Crop yield estimation method and device, electronic equipment and storage medium

Abstract

The invention provides a crop yield estimation method and device, electronic equipment and a storage medium, and relates to the technical field of agricultural information technology, remote sensing monitoring and artificial intelligence crossing, and the method comprises the steps: carrying out VOD space-time fusion processing based on multi-source microwave remote sensing data through a VOD space-time fusion model, and obtaining VOD fusion data; simulating crop growth based on the environment data and the crop planting data through a target crop growth model to obtain crop physiological process data; constructing multi-source feature data based on the VOD fusion data, the crop physiological process data and the environmental data; and through a double-flow space-time deep learning model, carrying out space-time feature extraction and fusion on the multi-source feature data to obtain a biomass estimated value, and further calculating a crop yield estimated value. According to the method, the crop yield is estimated based on fusion of the multi-source microwave remote sensing data, the crop growth model and the deep learning model, the restriction of a single technical normal form is broken through, and the accuracy of an estimation result can be improved.

Description

Technical Field

[0001] This invention relates to the fields of agricultural information technology, remote sensing monitoring and artificial intelligence, and in particular to a method, apparatus, electronic device and storage medium for crop yield prediction. Background Technology

[0002] Accurate crop yield forecasting is crucial for ensuring national food security, guiding agricultural production, stabilizing agricultural product markets, and addressing climate change. Traditional yield estimation methods mainly rely on manual field surveys and statistical reporting, which suffer from problems such as high subjectivity, poor timeliness, high costs, and insufficient spatial coverage.

[0003] To achieve accurate crop yield assessment, existing technologies mainly revolve around two core aspects: growth information acquisition and yield modeling and analysis. However, the technological paradigms currently used in these two aspects both have inherent limitations.

[0004] In the process of acquiring growth information, existing observation technologies all have their inherent limitations, resulting in irreconcilable contradictions between key indicators such as spatiotemporal resolution, data continuity and coverage, thus making it impossible to comprehensively, continuously and unbiasedly characterize the true growth status of crops throughout their entire growth period.

[0005] In the yield modeling and analysis phase, existing model building methods mainly follow two technical routes, each with significant shortcomings. One type is analytical models based on prior knowledge and mechanistic processes. These models attempt to predict yield by mathematically describing the physical and physiological mechanisms of crop growth. However, their inherent complexity and strong dependence on massive amounts of high-precision driving parameters greatly limit their adaptability and accuracy in large-scale applications. The other type is data-driven deep learning models. While these models can autonomously learn the mapping relationship between yield and various influencing factors from historical data using statistical or machine learning algorithms, their stability and generalization ability often face severe challenges in the context of limited sample data commonly found in agriculture.

[0006] In summary, existing methods for acquiring growth information and for yield modeling and analysis are limited in their paradigms, resulting in poor accuracy in crop yield prediction. Therefore, how to overcome the constraints of a single technological paradigm and provide a new method for crop yield prediction to improve the accuracy of the results is a pressing issue in this field. Summary of the Invention

[0007] This invention provides a method, apparatus, electronic device, and storage medium for crop yield prediction, which can predict crop yield based on the fusion of multi-source microwave remote sensing data, crop growth models, and deep learning models, breaking through the limitations of a single technical paradigm and improving the accuracy of crop yield prediction results.

[0008] This invention provides a method for predicting crop yield, comprising: By using the vegetation optical thickness (VOD) spatiotemporal fusion model, VOD spatiotemporal fusion processing is performed based on multi-source microwave remote sensing data to obtain VOD fused data. By using a target crop growth model, crop growth is simulated based on environmental and crop planting data to obtain crop physiological process data; Based on the VOD fusion data, the crop physiological process data, and the environmental data, multi-source feature data is constructed. Using a dual-stream spatiotemporal deep learning model, spatiotemporal features are extracted and fused from the multi-source feature data to estimate biomass and obtain a biomass estimate. Based on the estimated biomass, the estimated crop yield is calculated.

[0009] According to the crop yield prediction method provided by the present invention, the multi-source microwave remote sensing data includes low-frequency satellite microwave remote sensing data and high-frequency UAV microwave remote sensing data. The method involves performing VOD spatiotemporal fusion processing on the multi-source microwave remote sensing data using a vegetation optical thickness (VOD) spatiotemporal fusion model to obtain VOD fused data, including: The VOD inversion algorithm is used to perform VOD inversion on the low-frequency satellite microwave remote sensing data to obtain satellite VOD data, and the VOD inversion algorithm is used to perform VOD inversion on the high-frequency UAV microwave remote sensing data to obtain UAV VOD data. From the satellite VOD data, obtain satellite VOD sub-data corresponding to the collection time point of the UAV VOD data; The VOD spatiotemporal fusion model is trained using the satellite VOD sub-data and the UAV VOD data to obtain the VOD spatiotemporal fusion model. The satellite VOD data is input into the VOD spatiotemporal fusion model for VOD spatiotemporal fusion processing to obtain the VOD fused data output by the VOD spatiotemporal fusion model.

[0010] According to the crop yield prediction method provided by the present invention, the loss function used in the training process of the VOD spatiotemporal fusion model is determined based on spatial reconstruction loss, temporal smoothing constraint and regularization term.

[0011] According to a crop yield prediction method provided by the present invention, before obtaining crop physiological process data by simulating crop growth based on environmental data and crop planting data using a target crop growth model, the method further includes: The initial crop growth model is simplified and calibrated regionally to obtain the target crop growth model; wherein the initial crop growth model is constructed based on the light energy utilization rate model. The process involves simulating crop growth using a target crop growth model, based on environmental and crop planting data, to obtain crop physiological process data, including: The environmental data and the crop planting data are input into the target crop growth model to perform regionalized crop growth simulation and obtain the crop physiological process data. The crop physiological process data include at least one of potential biomass, water stress coefficient, heat stress coefficient, and phenological period.

[0012] According to a crop yield prediction method provided by the present invention, the multi-source feature data includes multi-channel feature maps of key phenological dates and multi-source feature time series data for each pixel. The method involves extracting and fusing spatiotemporal features from the multi-source feature data using a dual-stream spatiotemporal deep learning model to predict biomass and obtain a biomass prediction value. This includes: The multi-channel feature map of the key phenological date is input into the spatial feature extraction module of the dual-stream spatiotemporal deep learning model to perform spatial feature extraction, and the spatial feature vector output by the spatial feature extraction module is obtained. The multi-source feature time series data of each pixel is input into the time feature extraction module of the dual-stream spatiotemporal deep learning model to extract time features and obtain the time feature vector output by the time feature extraction module. The spatial feature vector and the temporal feature vector are input into the feature fusion module of the dual-stream spatiotemporal deep learning model to perform feature fusion, and the feature fusion vector output by the feature fusion module is obtained. The feature fusion vector is input into the fully connected layer of the dual-stream spatiotemporal deep learning model to perform biomass regression, thereby obtaining the biomass estimate output by the fully connected layer.

[0013] According to a crop yield prediction method provided by the present invention, the construction of multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data includes: Based on the crop growth cycle, the VOD fusion data is reconstructed into a time series to obtain the VOD time series for each pixel; VOD dynamic feature data is calculated based on the VOD time series of each pixel; wherein, the VOD dynamic feature data includes at least one of the following: maximum VOD value, time to reach maximum VOD, cumulative VOD, VOD growth rate, and VOD decay rate; Based on the phenological periods in the crop physiological process data, key phenological dates are determined; From the VOD fusion data, the VOD dynamic feature data of each pixel, and the crop physiological process data, the first target feature data corresponding to the key phenological date is obtained, and the data is stacked to obtain a multi-channel feature map of the key phenological date; Second target feature data within a time window from sowing to the key phenological date are obtained from the VOD fusion data, the crop physiological process data, and the environmental data to construct multi-source feature time series data for each pixel.

