Corn yield prediction method and system based on chlorophyll fluorescence and deep learning

Abstract

The invention discloses a corn yield prediction method and system based on chlorophyll fluorescence and deep learning, and the method comprises the steps: obtaining the yield prediction related data of a target region for many years in the past, and carrying out the preprocessing of the yield prediction related data; screening yield prediction related data of the hot and dry years according to a preset threshold value, and constructing a training data set under a stress condition based on the yield prediction related data of the hot and dry years; constructing a deep neural network model, and training, verifying and testing the deep neural network model by adopting the training data set to obtain a corn yield prediction model based on DNN; and inputting to-be-predicted data into the trained corn yield prediction model, outputting a corn yield prediction result, and generating a spatial distribution diagram. According to the method, the spatial distribution goodness of fit between a prediction result and official statistical data is remarkably superior to that of a traditional method, and a quantifiable and low-cost solution is provided for precise agricultural decision making under the extreme climate condition.

Description

Technical Field

[0001] This invention relates to the field of maize yield prediction technology, and in particular to a maize yield prediction method and system based on chlorophyll fluorescence and deep learning. Background Technology

[0002] As one of the world's three major food crops, maize plays a crucial role in global food security. Accurately predicting maize yields before harvest to support macro-level decision-making and field management is essential for ensuring food security. With technological advancements, obtaining high-precision crop yield information on a large regional scale using remote sensing technology has become possible, attracting increasing attention and becoming a hot research topic in agricultural remote sensing. Existing methods primarily rely on ground-based measurement data and variables related to crop growth, including remote sensing and meteorological data. However, the increasing frequency of extreme weather events such as high temperatures and droughts due to global climate change is increasing the constraints on food security production. Therefore, accurately predicting crop yields under abnormal climatic conditions is crucial for timely intervention and mitigation measures.

[0003] A comprehensive analysis of existing studies has revealed that, under abiotic stress conditions, solar-induced chlorophyll fluorescence (SIF) shows a stronger correlation with crop growth and yield than traditional vegetation indices (such as the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)). Especially under stress conditions, SIF demonstrates a more powerful monitoring capability and has been widely applied in phenological monitoring, net primary productivity analysis, drought assessment, and crop parameter retrieval. However, research using SIF for crop yield prediction under drought and other stresses remains relatively limited.

[0004] Traditional statistical models struggle to handle nonlinear relationships between independent and dependent variables. Machine learning (ML) and deep learning (DL) methods, as advanced extensions to these statistical regression models, effectively address complex nonlinear relationships and improve model generalization capabilities. Commonly used machine learning models for yield prediction, such as Random Forest Regression (RFR) and Support Vector Regression (SVR), have demonstrated strong predictive power in crop yield estimation. Furthermore, deep neural networks (DNNs), as one of the most representative deep learning algorithms, also perform well in yield prediction. However, given the variability introduced by regional climate change, crop type, and data sources, the optimal feature combination and model type for yield prediction remain undetermined. In addition, research on crop yield prediction under extreme climatic conditions is relatively rare. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting maize yield based on chlorophyll fluorescence and deep learning. The prediction results show a significantly better spatial distribution agreement with official statistical data than traditional methods, providing a quantifiable and low-cost solution for precision agriculture decision-making under extreme climatic conditions.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for predicting maize yield based on chlorophyll fluorescence and deep learning includes the following steps:

[0008] S1. Obtain relevant yield forecast data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and preprocess the above-mentioned yield forecast data.

[0009] S2. Filter the yield prediction data of hot and dry years according to the preset threshold, and construct the training dataset under stress conditions based on the yield prediction data of hot and dry years.

[0010] S3. Construct a deep neural network model, use the training dataset to train, validate and test the deep neural network model, determine the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtain a maize yield prediction model based on DNN.

[0011] S4. Input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

[0012] Preferably, in S1, the vegetation index is derived from the MOD13A1 dataset, including NDVI and EVI, and the calculation formula is as follows:

[0013]

[0014] Wherein, NIR, RED, and BLUE represent the reflectance of the near-infrared, red, and blue bands, respectively; G is the gain factor; C1 and C2 are atmospheric correction factors; and L is the soil adjustment factor.

