Crop yield prediction method based on time sequence remote sensing image

Abstract

The invention relates to the technical field of crop yield analysis, in particular to a crop yield prediction method based on a time sequence remote sensing image, and the method comprises the steps: obtaining an optical remote sensing image and a synthetic aperture radar image in a crop growth period, and constructing a time sequence image through complementary information fusion and under-cloud information recovery; the problems of cloud interference and data missing are solved, a multi-scale time regulation and control convolution module is used for processing a time sequence image, multi-scale time sequence features are extracted, feature fusion is carried out in combination with attention difference jump connection, and a rice distribution map is obtained; determining a rice planting area based on the rice distribution diagram, selecting an optimal assimilation time period in the rice planting area, and assimilating the leaf area index subjected to remote sensing inversion and a crop growth model simulation result to generate a simulated yield after assimilation; based on the multi-scale time sequence characteristics, a parameter dynamic response matrix related to crop growth model parameters is established, and residual error correction is performed on assimilated simulation yield, so that the rice yield prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of crop yield analysis, and specifically to a crop yield prediction method based on temporal remote sensing images. Background Art

[0002] Traditional crop yield prediction methods mainly rely on a single remote sensing data source (such as optical or radar images), and it is difficult to overcome problems such as cloud occlusion and spectral information loss, resulting in poor continuity of temporal data. In the prior art, optical images are easily interfered by clouds and rain, radar images have insufficient resolution, and multi-source data fusion methods fail to effectively extract multi-scale growth characteristics of crops; At the same time, fixed phenological parameters are often used in crop model assimilation, ignoring the spatio-temporal response differences between the dynamic changes in the rice growth period and remote sensing characteristics, resulting in biases in the selection of assimilation periods and the accumulation of yield residuals. In addition, existing models lack an effective mechanism for reconstructing the surface information under clouds and dynamically correcting model parameters, resulting in errors in planting area identification and a decline in yield prediction accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a crop yield prediction method based on temporal remote sensing images to solve the problems raised in the above background art. The core problems to be solved include how to construct high-precision time series images through complementary information fusion and under-cloud information restoration to solve the problems of cloud interference and data loss; how to achieve the collaborative optimization of crop growth models and remote sensing inversion data based on multi-scale temporal feature extraction and dynamic parameter correction, so as to improve the rice yield prediction accuracy.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A crop yield prediction method based on temporal remote sensing images, the method steps of which include: S1. Respectively extract the spectral features of optical images and the backscattering features of radar images through a dual-branch feature extraction module, calculate the complementary weight matrix using the cross-attention mechanism, supplement the optical spectral features in the radar missing areas, reconstruct the surface information in the optical cloud-covered areas, and integrate the shallow details and deep semantic features through a multi-scale feature fusion module to generate a fused image; This solution directly solves the data loss problems caused by cloud occlusion in optical images and insufficient resolution in radar images, and improves the integrity and spectral-texture consistency of temporal images; Generate a cloud binary mask to distinguish the completely thick-cloud occluded areas, thin-cloud covered areas, and cloud shadow areas, and use spectral unmixing technology to separate the cloud and surface reflection signals in the thin-cloud covered areas; For the completely thick-cloud occluded areas, combine the cloud-free images of the previous time phase and the radar images of the same period, and reconstruct the surface reflectance under the clouds through a spatio-temporal sequence prediction model; This solution ensures the restoration of surface information in cloud-covered areas and guarantees the continuity of time series images.

[0005] S2. Adopt a multi-scale time regulation convolutional module to extract short-term phenological features, medium-term trend features, and long-term periodic features using convolutional kernels with different time windows respectively, and model the periodic law of the rice growth period through a gated recurrent unit. This solution enhances the representation ability of crop growth dynamics. Introduce an attention difference jump connection to capture the growth rate difference (such as the growth rate difference between the tillering stage and the heading stage) through element-wise difference operations, generate channel attention weights to strengthen the feature response of key phenological periods (such as the heading stage), and fuse deep semantic features and shallow detail features across levels. This solution solves the problems of traditional convolutional networks ignoring crop growth rate differences and blurred field boundaries, and improves the spatial accuracy of the rice distribution map. S3. Based on the rice phenological period, divide the remote sensing observation window, calculate the determination coefficient and root mean square error of the assimilated yield and the measured yield in each period through parameter sensitivity analysis, and select the phenological period with the highest determination coefficient, the smallest error, and dynamic synchronization with the yield critical period as the optimal assimilation period. This solution solves the problem of time period selection deviation caused by traditional assimilation methods relying on fixed phenological periods. Invert the leaf area index based on the radiative transfer model at the heading stage, drive the crop growth model to simulate the spatio-temporal distribution, and iteratively update the model state variables through pixel-level assimilation. This solution realizes the dynamic coupling of remote sensing data and crop models and improves the assimilation accuracy.

