Large model training method for generating corn growth guidance

Abstract

The invention discloses a large model training method for generating corn growth guidance. The method belongs to the technical field of agricultural intelligent monitoring and prediction. The method solves the problem that when applied to a corn growth management scene, a large model method depends on multi-source heterogeneous information such as remote sensing images and meteorological data, although preliminary analysis and prediction based on the crop growth state can be achieved, the large model method still has limitation in the aspects of accuracy, interpretability and actual adaptability of decision suggestions. During data set construction, besides fusion of remote sensing data and multi-dimensional meteorological information, expert suggestions are added, more multivariate fusion features are constructed, and due to similarity between the expert suggestions and farmland management suggestions, correspondence between the features and tags is more accurate; according to the method, corn growth guidance based on a large model architecture can realize a technical span from'visible 'to'understandable' to'guidance ', and a scientific basis and decision support are provided for precise management of corn production.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of agricultural intelligent monitoring and prediction, and particularly relates to a large model training method for generating corn growth guidance. BACKGROUND

[0002] Corn, as one of the important food and feed crops in the world, has a strategic position in ensuring food security and agricultural economic development. With the rapid development of smart agriculture, how to realize the accurate monitoring, scientific management and intelligent decision-making of the growth process of corn has become a key issue to improve yield level, optimize resource utilization efficiency and cope with the challenges of climate change.

[0003] Traditional corn growth management mainly relies on manual field investigation and experience judgment, which has problems such as poor timeliness, strong subjectivity, limited coverage, etc., and is difficult to meet the development needs of modern agriculture in terms of scale, refinement and intelligence. In recent years, emerging technologies such as remote sensing, Internet of Things and artificial intelligence have gradually been applied in the field of agriculture, providing new means for dynamic perception of crop growth status. Among them, unmanned aerial vehicle remote sensing can quickly obtain multispectral, thermal infrared and visible light images at field scale with high spatio-temporal resolution, flexible deployment and low cost, effectively reflecting key growth parameters such as corn canopy structure, leaf area index, vegetation coverage and stress status. At the same time, meteorological data (such as temperature, precipitation, light, humidity, etc.) as the core environmental factors affecting corn growth and development play a decisive role in crop phenology, yield formation and disaster risk.

[0004] In recent years, large model technology has made breakthrough progress in the fields of computer vision and natural language processing, and has shown strong multi-modal fusion ability, context understanding ability and generalization reasoning ability. By introducing it into the corn growth management scene, it is expected to realize the unified modeling and semantic understanding of multi-source heterogeneous agricultural data, so as to build a more accurate, interpretable and adaptive crop growth intelligent guidance system. However, the current large model method usually relies on multi-source heterogeneous information such as remote sensing images and meteorological data when applied to the corn growth management scene, although it can realize preliminary analysis and prediction based on crop growth status, but it still has limitations in the accuracy, explainability and actual adaptability of decision-making suggestions. SUMMARY

[0005] In order to solve the technical problem that the current large model method usually relies on multi-source heterogeneous information such as remote sensing images and meteorological data when applied to the corn growth management scene, although it can realize preliminary analysis and prediction based on crop growth status, but it still has limitations in the accuracy, explainability and actual adaptability of decision-making suggestions, the application provides a large model training method for generating corn growth guidance.

[0006] The method specifically comprises S1, acquiring remote sensing image data of corn growth at different stages, and extracting features of remote sensing images at different stages through a CNN network with an adaptive instance normalization module; S2, acquiring meteorological data of corn growth at different stages, and extracting meteorological features at different stages through a bidirectional LSTM network; S3, constructing expert recommendations corresponding to different stages according to the remote sensing image data of corn growth at different stages and the meteorological data of corn growth at different stages, and encoding the expert recommendations at different stages by using a pre-trained Chinese BERT model to obtain expert recommendation features at different stages; S4, combining the remote sensing image features at different stages, the meteorological features at different stages and the expert recommendation features at different stages, and giving farmland management recommendations for the growth of corn at different stages; thereby generating a data set taking the remote sensing image features, the meteorological features and the expert recommendation features as samples and taking the farmland management recommendations as labels; S5, constructing a corn growth guidance large model, fine-tuning the large model by using the data set, and outputting corn growth guidance recommendations through the large model after fine-tuning.

