growth situation prediction method based on YOLO and Transform space-time collaboration

Abstract

The invention relates to the technical field of smart agriculture, and specifically provides a crop growth situation prediction method based on YOLO and Transform space-time collaboration. The method aims at solving the problem that in the prior art, prediction precision is insufficient due to splitting of spatial features and time sequence information, and crop phenotypic features (plant height, leaf area index, canopy coverage and the like) in a farmland video stream are extracted in real time through an improved YOLO model; constructing a spatial-temporal feature encoder, and modeling a coupling relationship between historical time sequence environmental data (temperature, illumination and soil moisture content) and phenotypic features through a Transform network; designing a feature collaborative fusion module, and integrating spatial visual features and time sequence environment features by adopting an adaptive weight distribution mechanism; and establishing a growth situation prediction model based on space-time cooperation characteristics, and outputting key parameter probability distribution of a future growth stage.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of agricultural information technology, and specifically relates to a crop growth trend prediction method based on YOLO (You Only Look Once) and Transformer spatiotemporal collaboration. BACKGROUND

[0002] Crop growth trend prediction is a key link in agricultural production management, and its accuracy directly affects the scientificity and rationality of agricultural production decisions, such as irrigation, fertilization, and pest control. Traditional crop growth trend prediction methods mainly rely on experience judgment, statistical analysis, or single machine learning models, which have many limitations.

[0003] Experience judgment is greatly influenced by human factors, is highly subjective, has low precision, and is difficult to adapt to large-scale and fine-grained agricultural production needs. Statistical analysis methods can to some extent mine the rules in the data, but for complex crop growth processes, they are difficult to capture the nonlinear relationship between growth trends and multiple factors.

[0004] Single machine learning models also have shortcomings when processing data. For example, some image-based prediction methods can only extract spatial features of crops, ignoring dynamic changes in the time dimension; while some time series-based prediction methods are difficult to fully utilize the rich spatial detail information in images. In addition, crop growth is influenced by environmental factors (such as temperature, humidity, and light) and its own growth rules, and existing methods have obvious defects in fusing these multi-source heterogeneous data and building spatiotemporal collaborative prediction models, resulting in low prediction accuracy and reliability.

[0005] With the development of computer vision and deep learning technology, the YOLO model shows high efficiency and accuracy in target detection and feature extraction, and can quickly extract rich spatial features from images; the Transformer model has significant advantages in processing sequence data and capturing long-distance dependencies with its self-attention mechanism, and is suitable for mining dynamic rules in the time dimension. However, there is currently no effective combination of YOLO and Transformer to achieve spatiotemporal collaborative prediction of crop growth trends. Therefore, there is an urgent need for a crop growth trend prediction method that can fully integrate spatiotemporal features and improve prediction accuracy. SUMMARY

[0006] The present application provides a crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration, which comprises the following steps:

[0007] 1. Data collection and preprocessing

[0008] Periodically acquire multi-view image data of crops at different growth stages using image acquisition devices such as drones, field cameras, etc., and simultaneously collect corresponding environmental data (including but not limited to temperature, humidity, light intensity, soil nutrient content, etc.) and timestamp information through sensors.

[0009] Perform preprocessing operations such as denoising, enhancement, etc. on the collected image data, unify the image size, and normalize the environmental data to ensure data consistency and usability.

[0010] 2. Spatial feature extraction of YOLO model

[0011] Construct and train a YOLO target detection model, input the preprocessed crop images into the trained YOLO model, and make it recognize and extract the key spatial features of the crops in the images, including the morphology of the crops (such as plant height, leaf shape and number, stem thickness, etc.), signs of diseases and pests (disease spot characteristics, pest species and quantity, etc.), and relative position relationships between crops, etc. Output feature vectors containing the above spatial features.

[0012] 3. Spatio-temporal data fusion and Transformer input preparation

[0013] Fuse the spatial feature vectors extracted by the YOLO model with the corresponding time series environmental data to form a dataset containing spatio-temporal dimensions. For time series data, perform sliding window processing, and treat the spatio-temporal data in each window as a sample unit.