[0006] S4. Use the random forest algorithm to quantify the correlation strength between multi-scale time series features and crop model parameters, screen key time series features, and construct a dynamic response matrix with model parameters as row vectors and time series features as column vectors. This solution solves the problem of parameter staticization in traditional assimilation and realizes the dynamic association between parameters and remote sensing features. By comparing the spatial residual distribution of the assimilated yield and the measured yield, inversely derive the parameter deviation combined with the dynamic response matrix, identify high-contribution parameters using the random forest feature contribution degree, input them as correction factors into the gated recurrent unit to predict the residual adjustment amount, and dynamically weight the correction contributions of different growth periods through the attention mechanism. This solution realizes the spatial differential correction of residuals. Generate a correction coefficient for the maximum growth rate parameter of the leaf area index at the pixel level, and use the alternating direction multiplier method to perform multiple rounds of collaborative iterative optimization of parameters and residuals until the determination coefficient converges. This solution significantly reduces the systematic residuals of yield estimation through parameter dynamic correction and model re-run.

[0007] Compared with the prior art, the beneficial effects of the present invention are: Through the complementary fusion of optical and radar data and the reconstruction of information under clouds, the integrity and usability of time-series images are improved; the discriminative ability of crop growth characteristics is enhanced by multi-scale time-series feature extraction and attention differential fusion mechanism; the dynamic parameter response matrix and residual correction optimization achieve the deep coordination between crop models and remote sensing inversion data, overcoming the problems of assimilation period selection bias and parameter solidification in traditional methods; this method is used to improve the spatial accuracy of rice planting area recognition and reduce the yield prediction residual, providing reliable technical support for agricultural precision management and food security decision-making. Description of the Drawings

[0008] Figure 1 It is a schematic diagram of the method steps of the present invention. Detailed Embodiments

[0009] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0010] Please refer to Figure 1 , the present invention provides a technical solution: a method for predicting crop yields based on time-series remote sensing images, including the following method steps: S1. Obtain optical remote sensing images and synthetic aperture radar images during the crop growth period, and construct time-series images through complementary information fusion and information restoration under clouds, specifically including: Obtain time-series optical remote sensing images (such as Sentinel-2 multispectral data) and synthetic aperture radar images (such as Sentinel-1 data) during the crop growth period through a satellite platform; perform radiometric calibration and atmospheric correction on the optical images to eliminate the influence of atmospheric scattering; perform radiometric calibration, multi-look processing, and terrain correction on the radar images to eliminate geometric distortions caused by terrain undulations; according to the crop phenological calendar, screen images covering the sowing period to the maturity period, and eliminate invalid images with cloud cover exceeding 50%; Construct a cross-modal fusion network based on the Swin-Transformer architecture to achieve feature-level complementary information fusion of optical and radar images. The complementary information fusion includes dual-branch feature extraction, complementary information enhancement module, and multi-scale feature fusion, where: Dual-branch feature extraction: Use convolutional layers to extract the spectral features of optical images and the backscattering features of radar images respectively, and align the spatial resolution differences of the two modalities through a spatial attention mechanism; Complementary Information Enhancement Module: Design a cross-attention mechanism in the Transformer encoder to calculate the complementary weight matrix of optical features and radar features. For areas where radar features are missing (such as water bodies), use optical spectral features for supplementation; for areas covered by optical clouds, utilize the radar penetration features to reconstruct surface information. Multi-scale Feature Fusion: In the decoder stage, fuse shallow detail features and deep semantic features to generate a fused image that retains both spectral characteristics and texture details.