[0014] According to a crop yield prediction method provided by the present invention, the step of calculating the crop yield prediction based on the biomass prediction value includes: The dynamic harvest index for each pixel is calculated based on the water stress coefficient and heat stress coefficient in the crop physiological process data. The dynamic harvest index of each pixel is multiplied by the corresponding biomass estimate to obtain the yield estimate of each pixel. Based on the estimated yield of each pixel, spatial aggregation calculation is performed to obtain the estimated crop yield.

[0015] The present invention also provides a crop yield prediction device, comprising: The VOD fusion module is used to perform VOD spatiotemporal fusion processing based on multi-source microwave remote sensing data through a VOD spatiotemporal fusion model to obtain VOD fused data. The growth simulation module is used to simulate crop growth based on environmental and crop planting data using a target crop growth model to obtain crop physiological process data. The data construction module is used to construct multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data; The biomass estimation module is used to extract and fuse spatiotemporal features from the multi-source feature data using a dual-stream spatiotemporal deep learning model in order to estimate biomass and obtain a biomass estimate. The yield estimation module is used to calculate the crop yield estimate based on the biomass estimate.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the crop yield prediction method as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the crop yield prediction method as described in any of the preceding claims.

[0018] The crop yield prediction method, apparatus, electronic device, and storage medium provided by this invention first use a VOD spatiotemporal fusion model to perform VOD spatiotemporal fusion processing on microwave remote sensing data from multiple sources (i.e., multi-source microwave remote sensing data), generating VOD fused data with high spatiotemporal resolution, long time series, and wide coverage. This solves the inherent defects of a single remote sensing data source and provides a high-quality data foundation for subsequent crop yield prediction. Then, crop growth is simulated using a target crop growth model to generate crop physiological process data with clear physical meaning. This data is used as prior knowledge of physical mechanisms and, together with VOD fused data and environmental data, constructs multi-source feature data. This data is then input into a dual-stream spatiotemporal deep learning model for spatiotemporal feature extraction and fusion to predict biomass, which is then used to calculate the crop yield prediction value. This deep learning paradigm guided by physical mechanisms not only overcomes the shortcomings of purely data-driven models, such as poor interpretability and overfitting in small samples, by utilizing physical constraints and enhancing the model's generalization ability, but also achieves simultaneous capture of the spatial heterogeneity and temporal dependence of crop growth through the collaborative extraction and fusion of spatiotemporal features of data. Compared with existing yield modeling and analysis methods, this invention makes fuller use of information and ultimately obtains more accurate and reliable biomass and crop yield estimates. In summary, this invention provides a crop yield prediction method based on the fusion of multi-source microwave remote sensing data, crop growth models, and deep learning models, breaking through the limitations of a single technical paradigm and improving the accuracy of crop yield prediction results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts of the crop yield prediction method provided by the present invention; Figure 2 This is the second flowchart illustrating the crop yield prediction method provided by the present invention; Figure 3 This is the third flowchart of the crop yield prediction method provided by the present invention; Figure 4 This is a schematic diagram of the crop yield prediction device provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Accurate crop yield forecasting is crucial for ensuring national food security, guiding agricultural production, stabilizing agricultural product markets, and addressing climate change. Traditional yield estimation methods mainly rely on manual field surveys and statistical reporting, which suffer from problems such as high subjectivity, poor timeliness, high costs, and insufficient spatial coverage.

[0023] To achieve accurate crop yield assessment, existing technologies mainly revolve around two core aspects: growth information acquisition and yield modeling and analysis. However, the technological paradigms currently used in these two aspects both have inherent limitations.

[0024] In the growth information acquisition stage, with the development of remote sensing technology, especially the application of optical remote sensing (such as MODIS (Moderate-resolution Imaging Spectroradiometer), Landsat, Sentinel-2) and its derived spectral vegetation indices (such as NDVI (Normalized Difference Vegetation Index) and EVI (Enhanced Vegetation Index)), a data foundation has been provided for large-scale crop monitoring.

[0025] However, analysis shows that optical remote sensing is easily affected by atmospheric conditions such as clouds, rain, and fog, making it difficult to obtain continuous time series data. Furthermore, in the middle and late stages of crop growth, spectral indices are prone to saturation, making it difficult to accurately reflect the true accumulation of biomass.

[0026] In comparison, microwave remote sensing can penetrate clouds, enabling all-day, all-weather Earth observation, and has a stronger penetrating ability to vegetation canopy. VOD (Vegetation Optical Depth), as a key indicator for microwave remote sensing inversion, has a clear physical correlation with the water content and biomass within vegetation, and is more suitable for characterizing the dynamic changes in crop biomass in the middle and late stages compared to the aforementioned spectral vegetation indices.

[0027] Further analysis reveals that microwave remote sensing data from a single platform also has inherent limitations in terms of spatiotemporal capabilities. For example, while low-frequency satellite microwave remote sensing has long time series and wide coverage, its spatial resolution is usually coarse; while emerging UAV-borne or near-space platform microwave remote sensing has higher spatial resolution, but its time series is shorter and its coverage is limited, making it difficult to support continuous monitoring of large areas and long time series.

[0028] In the yield modeling and analysis stage, existing model construction methods mainly follow two technical routes, each with significant shortcomings. One type is based on analytical models grounded in prior knowledge and mechanistic processes, such as existing crop growth models like DSSAT (Decision Support System for Agrotechnology Transfer) and WOFOST (World Food Studies). These models can simulate crop growth and yield formation processes based on crop physiological and ecological mechanisms, but their accuracy heavily depends on the accuracy of initial parameters and driving data, and parameter calibration is cumbersome when applied regionally. The other type is data-driven deep learning models, such as CNN (Convolutional Neural Network) models and LSTM (Long Short-Term Memory) networks. These models excel at automatically learning complex features from massive amounts of data, but they typically have poor physical interpretability, and their stability and generalization ability often face severe challenges in small-sample agricultural scenarios.

[0029] In summary, existing methods for acquiring crop growth information and for yield modeling and analysis suffer from paradigmatic limitations and poor accuracy in crop yield prediction. Based on the above analysis, a new crop yield prediction method can be developed by comprehensively utilizing the advantages of multi-platform microwave remote sensing and effectively integrating the physical mechanisms of crop growth with the data-driven capabilities of deep learning. This method offers physical interpretability, high spatiotemporal accuracy, and strong generalization ability, thereby improving the accuracy of crop yield prediction results.

[0030] Based on the above, this invention proposes a crop yield prediction method, device, electronic device, and storage medium, which are described below in conjunction with... Figures 1-5 Describe it.

[0031] Figure 1 This is one of the flowcharts illustrating the crop yield prediction method provided by this invention, such as... Figure 1 As shown, the crop yield prediction method includes steps S110, S120, S130, S140 and S150.

[0032] Step S110: Using the vegetation optical thickness (VOD) spatiotemporal fusion model, VOD spatiotemporal fusion processing is performed based on multi-source microwave remote sensing data to obtain VOD fused data.

[0033] In this embodiment, the multi-source microwave remote sensing data includes low-frequency satellite microwave remote sensing data and high-frequency UAV microwave remote sensing data.

[0034] It should be noted that when acquiring multi-source microwave remote sensing data, the spatial dimension is a target planting area and the temporal dimension is a crop growth cycle. For example, multi-source microwave remote sensing data of a target crop for an entire growth cycle can be collected from a farm.

[0035] The VOD spatiotemporal fusion model is trained using training samples constructed from multi-source microwave remote sensing data.