[0014] Encode each sample unit and convert it into a format suitable for Transformer model input, such as converting it into a fixed-length sequence vector, where each position encoding contains the spatial features and environmental data information at that time.

[0015] 4. Training of Transformer model and spatio-temporal feature learning

[0016] Build a prediction model based on the Transformer architecture, which includes components such as multi-head attention mechanism and feedforward neural network. Input the encoded spatio-temporal sample data into the Transformer model for training.

[0017] During the training process, the Transformer model automatically learns the complex dependencies between spatial features at different time points and between spatial features and environmental factors in the crop growth process through the multi-head attention mechanism, captures the spatio-temporal dynamic patterns of crop growth, and optimizes the model parameters to minimize the loss function (such as mean square error loss function) between the prediction results and the real growth trend data.

[0018] 5. Crop growth trend prediction

[0019] The pre-processed and encoded spatio-temporal data to be predicted is input into the trained Transformer model, and the model predicts the growth trend of the crop at a specific future time point or time period based on the learned spatio-temporal feature patterns, including but not limited to plant height growth prediction, yield estimation, disease and pest occurrence probability prediction, etc.

[0020] The prediction results are post-processed, such as scale conversion, threshold judgment, etc. according to the actual agricultural needs, and the intuitive and practical crop growth trend prediction report is output, which provides scientific basis for agricultural production decision-making.

[0021] In the above data collection and preprocessing steps, the setting parameters of the image collection device (such as shooting resolution, shooting angle, shooting interval time, etc.) are dynamically adjusted according to the type of crop, planting area, and growth cycle characteristics to ensure that the collected image data can fully and accurately reflect the growth state of the crop; an adaptive denoising algorithm is used to denoise the image data, which can automatically adjust the denoising strength according to the local features of the image, and at the same time, the details of the image are preserved to the greatest extent; histogram equalization and other enhancement algorithms are used to enhance the contrast of the image and highlight the features of the crop; for the normalization of environmental data, a normalization method based on the statistical characteristics of the data set is used, such as mapping the value of each environmental variable to the interval [0, 1] or [-1, 1], the specific calculation formula is: , and are the minimum and maximum values of the environmental variable in the data set, respectively.

[0022] In the spatial feature extraction step of the YOLO model, the network structure of the YOLO model is customized and improved according to the needs of crop target detection, such as adding convolution layers or feature fusion modules sensitive to crop features in the backbone network to improve the extraction ability of key crop features; when training the YOLO model, a multi-scale training strategy is adopted, that is, training on images of different resolutions, so that the model can adapt to the presentation of crop targets of different sizes in the image, improve the robustness and generalization ability of the model; using transfer learning technology, fine-tuning the YOLO model pre-trained on a large-scale general image dataset for the crop image dataset, accelerating the model convergence process, and reducing the number of samples and training time required for training.

[0023] In the spatio-temporal data fusion and Transformer input preparation step, an attention mechanism-based fusion method is used to fuse spatial feature vectors and time series environmental data, different weights are assigned to different types of data to highlight the data features that have a greater impact on crop growth trends; when processing time series data with a sliding window, the window size and step are dynamically adjusted according to the time correlation of crop growth and the time span of the prediction, for example, for stages with rapid growth changes, a smaller window size and step are used to capture more detailed time change information; for relatively stable growth stages, the window size and step are appropriately increased to reduce the amount of calculation; when encoding the sample unit, in addition to converting spatio-temporal data into sequence vectors, position encoding information is also introduced to identify the position of each data point in the time series, helping the Transformer model better understand the time order relationship of the data. The position encoding is generated by combining sine and cosine functions, and the formula is:

[0024]

[0025]

[0026] wherein is the position index, is the encoding dimension index, is the dimension of the model.