[0011] A cloud-aware spatio-temporal information reconstruction mechanism is designed to address the cloud occlusion problem in optical images for cloud-bottom information restoration, specifically including: Generate a cloud binary mask using the shortwave infrared band of the optical image and a cloud detection algorithm to distinguish thick cloud completely occluded areas, thin cloud covered areas, and cloud shadow areas. For thin cloud covered areas, use spectral unmixing technology to separate the cloud and surface reflection signals; for thick cloud completely occluded areas, combine the cloud-free image of the previous time phase and the contemporaneous radar image, and reconstruct the cloud-bottom surface reflectance through a spatio-temporal sequence prediction model (based on gated recurrent unit); for cloud shadow areas, based on the backscattering coefficient of the radar image and the shortwave infrared band characteristics of the historical cloud-free image, use a spatio-temporal adaptive filtering algorithm to eliminate the reflectance attenuation caused by shadows. Specifically, detect the cloud shadow boundary by the ratio of the shortwave infrared band to the near-infrared band to generate a shadow mask, establish a linear regression model using the VV polarization backscattering coefficient of the contemporaneous radar image and the reflectance of the corresponding time phase of the historical cloud-free image to correct the reflectance attenuation in the visible light band of the shadow area, and use median filtering of the time series sliding window to eliminate residual noise. Introduce a cloud mask weighting term in the loss function, and the calculation formula is: total loss = image reconstruction loss + λ × cloud area reconstruction loss, where λ is dynamically adjusted according to the cloud thickness (for example, λ = 2.0 in thick cloud areas, λ = 0.5 in thin cloud areas) to ensure the spectral continuity at the cloud edge.

[0012] For time nodes where data is still missing, use an adaptive weighted time series filtering algorithm. Based on the cubic spline function, dynamically adjust the interpolation weight according to the phenological characteristics of the crop growth curve (such as rapid growth during the tillering stage and stable change during the heading stage) to ensure the smoothness of the growth curve; detect and correct abnormally fluctuating pixels by calculating the difference in the NDVI change rate (normalized difference vegetation index change rate) of adjacent pixels in the time series. For areas with abnormally fluctuating pixels for 3 consecutive time phases, start the radar image assisted verification mechanism, and use the temporal correlation of the backscattering coefficient for data rectification, and finally construct a time series image.

[0013] Step S1 synchronously acquires Sentinel-2 multispectral images and Sentinel-1 radar images covering the entire growth period of crops. First, atmospheric correction and geometric registration are performed on the optical images, and a cloud mask is generated using the short-wave infrared band to mark cloud-contaminated areas. At the same time, radiometric calibration and terrain correction are carried out on the radar images. Based on a dual-branch deep learning network, spectral features of the optical images and scattering features of the radar images are extracted respectively. Cross-modal associations are established through a cross-attention mechanism. In vegetated areas, the spectral features of the optical images are dominant, supplemented by the backscattering features of the radar images. In water areas, the backscattering features of the radar images are dominant, supplemented by the short-wave infrared band features of the optical images, realizing the complementary advantages of cross-modal features. For cloud-obscured areas, thin cloud spectral unmixing and thick cloud spatio-temporal prediction (combining historical cloud-free images and contemporaneous radar data) are used to hierarchically repair the surface reflectance, and a cloud mask weighted loss function is designed to optimize the reconstruction accuracy. Finally, adaptive cubic spline interpolation is used to fill in missing time phases. Combining NDVI temporal consistency checks and radar backscattering verification, a time series image with complete space and continuous time is generated, with a time resolution of 10 days, a reconstruction accuracy of more than 90% for the cloud-free area, and complete coverage of the key growth stages from crop sowing to maturity, providing data support for precision agriculture monitoring.

[0014] S2. Use a multi-scale time-regulating convolutional module to process the time series image, extract multi-scale temporal features, and perform feature fusion in combination with attention differential skip connections to obtain a rice distribution map, specifically including: Based on the time series image generated in S1, a three-level temporal feature extraction network is constructed using a multi-scale time-regulating convolutional module, including short-term phenological feature extraction, mid-term trend capture, and long-term cycle modeling, where: Short-term phenological feature extraction: Use a 1×1 time convolutional kernel to process adjacent 1-frame images (10-day window) to capture short-term change features such as the rapid growth during the tillering stage of rice, and output a 32-channel feature map. Mid-term trend capture: Use a 3×3 time convolutional kernel to process adjacent 3-frame images (30-day window) to extract progressive growth features from the jointing stage to the heading stage. The receptive field is expanded through dilated convolution (dilation rate = 2), and a 64-channel feature map is output. Long-term cycle modeling: Use a 5×5 time convolutional kernel to process adjacent 5-frame images (50-day window), and combine a gated recurrent unit (GRU) to memorize the periodic laws of the complete growth period of rice (from sowing to maturity), and output a 128-channel feature map.

