Corn yield prediction method based on cooperation of unmanned aerial vehicle and satellite remote sensing

Abstract

The invention provides a corn yield prediction method based on cooperation of an unmanned aerial vehicle and satellite remote sensing, and relates to the technical field of deep learning, and the method comprises the steps: S1, constructing a multi-source basic data set which comprises a ground real data set, an unmanned aerial vehicle remote sensing data set and a satellite remote sensing data set; s2, key feature images in the remote sensing data set are extracted for the phenological period of corn yield estimation, and a feature data set is constructed; s3, constructing an improved corn yield prediction model; and S4, introducing a self-adaptive penalty loss function into the improved corn yield prediction model, quantifying the consistency and conflict in the model by using the internal state of the improved corn yield prediction model, and adaptively adjusting the penalty intensity to obtain an adjusted loss function. According to the method, large-area and high-resolution corn yield spatial distribution prediction can be realized, the prediction precision is high, the timeliness is strong, and key data support is provided for agricultural precise management.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a corn yield prediction method based on collaboration between unmanned aerial vehicle (UAV) and satellite remote sensing. Background Art

[0002] Drone remote sensing is a new monitoring method that combines drone technology with remote sensing technology. It uses drones equipped with various sensors (such as optical cameras, infrared sensors, lidar, etc.) to obtain images and data of target areas from the air, thereby realizing the collection and analysis of information on the surface environment and target objects. It has the advantages of high flexibility, high resolution, and low analysis cost. Currently, in the agricultural sector, existing technologies already use drone remote sensing technology for crop classification and identification, crop yield estimation, and pest and disease control. Crop yield estimation can not only achieve early identification of growth anomalies and reduce yield losses, but also improve resource utilization, promote sustainable planting, and strengthen the management of various factors such as disasters and risks, and reduce the impact of uncontrollable factors. However, existing large-area crop yield estimation methods have limitations in practical applications. For example, current large-area yield prediction methods mostly rely on satellite remote sensing data, but satellite imagery is greatly affected by clouds and rain and has relatively low spatial resolution, making it difficult to capture subtle crop growth differences within fields. As a result, prediction models constructed solely based on satellite data have difficulty learning the complex relationship between crop growth and yield, resulting in large prediction bias, low accuracy, and poor generalization ability. At the same time, existing models have poor adaptability to complex and heterogeneous agricultural planting environments and find it difficult to mine deep information in spatiotemporal data, resulting in a time lag in yield estimation and an inability to meet the needs of real-time agricultural management. Summary of the Invention

[0003] In response to the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a corn yield prediction method based on the collaboration of drones and satellite remote sensing. The method uses ultra-high-resolution remote sensing data acquired by drones in representative sample plots as "bridge data" or "calibration data" and combines it with ground-measured yield data to train and optimize an improved deep learning yield prediction model; then, the trained high-performance model is applied to satellite remote sensing data covering a large area but with low resolution, thereby solving the problem of insufficient model accuracy and poor generalization ability caused by relying solely on satellite data for modeling, and ultimately achieving large-area, high-resolution, and high-precision spatial distribution prediction of corn yield with high prediction accuracy and strong timeliness.

[0004] The purpose of the present invention is achieved through the following technical solutions: A corn yield prediction method based on the collaboration of UAV and satellite remote sensing includes: Step S1: construct a multi-source basic dataset, including a ground truth dataset, a UAV remote sensing dataset, and a satellite remote sensing dataset; Step S2: extracting key feature images from the remote sensing dataset for the phenological period of corn yield estimation to construct a feature dataset; Step S3: constructing an improved corn yield prediction model, and performing model training, verification, and evaluation in sequence; Step S4: introducing an adaptive penalty loss function into the improved corn yield prediction model, utilizing the internal state of the improved corn yield prediction model to quantify the consistency and conflict within the model, adaptively adjusting the penalty intensity, and obtaining the adjusted loss function.

[0005] Based on further optimization of the above scheme, in step S1, the method for obtaining the real data set is as follows: first, a set of representative sampling points and their high-precision geographic coordinates are planned and determined within the target area; then, during the corn crop maturity and harvest period, field measurements are conducted at the planned sampling points to obtain corresponding yield data, thereby constructing a ground truth data set.