[0027] In the training and spatio-temporal feature learning steps of the Transformer model, the attention outputs of different heads in the multi-head attention mechanism of the Transformer model are weighted and fused, and the weights are learned through training according to the importance of each head in capturing different types of spatio-temporal dependencies, so as to optimize the learning effect of the model on complex spatio-temporal features; In the feedforward neural network part, residual connection and layer normalization technology are used to alleviate the problems of gradient disappearance and gradient explosion in deep neural networks, accelerate the training convergence process of the model, and improve the training stability and learning ability of the model; In the training process, the early stopping method is used to prevent model overfitting, that is, when the loss function value of the model on the validation set no longer decreases for consecutive multiple training periods, the training is stopped and the current optimal model parameters are saved; At the same time, learning rate adjustment strategies are combined, such as using a larger learning rate in the early stage of training to quickly converge, and gradually reducing the learning rate as the training progresses to fine-tune the model parameters and improve the generalization performance of the model.

[0028] In the crop growth trend prediction step, for yield estimation, the prediction results output by the Transformer model are combined with historical yield data and actual planting area information, and further calibrated using empirical formulas or machine learning regression models to obtain more accurate crop yield prediction values; For disease and pest occurrence probability prediction, the prediction results of the Transformer model are compared with the threshold value of disease and pest occurrence, and when the prediction value exceeds the threshold value, the system automatically issues a disease and pest warning information, and combines the early signs of disease and pest detected by the YOLO model to provide preliminary judgment of disease and pest types and possible impact range; The prediction report is presented in a visual manner, including but not limited to generating crop growth trend change curve, prediction result heat map, risk level identification map, etc., to facilitate agricultural producers to intuitively understand the growth status and future development trend of crops, and to assist them in making scientific and reasonable agricultural production management decisions, such as timing and intensity arrangement of irrigation, fertilization, disease and pest control measures, etc.

[0029] In addition, the present application also provides a system for implementing the above-mentioned crop growth trend prediction method based on YOLO and Transformer spatio-temporal collaboration, which comprises:

[0030] 1. Data acquisition module: used for deploying image acquisition equipment and sensors to collect crop image data, environmental data and corresponding timestamp information according to the set period and parameters.

[0031] 2. Data preprocessing module: performs denoising, enhancement, normalization and other preprocessing operations on the collected image data and environmental data, unifies the data format and scale, and prepares for subsequent model processing.

[0032] 3. YOLO spatial feature extraction module: build, train and run the YOLO target detection model to extract key spatial features from the pre-processed crop images and output feature vectors.

[0033] 4. Spatio-temporal data fusion and encoding module: fuse the spatial feature vectors extracted by YOLO with environmental data, and convert spatio-temporal data into sequence vectors suitable for Transformer model input according to specific sliding window and encoding method.

[0034] 5. Transformer model training and prediction module: build and train the Transformer model to learn the spatio-temporal feature patterns of crop growth, and use the trained model to predict the growth trend of input spatio-temporal data, output the prediction results.

[0035] 6. Result post-processing and report generation module: post-process the prediction results of the Transformer model, convert and analyze the data according to the actual needs of agriculture, generate visual crop growth trend prediction reports, and provide early warning information and decision-making suggestions.

[0036] In this system, the image acquisition devices and sensors in the data acquisition module transmit the collected data in real time to the data preprocessing module through wireless communication technology (such as Wi-Fi, 4G / 5G, etc.), ensuring the timeliness and integrity of the data; the data preprocessing module, YOLO spatial feature extraction module, spatio-temporal data fusion and encoding module, Transformer model training and prediction module, and result post-processing and report generation module are all based on high-performance computing devices (such as GPU clusters, cloud computing platforms, etc.) to meet the efficiency requirements of large-scale data processing and complex model operations; the system also includes a user interaction interface for agricultural producers to input prediction task parameters (such as prediction time span, crop variety, etc.), view prediction reports and early warning information, and adjust and optimize the system's setting parameters (such as image acquisition frequency, model training parameters, etc.) according to actual conditions.

[0037] The present application has the following advantages:

[0038] 1. Fully integrated spatio-temporal features: the present application fuses the spatial features extracted by the YOLO model with time series environmental data through the Transformer model, realizing the collaborative use of spatio-temporal features in crop growth, and being able to more comprehensively capture the dynamic rules of crop growth, improving the accuracy of prediction.