[0007] Further, the different stages refer to different stages of corn growth, including the seedling stage, the jointing stage, the tasseling stage, the grain filling stage and the mature stage.

[0008] Further, the method for acquiring remote sensing image data of corn growth at different stages comprises: acquiring remote sensing image data by a UAV carrying a multi-spectral camera and an RGB camera, dividing the collected image data, and calculating remote sensing indexes of all pictures 、 and , forming three-channel data according to the remote sensing indexes, and performing data enhancement on the three-channel data.

[0009] Further, the CNN network comprises, in sequence from input to output, a convolutional layer, 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 of each channel of each sample respectively, and uses a learnable scaling coefficient and an offset coefficient to recalibrate the normalized features.

[0010] Further, the meteorological data of corn growth is continuous N-day meteorological data, including temperature data, humidity data and precipitation, and the continuous N-day meteorological data is normalized to obtain daily average meteorological indexes.

[0011] Further, the expert recommendations are pre-arranged according to different stages of corn growth and management types, and the management types include water management, nutrient management, disease and pest control and disaster prevention.

[0012] Further, the farmland management suggestions of each stage are given by building a knowledge base, the knowledge base is integrated from various existing technical documents, and the farmland management suggestions suitable for the remote sensing image data and the meteorological data of each stage are selected from the knowledge base through manual analysis.

[0013] Further, the large model comprises a weighted feature fusion part and an encoding and decoding part, the remote sensing image features, the meteorological features and the expert suggestion features are weighted and fused by the weighted feature fusion part to obtain fused features; the encoding and decoding part adopts a Transformer architecture, and the farmland management suggestions corresponding to the fused features are obtained by sequentially performing encoding and decoding operations on the fused features.

[0014] Further, the fine-tuning training of the large model is specifically: using a pre-trained Chinese BERT model as an initialization model of the decoder, injecting a low-rank adapter into the self-attention layer of the BERT model through the LoRA method, only updating the adapter parameters, the weighted feature fusion part parameters and the output layer parameters of the encoding and decoding part, and keeping the original weights of the BERT model frozen, so as to reduce the fine-tuning parameter amount; In the training stage, the remote sensing image features, the meteorological features and the expert suggestion features are used as conditional inputs, the farmland management suggestions corresponding thereto are used as target sequences, the paired samples of input multi-modal features-output agricultural suggestions are constructed, the difference between the generated sequence and the real text is constrained by using a cross-entropy loss function, an AdamW optimizer is selected, and the data set is divided by a training / validation / testing ratio of 7:2:1.

[0015] The method has the beneficial effects that: the large model training method can more accurately guide the growth of corn at different stages, the large model after training can more accurately guide the growth of corn at different stages, the expert suggestions are added in addition to the fusion of remote sensing data and multi-dimensional meteorological information during the construction of the data set, more diversified fusion features are constructed, the correspondence between the features and the labels is more accurate due to the similarity between the expert suggestions and the farmland management suggestions, and the corn growth guidance based on the large model architecture realizes the technical leap from "seeing" to "understanding" to "guiding", thereby providing a scientific basis and decision support for the precision management of corn production. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The method flowchart in the embodiment of the application; Figure 2 The meteorological humidity data collected in the embodiment of the application; Figure 3 The meteorological temperature data collected in the embodiment of the application; Figure 4This is a diagram of the CNN network structure in an embodiment of the present invention; Figure 5 This is a diagram of the bidirectional LSTM network structure in an embodiment of the present invention; Figure 6 This is a structural diagram of the large model in an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1 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: S1. Acquire remote sensing image data of maize growth at different stages and meteorological data of maize growth at different stages; S2. Perform preprocessing on the above data respectively; 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. 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. 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. 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.

[0019] 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.

[0020] Example 2 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: 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.

[0021] Select different areas, different time, light, select multi-rotor unmanned aerial vehicle carrying multi-spectral camera (including red, green, blue, near infrared band) and RGB camera, ensure that the band resolution is not less than 10 nm, the spatial resolution is better than 5 cm / pixel. According to the growth stage of corn and the size of the field, set the flight height to 1.5*plant height+20 meters, the flight speed is 3-5 m / s.

[0022] Set the flight route to ensure that the adjacent image overlap rate is ≥70% (heading) and ≥50% (lateral), covering the entire corn planting area. If there are water pits and ditches in the field, increase the density of this route by 10%.