[0036] Specifically, VOD inversion algorithms are used to perform VOD inversion on low-frequency satellite microwave remote sensing data to obtain satellite VOD data, and VOD inversion is also performed on high-frequency UAV microwave remote sensing data to obtain UAV VOD data. Then, satellite VOD sub-data corresponding to the acquisition time points of UAV VOD data are obtained from the satellite VOD data. The preset VOD spatiotemporal fusion model is trained using the satellite VOD sub-data and UAV VOD data to obtain the VOD spatiotemporal fusion model.

[0037] Then, the satellite VOD data is input into the VOD spatiotemporal fusion model for VOD spatiotemporal fusion processing to obtain the VOD fused data output by the VOD spatiotemporal fusion model.

[0038] The specific execution process can be found in the following examples, which will not be elaborated here.

[0039] Step S120: Using the target crop growth model, crop growth is simulated based on environmental data and crop planting data to obtain crop physiological process data.

[0040] Environmental data includes meteorological and soil data. The meteorological data is diurnal, specifically including maximum and minimum temperatures, precipitation, solar radiation, and humidity, and can be obtained from reanalysis data products such as ERA5 (ECMWF Reanalysis v5). The soil data is static, including soil type, percentage of clay content, percentage of sand content, soil organic carbon content, field capacity, and wilting point.

[0041] Crop planting data includes the planting area and the crop growth cycle. The planting area can be derived from officially published maize planting distribution maps. These maps are spatial maps that identify which pixels within a region are planted with the target crop; they are typically represented by a binary mask, where "1" represents a target crop pixel and "0" represents a non-target crop pixel. The crop growth cycle defines the sowing and harvesting dates of the target crop within the target planting area. It should be understood that when the target planting area is large, it may involve multiple sowing and harvesting dates; therefore, the average sowing and harvesting dates can be used.

[0042] Environmental and crop planting data are used as drivers and input into a target crop growth model to simulate the crop's daily potential growth and stress conditions, thereby obtaining crop physiological process data. This data includes at least one of potential biomass, water stress coefficient, heat stress coefficient, and phenological stage.

[0043] Step S130: Construct multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data.

[0044] The multi-source feature data includes multi-channel feature maps of key phenological dates and multi-source feature time series data for each pixel.

[0045] The construction process of multi-source feature data is as follows: Based on the crop growth cycle, the VOD fusion data is reconstructed into a time series to obtain the VOD time series for each pixel; based on the VOD time series of each pixel, VOD dynamic feature data is calculated; wherein, the VOD dynamic feature data includes at least one of the following: maximum VOD value, time to reach maximum VOD, cumulative VOD, VOD growth rate, and VOD decay rate; based on the phenological periods in the crop physiological process data, key phenological dates are determined; from the VOD fusion data, the VOD time series of each pixel, and the crop physiological process data, first target feature data corresponding to the key phenological dates are obtained, and stacked to obtain a multi-channel feature map of the key phenological dates; from the VOD time series of each pixel, the crop physiological process data, and the environmental data, second target feature data within the time window from sowing to the key phenological date are obtained to construct the multi-source feature time series data for each pixel. The specific construction process can be referred to in the following embodiments, which will not be elaborated here.

[0046] Step S140: Using a dual-stream spatiotemporal deep learning model, spatiotemporal features are extracted and fused from the multi-source feature data to estimate biomass and obtain a biomass estimate.

[0047] The dual-stream spatiotemporal deep learning model adopts a dual-stream network architecture, which includes two extraction streams: a spatial feature extraction stream and a temporal feature extraction stream. These two extraction streams extract spatial and temporal features from multi-source feature data respectively, and then fuse and regress to output a biomass prediction value.

[0048] Specifically, the multi-channel feature maps of key phenological dates are input into the spatial feature extraction module of the dual-stream spatiotemporal deep learning model for spatial feature extraction, yielding a spatial feature vector output by the module. Simultaneously, the multi-source feature time series data of each pixel is input into the temporal feature extraction module of the model for temporal feature extraction, yielding a temporal feature vector output by the module. Then, the spatial and temporal feature vectors are input into the feature fusion module of the model for feature fusion, yielding a fused feature vector output by the module. Finally, the fused feature vector is input into the fully connected layer of the model for biomass regression, obtaining the biomass prediction value output by the fully connected layer. The specific execution process can be found in the following embodiment, and will not be elaborated upon here.

[0049] Biomass can be learned and predicted from complex spatiotemporal features using a dual-stream spatiotemporal deep learning model.

[0050] Furthermore, the dual-stream spatiotemporal deep learning model is trained using multi-source feature data as sample data and ground-measured crop biomass data as ground truth. During training, mean squared error (MSE) is used as the loss function.

[0051] Step S150: Calculate the crop yield estimate based on the biomass estimate.

[0052] In one embodiment, a preset harvest index is obtained, which is a preset fixed value. Then, the preset harvest index is multiplied by the estimated biomass value of each pixel to obtain the estimated yield value of each pixel. Based on the estimated yield value of each pixel, spatial aggregation calculation is performed to obtain the estimated crop yield value.

[0053] In another embodiment, the dynamic harvest index of each pixel is calculated based on the water stress coefficient and heat stress coefficient in the crop physiological process data; the dynamic harvest index of each pixel is multiplied by the corresponding biomass estimate to obtain the yield estimate of each pixel; and spatial aggregation calculation is performed based on the yield estimate of each pixel to obtain the crop yield estimate.

[0054] The crop yield prediction method provided in this invention first uses a VOD spatiotemporal fusion model to perform VOD spatiotemporal fusion processing on microwave remote sensing data from multiple sources (i.e., multi-source microwave remote sensing data), generating VOD fused data with high spatiotemporal resolution, long time series, and wide coverage. This solves the inherent defects of a single remote sensing data source and provides a high-quality data foundation for subsequent crop yield prediction. Then, crop growth is simulated using a target crop growth model to generate crop physiological process data with clear physical meaning. This data is used as prior knowledge of physical mechanisms and, together with the VOD fused data and environmental data, constructs multi-source feature data. This data is then input into a dual-stream spatiotemporal deep learning model for spatiotemporal feature extraction and fusion to predict biomass, which is then used to calculate the crop yield prediction value. This deep learning paradigm guided by physical mechanisms not only overcomes the shortcomings of purely data-driven models, such as poor interpretability and overfitting in small samples, by utilizing physical constraints and enhancing the model's generalization ability, but also achieves simultaneous capture of the spatial heterogeneity and temporal dependence of crop growth through the collaborative extraction and fusion of spatiotemporal features of data. Compared with existing yield modeling and analysis methods, this invention makes fuller use of information, ultimately obtaining more accurate and reliable biomass and crop yield estimates. In summary, this invention provides a crop yield prediction method based on the fusion of multi-source microwave remote sensing data, crop growth models, and deep learning models, breaking through the limitations of a single technical paradigm and improving the accuracy of crop yield prediction results.

[0055] Based on any of the above embodiments, the multi-source microwave remote sensing data includes low-frequency satellite microwave remote sensing data and high-frequency UAV microwave remote sensing data. Figure 2 This is the second flowchart illustrating the crop yield prediction method provided by this invention, as shown below. Figure 2 As shown, step S110 includes: step S111, step S112, step S113 and step S114.

[0056] Step S111: VOD inversion is performed on the low-frequency satellite microwave remote sensing data using a VOD inversion algorithm to obtain satellite VOD data, and VOD inversion is performed on the high-frequency UAV microwave remote sensing data to obtain UAV VOD data.

[0057] Microwave remote sensing data acquired from low-frequency satellite platforms, such as SMOS (Soil Moisture and Ocean Salinity) or SMAP (Soil Moisture Active Passive), are denoted as low-frequency satellite microwave remote sensing data. These data are characterized by long time series and wide coverage, but lower spatial resolution and higher temporal resolution. Furthermore, low-frequency satellite microwave remote sensing data is in the L-band, where the microwave wavelength is relatively long (approximately 21 cm), exhibiting a stronger physical correlation with biomass.