[0039] 2. Effective utilization of multi-source data: The method comprehensively utilizes crop image data and environmental data and other multi-source information, fully plays the role of different types of data in prediction through reasonable preprocessing and fusion methods, and solves the problem of insufficient data utilization in traditional methods.

[0040] 3. High prediction accuracy: The YOLO model can efficiently and accurately extract the spatial features of crops, and the Transformer model can effectively learn the spatio-temporal correlation through the self-attention mechanism, and the combination of the two makes the prediction accuracy of the method significantly higher than that of traditional methods and single models.

[0041] 4. Strong generalization ability: Through multi-scale training strategies, transfer learning techniques, and reasonable model optimization methods, the robustness and generalization ability of the model are improved, which can adapt to the growth trend prediction of different crop types and different planting environments.

[0042] 5. High application value: It can provide scientific and accurate basis for agricultural production decision-making, help agricultural producers to reasonably arrange irrigation, fertilization, pest control and other measures, improve agricultural production efficiency, reduce production cost, and has strong practical application value and popularization prospect. DETAILED DESCRIPTION

[0043] This embodiment takes the wheat growth trend prediction as an example to illustrate the specific implementation process of the crop growth trend prediction method based on YOLO and Transformer spatio-temporal collaboration.

[0044] (I) Data collection and preprocessing

[0045] Data collection: Select a piece of wheat planting field as the test area, deploy unmanned aerial vehicles and field cameras as image collection equipment, the unmanned aerial vehicles take pictures once a week to obtain the overall image of the wheat field, the field cameras take pictures of the local close-up image of the wheat every 2 days, the shooting resolution is set to 1920×1080, and the shooting angle is adjusted according to the growth height of the wheat to ensure that the leaves, stems and other features of the wheat can be clearly photographed. At the same time, sensors are uniformly arranged in the test area to collect temperature, humidity, light intensity, soil nitrogen, phosphorus and potassium content and other environmental data, the collection frequency is once an hour, and the corresponding timestamp information is recorded.

[0046] Image preprocessing: adopt an adaptive denoising algorithm to denoise the collected wheat images, the algorithm automatically adjusts the denoising parameters according to the noise intensity of different regions in the image, removes the noise while retaining the details such as the texture of the wheat leaves, and uses a histogram equalization algorithm to enhance the contrast of the image, making the morphological features of the wheat more prominent; all images are uniformly adjusted to 640×640 size.

[0047] Environment data preprocessing: The collected temperature, humidity and other environmental data are normalized to map the value of each environmental variable to the interval [0, 1]. The formula is: , and are the minimum and maximum values of the environmental variable in the dataset, respectively.

[0048] (B) Spatial feature extraction of YOLO model

[0049] Model construction and training: YOLOv5 is selected as the base model. According to the requirements of wheat target detection, two convolution layers sensitive to wheat features are added to the backbone network to improve the extraction ability of wheat height, leaf number and other features. A large number of image datasets of different growth stages of wheat are collected, and the key features such as wheat height, leaf shape and number, disease and pest signs are labeled. The dataset is divided into training set and validation set according to the ratio of 8:2. Using transfer learning technology, the YOLOv5 model pre-trained on the COCO general image dataset is fine-tuned using the wheat image training set. A multi-scale training strategy is adopted to train alternately on images with resolutions of 416x416, 640x640 and 800x800. The training rounds are set to 100, and the initial learning rate is set to 0.01, which gradually decreases with training.

[0050] Spatial feature extraction: The preprocessed wheat images are input into the trained YOLO model, which outputs a spatial feature vector containing information such as wheat height, leaf number, and presence of disease and pests. The vector dimension is 512.

[0051] (Three) Spatio-temporal data fusion and Transformer input preparation

[0052] Data fusion: An attention mechanism-based fusion method is used to fuse the 512-dimensional spatial feature vector extracted by the YOLO model with the corresponding time series environmental data (including temperature, humidity and other 6 environmental variables). Through training, the weights of different data types are learned, and features such as wheat leaf number and temperature that have a greater impact on growth trend are given higher weights.