[0023] Obtain different stage corn growth weather data, which is temperature data, humidity data and precipitation in the last N days. Read the weather data and check if there is any missing value. If there is missing value, use the mean method to fill in the missing value with the last two days and the next two days.

[0024] As Figure 2 and 3 are the weather humidity data and weather temperature data for 30 consecutive days, and the image drawing is 30 days of data. In this embodiment, the input model is the weather data of the last 7 days. When modeling, the last 7 days of weather observation sequence is used as an input sample to input into the bidirectional LSTM network. The 30-day temperature and humidity curve is only used to show the overall trend of the meteorological elements in this period, which is convenient for artificial understanding of data distribution, and corresponds to the 7-day sliding time window length used by the model.

[0025] Embodiment 3, This embodiment is a further limitation of embodiment 1, and further describes step S2.

[0026] The preprocessing of corn growth remote sensing image data is as follows: Divide the unmanned aerial vehicle data, cut the image into 512*512 size pictures; delete the image data with less than 75% of corn coverage; calculate the remote sensing index of all pictures: ; wherein, NIR is the near-infrared reflectance, R is the red light reflectance.

[0027] ; wherein, G is the green light reflectance.

[0028] ; wherein, is the red edge band (700-750 nm) reflectance.

[0029] According to the remote sensing index, three-channel data is formed, and finally the generated data style is: the size of the remote sensing image data is 512*512, each pixel contains three-channel data, and the three-channel data is 、 and , which maps the remote sensing index into a picture pixel.

[0030] Data augmentation is performed, and the Augmentor module is used for augmentation, including rotation, scaling, flipping, small block deformation, etc.

[0031] The preprocessing of the corn growth meteorological data is specifically: Data cleaning of meteorological outliers: for temperature data: eliminate outliers below-40℃ or above 60℃, for humidity data: eliminate outliers below 0% or above 100%, and for precipitation: eliminate extreme values of single-day precipitation exceeding 500mm.

[0032] Different meteorological data have different magnitudes and distribution characteristics. The meteorological data is normalized using Z-Score normalization, as follows: ; wherein, is the data point, is the data set mean, is the data set variance.

[0033] 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 feature dimension of each time step, including daily mean temperature, daily maximum temperature, daily minimum temperature, daily average humidity, precipitation, sunshine duration and wind speed. In Figure 5 , each row corresponds to a day of meteorological records, and each column corresponds to a specific meteorological element, forming a 7*7 time sequence feature matrix for bidirectional LSTM encoding.

[0034] Embodiment 4, This embodiment is a further limitation of embodiment 1, and further describes step S3.

[0035] The CNN network with an adaptive instance normalization module is shown in Figure 4 , and the input data format is three-channel remote sensing image data of (B, 3, 512, 512), where B represents the sample batch size of the CNN sent 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.

[0036] The three-channel remote sensing image data is subjected to feature extraction by using a convolution layer with a size of 7*7, wherein BN (Batch Normalization) is a batch normalization layer, which normalizes the convolution output in the channel dimension to stabilize the feature distribution; ReLU (Rectified Linear Unit) is a linear rectification activation function, which sets negative values to zero to enhance the non-linear expression ability; and MP (Max Pooling) is a maximum pooling layer, which takes the maximum value in a local window to realize down-sampling and highlight significant responses. After the features extracted by the 7*7 convolution are sequentially subjected to the BN, ReLU and MP processing, a remote sensing index feature map with reduced spatial resolution but more concentrated semantic information is obtained, which provides stable and discriminative high-level features for the subsequent residual module and global average pooling.

[0037] The remote sensing index feature map is input into an added adaptive instance normalization module, Figure 4 wherein 64 represents the number of channels processed by the adaptive instance normalization module, i.e., 64 feature channels output by the first convolution layer. The adaptive instance normalization module respectively calculates the mean and standard deviation of each channel of each sample, and uses a learnable scaling coefficient and an offset coefficient to rescale the normalized features, thereby weakening the influence of imaging style differences such as lighting conditions and atmospheric visibility on feature distribution, making the remote sensing features of the same plot at different shooting times and in different external environments more consistent in a statistical sense, and improving the robustness and transferability of the CNN extracted features. The data processing in the adaptive instance normalization module is as follows: ; wherein, is the extracted image feature, is the mean of the image feature, is the variance of the image feature, is a learnable scaling parameter, is a learnable offset parameter.