[0006] Based on further optimization of the above scheme, in step S1, the construction of the UAV remote sensing dataset is to fly a UAV equipped with a multispectral sensor over the planned sampling points during the critical growth period of the corn crop to obtain UAV remote sensing images with ultra-high spatial resolution, which serve as a high-quality data source for subsequent model training; the construction of the satellite remote sensing dataset is to obtain multispectral satellite remote sensing images such as Sentinel-2, covering the entire target area and the time range covering the critical growth period of corn, through cloud platforms such as Google Earth Engine (GEE), and perform standardized preprocessing.

[0007] Based on further optimization of the above scheme, step S2 is specifically as follows: extracting remote sensing characteristic values ​​of each sampling point in the key time window of the entire corn growing season; the remote sensing characteristic values ​​include the maximum and average values ​​of the normalized vegetation index NDVI, the maximum and average values ​​of the vegetation near-infrared reflectance index NIRv, the maximum and average values ​​of the enhanced vegetation index EVI, the maximum and average values ​​of the kernel normalized vegetation index kNDVI, and the maximum and average values ​​of the normalized difference water index NDWI; matching the yield data of the real data set with the corresponding remote sensing characteristic values ​​for subsequent model training.

[0008] Based on the further optimization of the above scheme, the improved corn yield prediction model includes a primary multi-scale feature extractor, a two-layer feature refinement block and an output MLP layer; Among them, the primary multi-scale feature extractor obtains input features and outputs them to the two-layer feature refinement block. The input features are first embedded through a 1x1 convolution layer.Input_Dim Map to B ase_Channels , followed by batch normalization and GELU activation:

[0009] Where: Represents input features; Indicates output; B represents the batch size, T represents the time step, D in Input feature dimension, C base The number of base channels of the embedding layer; Indicates 1x1 convolution processing; Indicates batch normalization processing; GELU( ) express GELU Activation function; After that, the embedded features are fed into a multi-scale gated convolution branch, i.e., multiple parallel gated convolution branches, each using a different convolution kernel to capture temporal limitations at different scales. Finally, the outputs of each scale branch are spliced ​​in the channel dimension to form fused multi-scale features.

[0010] Based on the further optimization of the above scheme, the multi-scale gated convolution branch includes three parallel gated convolution branches: 3x3 gated convolution, 5x5 gated convolution and 7x7 gated convolution; Gated convolution is a neural network component that combines a convolution operation with a gating mechanism: For input features 、 C in Represents the number of channels of the input layer, which is passed through a 1D convolutional layer whose output is twice the number of desired output channels C out Channel:

[0011] Then, H is divided into two parts along the channel dimension, namely the feature part With the gate control part ; Finally, the output of the gated convolution is achieved:

[0012] Where: represents element-wise product; Represents the Sigmoig function.

[0013] Based on the further optimization of the above solution, the feature refining block is specifically as follows: Input features First replaced by , to adapt to the input format of the multi-head self-attention module; then, the permuted features are layer-normalized and fed into the multi-head self-attention module:

[0014]

[0015] Where: Normalization of the representation layer; h Indicates the number of attention heads; Represents the output projection matrix; Respectively represent the corresponding projection matrices; d k Dimensions representing queries and keys, ; The output of the multi-head self-attention module is regularized ( Dropout ) and dimension replacement, we get ; Afterwards, by the output of the multi-head self-attention module X at After concatenating with the original input X of the block, perform a 1x1 convolution, and then apply the Sigmoid function to obtain the gating vectors of the update gate and reset gate z t 、 r t :

[0016] At the same time, the current multi-head self-attention module output is integrated X at Together with the block input conditioned by the reset gate, we obtain the candidate feature states:

[0017] Where: tanh represents the activation function; Then, by updating the gate control vector z t Linearly combine input features X With candidate features h t , get the output features of the refined block:

[0018] Finally, the output features X re-in Batch normalization, GELU activation and regularization are performed in sequence ( Dropout ), obtain features Xre-fi ; and connect the features through a residual connection X re-fi With the original input of the block X Add together to get the final output of the feature refinement block: .

[0019] Based on the further optimization of the above scheme, the output MLP layer is specifically processed as follows: the final output of the feature refinement block is first processed by N refinement blocks to obtain the refined features. , and then the refined features are globally averaged pooled in the time dimension through the adaptive average pooling layer (AdaptiveAvgPool1d), and the time series features are aggregated into a vector of fixed length :

[0020] Finally, the pooled feature vector is fed into a multi-layer perceptron (MLP) for yield regression prediction:

[0021] The multi-layer perceptron (MLP) contains 1 to 2 hidden layers and uses the GELU activation function and regularization ( Dropout ) to prevent overfitting.