[0053] Sliding window processing: According to the growth characteristics of wheat, a sliding window with a size of 7 days and a step of 1 day is used to process time series data during the rapid growth stages such as the return green period and the jointing period. In the relatively stable growth stage such as the filling period, a sliding window with a size of 14 days and a step of 2 days is used. The spatio-temporal data in each window is taken as a sample unit.

[0054] Sample encoding: Each sample unit is converted into a sequence vector with a length of 128, while introducing position encoding information. The position encoding is generated by a combination of sine and cosine functions, with the formula:

[0055]

[0056]

[0057] wherein is the position index, is the encoding dimension index, is the dimension of the model, set to 128.

[0058] (4) Training of the Transformer model and learning of spatio-temporal features

[0059] Model building: A prediction model based on the Transformer architecture is built, containing 6 encoder layers, each composed of a multi-head attention mechanism (with 8 heads) and a feedforward neural network (with a hidden layer dimension of 512). In the feedforward neural network part, residual connection and layer normalization techniques are used.

[0060] Model training: The encoded spatio-temporal sample data is divided into training and validation sets in a ratio of 7:3, and input into the Transformer model for training. The loss function is set as the mean square error loss function, and the Adam optimizer is used with an initial learning rate of 0.001. During training, the early stopping method is used, and when the validation set loss does not decrease for 5 consecutive rounds, training is stopped and the optimal model parameters are saved; at the same time, the learning rate adjustment strategy is used, and the learning rate is halved every 20 rounds.

[0061] (5) Prediction of crop growth trend

[0062] Prediction process: The preprocessed and encoded spatio-temporal data before the prediction period is input into the trained Transformer model, and the model predicts the growth of wheat plant height, yield estimation, and the probability of disease and pest occurrence in the next 15 days.

[0063] Result post-processing:

[0064] Plant height growth prediction: The plant height prediction value output by the model is scaled and restored to the actual plant height unit (centimeters).

[0065] Yield estimation: The yield prediction result output by the Transformer model is combined with the historical wheat yield data of the region in the past 5 years and the actual planting area, and a linear regression model is used for calibration to obtain the final yield prediction value.

[0066] Disease and pest prediction: Set the disease and pest occurrence probability threshold to 0.3, when the model predicts the disease and pest occurrence probability exceeds the threshold, the system automatically sends a warning message, and combines the early signs such as wheat leaf spot detected by the YOLO model to determine the type of possible disease and pest as powdery mildew, and the influence range is about 15% of the test area.

[0067] Report generation: Generate a visual prediction report, including wheat plant height growth curve, yield prediction column chart, disease and pest risk level heat map, etc., to intuitively show the future growth trend of wheat.

[0068] Through the method of the embodiment, the growth trend of wheat is predicted, the plant height prediction error is within 5%, the yield prediction error is within 8%, and the occurrence of disease and pest can be accurately warned, which provides effective decision support for wheat planting management.