[0058] Meanwhile, during key phenological stages of crop growth (such as the jointing stage and heading stage), high-frequency microwave remote sensing data is acquired by UAVs carrying K / Ka-band UAV payloads. This data is referred to as high-frequency UAV microwave remote sensing data. Its characteristics include high spatial resolution (meter level) and controllable temporal resolution, but short time series and small coverage area.

[0059] Multi-source microwave remote sensing data, combining low-frequency satellite microwave remote sensing data and high-frequency UAV microwave remote sensing data, was chosen because the long wavelengths of low-frequency microwave remote sensing have strong canopy penetration capabilities, enabling the perception of overall vegetation moisture status. The VOD (Volatile Organic Compounds) derived from this data shows a strong physical correlation with vegetation biomass, providing a robust but coarse macroscopic measurement of the overall vegetation condition. Conversely, the short wavelengths of high-frequency microwave remote sensing primarily interact with the upper canopy and leaves, being more sensitive to the fine spatial structures of leaves, such as geometry and density, thus providing a fine but shallow microscopic measurement. The two sources are highly complementary in terms of physical sensing scale, providing clear and reliable monitoring signals for the VOD spatiotemporal fusion model. This makes it easier for the VOD spatiotemporal fusion model to learn the mapping relationship from macro to micro, and thus be used for VOD spatiotemporal fusion.

[0060] The VOD (Voice of Demand) data is obtained by performing VOD inversion on low-frequency satellite microwave remote sensing data using a VOD inversion algorithm, and is denoted as VOD.LR (t,x,y). Where LR represents low frequency, t represents the time dimension, and x and y represent the spatial dimensions.

[0061] Simultaneously, VOD inversion algorithms are used to perform VOD inversion on high-frequency UAV microwave remote sensing data to obtain UAV VOD data, denoted as VOD. HR (t i (x,y). Where HR represents high frequency, t i This represents the specific moment of the i-th flight of the drone.

[0062] It should be understood that different types of microwave remote sensing data can be retrieved using their corresponding VOD inversion algorithms. By using VOD inversion algorithms, raw observations from different platforms and sensors are unified to VOD, a core physical quantity, ensuring the consistent physical meaning of subsequent data source fusion and avoiding the confusion of physical meaning that can result from directly fusing raw observations.

[0063] Step S112: Obtain satellite VOD sub-data from the satellite VOD data corresponding to the collection time point of the UAV VOD data.

[0064] VOD from satellite VOD data LR Extract the time point t from (t,x,y) corresponding to the data collection time of the UAV. i The corresponding data is denoted as satellite VOD sub-data, i.e., VOD. LR (t i (x,y).

[0065] Step S113: Train the preset VOD spatiotemporal fusion model using the satellite VOD sub-data and the UAV VOD data to obtain the VOD spatiotemporal fusion model.

[0066] The one-to-one correspondence of VODs obtained above LR (t i (x,y) and VOD HR (t i The data pairs (x, y) constitute the training samples for training the VOD spatiotemporal fusion model.

[0067] The preset VOD spatiotemporal fusion model is preferably a deep learning model, specifically, an improved SRCNN (Super-Resolution Convolutional Neural Network) model. The preset VOD spatiotemporal fusion model is essentially a mapping model that learns how to map low-spatial-resolution satellite VOD data to high-spatial-resolution UAV VOD data.

[0068] VODLR (t i Input (x, y) into the preset VOD spatiotemporal fusion model (denoted as F) θ The model is based on VOD. LR (t i The output of the predicted drone VOD data is denoted as VOD(x,y). Fused (t i (x,y). VOD based on predicted drone VOD data. Fused (t i (x,y) and real drone data VOD HR (t i Given the model parameters θ, we determine the total loss (x, y), minimize the total loss using optimization algorithms such as gradient descent, and continuously adjust the model parameters θ to finally obtain the trained VOD spatiotemporal fusion model F. θ *

[0069] Furthermore, during the training process of the VOD spatiotemporal fusion model, the loss function used is determined based on spatial reconstruction loss, temporal smoothing constraints, and a regularization term. The specific formula is as follows: L=α×L spatial +β×L temporal +γ×L regularization ; Where L represents the total loss, L spatial L represents the spatial reconstruction loss. temporal L represents the time smoothing constraint. regularization represents the regularization term, and α, β, and γ represent the preset weight coefficients.

[0070] Furthermore, the spatial reconstruction loss is based on predicted drone VOD data. Fused (t i (x,y) and real drone data VOD HR (t i The values ​​of x, y, and y are determined and can be characterized by mean square error. The specific formula is as follows: L spatial =|VOD Fused (t i (x,y)-VOD HR (t i ,x,y)| 2 .

[0071] The time smoothing constraint is based on predicted drone VOD data. Fused (t i The values ​​(x, y) are determined. Specifically, the model can be used to calculate the predicted VOD data for drones. Fused (t iThe variation of x, y between adjacent time points (such as the first difference) is given by the following formula: L temporal =|Δt(VOD Fused (t i ,x,y))| 2 ; Where Δt represents the first-order difference.

[0072] The regularization term is the L2 regularization term for the model parameters θ. The specific formula is as follows: L regularization =|θ| 2 .

[0073] The loss function in this embodiment of the invention is determined based on spatial reconstruction loss, temporal smoothing constraints, and a regularization term. By introducing temporal smoothing constraints, the fused time series changes are ensured to be smooth and continuous, conforming to the natural laws of vegetation biomass accumulation and decay, and avoiding unreasonable temporal fluctuations. The introduction of a regularization term prevents the model from overfitting on the training data, thereby improving the model's generalization ability to unseen data. Training the model using the above loss function not only improves the spatial reconstruction accuracy of the model but also ensures the physiological rationality of the generated data in the temporal dimension, solving the problems of poor interpretability and overfitting in small sample sizes inherent in pure deep learning models.

[0074] Step S114: Input the satellite VOD data into the VOD spatiotemporal fusion model, perform VOD spatiotemporal fusion processing, and obtain the VOD fused data output by the VOD spatiotemporal fusion model.

[0075] All satellite VOD data is input into the pre-trained VOD spatiotemporal fusion model for processing, resulting in VOD fused data output by the model. This method enables super-resolution reconstruction of low spatial resolution satellite VOD data, yielding high spatiotemporal resolution VOD fused data.

[0076] The crop yield prediction method provided in this invention successfully overcomes the spatiotemporal scale limitations of a single remote sensing platform by synergistically utilizing the long-term temporal advantages of low-frequency satellite data and the high-resolution advantages of high-frequency UAV data, generating VOD fusion data with both high spatiotemporal resolution. Furthermore, in this invention, only a small amount of UAV VOD data from key periods is needed to drive the model to perform super-resolution reconstruction of satellite VOD data for the entire growing season. This achieves the acquisition of high spatiotemporal resolution VOD data throughout the entire growing cycle at a lower cost. This high-quality, physically meaningful VOD fusion data provides a high-quality data foundation for crop yield prediction.

[0077] Based on any of the above embodiments, before step S120, step S100 is further included.

[0078] Step S100: Perform parameter simplification and regional calibration on the initial crop growth model to obtain the target crop growth model; wherein, the initial crop growth model is constructed based on the light energy utilization model.

[0079] An initial crop growth model was constructed based on the core mechanism of the Light Use Efficiency (LUE) model. The core formula of this initial crop growth model is: Dry matter production = Incident photosynthetically active radiation × Light use efficiency × Stress factors. The model retains the most critical processes driving crop growth, such as stress effects and phenological development. Stress effects are addressed by introducing stress factors such as water and temperature to modify the potential growth rate, while phenological development can be divided into phenological stages using an accumulated temperature model.

