A large model training method for generating corn growth guidance

Abstract

A large model training method for generating corn growth guidance. It belongs to the field of agricultural intelligent monitoring and prediction technology. It solves the problem that when the large model method is applied to the corn growth management scene, it depends on remote sensing images and meteorological data and other multi-source heterogeneous information. Although it can realize preliminary analysis and prediction based on the growth state of crops, there are still limitations in the accuracy, explainability and actual adaptability of decision-making suggestions. When constructing the data set, in addition to the fusion of remote sensing data and multi-dimensional meteorological information, expert suggestions are also added to construct more diverse fusion features. Due to the similarity between expert suggestions and farmland management suggestions, the correspondence between features and labels is more accurate, which can make the corn growth guidance based on the large model architecture realize the technical leap from "seeing" to "understanding" to "guiding", and provide scientific basis and decision support for the precision management of corn production.

Description

Technical Field

[0001] This invention belongs to the field of agricultural intelligent monitoring and prediction technology, specifically relating to a large model training method for generating maize growth guidance. Background Technology

[0002] As one of the world's most important food and feed crops, maize holds a strategic position in ensuring food security and agricultural economic development. With the rapid development of smart agriculture, how to achieve precise monitoring, scientific management, and intelligent decision-making in the maize growth process has become a key issue in improving yield per unit area, optimizing resource utilization efficiency, and addressing the challenges of climate change.

[0003] Traditional maize growth management relies primarily on manual field surveys and experience-based judgments, which suffers from poor timeliness, strong subjectivity, and limited coverage, making it difficult to meet the demands of modern agriculture's large-scale, precise, and intelligent development. In recent years, emerging technologies such as remote sensing, the Internet of Things (IoT), and artificial intelligence (AI) have been gradually applied in agriculture, providing new means for the dynamic perception of crop growth status. Among these, unmanned aerial vehicle (UAV) remote sensing, with its advantages of high spatiotemporal resolution, flexible deployment, and low cost, can quickly acquire multispectral, thermal infrared, and visible light images at the field scale, effectively reflecting key growth parameters such as maize canopy structure, leaf area index, vegetation cover, and stress conditions. Meanwhile, meteorological data (such as temperature, precipitation, light intensity, and humidity), as core environmental factors affecting maize growth and development, play a decisive role in crop phenology, yield formation, and disaster risk.

[0004] In recent years, large model technology has made groundbreaking progress in fields such as computer vision and natural language processing, demonstrating powerful multimodal fusion capabilities, contextual understanding capabilities, and generalized reasoning capabilities. Introducing it into the maize growth management scenario holds promise for achieving unified modeling and semantic understanding of multi-source heterogeneous agricultural data, thereby constructing a more accurate, interpretable, and adaptive intelligent guidance system for crop growth. However, current large model methods, when applied to maize growth management scenarios, typically rely on multi-source heterogeneous information such as remote sensing imagery and meteorological data. While they can achieve preliminary analysis and prediction based on crop growth status, they still have limitations in terms of the accuracy, interpretability, and practical adaptability of decision-making recommendations. Summary of the Invention

[0005] To address the limitations of current large-scale model methods applied to maize growth management scenarios, which typically rely on heterogeneous information from multiple sources such as remote sensing imagery and meteorological data, although they can achieve preliminary analysis and prediction based on crop growth status, they still have limitations in terms of the accuracy, interpretability, and practical adaptability of decision recommendations, this invention provides a large-scale model training method for generating maize growth guidance.

[0006] The method specifically involves S1: acquiring remote sensing image data of maize growth at different stages, and extracting remote sensing image features at different stages through a CNN network with an adaptive instance normalization module.

[0007] S2. Obtain meteorological data on corn growth at different stages, and extract meteorological characteristics of different stages through a bidirectional LSTM network;

[0008] S3. Based on remote sensing image data and meteorological data of maize growth at different stages, construct corresponding expert recommendations for different stages, and use a pre-trained Chinese BERT model to encode the expert recommendations for different stages to obtain the features of expert recommendations for different stages.