[0015] Preferably, in S1, the meteorological data includes actual evapotranspiration, minimum temperature, maximum temperature, cumulative precipitation, water vapor pressure, water vapor pressure difference, reference evapotranspiration, and Palmer drought severity index.

[0016] Preferably, in S1, the preprocessing of SIF data, vegetation index, meteorological data and historical yield data includes data format conversion, raster masking, regional statistics and spatiotemporal resolution unification.

[0017] Preferably, in S3, the training, validation, and testing of the corn yield prediction model based on the training dataset specifically includes:

[0018] The optimal hyperparameters of the corn yield prediction model were determined using 10-fold cross-validation and grid search on the training dataset. The performance of the corn yield prediction model on different data splits was evaluated using 10-fold cross-validation to ensure the robustness of parameter selection. The grid search method was used to systematically explore the parameter space of the corn yield prediction model.

[0019] Preferably, step S3 also includes an evaluation of the corn yield prediction model, with the evaluation index being the coefficient of determination R. 2 The root mean square error (RMSE) is calculated using the following formula:

[0020]

[0021] Where n is the number of observations, y i , These represent the measured value, the predicted value, and the mean of the measured data, respectively.

[0022] A maize yield prediction system based on chlorophyll fluorescence and deep learning, applying any one of the above-mentioned maize yield prediction methods based on chlorophyll fluorescence and deep learning, including:

[0023] The data acquisition module is used to acquire yield forecasting data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and to preprocess the above-mentioned yield forecasting data.

[0024] The data filtering module is used to filter data from hot and dry years according to a preset threshold, and to construct a training dataset under stress conditions based on the data from hot and dry years.

[0025] The model building module constructs a deep neural network model, uses a training dataset to train, validate, and test the deep neural network model, determines the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtains a maize yield prediction model based on DNN.

[0026] The yield prediction module is used to input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

[0027] 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 a corn yield prediction method based on chlorophyll fluorescence and deep learning as described above.

[0028] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] This invention achieves a significant technological breakthrough in maize yield prediction under extreme conditions such as high temperature and drought by innovatively integrating solar-induced chlorophyll fluorescence (SIF) data with a deep learning model. Its core effects are reflected in three aspects: First, SIF data is highly sensitive to abiotic stresses such as high temperature and drought, directly reflecting changes in crop photosynthetic efficiency and significantly improving prediction accuracy compared to traditional vegetation indices (NDVI / EVI). Second, the proposed multi-source data fusion framework (combining meteorological data, SIF, and vegetation indices) enhances the accuracy of deep neural network (DNN) models in county-level predictions. Finally, the specialized feature combination (meteorological data + SIF + VI) developed for hot and dry years exhibits strong robustness, with its prediction results showing significantly better spatial distribution agreement with official statistics than traditional methods, providing a quantifiable and low-cost solution for precision agriculture decision-making under extreme climate conditions. These innovations give this technology outstanding application value in the context of intensifying global climate change. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A schematic diagram of the maize yield prediction method based on chlorophyll fluorescence and deep learning provided by this invention;

[0032] Figure 2 This is a normal distribution map of temperature and precipitation in 2020, as shown in this embodiment of the invention.

[0033] Among them, (a) is the normal distribution of temperature and precipitation in July 2020, (b) is the normal distribution of temperature and precipitation in August 2020, and (c) is the normal distribution of temperature and precipitation in September 2020.

[0034] Figure 3 This is a correlation analysis diagram of the variables and output in the embodiments of the present invention;

[0035] Figure 4 This is a performance comparison chart of four models in yield prediction in the embodiments of the present invention;

[0036] (a) shows the comparison of the root mean square error (RMSE) of the four models, and (b) shows the coefficient of determination (R²) of the four models. 2 contrast;

[0037] Figure 5This is a distribution diagram of corn yield prediction error based on a DNN model in an embodiment of the present invention;

[0038] Figure 6 This is a comparison chart of the accuracy of prediction models with different parameter combinations in the embodiments of the present invention;

[0039] Figure 7 This is a spatial distribution map of maize yield in a hot and dry year at the county scale, based on a DNN model.