[0080] After constructing the initial crop growth model, the parameters of the initial crop growth model are simplified. Specifically, from complex complete mechanistic models (such as DSSAT and WOFOST), parameters that are not sensitive to the prediction of final biomass or yield, or that are difficult to obtain accurately in large-scale applications, are identified and set as fixed constants or empirical tables. Parameters that are crucial to growth simulation are retained, such as crop maximum light energy utilization rate, radiation interception coefficient, and development rate. After the above simplification, the model input only requires radiation, temperature, moisture index, phenological period / sowing / harvest date, and several basic soil properties, i.e., the environmental data and crop planting data obtained above.

[0081] By employing the above methods, an optimal balance can be achieved between computational complexity, parameter availability, and the degree of preservation of physical mechanisms, enabling the model to be applicable to large-scale, distributed simulations based on remote sensing raster data.

[0082] Then, the simplified initial crop growth model underwent regional calibration. Specifically, key parameters were fitted and optimized using historical yield, ground biomass, and phenological observations to make the simulated intermediate physical quantities (i.e., crop physiological process data) more closely reflect the actual conditions of the crop planting area. This regional calibration ensured the accuracy of the target crop growth model within a specific region, enabling the output crop physiological process data to accurately reflect the current crop growth status in the planting area and providing reliable prior physical mechanisms for subsequent analysis.

[0083] At this point, step S120 includes: inputting the environmental data and the crop planting data into the target crop growth model to perform regionalized crop growth simulation and obtain the crop physiological process data.

[0084] The crop physiological process data include at least one of potential biomass, water stress coefficient, heat stress coefficient, and phenological period.

[0085] Environmental and crop planting data are used as driving data and input into the target crop growth model. The model simulates the crop growth process daily at each pixel to obtain crop physiological process data. Crop physiological process data includes, but is not limited to: potential biomass, water stress coefficient, heat stress coefficient, and phenological stage.

[0086] Among them, potential biomass represents the potential biomass growth under light and temperature limitations, which is the cumulative biomass of crops under ideal stress-free conditions; water stress coefficient, ranging from 0 to 1, represents the degree to which water inhibits photosynthetic potential; heat stress coefficient, also ranging from 0 to 1, represents the degree to which high temperature inhibits growth rate; phenological stages include sowing, jointing, tasseling, grain-filling, and maturity stages, and also include the corresponding dates to distinguish different growth stages. It should be noted that the water stress coefficient and heat stress coefficient are calculated daily.

[0087] The crop yield prediction method provided in this invention simplifies parameters and performs regional calibration, resulting in a lightweight target crop growth model suitable for large-scale raster data-driven applications. Compared to existing complex crop growth mechanism models, this invention resolves the core contradiction between mechanism models and big data applications. Then, by generating crop physiological process data with clear physical meaning through regional simulation, and inputting this data into a subsequent dual-stream spatiotemporal deep learning model, the interpretability, generalization ability, and biomass prediction reliability of the model can be enhanced. This achieves effective complementarity and deep integration between physical mechanisms and data-driven approaches, improving the accuracy of crop yield prediction results.

[0088] Based on any of the above embodiments, the multi-source feature data includes multi-channel feature maps of key phenological dates and multi-source feature time series data for each pixel. Figure 3 This is the third flowchart of the crop yield prediction method provided by the present invention, as shown below. Figure 3 As shown, step S140 includes: step S141, step S142, step S143 and step S144.

[0089] Step S141: Input the multi-channel feature map of the key phenological date into the spatial feature extraction module of the dual-stream spatiotemporal deep learning model to perform spatial feature extraction and obtain the spatial feature vector output by the spatial feature extraction module.

[0090] Step S142: Input the multi-source feature time series data of each pixel into the time feature extraction module of the dual-stream spatiotemporal deep learning model to extract time features and obtain the time feature vector output by the time feature extraction module.

[0091] The dual-stream spatiotemporal deep learning model includes a spatial feature extraction module, a temporal feature extraction module, a feature fusion module, and a fully connected layer. The spatial feature extraction module can be a spatial feature extraction deep learning network, such as CNN (Convolutional Neural Network), ResNet (Residual Network), or U-Net (U-shaped Convolutional Network), used to extract high-level spatial context information and output a spatial feature vector. The temporal feature extraction module can be a time-series feature extraction model, such as LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), or Transformer, used to learn dynamic patterns and temporal dependencies during the growth process and output a temporal feature vector.

[0092] The multi-source feature data includes multi-channel feature maps of key phenological dates and multi-source feature time series data for each pixel.

[0093] The multi-channel feature map of key phenological dates is input into the spatial feature extraction module to extract spatial features, and the spatial feature vector output by the spatial feature extraction module is obtained.

[0094] Simultaneously, the multi-source feature time series data of each pixel is input into the time feature extraction module to extract time features, learn its dynamic pattern of change over time, and obtain the time feature vector output by the time feature extraction module.

[0095] Step S143: Input the spatial feature vector and the temporal feature vector into the feature fusion module of the dual-stream spatiotemporal deep learning model to perform feature fusion and obtain the feature fusion vector output by the feature fusion module.

[0096] Spatial and temporal feature vectors are input together into the feature fusion module for feature fusion, resulting in a fused feature vector output by the module. Feature fusion can be achieved through direct concatenation or weighted summation.

[0097] Step S144: Input the feature fusion vector into the fully connected layer of the dual-stream spatiotemporal deep learning model to perform biomass regression and obtain the biomass estimate output by the fully connected layer.

[0098] Finally, the feature fusion vector is input into the fully connected layer, and nonlinear transformation and regression calculation are performed through the fully connected layer to obtain the final biomass estimate.

[0099] The crop yield prediction method provided in this invention constructs a dual-stream spatiotemporal deep learning model that couples a spatial feature extraction module and a temporal feature extraction module. This model can simultaneously and fully capture the spatial heterogeneity and temporal dependence of yield formation. Compared with a single-structure feature extraction model, this invention makes fuller use of spatiotemporal features, thereby improving the estimation accuracy of biomass.

[0100] Based on any of the above embodiments, step S130 includes: step S131, step S132, step S133, step S134 and step S135.

[0101] Step S131: Based on the crop growth cycle, the VOD fusion data is reconstructed into a time series to obtain the VOD time series for each pixel.

[0102] Step S132: Calculate VOD dynamic feature data based on the VOD time series of each pixel; wherein the VOD dynamic feature data includes at least one of the following: maximum VOD value, time to reach maximum VOD, cumulative VOD, VOD growth rate, and VOD decay rate.

[0103] The VOD fusion data is formatted as a three-dimensional data cube, encompassing the time dimension t of the entire crop growth cycle, as well as a fine spatial grid (x,y).

[0104] The crop growth cycle includes the sowing date and the harvest date. Based on the sowing date and the harvest date, all VOD fusion data within the corresponding time window is extracted from the VOD fusion data cube.

[0105] Then, for each cell, all VOD values ​​from the sowing date to the harvest date are extracted along the time dimension t. This series of VOD values ​​arranged in chronological order constitutes the VOD time series of that cell.

[0106] For each pixel's VOD time series, calculate VOD dynamic feature data; wherein, VOD dynamic feature data includes at least one of the following: maximum VOD, time to reach maximum VOD, cumulative VOD, VOD growth rate, and VOD decay rate.

[0107] The maximum VOD value, which is the maximum value in the entire VOD time series, represents the peak value of vegetation biomass and water content and is strongly correlated with the maximum photosynthetic capacity of crops.

[0108] The time when the maximum VOD value is reached, i.e., the specific date corresponding to the maximum VOD value mentioned above, is then converted into accumulated temperature or growing days. This feature reflects the phenological progress of the crop; for example, whether the peak occurs early or late can indicate the impact of varietal characteristics, environmental stress, or management practices.