[0009] S4. Combine the remote sensing image features, meteorological features, and expert recommendation features of each stage, and provide farmland management recommendations for each stage based on the corn growth status; thereby generating a dataset with remote sensing image features, meteorological features, and expert recommendation features as samples and farmland management recommendations as labels.

[0010] S5. Construct a large-scale model for corn growth guidance, fine-tune the large-scale model using a dataset, and output corn growth guidance suggestions through the fine-tuned large-scale model.

[0011] Furthermore, the different stages refer to different stages of corn growth, including the seedling stage, jointing stage, tasseling stage, grain-filling stage, and maturity stage.

[0012] Furthermore, the method for acquiring remote sensing image data of maize growth at different stages is as follows: remote sensing image data is acquired using a drone equipped with a multispectral camera and an RGB camera; the acquired image data is divided; and the remote sensing index of all images is calculated. , and Based on the aforementioned remote sensing indices, three-channel data is generated, and data augmentation is performed on the three-channel data.

[0013] Furthermore, the CNN network, from input to output, sequentially includes convolutional layers, an adaptive instance normalization module, a four-layer residual module group, a global average pooling layer, and a fully connected layer. The adaptive instance normalization module calculates the mean and standard deviation for each channel of each sample, and utilizes a learnable scaling factor. and offset coefficient The normalized features are recalibrated.

[0014] Furthermore, the meteorological data for corn growth consists of meteorological data for N consecutive days, including meteorological indicators such as temperature, humidity, and precipitation. The meteorological data for N consecutive days is normalized to obtain the daily average meteorological indicators.

[0015] Furthermore, experts recommend that the corn be prepared in advance according to different growth stages and management types, including water management, nutrient management, pest and disease control, and disaster prevention.

[0016] Furthermore, farmland management recommendations for each stage are provided by constructing a knowledge base, which is obtained by integrating various existing technical documents. Through manual analysis, farmland management recommendations that are compatible with remote sensing image data and meteorological data for each stage are selected from the knowledge base.

[0017] Furthermore, the large model includes a weighted feature fusion part and an encoding and decoding part. The weighted feature fusion part weights and fuses remote sensing image features, meteorological features, and expert suggestion features to obtain fused features. The encoding and decoding part adopts the Transformer architecture, and obtains farmland management suggestions corresponding to the fused features by sequentially encoding and decoding the fused features.

[0018] Furthermore, the fine-tuning training of the large model is specifically as follows: a pre-trained Chinese BERT model is used as the initialization model of the decoder. A low-rank adapter is injected into the self-attention layer of the BERT model through LoRA. Only the adapter parameters, the weighted feature fusion part parameters, and the output layer parameters of the encoding and decoding parts are updated, while the original weights of the BERT model are kept frozen, so as to reduce the number of fine-tuning parameters.

[0019] During the training phase, remote sensing image features, meteorological features, and expert suggestion features were used as conditional inputs, and the corresponding farmland management suggestions were used as target sequences. Paired samples of input multimodal features and output agricultural suggestions were constructed. The cross-entropy loss function was used to constrain the difference between the generated sequence and the real text. AdamW was selected as the optimizer, and the dataset was divided into training / validation / test ratios of 7:2:1.

[0020] The beneficial effects of the method described in this invention are as follows: The large-scale model training method described in this invention aims to enable the trained large-scale model to provide more accurate guidance for corn growth at different stages. When constructing the dataset, in addition to the fusion of remote sensing data and multidimensional meteorological information, expert suggestions are added to construct more diverse fusion features. Due to the similarity between expert suggestions and farmland management suggestions, the correspondence between features and labels is more accurate. This allows corn growth guidance based on the large-scale model architecture to achieve a technological leap from "visible" to "understandable" and then to "able to guide," providing a scientific basis and decision support for the precision management of corn production. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method described in an embodiment of the present invention;

[0022] Figure 2The above refers to the meteorological humidity data collected in this embodiment of the invention.