[0040] Among them, (a) is the distribution of corn yield according to official county-level statistics in Shandong Province in 2021, (b) is the spatial distribution of corn yield predicted based on the SIF model, and (c) is the spatial distribution of corn yield predicted based on the VI model.

[0041] Figure 8 This is a spatial distribution map of the accuracy of maize yield prediction results in a hot and dry year at the county scale based on the DNN model.

[0042] Among them, (a) and (c) are the spatial distribution map and scatter plot comparison of the deviation between the predicted corn yield in 2021 and the actual statistical yield using SIF data, respectively; (b) and (d) are the spatial distribution map and scatter plot comparison of the deviation between the predicted corn yield in 2021 and the actual statistical yield using only traditional VIs, respectively. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, the present invention provides a method for predicting maize yield based on chlorophyll fluorescence and deep learning, comprising the following steps:

[0046] S1. Obtain relevant yield forecast data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and preprocess the above-mentioned yield forecast data.

[0047] S2. Filter the yield prediction data of hot and dry years according to the preset threshold, and construct the training dataset under stress conditions based on the yield prediction data of hot and dry years.

[0048] S3. Construct a deep neural network model, use the training dataset to train, validate and test the deep neural network model, determine the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtain a maize yield prediction model based on DNN.

[0049] S4. Input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

[0050] In a specific embodiment, the present invention includes the following steps:

[0051] 1. Dataset and Preprocessing

[0052] 1.1VI Data

[0053] NDVI and EVI are the most commonly used VIs in crop parameter inversion and identification experiments because they demonstrate powerful performance in crop inversion and identification. This invention utilizes the MOD13A1 dataset (500-meter resolution, 16-day interval) provided by the U.S. Geological Survey (https: / / lpdaac.usgs.gov / products / mod13a1v061 / ), which provides nearly 24 years of global NDVI and EVI data and multi-band information. NDVI and EVI are calculated using the following formula:

[0054]

[0055] Here, NIR, RED, and BLUE represent the reflectance in the near-infrared, red, and blue bands, respectively, corresponding to MODIS bands 7, 1, and 3. Key coefficients for EVI calculation include the gain factor G (=2.5), atmospheric correction factors C1 (=6) and C2 (=7.5), and the soil adjustment factor L (=1).

[0056] The MOD13A1 dataset is compatible with Google Earth Engine (GEE). By importing maize planting distribution maps and county boundary maps of Shandong Province, the MOD13A1 band data underwent masking and statistical processing by region.

[0057] 1.2 Meteorological Data

[0058] Meteorological data were obtained from the TerraClimate dataset provided in the GEE data catalog (https: / / developers.google.cn / earth-engine / ). This dataset contains features closely related to yield and effectively reflecting crop growth conditions, such as actual evapotranspiration (Aet), minimum temperature (Tmmn), maximum temperature (Tmmx), cumulative precipitation (Pr), vapor pressure (Vap), vapor pressure difference (Vpd), reference evapotranspiration (Pet), and the Palmer Drought Severity Index (Pdsi). In addition, a precipitation (Pre) dataset was generated using 1-kilometer monthly precipitation data from China (Peng et al., 2019), with a spatial resolution of 0.0083° (approximately 1 kilometer) and a time range from 1901 to 2022. All the above data were converted to commonly used raster formats and uploaded to the GEE platform. GEE's computing power was used to perform operations such as masking and zonal statistics to generate the tabular data required for model training.

[0059] 1.3 SIF Data

[0060] The global solar-induced chlorophyll fluorescence (GOSIF) dataset is available at https: / / globalecology.unh.edu / data / GOSIF.html, covering the period from 2000 to 2022. The spatial resolution is 0.05°*0.05°.

[0061] The process of downloading and processing SIF data is similar to that of the precipitation dataset. FileZilla software was used to remotely transfer NC files, Panoply software was used to check basic information within the NC files, and ArcGIS software was used to batch convert the NC files to TIF files. Finally, the data was uploaded to GEE for further processing. The processing method was similar to the batch processing method for the MOD13A1 dataset, and SIF data for all districts and counties in Shandong over the past 20 years were exported in tabular form.