[0109] Cumulative VOD is obtained by calculating the area under the VOD time series curve over the entire growth cycle. The most commonly used calculation method is numerical integration, such as the trapezoidal method. This feature approximates the total biomass / water accumulation of vegetation throughout the growing season and is a comprehensive indicator reflecting the total workload of vegetation.

[0110] VOD growth rate is calculated by linearly fitting a VOD time series within a selected period in the early stages of growth (usually between jointing and tasseling). The slope of this linear fit is the VOD growth rate. This characteristic characterizes the vigor and speed of vegetation growth; rapid growth typically indicates favorable environmental conditions.

[0111] VOD decay rate is calculated by fitting a VOD time series during a selected period in the later stages of growth (the senescence period, such as from peak yield to harvest). The slope of the fitted data is the VOD decay rate. This feature characterizes the rate of crop senescence. Excessive senescence may stem from drought, heat stress, or disease, and can affect the final yield.

[0112] Furthermore, VOD dynamic characteristic data can also include: VOD values ​​during specific phenological periods (such as the tasseling period and the grain-filling period), the starting point of the time series (the date when VOD exceeds a certain threshold), and the magnitude of seasonal changes.

[0113] Step S133: Determine the key phenological dates based on the phenological periods in the crop physiological process data.

[0114] Crop physiological data includes phenological stages, specifically each phenological stage and its date. The key phenological stages of the target crop are obtained, and then the key phenological dates are determined.

[0115] For example, the tasseling stage of maize is a critical phenological stage, which is crucial for final yield formation. The date corresponding to the tasseling stage is then identified as the critical phenological date.

[0116] Step S134: Obtain first target feature data corresponding to the key phenological date from the VOD fusion data, the VOD dynamic feature data of each pixel and the crop physiological process data, and stack them to obtain a multi-channel feature map of the key phenological date.

[0117] From VOD fusion data, VOD dynamic feature data of each pixel, and crop physiological process data, feature data corresponding to key phenological dates are obtained, i.e., the feature data of the key phenological date itself, denoted as the first target feature data. These are then stacked to obtain a multi-channel feature map of the key phenological date. The channels of the multi-channel feature map can be: VOD fusion data corresponding to the key phenological date, VOD dynamic feature data, potential biomass, water stress coefficient, and heat stress coefficient, etc.

[0118] It should be noted that after obtaining the first target feature data, the first target feature data can be aligned in the spatiotemporal dimension. The alignment method can be nearest neighbor interpolation or bilinear interpolation to ensure spatial consistency.

[0119] Step S135: Obtain second target feature data within the time window from sowing to the key phenological date from the VOD fusion data, the crop physiological process data, and the environmental data, so as to construct multi-source feature time series data for each pixel.

[0120] From the VOD time series, crop physiological process data, and environmental data of each pixel, characteristic data within the time window from sowing to the key phenological date are extracted and denoted as the second target characteristic data. The second target characteristic data specifically includes VOD time series, potential biomass, water stress coefficient, heat stress coefficient, meteorological data, and soil data. Correspondingly, the multi-source characteristic time series data of each pixel includes six feature dimensions.

[0121] The crop yield prediction method provided in this invention integrates multi-source remote sensing data, crop physiological process data, and environmental data to construct spatial features (multi-channel feature maps of key phenological dates) and temporal features (multi-source feature time series data of each pixel) with clear physical meaning. Compared with single feature data, this invention provides reliable and information-rich data input for subsequent dual-stream spatiotemporal deep learning models, thereby improving the accuracy of biomass prediction results and thus improving the accuracy of crop yield prediction results.

[0122] Based on any of the above embodiments, step S150 includes: step S151, step S152 and step S153.

[0123] Step S151: Calculate the dynamic harvest index for each pixel based on the water stress coefficient and heat stress coefficient in the crop physiological process data.

[0124] Crop physiological data include daily water stress coefficients and heat stress coefficients.

[0125] Specifically, the water stress coefficient for the critical period of yield formation is first obtained from the daily water stress coefficients. A representative value is then extracted from the water stress coefficient for the critical period of yield formation using the average method, minimum method, or cumulative stress method, and used as the target water stress coefficient. The critical period of yield formation refers to the key period affecting yield. For example, for maize, the critical period of yield formation is the post-flowering stage, specifically corresponding to the period from silking to physiological maturity; for rice, the critical period of yield formation is the period from heading to milk stage. Then, the target water stress coefficient is substituted into a preset water stress response function to obtain the water stress adjustment factor.

[0126] Similarly, the thermal stress coefficient for the critical period of yield formation is obtained from the daily thermal stress coefficient. A representative value is extracted from the thermal stress coefficient for the critical period of yield formation using the average value method, minimum value method, or cumulative stress method, and denoted as the target thermal stress coefficient. Then, the target thermal stress coefficient is substituted into the preset thermal stress response function to obtain the thermal stress adjustment factor.

[0127] It should be noted that the preset water stress response function and the preset heat stress response function are both empirical functions obtained by empirical fitting based on historical experimental data or observation data. They are usually linear or exponential decay functions, used to quantify the adverse effects of stress on yield formation during the critical period of yield formation.

[0128] The commonly used function form is a linear function, for example, f(Water Stress Target) = 1 - k × (1 - Water Stress Target), where f() is the water stress response function, Water Stress Target is the target water stress coefficient, which is a value between 0 and 1, and k is a preset fitting parameter. When there is no water stress, Water Stress Target = 1, f(Water Stress Target) = 1; when under complete stress, Water Stress Target = 0, f(Water Stress Target) = 1 - k.

[0129] The thermal stress response function g() treats the target thermal stress coefficient in a similar way.

[0130] After calculating the moisture stress adjustment factor and the heat stress adjustment factor, a multiplication operation is performed based on these factors and the preset basic harvest index to obtain the dynamic harvest index for each pixel. The specific formula is as follows: HI dynamic =HI base ×f(Water Stress Target)×g(Heat Stress Target) Among them, HI dynamic HI represents the dynamic harvest index. base This represents the basic harvest index.

[0131] It should be noted that different base harvest indices can be set for different pixels.

[0132] Step S152: Multiply the dynamic harvest index of each pixel with the corresponding biomass estimate to obtain the yield estimate of each pixel.

[0133] For each pixel, the dynamic harvest index of that pixel is multiplied by the corresponding biomass estimate to obtain the yield estimate for that pixel. The yield estimate for each pixel can form a yield projection map.

[0134] Step S153: Based on the estimated yield of each pixel, perform spatial aggregation calculation to obtain the estimated crop yield.

[0135] Specifically, a crop planting distribution map can be used as a mask to filter out all pixels planted with the target crop from the yield prediction map. The yield prediction values ​​of each selected pixel are then summed to obtain the crop yield prediction value.

[0136] The crop yield prediction method provided in this invention calculates a dynamic harvest index using water stress and heat stress coefficients, and then uses this index to estimate crop yield. This approach comprehensively considers the dynamic impacts of water and heat stress on crop yield formation, effectively avoiding the systematic biases inherent in using a fixed harvest index. This method makes the conversion process from biomass to yield more consistent with agronomic mechanisms, contributing to improved accuracy and reliability of yield prediction results.

[0137] The following section uses a major corn-producing area in China as an example to provide a detailed explanation of the crop yield prediction method of this invention.

[0138] Step 1: Data Preparation. Data includes: satellite VOD data, drone VOD data, environmental data, crop planting data, and ground-based ground truth data.

[0139] Satellite VOD data: Acquired daily-scale VOD data at the SMAP L3 level in 2023, with a spatial resolution of 36 km.

[0140] UAV VOD data: In July (jointing stage), August (tasseling stage), and September (grouting stage) of 2023, UAVs equipped with microwave payloads were used to fly in typical sample areas to acquire VOD data with a resolution of 10 meters.