[0015] Perform element-wise difference operations on feature maps of different scales (multi-scale temporal features) to capture the differences in crop growth rates. Use the Sigmoid function to generate channel attention weights to dynamically strengthen the feature responses of key phenological periods, and combine skip connections to cross-level fuse the weighted deep semantic features and shallow detail features, enhancing the spectral discriminability of highly sensitive periods such as the heading stage while retaining the spatial details of the field boundaries. Finally, output the rice distribution map through feature stitching and upsampling to achieve double optimization of classification accuracy and phenological logic.

[0016] S3. Determine the rice planting area based on the rice distribution map, and select the optimal assimilation period within the rice planting area. Assimilate the remotely sensed leaf area index and the simulation results of the crop growth model to generate the assimilated simulated yield, specifically including: Based on the rice distribution map output by S2, extract a binary mask of rice planting with a spatial resolution of 10 meters, exclude non-planting areas (such as forests, bare lands, water bodies), and generate a spatially continuous vector boundary of rice planting; combine the administrative division data of the Chengdu Plain Economic Region, and count the rice planting area by county unit as the basic spatial unit for regional yield estimation; According to the rice phenological periods (transplanting period, tillering period, jointing period, heading period, milk ripening period), use the time series images constructed by S1 to extract the corresponding remote sensing observation windows for each phenological period (such as the heading period corresponding to early to mid-August); Assimilate the remotely sensed leaf area index and the simulation results of the crop growth model for different phenological periods. Calculate the determination coefficient of the assimilated yield and the measured yield for each period through parameter sensitivity analysis, and evaluate the goodness of fit of the model by combining the root mean square error; According to the quantitative indicators of the determination coefficient and the root mean square error, select the phenological period with the highest determination coefficient, the smallest model error, and the strongest synchronization with the dynamic change of the leaf area index during the key period of yield formation as the optimal assimilation period; for example, determine the heading period as the optimal assimilation period because its leaf area index is highly correlated with the key period of yield formation; Based on the radiative transfer model, input Sentinel-2 multi-spectral images and sun-sensor geometric parameters, and invert the leaf area index at the rice heading stage with a spatial resolution of 10 meters through the look-up table method; Drive the locally calibrated crop growth model, input meteorological data (sunshine, temperature and humidity), soil parameters (water retention, water conductivity) and management parameters (irrigation, fertilization), and simulate the spatio-temporal distribution of the leaf area index at the heading stage; Use the remotely sensed leaf area index as the observed value, and perform pixel-level assimilation with the simulation results of the crop growth model. Generate the assimilated set of rice growth parameters by iteratively updating the model state variables (leaf area index, biomass); Input the assimilated leaf area index into the crop growth model to drive the model to complete the growth simulation from the heading stage to the maturity stage, and output the yield per unit area; based on the spatial weights of the rice distribution map, aggregate pixel by pixel to generate the total yield at the county scale as the simulated yield after assimilation.