[0062] 1.4 Maize Planting Distribution and Yield Data

[0063] The Shandong Province maize planting distribution map used in this invention is from Peng et al. (2023), with an overall average accuracy of 80.06% at the provincial scale. Maize yield data for each county in Shandong Province are from the *Shandong Statistical Yearbook* (http: / / www.stats.gov.cn / ). To optimize the outlier values ​​in the yield data from 2002 to 2021, the method of Chen et al. (2018) was adopted, which excludes values ​​outside the reasonable production range and values ​​exceeding two standard deviations of the normal distribution. This method yielded 1332 samples, of which 105 samples from 2021 were used as the test dataset, and the remaining samples were used for training and the dataset.

[0064] 1.5 Data Preprocessing

[0065] The main preprocessing steps for raster data include data format conversion (e.g., from NetCDF to TIFF), raster masking, and regional statistics. All raster datasets are processed uniformly at a one-month temporal resolution and a 1-kilometer spatial resolution to ensure consistency across different data sources.

[0066] Given the compatibility between the MOD13A1 dataset and the GEE platform, raster data from different time periods were extracted directly from the dataset using specific Earth Engine commands, eliminating the need to download the entire dataset. The specific steps for MODIS data processing are as follows: First, the maize planting distribution map and county-level boundary files for Shandong Province were imported into the GEE platform. Then, using a JavaScript code editor, specific API commands were used to apply masks and perform regional statistics to process the MOD13A1 band data. Subsequently, 20 years of NDVI and EVI data were processed in batches using the same method. The results were exported in tabular form to meet the input requirements of the yield estimation model.

[0067] Feature selection for yield forecasting models is based on the correlation between latent predictor variables and maize yield. Variables highly correlated with yield are considered beneficial for model training, while variables unrelated to yield may degrade model performance. Exploratory data analysis (EDA) is conducted to quantify the correlation between feature variables and yield, identifying the most effective variables for training the forecasting model, and the Pearson correlation coefficient is used to assess the non-linear relationship between each variable and yield.

[0068] 2. Criteria for classifying hot and dry years

[0069] Between 2002 and 2021, Shandong Province experienced several periods of high temperatures and droughts, which severely impacted the normal growth of crops. First, the hot-dry year classification method provided by Liu et al. (2017) was used to classify the years from 2002 to 2021 as hot-dry years. Specifically, years with temperatures exceeding μ+0.43σ were classified as high-temperature years, and years with precipitation below μ-0.43σ were classified as drought years, where μ and σ represent the average temperature and standard deviation during the maize growing season, respectively.

[0070] The normal distribution of temperature and precipitation from July to September over the past 20 years is as follows: Figure 2 As shown, the data visualization uses kernel density estimation (KDE) curves superimposed with histograms. This non-parametric method, rather than the standard normal distribution, is chosen to better capture the actual shape of the data distribution, including potential skewness, multimodality, or heavy-tailed characteristics.

[0071] A year is classified as a hot year if more than 50% of county-level cities in Shandong Province experience temperatures above a threshold for at least one month during the corn growing season; similarly, a year is classified as a dry year if more than 50% of county-level cities experience precipitation below a threshold for at least one month; years that meet both conditions are classified as hot-dry years. Between 2002 and 2021, hot-dry years included 2002, 2006, 2013, 2015, 2016, 2018, 2019, and 2021. Data from these years was used for model training and testing, and model performance was evaluated.