[0038] The four-layer residual module is used to perform step-by-step residual coding on the feature map after adaptive instance normalization; the residual coding adopts a BasicBlock structure, each BasicBlock includes two 3*3 convolution operations, and element-wise addition is performed between the main branch and the shortcut branch to form a residual connection, thereby preserving the original information and enhancing the high-level semantic representation when deepening the network.

[0039] After completing the step-by-step encoding of the four-layer residual module group, the global average pooling is used to globally average and aggregate each channel in the spatial dimension, compressing the high-level feature map into a channel-level representation; then the linear mapping is performed through the fully connected layer, and the fixed-dimension tif feature vector is output, which has a data form of (B, 128), wherein B is the sample batch size (the number of land samples processed simultaneously in one forward calculation), and 128 is the final feature dimension set by the network, which is used to compactly represent the comprehensive growth and stress state of the corresponding land.

[0040] The bidirectional LSTM network uses existing modules, and its structure diagram is as shown in Figure 5 In Figure 5 , the configuration of each level is given, and in the two LSTM layers, return_sequence represents the return parameter, which is set to True mode, and dropout represents the regularization parameter, which is set to 0.2; in the attention mechanism layer, first, the 256-dimensional hidden state of each time step is mapped to a scalar score through Dense(1, tanh), and the tanh activation is used to limit its range; then, the Softmax operation is applied to the scores of all time steps to obtain the normalized attention weight distribution; finally, the Weighted Sum is used to weight and sum the hidden states of the 7 time steps using the weights, and the original 7x256 sequence is compressed into a 1x256 global weather feature vector, thereby automatically highlighting the most critical time period for the current farming recommendation and weakening the time points with less contribution.

[0041] Experts recommend that the different stages of corn growth and management types be sorted in advance, including water management, nutrient management, pest control, and disaster prevention.

[0042] Experts recommend that relevant agricultural guidance experts analyze the corn growth remote sensing image data and corn growth meteorological data according to the corresponding stage, and give recommendations.

[0043] For each Chinese expert recommendation text, a pre-trained Chinese BERT model is used for encoding to convert it into a continuous vector representation, and it is compressed to a 64-dimensional feature space through linear mapping; the lightweight LoRA adapter is used to fine-tune part of the attention layers of BERT, so that the 64-dimensional features not only retain general language semantics, but also better fit the corn growth management scenario.

[0044] Embodiment 5, This embodiment is a further limitation of embodiment 1, and step S4 is further described.

[0045] The farmland management suggestions of each stage are given by building a knowledge base, the knowledge base is obtained by integrating various existing technical documents, and the farmland management suggestions suitable for the remote sensing image data and the meteorological data of each stage are selected from the knowledge base through manual analysis.

[0046] Specifically, the indexes contained in the remote sensing image data are remote sensing indexes 、 and ; the indexes contained in the meteorological data are daily mean temperature, daily maximum temperature, daily minimum temperature, daily mean humidity, precipitation, sunshine duration and wind speed; the standardized suggestions for farmland management of corn according to the remote sensing image data and the meteorological data are sorted out from the existing public technical documents as the knowledge base for manual reference.

[0047] When building a data set with remote sensing image features, meteorological features and expert suggestion features as samples and farmland management suggestions as labels, labels can also be set according to the growth state of crops, for example, six categories of normal, water shortage, fertilizer deficiency, low temperature, high temperature, disease and pest are embedded into the corresponding labels as growth stage indicators.

[0048] Embodiment 6, This embodiment is a further limitation of embodiment 1, and step S5 is further described.

[0049] The large model includes a weighted feature fusion part and an encoding and decoding part, the weighted feature fusion part is used to weight and fuse the remote sensing image features, the meteorological features and the expert suggestion features to obtain the fused features; the encoding and decoding part adopts the Transformer architecture, and the farmland management suggestions corresponding to the fused features are obtained by sequentially encoding and decoding the fused features.

[0050] As shown in Figure 6 , 128-dimensional remote sensing image features, 256-dimensional meteorological features and 64-dimensional expert suggestion features (denoted as knowledge features in Figure 6 ) are input into the corn growth guidance large model, wherein Squeeze(dim=1) means removing the time / sequence dimension with a length of 1 in the tensor shape, and compressing the three-dimensional tensors of [B,1,128], [B,1,256], [B,1,64] into two-dimensional vectors of [B,128], [B,256], [B,64]; Concat(dim=-1) means concatenating the above three vectors in the last feature dimension, combining them into a fused feature with a size of [B,448], which is used for subsequent dynamic attention and Transformer encoding.