[0022] Based on the further optimization of the above solution, the step S4 is specifically as follows: first, an auxiliary prediction head is introduced to predict the yield output by the primary multi-scale feature extractor; Then, the main loss of the multi-layer perceptron (MLP) output prediction is calculated separately L ba Auxiliary loss predicted by the auxiliary prediction head output L au :

[0023]

[0024] Where: Y tr represents the actual production, obtained through the ground truth dataset; Y au represents the output of the prediction by the auxiliary prediction head through the primary multi-scale feature extractor; 、 Respectively represent the corresponding control error switching thresholds; Afterwards, the main prediction errors are obtained respectively E c , auxiliary prediction errorE au and their corresponding absolute values ​​| E c |、| E au |:

[0025] Use Exponential Moving Average (EMA) to smooth the current batch| E c | p Quantiles to obtain the main forecast error scale e scale ; Then, initialize the penalty weight w st , and check whether the main error is significant. The significant error threshold is:

[0026] Where: s main represents the scaling factor of the main error prediction threshold; like , then the penalty weight is not updated w st ; like , then the penalty weight is determined by the following two situations w st Update: Case A: Collaborative extreme error: Condition 1: Main prediction error E c and auxiliary prediction error E au Same direction, that is sign ( E c )= sign ( E au ), Condition 2: Main prediction error E c and auxiliary prediction error E au The absolute error is large, that is: , where: s syn represents the error trigger factor; If the above two conditions are met at the same time, the collaborative penalty enhancement factor is obtained :

[0027] Where: represents the collaborative penalty strength, which is based on the main prediction error E c Make corresponding choices based on the direction (overestimation or underestimation); Represents the penalty index, which is used to regulate the rate of change of the synergistic penalty enhancement factor; Case B: Judgment changes dramatically: Main prediction Y c and auxiliary prediction Y au The absolute difference between , where: s shift represents the prediction difference threshold factor; If the above conditions are met, the transition penalty enhancement factor is obtained :

[0028] Where: Indicates the transition penalty strength, which is based on the main prediction error E c Make corresponding choices based on the direction (overestimation or underestimation); Selecting collaborative penalty enhancement factors and the transition penalty enhancement factor The larger of the two (if one of the situations is not triggered, the corresponding penalty enhancement factor is considered to be 1), and the final penalty enhancement factor is obtained:

[0029] Finally, the total loss function is obtained:

[0030] Where: mean( ) represents the mean function; Represents the weight coefficient of auxiliary loss.

[0031] The following are the technical effects of the solution of the present invention: The improved corn yield prediction model of the present invention adopts a primary multi-scale feature extractor and utilizes parallel gated convolution models with different convolution kernel sizes, so that the prediction model can simultaneously capture the local patterns of the input time series at different scales, while ensuring that each convolution branch can selectively focus on and transmit information, enhance the flexibility and effectiveness of feature extraction, and thus improve the dynamic adjustment ability of the prediction model; at the same time, the improved corn yield prediction model of the present invention can also adaptively decide whether to rely more on the contextual information captured by the self-attention mechanism, or to retain more of the original input information, or to fuse the two according to the characteristics of the current features. It has strong self-learning ability and good environmental adaptability, and provides more refined control for the flow and transformation of features, so that the model can effectively filter noise, enhance important signals, and has stronger robustness and expression ability. In addition, the present invention introduces an adaptive penalty loss function. Through the synergistic effect of the adaptive penalty loss function and the improved corn yield prediction model, the loss function can perceive and utilize information of different processing depths within the model, dynamically combine the performance of the auxiliary prediction with the error level of the main prediction, and form a conditional and interactive penalty. Therefore, for the model with a clear staged processing flow, that is, the improved corn yield prediction model, the loss function is directly obtained by utilizing the dynamic relationship between the prediction values ​​of different stages to adapt to the asymmetric risks of different error directions in the yield prediction, while taking into account the sensitivity to small errors and the robustness to large errors, thereby improving the prediction accuracy and real-time responsiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a structural block diagram of the improved corn yield prediction model in an embodiment of the present invention.

[0033] Figure 2 This is a scatter plot of corn yield prediction accuracy in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present invention will be described clearly and completely below. In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are proposed to facilitate a thorough understanding of the embodiments of the present invention.