[0072] 3. Model Construction and Training

[0073] 3.1 Model Construction

[0074] Linear regression, deep learning, and machine learning techniques are used to analyze the relationship between multiple factors and corn yield. For example, lasso regression is used to predict crop yield; its core principle is to eliminate or reduce noisy variables in the model to reduce model complexity and improve prediction accuracy. Three commonly used machine learning and deep learning algorithms, Random Forest Regression (RFR), Support Vector Regression (SVR), and Deep Neural Networks (DNN), are also widely used for crop yield prediction. RFR is an ensemble learning method based on decision trees, capable of effectively handling large datasets and exhibiting strong predictive performance. RFR parameter tuning includes the number of trees (n_estimators = 100 to 1000, increasing by 100 each time), the maximum depth (max_depth = 3 to 15), and the minimum number of samples per leaf node (min_samples_leaf = 1 to 10). SVR constructs a hyperplane in a high-dimensional space to model the relationships in the training data, making it well-suited for capturing nonlinear relationships. Key parameters for SVR tuning include kernel type (kernel = 'rbf', 'linear', 'poly'), regularization parameter (C = 0.1 to 1000, logarithmic scale), and kernel coefficient (gamma = 0.001 to 10, logarithmic scale). DNNs are based on artificial neural networks and learn complex feature representations through multiple neural layers. DNN tuning parameters include the number of hidden layers (n_layers = 2 to 5), the number of neurons per layer (n_neurons = 16 to 128, doubling with each layer), the learning rate (learning_rate = 0.0001 to 0.1, logarithmic scale), and the dropout rate (dropout_rate = 0.1 to 0.5).

[0075] These algorithms were used to estimate corn yield, and their effectiveness and reliability were validated. Some models, such as DNN, SVR, and lasso regression, require dataset standardization. Standardization ensures that the penalty term is applied evenly among the variables, thereby improving model performance. All feature and yield data were standardized to have a mean of 0 and a standard deviation of 1.

[0076] The dataset (excluding the 2021 data) was used to determine the optimal hyperparameters of the models using 10-fold cross-validation and grid search. 10-fold cross-validation was used to evaluate the model performance across different data splits, ensuring the robustness of the parameter selection. Grid search was used to systematically explore the parameter space of each model, with the parameter ranges as described above.

[0077] Data from each county between 2002 and 2020 were randomly allocated, with 80% used for training and 20% for validation. The 2021 dataset was reserved as a separate test set and not used for training or validation. This approach ensures an unbiased evaluation of the model's generalization ability. The process simulates a real-world scenario by separating the test data, where the model predicts future yields based on historical data. After training, the performance of each model was evaluated on the 2021 test set to assess its ability to predict maize yields under unforeseen conditions.

[0078] 3.2 Model Evaluation Indicators

[0079] Evaluating the accuracy of the estimation model is crucial. This invention employs two commonly used evaluation metrics: the coefficient of determination (R²). 2 R and root mean square error (RMSE). 2 R measures the proportion of the variance in the dependent variable that can be explained by the independent variables. Its value ranges from 0 to 1; a higher value, closer to 1, indicates a better model fit. RMSE measures the average magnitude of the prediction error; a lower error indicates better simulation accuracy. 2 The formula for calculating RMSE is:

[0080]

[0081] The measured value is y i The predicted value is is the mean of the measured values, and n is the number of observations.

[0082] These metrics comprehensively evaluated the accuracy and stability of the model's maize yield estimation. Optimal hyperparameters were determined through 10-fold cross-validation to achieve the highest R-value. 2 And the lowest RMSE.

[0083] 4. Production Forecast

[0084] Multiple feature datasets obtained from preprocessing, including solar-induced fluorescence (SIF), normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), minimum temperature (Tmmn), maximum temperature (Tmmx), precipitation (Pre), potential evapotranspiration (Pet), vapor pressure gradient (Vpd), and maize yield data from 2002 to 2020, were used as the sample dataset. To comprehensively compare the predictive power of each model, different feature combinations were used as input data for model training to evaluate the predictive power of SIF relative to traditional vegetation indices such as NDVI and EVI. Furthermore, the potential to combine meteorological data with SIF or VI to improve predictive accuracy was also evaluated.

[0085] By analyzing the impact of different variable combinations on the prediction results, the most suitable variable combinations for yield forecasting in abnormal years were determined. For each model and input feature combination, Ra was calculated. 2 RMSE is used to evaluate the predictive accuracy of the model.

[0086] 5 Results

[0087] 5.1 Correlation analysis of characteristic variables and their combinations with output

[0088] Improving agricultural productivity requires a comprehensive understanding of the complex interactions between environmental factors and crop yields. Figure 3 A heatmap showing the correlation between each variable and maize yield is presented. Satellite-derived variables (NDVI, EVI, and SIF) show correlations with yield exceeding 0.3, indicating a significant positive correlation (p < 0.001). In contrast, the absolute correlations between Aet, Pdsi, Vap, and Vpd and yield are below 0.1, indicating poor statistical significance. Using these variables as model inputs may negatively impact model performance and increase the risk of overfitting.