[0017] S4. Based on the multi-scale temporal and spatial features in step S2, establish a parameter dynamic response matrix related to the crop growth model parameters, and perform residual correction on the simulated yield after assimilation in step S3 to output an optimized estimated yield of the rice planting area, specifically including: Based on the multi-scale temporal and spatial features extracted in step S2, by fusing the temporal and spatial features in the short term (tillering stage), medium term (jointing stage), and long term (heading stage), establish the dynamic correlation between the sensitive parameters of the crop growth model and the remote sensing temporal and spatial features. Specifically, use the random forest algorithm to quantify the non-linear response intensity between the temporal and spatial features at different scales (such as the spectral coefficient of variation at the tillering stage, the texture entropy value at the jointing stage) and the crop growth model parameters (such as the maximum growth rate of leaf area index, the accumulated temperature threshold). Through feature importance analysis and statistical significance tests, screen out the key temporal and spatial feature sets sensitive to parameter changes, and construct a multi-dimensional parameter-feature response matrix; the matrix uses the crop growth model parameters as row vectors and the multi-scale temporal and spatial features as column vectors, and the matrix element values represent the dynamic response intensity of specific parameters to the corresponding features; in the process of constructing the matrix, the gated recurrent unit network is introduced to encode the dynamic evolution mode of the temporal and spatial features, capture the temporal and spatial dependence between parameters and features, so as to form a parameter regulation knowledge base supporting dynamic optimization and provide a theoretical basis for subsequent residual correction. Based on the spatial residual distribution of the simulated yield after assimilation and the measured yield in step S3, combined with the parameter dynamic response matrix, screen out the key correction factors; first, calculate the yield residual value of each pixel by spatial statistical methods to generate a residual spatial distribution map; then, use the parameter-feature correlation relationship recorded in the parameter dynamic response matrix to inversely deduce the potential model parameter deviation causing the residual; specifically, extract the corresponding multi-scale temporal and spatial features (such as abnormal spectral variation at the heading stage, sudden change of texture entropy value at the jointing stage) in the high-residual value area, and identify the crop growth model parameters (such as the parameter error of the maximum growth rate of leaf area index) that significantly match the residual spatial distribution pattern through random forest feature contribution analysis. The selected key correction factors (i.e., high-contribution parameters and their associated temporal and spatial features) are input into the gated recurrent unit network, which predicts the spatially differentiated residual adjustment amount by learning the spatio-temporal coupling relationship between the historical residual adjustment amount and the multi-scale temporal and spatial features; in this process, the network dynamically weights the feature contributions at different growth stages through the attention mechanism to ensure the differential correction logic of the fast growth features at the tillering stage and the stable features at the heading stage, and finally outputs a residual adjustment amount field matching the crop growth spatial heterogeneity.

[0018] Apply the predicted residual adjustment amount to the sensitive parameters of the crop growth model to achieve iterative optimization of the model parameters and dynamic correction of yield estimation. For the parameter of the maximum growth rate of leaf area index, according to the spatial distribution of the residual adjustment amount field, generate a parameter correction coefficient at the pixel scale: reduce the growth rate parameter in the region with overestimated residuals to inhibit overgrowth simulation, and increase the parameter in the region with underestimated residuals to enhance biomass accumulation. The corrected parameter set is re-input into the crop growth model to drive the model to re-run from the heading stage to the maturity stage, generating the growth simulation results after parameter optimization. This process uses the alternating direction method of multipliers for parameter-residual collaborative optimization, and gradually approaches the true value of the parameter through multiple rounds of iteration until the determination coefficient between the simulated yield after assimilation and the measured yield reaches the convergence threshold. Finally, aggregate the optimized pixel-level yield estimation values by county unit, and combine the administrative division weights to generate the zonal yield forecast results that meet the agricultural management requirements. At the same time, evaluate the generalization ability of the parameter dynamic response matrix through leave-one-out cross-validation to ensure the stability of the optimization mechanism under different climate years and planting systems.

[0019] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A method for predicting crop yields based on temporal remote sensing images, characterized in that, The method steps are as follows: S1. Obtain optical remote sensing images and synthetic aperture radar images during the growth period of crops, and construct time series images through complementary information fusion and under-cloud information restoration; S2. Process the time series images using a multi-scale time regulation convolution module, extract multi-scale temporal features, and perform feature fusion in combination with attention differential skip connections to obtain a rice distribution map; S3. Determine the rice planting area based on the rice distribution map, select the optimal assimilation period within the rice planting area, assimilate the remotely sensed leaf area index with the simulation results of the crop growth model, and generate the assimilated simulated yield; S4. Based on the multi-scale temporal features in step S2, establish a parameter dynamic response matrix related to the crop growth model parameters, perform residual correction on the assimilated simulated yield in step S3, and output the optimized estimated value of the rice planting area yield.

2. The crop yield prediction method based on temporal remote sensing images according to claim 1, wherein The process of the complementary information fusion specifically includes: Extract the spectral features of the optical image and the backscattering features of the radar image respectively through a dual-branch feature extraction module; Use a cross-attention mechanism to calculate the complementary weight matrix of the optical features and the radar features, supplement the optical spectral features in the radar feature missing areas, and reconstruct the surface information in the optical cloud-covered areas; Adopt a multi-scale feature fusion module to fuse the shallow detail features and the deep semantic features to generate a fused image that retains spectral characteristics and texture details.