[0035] Example 1: A corn yield prediction method based on the collaboration of UAV and satellite remote sensing includes: Step S1: construct a multi-source basic dataset, including a ground truth dataset, a UAV remote sensing dataset, and a satellite remote sensing dataset; The method for obtaining the real data set is as follows: first, a set of representative sampling points and their high-precision geographic coordinates are planned and determined within the target area; then, during the corn crop maturity and harvest period, field measurements are conducted at the planned sampling points to obtain the corresponding yield data, thereby constructing a ground truth data set.

[0036] The construction of the UAV remote sensing dataset is to fly UAVs equipped with multispectral sensors over planned sampling points during the critical growth period of corn crops to obtain UAV remote sensing images with ultra-high spatial resolution, which serve as a high-quality data source for subsequent model training; the construction of the satellite remote sensing dataset is to obtain multispectral satellite remote sensing images such as Sentinel-2, covering the entire target area and the time range covering the critical growth period of corn, through cloud platforms such as Google Earth Engine (GEE), and perform standardized preprocessing (including cloud screening, cloud removal, atmospheric correction, etc.).

[0037] Step S2: Extract key feature images from the remote sensing dataset (initially, mainly UAV remote sensing dataset) for the phenological period of corn yield estimation and construct a feature dataset; specifically, extract remote sensing feature values ​​of each sampling point (i.e., remote sensing image) in the key time window of the entire corn growing season; the remote sensing feature values ​​include the maximum and average values ​​of the normalized vegetation index NDVI, the maximum and average values ​​of the vegetation near-infrared reflectance index NIRv, the maximum and average values ​​of the enhanced vegetation index EVI, the maximum and average values ​​of the kernel normalized vegetation index kNDVI, and the maximum and average values ​​of the normalized difference water index NDWI; match the yield data of the real dataset with the corresponding remote sensing feature values ​​for subsequent model training.

[0038] NDVI:

[0039] Where: 、 Represent the surface reflectance in the near-infrared band and the red light band respectively; NIRv:

[0040] EVI:

[0041] Where: Indicates the surface reflectance in the blue light band; G represents the gain factor (usually 2.5); C 1. C 2 represents the atmospheric correction factor (usually C 1=6, C2=7.5); L represents the soil adjustment factor (usually 1); kNDVI:

[0042] Where: K(x,y) represents the kernel function. In this embodiment, radial basis function is used:

[0043] Where: represents the length scale constant, usually taken from all samples and the mean of the differences; NDWI:

[0044] Where: Represents the surface reflectance in the green light band.

[0045] Step S3: constructing an improved corn yield prediction model, and performing model training, verification, and evaluation in sequence; The improved corn yield prediction model consists of a primary multi-scale feature extractor, a two-layer feature refinement block, and an output MLP layer; Among them, the primary multi-scale feature extractor obtains input features and outputs them to the two-layer feature refinement block. The input features are first embedded through a 1x1 convolution layer. Input_Dim Map to B ase_Channels , followed by batch normalization and GELU activation:

[0046] Where: Represents input features; Indicates output; B represents the batch size, T represents the time step, D in Input feature dimension, C base The number of base channels of the embedding layer; Indicates 1x1 convolution processing; Indicates batch normalization processing; GELU( ) express GELU Activation function; After that, the embedded features are fed into a multi-scale gated convolution branch, i.e., multiple parallel gated convolution branches. Each branch uses a different convolution kernel to capture temporal limitations at different scales. The multi-scale gated convolution branch includes three parallel gated convolution branches: 3x3 gated convolution, 5x5 gated convolution, and 7x7 gated convolution. Gated convolution is a neural network component that combines a convolution operation with a gating mechanism: For input features 、 C in Represents the number of channels of the input layer, which is passed through a 1D convolutional layer whose output is twice the number of desired output channels C out Channel:

[0047] Then, H is divided into two parts along the channel dimension, namely the feature part With the gate control part ; Finally, the output of the gated convolution is achieved:

[0048] Where: represents element-wise product; Represents the Sigmoig function.

[0049] Finally, the outputs of each scale branch are spliced ​​in the channel dimension to form fused multi-scale features.