[0089] Meanwhile, Tmmn, Pre, Pet, and Tmmx were significantly correlated with yield (p<0.001), with Tmmx (0.16) showing a positive correlation and Tmmn (-0.21) showing a negative correlation. Variables related to moisture, including Pre (-0.15) and Pet (-0.21), also showed a negative correlation with yield.

[0090] 5.2 Performance of different models for county-level output forecasting

[0091] Figure 4Tables (a)-(b) demonstrate the accuracy of four models (LASSO, RFR, SVR, and DNN) in predicting county-level maize yield based on different input features. These features include remote sensing data (NDVI, EVI, SIF) and meteorological data (Tmmn, Tmmx, Pre, Pet, Vpd). Overall, the DNN model performs slightly better than the RFR model, followed by the SVR model, while the LASSO model performs the worst. The best result is obtained by the DNN model trained using a multi-source dataset that combines satellite and meteorological data. The above results indicate that the DNN model is suitable for predicting county-level maize yield. Figure 5 The error distribution of a study on maize yield prediction based on a DNN model is presented.

[0092] 5.3 Characteristic Combinations for Yield Forecasting in Hot and Drought Years

[0093] Previous experiments have shown that the DNN algorithm is well-suited for building maize yield prediction models. Therefore, subsequent studies primarily employed a prediction model based on the DNN algorithm. This DNN model was developed using a unified framework and trained with input data containing six different feature combinations: C1 (climate), C2 (climate + NDVI), C3 (climate + EVI), C4 (climate + SIF), C5 (climate + NDVI + EVI), and C6 (climate + NDVI + EVI + SIF). Figure 5 This study demonstrates the differences in accuracy in predicting maize yield in hot, dry years using these feature combinations.

[0094] The accuracy of different feature combinations in predicting county-level maize yield, ranked from highest to lowest, is as follows: C6(R 2 :0.81; RMSE: 452.31kg / ha)>C4(R2: 0.78; RMSE: 485.62kg / ha)>C5(R 2 :0.77; RMSE:484.66kg / ha)>C2(R 2 :0.75; RMSE: 511.71kg / ha)>C3(R2: 0.74; RMSE: 517.44kg / ha)>C1(R 2 (0.61; RMSE: 638.61 kg / ha). Among the models trained using feature combinations C2, C3, and C4, the C4 combination, which includes SIF, showed superior prediction accuracy. This indicates that SIF is more effective than traditional vegetation indices (NDVI and EVI) in predicting maize yield in hot and dry years.

[0095] like Figure 6As shown, feature combination C6 achieved the highest prediction accuracy, indicating that adding more features strongly correlated with output can improve model performance. This highlights the importance of selecting diverse and highly correlated input data to build accurate prediction models.

[0096] 5.4 Final Yield Prediction Models at Different Spatial Scales in Hot and Drought Years

[0097] The validation dataset contains multiple features and statistics of maize yield in 2021. Results of 2021 maize yield forecasts using two different models are presented: one incorporating SIF data, and the other relying on the vegetation index (VI). Figure 7 (a) shows the distribution of maize yield in Shandong Province in 2021 according to official county-level statistics. Spatial distribution of maize yield predicted based on the SIF model ( Figure 7 (b) is closer to official statistics than the VI model-predicted corn yield. Figure 7 (c) This is not the case. This indicates that SIF data has stronger predictive power under high temperature and drought conditions. Furthermore, the scatter plot of model accuracy evaluation shows that the SIF-based model achieved a low root mean square error (RMSE) of 613.32 kg / ha. Figure 8 In (a) and (c), the RMSE of the VI-based model is 701.11 kg / ha. Figure 8 (b) and (d) highlight the effectiveness of SIF in predicting yields under adverse climatic conditions.