[0023] Figure 3 The meteorological temperature data collected in this embodiment of the invention;

[0024] Figure 4 This is a diagram of the CNN network structure in an embodiment of the present invention;

[0025] Figure 5 This is a diagram of the bidirectional LSTM network structure in an embodiment of the present invention;

[0026] Figure 6 This is a structural diagram of the large model in an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0028] Example 1

[0029] This embodiment provides a method for training a large model to generate corn growth guidance. For example... Figure 1 As shown, it includes the following steps:

[0030] S1. Acquire remote sensing image data of maize growth at different stages and meteorological data of maize growth at different stages;

[0031] S2. Perform preprocessing on the above data respectively;

[0032] S3. Extract remote sensing image features at different stages using a CNN network with an adaptive instance normalization module; extract meteorological features at different stages using a bidirectional LSTM network.

[0033] Based on remote sensing image data and meteorological data of maize growth at different stages, corresponding expert recommendations for different stages were constructed, and a pre-trained Chinese BERT model was used to encode the expert recommendations for different stages to obtain the features of expert recommendations for different stages.

[0034] S4. Combine the remote sensing image features, meteorological features, and expert recommendation features of each stage, and provide farmland management recommendations for each stage based on the corn growth status; thereby generating a dataset with remote sensing image features, meteorological features, and expert recommendation features as samples and farmland management recommendations as labels.

[0035] S5. Construct a large-scale model for corn growth guidance, fine-tune the large-scale model using a dataset, and output corn growth guidance suggestions through the fine-tuned large-scale model.

[0036] The different stages refer to the different stages of corn growth, including the seedling stage, jointing stage, tasseling stage, grain-filling stage, and maturity stage.

[0037] Example 2

[0038] This embodiment further defines embodiment 1 and provides a further explanation of step S1. Specifically, acquiring remote sensing image data of maize growth at different stages involves:

[0039] Observe the crop growth. If the corn seedlings are oriented north-south, set the flight path to east-west to avoid interference from leaf shadows.

[0040] Select different regions and times with good lighting conditions, and use multi-rotor drones equipped with multispectral cameras (including red, green, blue, and near-infrared bands) and RGB cameras to ensure a band resolution of no less than 10 nm and a spatial resolution better than 5 cm / pixel. Based on the corn's growth stage and field size, set the flight altitude to 1.5 * plant height + 20 meters and the flight speed to 3-5 m / s.

[0041] Set flight paths to ensure an overlap of ≥70% (heading) and ≥50% (lateral) between adjacent images, covering the entire corn planting area. If the plot has water pits or ditches, increase the density of this flight path by 10%.

[0042] The specific steps for obtaining meteorological data on corn growth at different stages are as follows: The meteorological data consists of temperature data, humidity data, and precipitation data over N consecutive days. The meteorological data is read, and any missing values ​​are checked. If any are missing, the mean method is used to fill in the missing values ​​using the values ​​from the two days before and after the missing values.

[0043] like Figure 2 and 3 The data consists of 30 consecutive days of meteorological humidity and temperature data, respectively. The image is drawn based on 30 days of data. In this embodiment, the input model is 7 consecutive days of meteorological data. In actual modeling, the 7 consecutive days of meteorological observation sequence is used as an input sample and fed into the bidirectional LSTM network. The 30-day temperature and humidity curves are only used to show the overall trend of meteorological elements during this period, which is convenient for human interpretation of the data distribution and corresponds to the 7-day sliding time window length used by the model.

[0044] Example 3

[0045] This embodiment further defines embodiment 1 and provides further explanation of step S2.

[0046] The preprocessing of remote sensing image data of maize growth specifically involves:

[0047] The drone data was segmented, with images cut into 512*512 pixels; images with corn cover below 75% were removed; and the remote sensing index of all images was calculated.

[0048] Where NIR is the reflectance in the near-infrared band and R is the reflectance in the red band.

[0049] Where G is the reflectivity of the green light band.

[0050] ;in, The reflectance is in the red-edge band (700-750nm).

[0051] Based on the aforementioned remote sensing indices, three-channel data is generated. The final data format is: remote sensing image data of size 512*512, where each pixel contains three channels of data. The three channels are as follows: , and This maps remote sensing indices to image pixels.