[0069] Although the embodiments of the present application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration, characterized in that, The method comprises the following steps: 1) Data collection and preprocessing: i. Periodically acquire multi-view image data of crops at different growth stages using image acquisition devices (such as drones, field cameras, etc.), and simultaneously collect corresponding environmental data (including but not limited to temperature, humidity, light intensity, soil nutrient content, etc.) and timestamp information through sensors; ii. Perform preprocessing operations such as denoising and enhancement on the collected image data, unify the image size, and normalize the environmental data to ensure data consistency and usability; 2) Spatial feature extraction of YOLO model: i. Build and train a YOLO target detection model, input the preprocessed crop images into the trained YOLO model, and make it recognize and extract the key spatial features of the crops in the images, including the morphology of the crops (such as plant height, leaf shape and number, stem thickness, etc.), signs of diseases and pests (disease spot characteristics, pest species and quantity, etc.), and the relative position relationship between crops, etc., output a feature vector containing the above spatial features; 3) Spatio-temporal data fusion and Transformer input preparation: i. Fuse the spatial feature vectors extracted by the YOLO model with the corresponding time series environmental data to form a data set containing spatio-temporal dimensions. For time series data, perform sliding window processing, and take the spatio-temporal data in each window as a sample unit; ii. Encode each sample unit and convert it into a format suitable for Transformer model input, such as converting it into a fixed-length sequence vector, where each position encoding contains the spatial features and environmental data information at that time; 4) Training of Transformer model and spatio-temporal feature learning: i. Build a prediction model based on the Transformer architecture, which includes components such as multi-head attention mechanism and feedforward neural network. Input the encoded spatio-temporal sample data into the Transformer model for training; ii. During training, the Transformer model automatically learns the complex dependencies between spatial features at different time points and between spatial features and environmental factors in the growth process of crops through the multi-head attention mechanism, captures the spatio-temporal dynamic patterns of crop growth, and optimizes the model parameters to minimize the loss function (such as mean square error loss function) between the predicted results and the true growth trend data; 5) Crop growth trend prediction: i. Input the preprocessed and encoded spatio-temporal data to be predicted into the trained Transformer model, and the model predicts the growth trend of crops at a specific future time point or time period based on the learned spatio-temporal feature patterns, including but not limited to plant height growth prediction, yield estimation, disease and pest occurrence probability prediction, etc.; ii. Post-process the prediction results, such as performing scale conversion, threshold judgment, etc. according to actual agricultural needs, and output intuitive and practically valuable crop growth trend prediction reports to provide scientific basis for agricultural production decision-making.

2. The crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to claim 1, characterized in that, In the data collection and preprocessing step: 1) The setting parameters of the image collection device (such as shooting resolution, shooting angle, shooting interval time, etc.) are dynamically adjusted according to the type of crops, planting area, and growth cycle characteristics to ensure that the collected image data can fully and accurately reflect the growth status of crops; 2) An adaptive denoising algorithm is used to denoise the image data, which can automatically adjust the denoising strength according to the local features of the image, and retain the details of the image to the greatest extent while removing noise; histogram equalization and other enhancement algorithms are used to enhance the contrast of the image and highlight the features of the crops; 3) For the normalization of environmental data, a normalization method based on the statistical characteristics of the dataset is adopted, such as mapping the value of each environmental variable to the interval of [0, 1] or [-1, 1], and the specific calculation formula is: . Wherein is the original environmental data value, and are the minimum and maximum values of the environmental variable in the dataset, respectively.

3. The crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to claim 1, characterized in that, In the spatial feature extraction step of the YOLO model: 1) The network structure of the YOLO model is customized and improved according to the needs of crop target detection, such as adding convolution layers or feature fusion modules sensitive to crop features in the backbone network to improve the extraction ability of key crop features; 2) When training the YOLO model, a multi-scale training strategy is adopted, that is, training on images of different resolutions, so that the model can adapt to the presentation of crops of different sizes in the image, and improve the robustness and generalization ability of the model; Using transfer learning technology, the YOLO model pre-trained on a large-scale general image dataset is fine-tuned for the crop image dataset, which speeds up the convergence process of the model and reduces the number of samples and training time required.

4. The crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to claim 1, characterized in that, In the spatio-temporal data fusion and Transformer input preparation step: 1) An attention mechanism-based fusion method is used to fuse spatial feature vectors and time series environmental data, and different weights are assigned to different types of data to highlight the data features that have a greater impact on crop growth; 2) When processing time series data in a sliding window, the window size and step are dynamically adjusted according to the time correlation of crop growth and the time span of the prediction, for example, for stages with rapid growth changes, a smaller window size and step are used to capture more detailed time change information; for relatively stable growth stages, the window size and step are appropriately increased to reduce the amount of calculation; 3) When encoding the sample unit, in addition to converting spatio-temporal data into sequence vectors, position encoding information is also introduced to identify the position of each data point in the time series, helping the Transformer model better understand the time sequence relationship of the data, and the position encoding is generated using a combination of sine and cosine functions, with the formula: ； ； wherein is a position index, is an encoding dimension index, is a dimension of the model.