[0098] 6. Discussion

[0099] 6.1 Analysis of the Importance of Features in Production Forecasting

[0100] This invention emphasizes the crucial role of feature selection in constructing robust maize yield prediction models. Among the features evaluated, meteorological data proved to be the most influential, reflecting the significant impact of long-term environmental conditions on crop growth throughout the growing season. Key meteorological variables, such as rainfall and temperature, directly affect factors such as soil moisture and canopy development, thereby influencing yield outcomes.

[0101] During the mid-to-late growth stages of maize, the SIF (Self-Improving Fiber Indices) emerged as the most valuable yield performance characteristic, outperforming traditional vegetation indices such as NDVI (Negative Density Variable Index) and EVI (Employment Variable Index). This finding is consistent with the findings of Peng et al. (2020), demonstrating the superior ability of SIF to predict crop yield due to its sensitivity to photosynthetic activity. Unlike NDVI and EVI, which are affected by soil and shading disturbances, SIF effectively captures short-term stress responses, such as heat deficit and water shortage, thus more accurately reflecting crop health (Baker, 2008; Song et al., 2018a).

[0102] The fusion of meteorological and satellite data further improved prediction accuracy, demonstrating the synergistic value of integrating multiple data sources. Models utilizing these integrated inputs capture crop growth dynamics more comprehensively than models relying on single-variable datasets. These results are consistent with the findings of Cao et al. (2020), who found similar improvements in winter wheat yield prediction when combining vegetation indices and climate data. This highlights the importance of feature fusion in remote sensing applications, enabling a multi-dimensional reflection of crop growth.

[0103] 6.2 Comparison of Model Performance in Yield Prediction

[0104] The performance of various machine learning models in predicting maize yield was evaluated using different feature combinations. Figure 3 Among the tested models, the lasso regression model performed the worst, with its prediction accuracy and correlation with predicted yield ranging from 10% to 20%. This is mainly because lasso regression is based on linear regression with regularization techniques, which limits its ability to understand parameters with non-linear relationships.

[0105] In contrast, machine learning and deep learning models with stronger data mining capabilities, such as DNN, SVR, and RFR, outperformed corn yield predictions, demonstrating their ability to handle complex nonlinear relationships between crop yields and multiple variables. Among these models, DNNs performed particularly well, showcasing their ability to learn deep, complex relationships in high-dimensional data, which is especially important for yield prediction. While SVR also performs well with nonlinear data, DNNs typically provide superior results when dealing with large-scale datasets.

[0106] Random Forest Regression (RFR) is also effective, especially in assessing feature importance and handling nonlinearity. However, DNNs still outperform RFR in capturing deep, nonlinear relationships and perform exceptionally well with large, complex datasets. In the DNN model, a combination of SIF, NDVI, EVI, and climate variables as input features achieved the highest prediction accuracy, confirming the value of integrating multiple data sources for maize yield prediction. These findings are consistent with previous research, such as that of Cao et al. (2020), which demonstrated that combining different types of data can significantly enhance the predictive power of machine learning models.

[0107] 6.3 The Importance of SIF in Predicting Yields Under Pressure Conditions

[0108] This invention further explores the role of SIF in predicting maize yield under high temperature and drought stress. It utilizes eight years of high temperature and drought data from Shandong Province (…). Figure 5The study selected 2021 as a typical hot and dry year for model testing. Results showed that the DNN model trained with SIF data outperformed other models in predicting county-level maize yield accuracy. This aligns with the findings of Wang et al. (2023), who found that SIF is more sensitive in predicting summer maize yield in drought years, indicating that SIF can effectively capture the impact of high temperatures and drought on crops. This is because SIF provides richer crop growth information than traditional vegetation indices (VI). Unlike traditional visible and near-infrared reflectance data, SIF directly measures changes in light energy during crop photosynthesis, thus more accurately reflecting crop physiological states. Magney et al. (2019) pointed out that SIF can better describe the short-term stress response of crops under extreme weather conditions, such as drought and high temperatures. Therefore, SIF has great potential in predicting maize yield, especially in years with frequent extreme weather events.

[0109] A maize yield prediction system based on chlorophyll fluorescence and deep learning, applying any one of the above-mentioned maize yield prediction methods based on chlorophyll fluorescence and deep learning, including:

[0110] The data acquisition module is used to acquire yield forecasting data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and to preprocess the above-mentioned yield forecasting data.