[0052] Perform data augmentation using the Augmentor module, including rotation, scaling, flipping, and block deformation.

[0053] The preprocessing of meteorological data for maize growth is as follows:

[0054] Data cleaning removes meteorological outliers. For temperature data, outliers below -40℃ or above 60℃ are removed. For humidity data, outliers below 0% or above 100% are removed. For precipitation, extreme values ​​with daily precipitation exceeding 500mm are removed.

[0055] Different meteorological elements have different magnitudes and distribution characteristics. Normalization of the meteorological data is performed using Z-Score, as detailed below:

[0056] ;in, For data points, The mean of the dataset. Let V be the variance of the dataset.

[0057] The normalized meteorological data is in the form of (B, 7, 7), where B represents the batch size; the first 7 represents the time step, corresponding to 7 consecutive days; and the second 7 represents the meteorological characteristic dimensions of each time step, including 7 elements: daily average temperature, daily maximum temperature, daily minimum temperature, daily average humidity, precipitation, sunshine duration, and wind speed. Figure 5 In the matrix, each row corresponds to a day's weather record, and each column corresponds to a specific meteorological element. Together, they form a 7×7 time-series feature matrix for bidirectional LSTM encoding.

[0058] Example 4

[0059] This embodiment further defines embodiment 1 and provides further explanation of step S3.

[0060] CNN networks with adaptive instance normalization modules, such as Figure 4 As shown, the input data format is (B, 3, 512, 512) three-channel remote sensing image data, where B represents the batch size of samples fed into the CNN in each iteration, which is set to 4 in the specific implementation, 3 represents the number of pixel channels, and 512, 512 represents the size of the remote sensing image data as 512 pixels * 512 pixels.

[0061] A 7x7 convolutional layer was used to extract features from three-channel remote sensing image data. Batch Normalization (BN) normalizes the convolutional output along the channel dimension to stabilize the feature distribution. Rectified Linear Unit (ReLU) sets negative values ​​to zero to enhance nonlinear expressiveness. Max Pooling (MP) downsamples by taking the maximum value within a local window, highlighting significant responses. After processing the features extracted by the 7x7 convolution with BN, ReLU, and MP sequentially, a remote sensing index feature map with reduced spatial resolution but more concentrated semantic information was obtained. This provides stable and discriminative high-level features for subsequent residual modules and global average pooling.

[0062] Input the remote sensing index feature map into the added adaptive instance normalization module. Figure 4 The 64 in the figure represents the number of channels processed by the adaptive instance normalization module, i.e., the 64 feature channels output by the first convolutional layer. The adaptive instance normalization module calculates the mean and standard deviation for each channel of each sample, and utilizes a learnable scaling factor. and offset coefficient The normalized features are recalibrated to reduce the impact of imaging style differences such as lighting conditions and atmospheric visibility on feature distribution. This makes the remote sensing features of the same plot of land more statistically consistent across different shooting times and external environments, improving the robustness and transferability of CNN-extracted features. The data processing performed in the adaptive instance normalization module is as follows:

[0063] ;in, For the extracted image features, The mean of the image features. The variance of image features, For learnable scaling parameters, This is a learnable offset parameter.

[0064] The four-layer residual module is used to perform progressive residual encoding on the feature maps after adaptive instance normalization. The residual encoding adopts the BasicBlock structure, in which each residual block contains two 3×3 convolution operations, and element-wise addition is performed between the main branch and the shortcut branch to form residual connections, thereby preserving the original information and enhancing the high-level semantic representation when deepening the network.

[0065] After completing the hierarchical encoding of the four residual module groups, "global average pooling" is used to perform global average aggregation of each channel in the spatial dimension, compressing the high-level feature map into a channel-level representation; then, a "fully connected layer" is used for linear mapping to output a fixed-dimensional TIFF feature vector, whose data form is (B, 128), where B is the sample batch size (the number of plot samples processed simultaneously in one forward computation), and 128 is the final feature dimension set by the network, used to compactly represent the comprehensive growth and stress status of the corresponding plot.