[0141] Environmental data: ERA5-day meteorological data were obtained from ECMWF (European Centre for Medium-Range Weather Forecasts); soil data at 250-meter resolution were obtained from the SoilGrids system.

[0142] Crop planting data: Obtain county-level maize planting distribution maps and average sowing date data from the agricultural and rural affairs departments. The average sowing date data includes the average sowing date and the average harvest date.

[0143] Ground-based true data: During the drone flight synchronization period, ground surveys were conducted in the sample area to measure maize canopy biomass (dry weight); yield was measured during the harvest period, and yield and harvest index were recorded.

[0144] Step 2: Collaborative Fusion of Multi-Source Microwave Remote Sensing VOD Data. Using a VOD spatiotemporal fusion model, low-resolution satellite VOD data is fused with high-resolution UAV VOD data to generate fused VOD data, i.e., daily-scale, 10-meter resolution VOD. Fused Data cube.

[0145] Specifically, satellite VOD data for the entire year and drone VOD data for three periods are input into a pre-defined VOD spatiotemporal fusion model, which is an improvement on the SRCNN model. First, drone VOD data is used as the HR (High Resolution) ground truth, and the spatially downsampled result of the corresponding date's satellite VOD data is used as the LR (Low Resolution) input to train the pre-defined VOD spatiotemporal fusion model, resulting in a trained VOD spatiotemporal fusion model. This trained model is then used for super-resolution reconstruction of satellite VOD data for all dates, generating daily-scale VOD data with a 10-meter resolution for the entire year of 2023. Fused Data cube.

[0146] Step 3: Construction of Multi-Source Feature Dataset. Extract VOD dynamic features from the VOD fusion data and combine them with environmental data and the output of the target crop growth model to construct a multi-feature data cube.

[0147] From VOD Fused Extract the VOD time series of each pixel and calculate the VOD dynamic feature data such as the maximum VOD value, the time to reach the maximum VOD, the cumulative VOD value, the VOD growth rate, and the VOD decay rate.

[0148] Using environmental data (including meteorological and soil data) and crop planting data to drive the target crop growth model, the daily-scale crop physiological process data for each pixel are simulated, including potential biomass, water stress coefficient, heat stress coefficient, and phenological stage.

[0149] Based on the aforementioned VOD fusion data, VOD dynamic features, meteorological data, soil data, and crop physiological process data, spatiotemporal alignment and resampling (10-meter resolution) were performed and unified to a daily scale, forming a basic (T,H,W,C) multi-feature data cube. Here, T represents the time length, H represents the image height, W represents the image width, and C represents the number of feature channels.

[0150] Subsequently, based on the key phenological dates, multi-channel feature maps of the key phenological dates and multi-source feature time series data of each pixel are extracted from the underlying multi-feature data cube.

[0151] The process of extracting multi-channel feature maps for key phenological dates is as follows: On the determined key phenological dates (the key phenological period for maize is the tasseling stage, denoted as T),... k Extract data from all feature channels, and then... k The VOD fusion data, VOD dynamic characteristics, potential biomass, water stress coefficient, and heat stress coefficient for this day are spatially stacked to generate a multi-channel feature map of key phenological dates. This feature map aims to capture the spatial distribution and pattern of crop growth during specific key growth stages.

[0152] The extraction process of multi-source feature time series data for each pixel is as follows: For each pixel, extract the data from the sowing date to the key phenological date T. k Time series data, such as VOD fusion data, potential biomass, water stress coefficient, heat stress coefficient, meteorological data, and soil data, are used to construct multi-source feature time series data for each pixel. This time series aims to reflect the continuous growth dynamics of crops from emergence to critical stages.

[0153] Step 4: Biomass estimation based on a dual-stream spatiotemporal deep learning model. Using the aforementioned multi-source feature data as input, the biomass of key phenological dates (masting period) is estimated using a dual-stream spatiotemporal deep learning model.

[0154] The training process of the two-stream spatiotemporal deep learning model is as follows: (1) Obtaining the training target (true value): Since the growth status of maize before the tasseling period determines the biomass accumulation, and the tasseling period mainly affects the grain filling, the maize canopy biomass (dry weight) obtained synchronously through ground survey during the tasseling period (e.g., August 15) is used as the true value for model training.

[0155] (2) Obtaining model input features: For each ground sample point, obtain its corresponding two multi-source feature data in the manner described above: the sample point during the male extraction period T k The multi-channel feature map, and the sample point from sowing to tasseling stage T k Multi-source feature time series data.

[0156] (3) Model training: 80% of all samples were used to train the pre-set dual-stream spatiotemporal deep learning model, and 20% were used for validation. Through end-to-end training, the model learns the complex mapping relationship from spatiotemporal features to masculinization biomass, and finally obtains the trained dual-stream spatiotemporal deep learning model.

[0157] For each pixel, inference can be performed on the entire region, while simultaneously inputting its corresponding multi-channel feature map and multi-source feature time series data. The model outputs the biomass estimate for each pixel during the masculinization period, ultimately generating a spatially continuous, 10-meter resolution biomass distribution map during the masculinization period. est This figure reveals the spatial heterogeneity of biomass.

[0158] Step 5: Yield conversion and estimation based on dynamic harvest index. The impact of environmental stress on yield formation during the post-flowering stage (from silking stage to physiological maturity) is considered using the dynamic harvest index, and biomass is converted into yield.

[0159] Analyzing historical data to determine the HI (highest fertility) of summer maize in this region. base The value was set to 0.45, and the water stress response function was fitted as f(x) = 1 - 0.3(1-x), and the heat stress response function was set as g(x) = 1 - 0.2(1-x). Using the daily water and heat stress factors simulated by the target crop growth model during the post-flowering stage, the target water stress coefficient and target heat stress coefficient for each pixel were calculated using the average value method. Then, the dynamic harvest index HI for each pixel was calculated. dynamic Finally, the biomass distribution map of the male ejaculation period obtained in step 4 is used. est With Dynamic Harvest Index (HI) dynamic Pixel-by-pixel multiplication yields a spatially continuous yield estimation map. The yield estimates for all maize-planted pixels within the target planting area are spatially aggregated to obtain the total estimated yield for that area. Specifically, a maize planting distribution map can be used as a mask to select corresponding pixels from the yield estimation map. The yield estimates for each selected pixel are then summed to obtain the total estimated yield.

[0160] Furthermore, the total estimated yield is compared with the actual yield data obtained from the harvest period, and evaluation indicators are calculated, including but not limited to: the coefficient of determination (R²). 2The root mean square error (RMSE, expressed as a percentage of average yield) and relative error (RE%) are used to arrive at the final result: R 2 With an RMSE of 0.85, an RE% of 10%, and a RE% of 5%, the estimation accuracy of the method in the embodiments of the invention is verified to be high.

[0161] The crop yield prediction device provided by the present invention is described below. The crop yield prediction device described below can be referred to in correspondence with the crop yield prediction method described above.

[0162] Figure 4 This is a schematic diagram of the crop yield prediction device provided by the present invention, as shown below. Figure 4 As shown, the device includes a VOD fusion module 410, a growth simulation module 420, a data construction module 430, a biomass prediction module 440, and a yield prediction module 450; wherein: VOD fusion module 410 is used to perform VOD spatiotemporal fusion processing based on multi-source microwave remote sensing data through a VOD spatiotemporal fusion model to obtain VOD fused data; The growth simulation module 420 is used to simulate crop growth based on environmental data and crop planting data using a target crop growth model to obtain crop physiological process data. Data construction module 430 is used to construct multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data; The biomass estimation module 440 is used to extract and fuse spatiotemporal features from the multi-source feature data using a dual-stream spatiotemporal deep learning model in order to estimate biomass and obtain a biomass estimate. The yield estimation module 450 is used to calculate the crop yield estimate based on the biomass estimate.