3. The crop yield prediction method based on temporal remote sensing images according to claim 1, characterized in that The process of the under-cloud information restoration specifically includes: Generate a cloud binary mask to distinguish the thick cloud completely blocked area, the thin cloud covered area and the cloud shadow area; Use spectral unmixing technology to separate the cloud layer and the surface reflection signal in the thin cloud covered area; For the thick cloud completely blocked area, combine the cloud-free image of the previous time phase and the contemporaneous radar image, and reconstruct the under-cloud surface reflectance through a spatio-temporal sequence prediction model; For the cloud shadow area, based on the backscattering coefficient of the radar image and the short-wave infrared band characteristics of the historical cloud-free image, adopt a spatio-temporal adaptive filtering algorithm to eliminate the reflectance attenuation caused by the shadow.

4. The crop yield prediction method based on temporal remote sensing images according to claim 1, wherein The multi-scale time regulation convolution module includes: Adopt convolution kernels with different time windows to extract short-term phenological features, medium-term trend features and long-term periodic features respectively; model the periodic law of the complete growth period of rice through a gated recurrent unit.

5. The crop yield prediction method based on temporal remote sensing images according to claim 1, wherein, The process of the attention differential skip connection specifically includes: Perform element-wise difference operations on the multi-scale temporal features to capture the crop growth rate differences; Generate channel attention weights to strengthen the feature responses of key phenological periods; Fuse the deep semantic features and the shallow detail features across levels, retain the spatial details of the field boundaries and enhance the spectral discriminability.

6. The crop yield prediction method based on temporal remote sensing images according to claim 1, wherein The process of selecting the optimal assimilation period includes: Extract the remotely sensed observation windows corresponding to each growth stage according to the rice phenological period; Assimilate the remotely sensed leaf area index of different phenological periods with the simulation results of the crop growth model, calculate the determination coefficient of the assimilated yield and the measured yield at each period through parameter sensitivity analysis, and evaluate the goodness of fit of the model in combination with the root mean square error; According to the quantitative indicators of the determination coefficient and the root mean square error, select the phenological period with the highest determination coefficient, the smallest model error and the strongest synchronization with the dynamic change of the leaf area index in the key period of yield formation as the optimal assimilation period.

7. The crop yield prediction method based on temporal remote sensing images according to claim 1, characterized in that, The assimilation process in step S3 includes: Inverting the leaf area index at the heading stage of rice based on the radiative transfer model; Driving the crop growth model to simulate the spatio-temporal distribution of the leaf area index at the heading stage; Performing pixel-level assimilation on the remote sensing inversion results and the simulation results of the crop growth model, and iteratively updating the state variables of the crop growth model.

8. The crop yield prediction method based on temporal remote sensing images according to claim 1, characterized in that, The construction process of the parameter dynamic response matrix specifically includes: Based on the multi-scale temporal features extracted in step S2, using the random forest algorithm to quantify the non-linear correlation strength between each temporal feature and the sensitive parameters of the crop growth model; screening out the key temporal feature set sensitive to parameter changes through feature importance analysis, and combining the gated recurrent unit network to encode the dynamic evolution pattern of the temporal features; constructing a parameter dynamic response matrix with the crop growth model parameters as row vectors and the multi-scale temporal features as column vectors, and the matrix element values represent the dynamic response strength of the parameters to the features.

9. The crop yield prediction method based on temporal remote sensing images according to claim 1, characterized in that, The process of residual correction specifically includes: By comparing the spatial residual distribution of the simulated yield and the measured yield after assimilation, and combining the parameter dynamic response matrix to inversely deduce the model parameter deviation causing the residual; extracting the multi-scale temporal features corresponding to the high residual value areas, and using the random forest feature contribution analysis to identify the sensitive parameters matching the residual; taking the selected high contribution parameters and their associated features as key correction factors, inputting them into the gated recurrent unit network to predict the spatially differentiated residual adjustment amount, and dynamically weighting the correction contributions of different growth stage features through the attention mechanism.

10. The crop yield prediction method based on temporal remote sensing images according to claim 1, wherein, The calculation of the optimized yield estimation value for the rice planting area includes: According to the spatial distribution of the residual adjustment amount, generating a correction coefficient for the maximum growth rate parameter of the leaf area index at the pixel level, reducing the parameter value for the overestimated residual area and increasing the parameter value for the underestimated area; inputting the corrected parameter set into the crop growth model to drive the re-run, and using the alternating direction multiplier method to perform multiple rounds of collaborative iteration optimization of the parameters and the residuals until the determination coefficient between the assimilated yield and the measured yield reaches the convergence threshold; finally generating the optimized yield estimation value for the rice planting area.

Images