[0050] The feature refining blocks are as follows: Input features First replaced by , to adapt to the input format of the multi-head self-attention module; then, the permuted features are layer-normalized and fed into the multi-head self-attention module:

[0051]

[0052] Where: Normalization of the representation layer; h Indicates the number of attention heads; Represents the output projection matrix; Respectively represent the corresponding projection matrices; d k Dimensions representing queries and keys, ; The output of the multi-head self-attention module is regularized ( Dropout ) and dimension replacement, we get ; Afterwards, by the output of the multi-head self-attention module X at After concatenating with the original input X of the block, perform a 1x1 convolution, and then apply the Sigmoid function to obtain the gating vectors of the update gate and reset gatez t 、 r t :

[0053] At the same time, the current multi-head self-attention module output is integrated X at Together with the block input conditioned by the reset gate, we obtain the candidate feature states:

[0054] Where: tanh represents the activation function; Then, by updating the gate control vector z t Linearly combine input features X With candidate features h t , get the output features of the refined block:

[0055] Finally, the output features X re-in Batch normalization, GELU activation and regularization are performed in sequence ( Dropout ), obtain features X re-fi ; and connect the features through a residual connection X re-fi With the original input of the block X Add together to get the final output of the feature refinement block: .

[0056] The specific processing of the output MLP layer is as follows: the final output of the feature refinement block is first processed by N refinement blocks to obtain the refined features. , and then the refined features are globally averaged pooled in the time dimension through the adaptive average pooling layer (AdaptiveAvgPool1d), and the time series features are aggregated into a vector of fixed length :

[0057] Finally, the pooled feature vector is fed into a multi-layer perceptron (MLP) for yield regression prediction:

[0058] The multi-layer perceptron (MLP) contains 1 to 2 hidden layers and uses the GELU activation function and regularization ( Dropout ) to prevent overfitting.

[0059] Step S4: introducing an adaptive penalty loss function into the improved corn yield prediction model, utilizing the internal state of the improved corn yield prediction model to quantify the consistency and conflict within the model, adaptively adjusting the penalty intensity, and obtaining an adjusted loss function. Specifically, first, introducing an auxiliary prediction head to predict the yield output by the primary multi-scale feature extractor; Then, the main loss of the multi-layer perceptron (MLP) output prediction is calculated separately L ba Auxiliary loss predicted by the auxiliary prediction head output L au :

[0060]

[0061] Where: Y tr represents the actual production, obtained through the ground truth dataset; Y au represents the output of the prediction by the auxiliary prediction head through the primary multi-scale feature extractor; 、 Respectively represent the corresponding control error switching thresholds; Afterwards, the main prediction errors are obtained respectively E c , auxiliary prediction error E au and their corresponding absolute values ​​| E c |、| E au |:

[0062] Use Exponential Moving Average (EMA) to smooth the current batch| E c | p Quantiles to obtain the main forecast error scale e scale ; Then, initialize the penalty weight w st , and check whether the main error is significant. The significant error threshold is:

[0063] Where: s main represents the scaling factor of the main error prediction threshold; like , then the penalty weight is not updated wst ; like , then the penalty weight is determined by the following two situations w st Update: Case A: Collaborative extreme error: Condition 1: Main prediction error E c and auxiliary prediction error E au Same direction, that is sign ( E c )= sign ( E au ), Condition 2: Main prediction error E c and auxiliary prediction error E au The absolute error is large, that is: , where: s syn represents the error trigger factor; If the above two conditions are met at the same time, the collaborative penalty enhancement factor is obtained :

[0064] Where: represents the collaborative penalty strength, which is based on the main prediction error E c Make corresponding choices based on the direction (overestimation or underestimation); Represents the penalty index, which is used to regulate the rate of change of the synergistic penalty enhancement factor; Case B: Judgment changes dramatically: Main prediction Y c and auxiliary prediction Y au The absolute difference between , where: s shift represents the prediction difference threshold factor; If the above conditions are met, the transition penalty enhancement factor is obtained :

[0065] Where: Indicates the transition penalty strength, which is based on the main prediction error E c Make corresponding choices based on the direction (overestimation or underestimation); Selecting collaborative penalty enhancement factors and the transition penalty enhancement factor The larger of the two (if one of the situations is not triggered, the corresponding penalty enhancement factor is considered to be 1), and the final penalty enhancement factor is obtained:

[0066] Finally, the total loss function is obtained:

[0067] Where: mean( ) represents the mean function; Represents the weight coefficient of auxiliary loss.