[0163] The crop yield prediction device provided in this invention first uses a VOD spatiotemporal fusion model to perform VOD spatiotemporal fusion processing on microwave remote sensing data from multiple sources (i.e., multi-source microwave remote sensing data), generating VOD fused data with high spatiotemporal resolution, long time series, and wide coverage. This solves the inherent defects of a single remote sensing data source and provides a high-quality data foundation for subsequent crop yield prediction. Then, crop growth is simulated using a target crop growth model to generate crop physiological process data with clear physical meaning. This data is used as prior knowledge of physical mechanisms and, together with the VOD fused data and environmental data, constructs multi-source feature data. This data is then input into a dual-stream spatiotemporal deep learning model for spatiotemporal feature extraction and fusion to predict biomass, which is then used to calculate the crop yield prediction value. This deep learning paradigm guided by physical mechanisms not only overcomes the shortcomings of purely data-driven models, such as poor interpretability and overfitting in small samples, by utilizing physical constraints and enhancing the model's generalization ability, but also achieves simultaneous capture of the spatial heterogeneity and temporal dependence of crop growth through the collaborative extraction and fusion of spatiotemporal features of data. Compared with existing yield modeling and analysis methods, this invention makes fuller use of information, ultimately obtaining more accurate and reliable biomass and crop yield estimates. In summary, this invention can predict crop yield based on the fusion of multi-source microwave remote sensing data, crop growth models, and deep learning models, breaking through the limitations of a single technical paradigm and improving the accuracy of crop yield prediction results.

[0164] It should be noted that the crop yield prediction device provided in this embodiment of the invention can implement all the method steps implemented in the above-mentioned crop yield prediction method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0165] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute the aforementioned crop yield prediction method.

[0166] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the crop yield estimation methods provided in the above embodiments.

[0168] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0169] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting crop yield, characterized in that, include: By using the vegetation optical thickness (VOD) spatiotemporal fusion model, VOD spatiotemporal fusion processing is performed based on multi-source microwave remote sensing data to obtain VOD fused data. By using a target crop growth model, crop growth is simulated based on environmental and crop planting data to obtain crop physiological process data; Based on the VOD fusion data, the crop physiological process data, and the environmental data, multi-source feature data is constructed. Using a dual-stream spatiotemporal deep learning model, spatiotemporal features are extracted and fused from the multi-source feature data to estimate biomass and obtain a biomass estimate. Based on the estimated biomass, the estimated crop yield is calculated.

2. The crop yield prediction method according to claim 1, characterized in that, The multi-source microwave remote sensing data includes low-frequency satellite microwave remote sensing data and high-frequency UAV microwave remote sensing data. The VOD spatiotemporal fusion processing based on the multi-source microwave remote sensing data, using a vegetation optical thickness (VOD) spatiotemporal fusion model, yields fused VOD data, including: The VOD inversion algorithm is used to perform VOD inversion on the low-frequency satellite microwave remote sensing data to obtain satellite VOD data, and the VOD inversion algorithm is used to perform VOD inversion on the high-frequency UAV microwave remote sensing data to obtain UAV VOD data. From the satellite VOD data, obtain satellite VOD sub-data corresponding to the collection time point of the UAV VOD data; The VOD spatiotemporal fusion model is trained using the satellite VOD sub-data and the UAV VOD data to obtain the VOD spatiotemporal fusion model. The satellite VOD data is input into the VOD spatiotemporal fusion model for VOD spatiotemporal fusion processing to obtain the VOD fused data output by the VOD spatiotemporal fusion model.

3. The crop yield prediction method according to claim 2, characterized in that, During the training process of the VOD spatiotemporal fusion model, the loss function used is determined based on spatial reconstruction loss, temporal smoothing constraint, and regularization term.

4. The crop yield prediction method according to claim 1, characterized in that, Before obtaining crop physiological process data by simulating crop growth based on environmental and crop planting data using a target crop growth model, the process also includes: The initial crop growth model is simplified and calibrated regionally to obtain the target crop growth model; wherein the initial crop growth model is constructed based on the light energy utilization rate model. The process involves simulating crop growth using a target crop growth model, based on environmental and crop planting data, to obtain crop physiological process data, including: The environmental data and the crop planting data are input into the target crop growth model to perform regionalized crop growth simulation and obtain the crop physiological process data. The crop physiological process data include at least one of potential biomass, water stress coefficient, heat stress coefficient, and phenological period.

5. The crop yield prediction method according to any one of claims 1 to 4, characterized in that, The multi-source feature data includes multi-channel feature maps of key phenological dates and multi-source feature time series data for each pixel. The multi-source feature data is then subjected to spatiotemporal feature extraction and fusion using a dual-stream spatiotemporal deep learning model to estimate biomass and obtain a biomass estimate, including: The multi-channel feature map of the key phenological date is input into the spatial feature extraction module of the dual-stream spatiotemporal deep learning model to perform spatial feature extraction, and the spatial feature vector output by the spatial feature extraction module is obtained. The multi-source feature time series data of each pixel is input into the time feature extraction module of the dual-stream spatiotemporal deep learning model to extract time features and obtain the time feature vector output by the time feature extraction module. The spatial feature vector and the temporal feature vector are input into the feature fusion module of the dual-stream spatiotemporal deep learning model to perform feature fusion, and the feature fusion vector output by the feature fusion module is obtained. The feature fusion vector is input into the fully connected layer of the dual-stream spatiotemporal deep learning model to perform biomass regression, thereby obtaining the biomass estimate output by the fully connected layer.

6. The crop yield prediction method according to claim 5, characterized in that, The construction of multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data includes: Based on the crop growth cycle, the VOD fusion data is reconstructed into a time series to obtain the VOD time series for each pixel; VOD dynamic feature data is calculated based on the VOD time series of each pixel; wherein, the VOD dynamic feature data includes at least one of the following: maximum VOD value, time to reach maximum VOD, cumulative VOD, VOD growth rate, and VOD decay rate; Based on the phenological periods in the crop physiological process data, key phenological dates are determined; From the VOD fusion data, the VOD dynamic feature data of each pixel, and the crop physiological process data, the first target feature data corresponding to the key phenological date is obtained, and the data is stacked to obtain a multi-channel feature map of the key phenological date; Second target feature data within a time window from sowing to the key phenological date are obtained from the VOD fusion data, the crop physiological process data, and the environmental data to construct multi-source feature time series data for each pixel.

7. The crop yield prediction method according to any one of claims 1 to 4, characterized in that, The calculation of crop yield forecast based on the biomass forecast includes: The dynamic harvest index for each pixel is calculated based on the water stress coefficient and heat stress coefficient in the crop physiological process data. The dynamic harvest index of each pixel is multiplied by the corresponding biomass estimate to obtain the yield estimate of each pixel. Based on the estimated yield of each pixel, spatial aggregation calculation is performed to obtain the estimated crop yield.

8. A crop yield prediction device, characterized in that, include: The VOD fusion module is used to perform VOD spatiotemporal fusion processing based on multi-source microwave remote sensing data through a VOD spatiotemporal fusion model to obtain VOD fused data. The growth simulation module is used to simulate crop growth based on environmental and crop planting data using a target crop growth model to obtain crop physiological process data. The data construction module is used to construct multi-source feature data based on the VOD fusion data, the crop physiological process data, and the environmental data; The biomass estimation module is used to extract and fuse spatiotemporal features from the multi-source feature data using a dual-stream spatiotemporal deep learning model in order to estimate biomass and obtain a biomass estimate. The yield estimation module is used to calculate the crop yield estimate based on the biomass estimate.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the crop yield estimation method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the crop yield prediction method as described in any one of claims 1 to 7.

Images