[0051] 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 different dimensions according to the comprehensive context of the current growth stage, so as to "amplify" the features highly related to the current growth diagnosis and suppress noise or redundant information.

[0052] Taking the 128-dimensional remote sensing image features as an example, the 448-dimensional fusion vector (including image, meteorological and expert recommendation codes) is input into the two-layer perceptron structure of Linear(448→128)→ReLU→Linear(128→128)→Sigmoid. First, the linear transformation (Linear) and ReLU are used to obtain a non-linear mapping, and then the Sigmoid is used to map each dimension to 0-1, obtaining an attention weight vector with a shape of [B, 128]. Then, element-wise multiplication is used to multiply the weight with the original 128-dimensional remote sensing image features to form the weighted image features. The meteorological features and expert recommendation features are processed in the same way, so as to realize "dynamic weighting" and feature selection based on the current context.

[0053] Since the encoder / decoder in the Transformer uses 256 dimensions uniformly, a linear layer is needed to compress the 448-dimensional feature obtained by splicing the three types of modal features to 256 dimensions to obtain [B, 256], which corresponds to the fusion output part in Figure 6 .

[0054] The input of the Transformer encoder is a sequence, generally with a shape of [B, T, 256]. Therefore, when assembling the sequence, the fusion vector is artificially added with a dimension of "sequence length = 1" to become a three-dimensional [B, 1, 256], which corresponds to the transformation sequence part in Figure 6 . The Transformer architecture part is composed of a section of multi-layer Transformer encoder and a section of Transformer decoder initialized based on the BERT model: the encoder end adopts a 2-layer encoder (num_layers=2), and the feature dimension d_model=256 and the number of attention heads n_head=8 in each layer. The hidden layer dimension of the feedforward network is 4*256=1024 (dim_feedforward=1024), the dropout is 0.1, and the sinusoidal position encoding with growth stage enhancement (Stage Enhanced Positional Encoding) is used to inject the sequence position information and the corn growth stage information into the token representation; the decoding end is also based on d_model=256 and nhead=8, and 4 layers of decoder layers (num_heads=4) are stacked, which contain self-attention, multi-head cross-attention and feedforward network structure, and through cross-attention, each time step in the decoding process is matched with the multi-modal memory sequence obtained by encoding, so as to comprehensively utilize the image features, weather features and expert suggestion features when generating agricultural guidance text.

[0055] The large model is fine-tuned, specifically: first, based on the constructed data set, each sample contains a UAV remote sensing image with 128-dimensional features extracted by CNN, a 7-day weather sequence with 256-dimensional features extracted by bidirectional LSTM in the corresponding time window, 64-dimensional expert suggestion features obtained by expert experience and encoded by BERT, and the corn growth stage label to which the sample belongs; then, the pre-trained Chinese BERT model is used as the initialization model of the text decoder, the low-rank adapter is injected into the self-attention layer through LoRA, only the adapter parameters and the parameters of the multi-modal fusion module, the output layer of the Transformer encoder and decoder are updated, while the original weights of the BERT model are basically frozen, so as to reduce the parameter quantity of fine-tuning. In the training stage, the above multi-modal features are used as the conditional input, the farmland management suggestion text corresponding to them is used as the target sequence, the paired samples of "input multi-modal features-output agricultural suggestions" are constructed, the cross-entropy loss function is used to constrain the difference between the generated sequence and the real text, the AdamW optimizer is selected, combined with the cosine annealing learning rate scheduling and gradient clipping strategy, and the optimal weight is selected through 7:2:1 training / validation / test division and early stopping mechanism. Different from the fine-tuning method in the general reference literature which only targets a single modality or only adjusts the decoder parameters, the embodiment jointly optimizes the multi-source feature gating fusion, stage-aware attention and BERT-based text decoding in the same framework, so that the generated agricultural guidance can accurately reflect the corn growth state and fully utilize the three types of information of UAV image, weather 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. 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 to generate 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, 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.

8. The method for training a large model for generating maize growth guidance according to claim 7, 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.

9. A method for training a large model for generating maize growth guidance according to claim 8, 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