5. The crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to claim 1, characterized in that, In the training and spatio-temporal feature learning step of the Transformer model: 1) In the multi-head attention mechanism of the Transformer model, the attention outputs of different heads are weighted and fused, and the weights are learned through training according to the importance of each head in capturing different types of spatio-temporal dependencies to optimize the learning effect of the model on complex spatio-temporal features; 2) In the feedforward neural network part, residual connection and layer normalization techniques are used to alleviate the gradient vanishing and gradient exploding problems in deep neural networks, accelerate the training convergence process of the model, and improve the training stability and learning ability of the model; 3) In the training process, the early stopping method is used to prevent model overfitting, that is, when the loss function value of the model on the validation set does not decrease for multiple training periods in a row, training is stopped and the current optimal model parameters are saved; At the same time, combined with the learning rate adjustment strategy, such as using a larger learning rate in the early stage of training to quickly converge, and gradually reducing the learning rate to fine-tune the model parameters as the training progresses, to improve the generalization performance of the model.

6. The crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to claim 1, characterized in that, In the crop growth trend prediction step: 1) For yield estimation, the prediction results output by the Transformer model are combined with historical yield data and actual planting area information, and further calibrated using empirical formulas or machine learning regression models to obtain more accurate crop yield prediction values; 2) For disease and pest occurrence probability prediction, compare the prediction results of the Transformer model with the threshold value of disease and pest occurrence. When the prediction value exceeds the threshold value, the system automatically issues a disease and pest warning message, and combines the early signs of disease and pests detected by the YOLO model to provide preliminary judgments of disease and pest types and possible impact ranges; The prediction report is presented in a visual manner, including but not limited to generating crop growth trend curves, prediction result heat maps, risk level identification maps, etc., to facilitate agricultural producers to intuitively understand the growth conditions and future trends of crops, and to assist them in making scientific and reasonable agricultural production management decisions, such as timing and intensity arrangements for irrigation, fertilization, disease and pest control measures, etc.

7. A system for implementing the crop growth trend prediction method based on YOLO and Transformer spatiotemporal collaboration according to any one of claims 1-6, characterized in that, It includes: 1) Data acquisition module, used to deploy image acquisition equipment and sensors to collect crop image data, environmental data and corresponding timestamp information according to the set period and parameters; 2) Data preprocessing module, which performs denoising, enhancement, normalization and other preprocessing operations on the collected image data and environmental data, unifies the data format and scale, and prepares for subsequent model processing; 3) YOLO spatial feature extraction module, which constructs, trains and runs the YOLO target detection model, extracts key spatial features from preprocessed crop images, and outputs feature vectors; 4) Spatio-temporal data fusion and encoding module, which fuses the spatial feature vectors extracted by YOLO with environmental data, and converts spatio-temporal data into sequence vectors suitable for input into the Transformer model according to specific sliding windows and encoding methods; 5) Transformer model training and prediction module, which builds and trains the Transformer model, learns the spatio-temporal feature patterns of crop growth, and uses the trained model to predict the growth trend of input spatio-temporal data, outputting the prediction results; The result post-processing and report generation module post-processes the prediction results of the Transformer model, converts and analyzes the data according to the actual needs of agriculture, generates a visual crop growth trend prediction report, and provides early warning information and decision-making suggestions.

8. The system for implementing the crop growth trend prediction method based on YOLO and Transformer spatio-temporal collaboration according to claim 7, characterized in that: 1) The image acquisition devices and sensors in the data acquisition module transmit the collected data in real time to the data preprocessing module through wireless communication technology (such as Wi-Fi, 4G / 5G, etc.), ensuring the timeliness and integrity of the data; 2) The data preprocessing module, YOLO spatial feature extraction module, spatio-temporal data fusion and encoding module, Transformer model training and prediction module, and result post-processing and report generation module are all based on high-performance computing devices (such as GPU clusters, cloud computing platforms, etc.) to meet the efficiency requirements of large-scale data processing and complex model operations; The system also includes a user interaction interface for agricultural producers to input prediction task parameters (such as prediction time span, crop variety, etc.), view prediction reports and early warning information, and adjust and optimize the system's setting parameters (such as image acquisition frequency, model training parameters, etc.) according to actual conditions.

Images