[0066] The bidirectional LSTM network uses existing modules, and its structure diagram is as follows: Figure 5 As shown. In Figure 5 The configuration for each layer is given. In the two LSTM layers, `return_sequence` represents the return parameter and is set to True mode, while `dropout` represents the regularization parameter and is set to 0.2. In the attention mechanism layer, the 256-dimensional hidden state of each time step is first mapped to a scalar score using `Dense(1, tanh)`, and its range is limited by `tanh` activation. Then, a Softmax operation is applied to the scores of all time steps to obtain a normalized attention weight distribution. Finally, the weighted sum is used to perform a weighted summation of the hidden states of the seven time steps, compressing the original 7×256 sequence into a 1×256 global meteorological feature vector, thereby automatically highlighting the time period most critical to the current agricultural recommendations and weakening the time points with smaller contributions.

[0067] Experts recommend preparing the corn in advance according to different growth stages and management types, including water management, nutrient management, pest and disease control, and disaster prevention.

[0068] The expert recommendations are suggestions given by relevant agricultural guidance experts based on the analysis of remote sensing image data and meteorological data on corn growth at the corresponding stage.

[0069] For each Chinese expert suggestion text, a pre-trained Chinese BERT model is used to encode it, converting it into a continuous vector representation, and then compressing it into a 64-dimensional feature space through linear mapping. A lightweight LoRA adapter is used to fine-tune some attention layers of BERT so that the 64-dimensional features retain the semantics of the general language and are more in line with the corn growth management scenario.

[0070] Example 5

[0071] This embodiment further defines embodiment 1 and provides further explanation of step S4.

[0072] Recommendations for farmland management at each stage are provided by constructing a knowledge base, which integrates various existing technical documents. Through manual analysis, farmland management recommendations that are compatible with remote sensing image data and meteorological data at each stage are selected from the knowledge base.

[0073] Specifically, the indicators included in remote sensing image data are remote sensing indices. , and The meteorological data includes indicators such as daily average temperature, daily maximum temperature, daily minimum temperature, daily average humidity, precipitation, sunshine duration, and wind speed. Standardized recommendations for farmland management of corn based on the remote sensing image data and meteorological data are compiled from existing publicly available technical literature and used as a knowledge base for manual reference.

[0074] When constructing a dataset that uses remote sensing image features, meteorological features, and expert advice features as samples and farmland management advice as labels, labels can also be set according to crop growth status. For example, six categories, such as normal, water shortage, fertilizer shortage, low temperature, high temperature, and pests and diseases, can be embedded as growth stage indicators into the corresponding labels.

[0075] Example 6

[0076] This embodiment further defines embodiment 1 and provides further explanation of step S5.

[0077] The large model includes a weighted feature fusion part and an encoding and decoding part. The weighted feature fusion part weights and fuses remote sensing image features, meteorological features and expert suggestion features to obtain fused features. The encoding and decoding part adopts the Transformer architecture, and obtains farmland management suggestions corresponding to the fused features by sequentially encoding and decoding the fused features.

[0078] like Figure 6 As shown, 128-dimensional remote sensing image features, 256-dimensional meteorological features, and 64-dimensional expert suggestion features are combined. Figure 6The input to the corn growth guidance model is represented by knowledge features. Here, Squeeze(dim=1) means removing the time / sequence dimension of length 1 from the tensor shape, compressing the three-dimensional tensor of the form [B,1,128], [B,1,256], [B,1,64] into two-dimensional vectors of the form [B,128], [B,256], [B,64]. Concat(dim=-1) means concatenating the above three vectors on the last feature dimension, combining them into a fusion feature of size [B,448], which is used for subsequent dynamic attention and Transformer encoding.

[0079] In the weighted feature fusion part, a dynamic attention mechanism is adopted. The core idea of ​​the dynamic attention mechanism is to adaptively assign weights to features of different modalities and dimensions based on the comprehensive context of the current multi-source features and the growth stage, thereby "amplifying" features that are highly relevant to the current growth diagnosis and suppressing noise or redundant information.