[0068] Example 2: As another preferred embodiment of the technical solution of the present invention, based on the solution of Example 1 above, in step S3, model training, verification and evaluation are carried out in sequence as follows: In order to obtain a robust evaluation of the generalization ability of the model, a k-fold cross-validation method is adopted. In this embodiment, a three-fold cross-validation method is adopted, which is specifically as follows: first, the complete feature data set is evenly divided into three subsets of similar size and non-overlapping by random sampling, and then three independent rounds of iterative training and validation are performed; in each round of validation, one of the subsets is used as the validation set of that round, and the union of the remaining two subsets is used as the training set.

[0069] The model is iteratively trained on the training set. After each training epoch, performance metrics, including loss, are calculated and recorded on the validation set. Model performance is then verified. After completing the cross-validation process, the model's predictive capabilities are comprehensively evaluated. In this example, the performance metrics achieved include a coefficient of determination (R²) of 0.7761, a root mean square error (RMSE) of 1855.0748 kg / ha, and a mean absolute error (MAE) of 1384.0231 kg / ha, demonstrating that the present invention can achieve highly accurate yield predictions.

[0070] Finally, save the final model; after the evaluation is completed, in order to obtain a single model that can be used for the final prediction application, the model can be retrained on all feature datasets and the trained model weight file can be saved.

[0071] In the above training process, the following training configuration is uniformly adopted: the total number of training rounds is 300 rounds; the loss function uses Huber Loss; the optimizer uses AdamW; the initial learning rate is set to 0.00005, and the DROPOUT regularization rate is set to 0.15; the learning rate scheduling adopts the ReduceLROnPlateau strategy. This strategy continuously monitors the loss value of the current validation set during each round of training. When the loss value of the validation set no longer decreases significantly within the preset number of rounds, the current learning rate is adaptively reduced.

Claims

1. A corn yield prediction method based on the collaboration of UAV and satellite remote sensing, characterized by: include: Step S1: construct a multi-source basic dataset, including a ground truth dataset, a UAV remote sensing dataset, and a satellite remote sensing dataset; Step S2: extracting key feature images from the remote sensing dataset for the phenological period of corn yield estimation to construct a feature dataset; Step S3: constructing an improved corn yield prediction model, and performing model training, verification, and evaluation in sequence; Step S4: introducing an adaptive penalty loss function into the improved corn yield prediction model, utilizing the internal state of the improved corn yield prediction model to quantify the consistency and conflict within the model, adaptively adjusting the penalty intensity, and obtaining the adjusted loss function.

2. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 1, characterized in that: In step S1, the method for obtaining the real data set is as follows: first, a set of representative sampling points and their high-precision geographic coordinates are planned and determined within the target area; then, during the corn crop maturity and harvest period, field measurements are performed at the planned sampling points to obtain corresponding yield data, thereby constructing a ground truth data set.

3. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 1 or 2, characterized in that: The construction of the UAV remote sensing dataset involves flying drones equipped with multispectral sensors over planned sampling points during the critical growth period of corn crops to obtain UAV remote sensing images with ultra-high spatial resolution, which serve as a high-quality data source for subsequent model training. The construction of the satellite remote sensing dataset involves obtaining Sentinel-2 multispectral satellite remote sensing images covering the entire target area and the time range covering the critical growth period of corn through the Google Earth Engine cloud platform, and performing standardized preprocessing.

4. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 1 or 3, characterized in that: The step S2 specifically comprises: extracting remote sensing characteristic values ​​of each sampling point in a key time window of the entire corn growing season; the remote sensing characteristic values ​​include the maximum and average values ​​of the normalized vegetation index NDVI, the maximum and average values ​​of the vegetation near-infrared reflectance index NIRv, the maximum and average values ​​of the enhanced vegetation index EVI, the maximum and average values ​​of the kernel normalized vegetation index kNDVI, and the maximum and average values ​​of the normalized difference water index NDWI; matching the yield data of the real data set with the corresponding remote sensing characteristic values ​​for subsequent model training.

5. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 4, characterized in that: The improved corn yield prediction model includes a primary multi-scale feature extractor, a two-layer feature refinement block and an output MLP layer; Among them, the primary multi-scale feature extractor obtains input features and outputs them to the two-layer feature refinement block. The input features are first embedded through a 1x1 convolution layer. Input_Dim Map to B ase_Channels , followed by batch normalization and GELU activation: Where: Represents input features; Indicates output; B represents the batch size, T represents the time step, D in Input feature dimension, C base The number of base channels of the embedding layer; Indicates 1x1 convolution processing; Indicates batch normalization processing; GELU( ) express GELU Activation function; After that, the embedded features are fed into a multi-scale gated convolution branch, i.e., multiple parallel gated convolution branches, each using a different convolution kernel to capture temporal limitations at different scales. Finally, the outputs of each scale branch are spliced ​​in the channel dimension to form fused multi-scale features.

6. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 5, characterized in that: The feature refining block is specifically: Input features First replaced by , to adapt to the input format of the multi-head self-attention module; then, the permuted features are layer-normalized and fed into the multi-head self-attention module: Where: Normalization of the representation layer; h Indicates the number of attention heads; Represents the output projection matrix; Respectively represent the corresponding projection matrices; d k Dimensions representing queries and keys, ; The output of the multi-head self-attention module is regularized and dimensionally permuted to obtain ; Afterwards, by the output of the multi-head self-attention module X at After concatenating with the original input X of the block, perform a 1x1 convolution, and then apply the Sigmoid function to obtain the gating vectors of the update gate and reset gate z t 、 r t : At the same time, the current multi-head self-attention module output is integrated X at Together with the block input conditioned by the reset gate, we obtain the candidate feature states: Where: tanh represents the activation function; Then, by updating the gate control vector z t Linearly combine input features X With candidate features h t , get the output features of the refined block: Finally, the output features X re-in Perform batch normalization, GELU activation and regularization processing in sequence to obtain features X re-fi ; and connect the features through a residual connection X re-fi With the original input of the block X Add together to get the final output of the feature refinement block: 。 7. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 5, characterized in that: The specific processing of the output MLP layer is as follows: the final output of the feature refinement block is first processed by N refinement modules to obtain the refined features , and then the refined features are globally averaged pooled in the time dimension through the adaptive average pooling layer to aggregate the time series features into a vector of fixed length : Finally, the pooled feature vector is fed into a multi-layer perceptron for yield regression prediction: The multilayer perceptron contains 1 to 2 hidden layers and uses the GELU activation function and regularization to prevent overfitting.

8. The corn yield prediction method based on the collaboration of UAV and satellite remote sensing according to claim 7, characterized in that: The step S4 specifically includes: first, introducing an auxiliary prediction head for predicting the yield output by the primary multi-scale feature extractor; Then, the main loss of the multilayer perceptron output prediction is calculated separately L ba Auxiliary loss predicted by the auxiliary prediction head output L au : Where: Y tr represents the actual production, obtained through the ground truth dataset; Y au represents the output of the prediction by the auxiliary prediction head through the primary multi-scale feature extractor; 、 Respectively represent the corresponding control error switching thresholds; Afterwards, the main prediction errors are obtained respectively E c , auxiliary prediction error E au and their corresponding absolute values ​​| E c |、| E au |: Use exponential moving average to smooth the current batch| E c | p Quantiles to obtain the main forecast error scale e scale ; Then, initialize the penalty weight w st , and check whether the main error is significant. The significant error threshold is: Where: s main represents the scaling factor of the main error prediction threshold; like , then the penalty weight is not updated w st ; like , then the penalty weight is determined by the following two situations w st Update: Case A: Collaborative extreme error: Condition 1: Main prediction error E c and auxiliary prediction error E au Same direction, that is sign ( E c )= sign ( E au ), Condition 2: Main prediction error E c and auxiliary prediction error E au The absolute error is large, that is: , where: s syn represents the error trigger factor; If the above two conditions are met at the same time, the collaborative penalty enhancement factor is obtained : Where: represents the collaborative penalty strength, which is based on the main prediction error E c Select the corresponding direction; Represents the penalty index, which is used to regulate the rate of change of the synergistic penalty enhancement factor; Case B: Judgment changes dramatically: Main prediction Y c and auxiliary prediction Y au The absolute difference between , where: s shift represents the prediction difference threshold factor; If the above conditions are met, the transition penalty enhancement factor is obtained : Where: Indicates the transition penalty strength, which is based on the main prediction error E c Select the corresponding direction; Selecting collaborative penalty enhancement factors and the transition penalty enhancement factor The larger of them gets the final penalty enhancement factor: Finally, the total loss function is obtained: Where: mean( ) represents the mean function; Represents the weight coefficient of auxiliary loss.

Images