[0111] The data filtering module is used to filter data from hot and dry years according to a preset threshold, and to construct a training dataset under stress conditions based on the data from hot and dry years.

[0112] The model building module constructs a deep neural network model, uses a training dataset to train, validate, and test the deep neural network model, determines the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtains a maize yield prediction model based on DNN.

[0113] The yield prediction module is used to input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

[0114] 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 a corn yield prediction method based on chlorophyll fluorescence and deep learning as described above.

[0115] 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.

[0116] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting maize yield based on chlorophyll fluorescence and deep learning, characterized in that, Includes the following steps: S1. Obtain relevant yield forecast data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and preprocess the above-mentioned yield forecast data. S2. Filter the yield prediction data of hot and dry years according to the preset threshold, and construct the training dataset under stress conditions based on the yield prediction data of hot and dry years. S3. Construct a deep neural network model, use the training dataset to train, validate and test the deep neural network model, determine the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtain a maize yield prediction model based on DNN. S4. Input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

2. The method for predicting maize yield based on chlorophyll fluorescence and deep learning according to claim 1, characterized in that, In S1, the vegetation index comes from the MOD13A1 dataset, including NDVI and EVI, and is calculated using the following formula: Wherein, NIR, RED, and BLUE represent the reflectance of the near-infrared, red, and blue bands, respectively; G is the gain factor; C1 and C2 are atmospheric correction factors; and L is the soil adjustment factor. Meteorological data include actual evapotranspiration, minimum temperature, maximum temperature, cumulative precipitation, water vapor pressure, water vapor pressure difference, reference evapotranspiration, and the Palmer drought severity index.

3. The method for predicting maize yield based on chlorophyll fluorescence and deep learning according to claim 1, characterized in that, In S1, the meteorological data includes actual evapotranspiration, minimum temperature, maximum temperature, cumulative precipitation, water vapor pressure, water vapor pressure difference, reference evapotranspiration, and Palmer drought severity index.

4. The method for predicting maize yield based on chlorophyll fluorescence and deep learning according to claim 1, characterized in that, In step S1, the preprocessing of SIF data, vegetation index, meteorological data and historical yield data includes data format conversion, raster masking, regional statistics and spatiotemporal resolution unification.

5. The method for predicting maize yield based on chlorophyll fluorescence and deep learning according to claim 1, characterized in that, In step S3, the training, validation, and testing of the corn yield prediction model based on the training dataset specifically include: The optimal hyperparameters of the corn yield prediction model were determined using 10-fold cross-validation and grid search on the training dataset. The performance of the corn yield prediction model on different data splits was evaluated using 10-fold cross-validation to ensure the robustness of parameter selection. The grid search method was used to systematically explore the parameter space of the corn yield prediction model.

6. The method for predicting maize yield based on chlorophyll fluorescence and deep learning according to claim 1, characterized in that, S3 also includes an evaluation of the corn yield prediction model, with the evaluation index being the coefficient of determination R. 2 The root mean square error (RMSE) is calculated using the following formula: Where n is the number of observations, y i , These represent the measured value, the predicted value, and the mean of the measured data, respectively.

7. A maize yield prediction system based on chlorophyll fluorescence and deep learning, employing the maize yield prediction method based on chlorophyll fluorescence and deep learning as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire yield forecasting data for the target area over the past several years, including SIF data, vegetation index, meteorological data and historical yield data, and to preprocess the above-mentioned yield forecasting data. The data filtering module is used to filter data from hot and dry years according to a preset threshold, and to construct a training dataset under stress conditions based on the data from hot and dry years. The model building module constructs a deep neural network model, uses a training dataset to train, validate, and test the deep neural network model, determines the correlation between SIF data, vegetation index, meteorological data and maize yield, and obtains a maize yield prediction model based on DNN. The yield prediction module is used to input the data to be predicted into the trained corn yield prediction model, output the corn yield prediction results, and generate a spatial distribution map.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a maize yield prediction method based on chlorophyll fluorescence and deep learning as described in any one of claims 1 to 6.

Images