[0080] Taking 128-dimensional remote sensing image features as an example, a 448-dimensional fused vector (containing image, meteorological, and expert suggestion codes) is input into a two-layer perceptron structure: Linear(448→128)→ReLU→Linear(128→128)→Sigmoid. First, a nonlinear mapping is obtained through linear transformation (Linear) and ReLU. Then, the Sigmoid function maps each dimension to a value between 0 and 1, resulting in an attention weight vector of shape [B, 128]. Subsequently, element-wise multiplication is used to multiply this weight with the original 128-dimensional remote sensing image features to form a weighted image feature. Meteorological features and expert suggestion features are processed similarly, thus achieving "dynamic weighting" and feature selection based on the current context.

[0081] Since the encoder / decoder in the Transformer uniformly uses 256 dimensions, a linear layer is needed to compress the 448 dimensions obtained by concatenating the three modal features to 256 dimensions, resulting in [B, 256]. This operation corresponds to... Figure 6 The fusion output section.

[0082] The input to a Transformer encoder is a sequence, typically of shape [B, T, 256]. Therefore, when assembling the sequence, the fusion vector is artificially modified from two-dimensional [B, 256] by adding a dimension of "sequence length = 1", transforming it into three-dimensional [B, 1, 256]. This operation corresponds to... Figure 6The transformation sequence part of the code. The Transformer architecture consists of a multi-layer Transformer encoder and a Transformer decoder based on the BERT model initialization: the encoder uses a 2-layer encoder (num_layers=2), with a feature dimension d_model=256 and an attention head number n_head=8 for each layer;

[0083] The feedforward network has a hidden layer dimension of 4×256=1024 (dim_feedforward=1024), dropout=0.1, and is equipped with stage-enhanced positional encoding to inject sequence position information and corn growth stage information into the token representation. The decoder is also based on d_model=256 and nhead=8, stacking 4 decoder layers (num_heads=4), which internally includes self-attention, multi-head cross-attention and feedforward network structure. Through cross-attention, each time step in the decoding process is matched with the encoded multimodal memory sequence, so as to comprehensively utilize image features, meteorological features and expert suggestion features when generating agricultural guidance text.

[0084] The large model is fine-tuned and trained as follows: First, based on the constructed dataset, each sample contains a UAV remote sensing image with 128-dimensional features extracted by CNN, a 7-day meteorological sequence with 256-dimensional features extracted by bidirectional LSTM within the corresponding time window, 64-dimensional expert suggestion features obtained from expert experience and encoded by BERT, and the label of the corn growth stage to which the sample belongs. Then, a pre-trained Chinese BERT model is used as the initialization model for the text decoder. A low-rank adapter is injected into its self-attention layer using LoRA. Only the adapter parameters and the parameters of the multimodal fusion module, Transformer encoder and decoder output layer are updated, while the original weights of the BERT model are kept basically frozen to reduce the number of fine-tuning parameters. During the training phase, the aforementioned multimodal features are used as conditional inputs, and the corresponding farmland management suggestion texts are used as target sequences to construct paired samples of "input multimodal features - output agricultural suggestions." The cross-entropy loss function is used to constrain the difference between the generated sequence and the real text. The optimizer is AdamW, combined with cosine annealing learning rate scheduling and gradient pruning strategies. Optimal weights are selected through a 7:2:1 training / validation / test split and an early stopping mechanism. Unlike the fine-tuning methods in general references that only target a single modality or adjust only the decoder parameters, this embodiment jointly optimizes multi-source feature gating fusion, stage-aware attention, and BERT-based text decoding within the same framework. This ensures that the generated agricultural guidance accurately reflects the corn growth status and fully utilizes three types of information: UAV imagery, meteorological time series, and knowledge rules.

Claims

1. A method for training a large model to generate maize growth guidance, characterized in that, The method includes the following steps: S1. Obtain remote sensing image data of maize growth at different stages, and extract the features of remote sensing images at different stages by adding an adaptive instance normalization module to the CNN network. S2. Obtain meteorological data on corn growth at different stages, and extract meteorological characteristics of different stages through a bidirectional LSTM network; S3. Based on remote sensing image data and meteorological data of maize growth at different stages, construct corresponding expert recommendations for different stages, and use a pre-trained Chinese BERT model to encode the expert recommendations for different stages to obtain the features of expert recommendations for different stages. S4. Combine the remote sensing image features, meteorological features, and expert recommendation features of each stage, and provide farmland management recommendations for each stage based on the corn growth status; thereby generating a dataset with remote sensing image features, meteorological features, and expert recommendation features as samples and farmland management recommendations as labels. The expert recommendations are suggestions given by relevant agricultural guidance experts after analyzing remote sensing image data and meteorological data on corn growth at the corresponding stage. Recommendations for farmland management at each stage are provided by constructing a knowledge base, which is obtained by integrating various existing technical documents. Through manual analysis, farmland management recommendations that are compatible with remote sensing image data and meteorological data at each stage are selected from the knowledge base. S5. Construct a large-scale model for corn growth guidance, fine-tune the large-scale model using a dataset, and output corn growth guidance suggestions through the fine-tuned large-scale model.

2. The method for training a large model for generating maize growth guidance according to claim 1, characterized in that, The different stages refer to the different stages of corn growth, including the seedling stage, jointing stage, tasseling stage, grain-filling stage, and maturity stage.

3. The method for training a large model for generating maize growth guidance according to claim 2, characterized in that, The method for acquiring remote sensing image data of maize growth at different stages is as follows: remote sensing image data is acquired by a drone equipped with a multispectral camera and an RGB camera; the acquired image data is divided; and the remote sensing index of all images is calculated. , and Based on the aforementioned remote sensing indices, three-channel data is generated, and data augmentation is performed on the three-channel data.

4. The method for training a large model for generating maize growth guidance according to claim 3, characterized in that, The CNN network, from input to output, includes convolutional layers, an adaptive instance normalization module, a four-layer residual module group, a global average pooling layer, and a fully connected layer. The adaptive instance normalization module calculates the mean and standard deviation for each channel of each sample, and utilizes a learnable scaling factor. and offset coefficient The normalized features are recalibrated.

5. The method for training a large model for generating maize growth guidance according to claim 4, characterized in that, The meteorological data for corn growth consists of meteorological data for N consecutive days, including meteorological indicators such as temperature, humidity, and precipitation. The meteorological data for N consecutive days is normalized to obtain the daily average meteorological indicators.

6. The method for training a large model for generating maize growth guidance according to claim 5, characterized in that, Experts recommend preparing the corn in advance according to different growth stages and management types, including water management, nutrient management, pest and disease control, and disaster prevention.

7. The method for training a large model for generating maize growth guidance according to claim 6, characterized in that, The large model includes a weighted feature fusion part and an encoding and decoding part. The weighted feature fusion part weights and fuses remote sensing image features, meteorological features and expert suggestion features to obtain fused features. The encoding and decoding part adopts the Transformer architecture. By sequentially encoding and decoding the fused features, farmland management suggestions corresponding to the fused features are obtained.

8. The method for training a large model for generating maize growth guidance according to claim 7, characterized in that, The fine-tuning training of the large model is as follows: a pre-trained Chinese BERT model is used as the initialization model of the decoder. A low-rank adapter is injected into the self-attention layer of the BERT model through LoRA. Only the adapter parameters, the weighted feature fusion part parameters, and the output layer parameters of the encoding and decoding parts are updated, while the original weights of the BERT model are frozen, so as to reduce the number of fine-tuning parameters. During the training phase, remote sensing image features, meteorological features, and expert suggestion features were used as conditional inputs, and the corresponding farmland management suggestions were used as target sequences. Paired samples of input multimodal features and output agricultural suggestions were constructed. The cross-entropy loss function was used to constrain the difference between the generated sequence and the real text. AdamW was selected as the optimizer, and the dataset was divided into training / validation / test ratios of 7:2:1.

Images