A crop yield prediction method based on feature clustering and convolutional neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2026-03-11
- Publication Date
- 2026-07-24
Smart Images
Figure CN122452828A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural remote sensing and crop yield prediction technology, and in particular to a crop yield prediction method based on feature clustering and convolutional neural networks. Background Technology
[0002] As an important grain and oil crop, soybean yield forecasting is of great significance for agricultural production management and macroeconomic decision-making. Traditional crop yield estimation methods mainly rely on large-scale ground surveys and statistical analysis, which have significant limitations in terms of timeliness and spatial coverage. In recent years, with the rapid development and widespread application of remote sensing technology, multi-source satellite remote sensing data has been increasingly incorporated into crop yield forecasting research, providing new technical means for large-scale, continuous yield monitoring and forecasting.
[0003] Crop yield prediction models can be broadly categorized into two types. The first type is based on crop growth mechanisms, such as WOFOST and DSSAT. These process-driven models simulate key biophysical processes during crop growth through a series of mathematical equations, including photosynthesis, respiration, dry matter accumulation and distribution, and phenological development, thereby achieving a dynamic characterization and quantitative representation of crop growth status. To further improve the accuracy of crop yield prediction, researchers assimilate or fuse remote sensing data into crop growth models to correct model state variables and key parameters. Although these models have strong mechanistic interpretability, they typically require extensive parameterization processes and have high computational costs. The second type is data-driven crop yield prediction models. These models predict crop yield by establishing direct mapping relationships between crop yield and various explanatory variables, which typically include remote sensing information and meteorological elements acquired during the growing season, such as vegetation indices, precipitation, and surface temperature. Machine learning methods used in crop yield prediction include both traditional algorithms such as random forests and support vector machines, as well as deep learning models that have developed rapidly in remote sensing applications in recent years, such as CNN, LSTM, and Transformer.
[0004] To achieve spatial consistency between remote sensing data and ground-based crop yield data, existing studies often employ spatial aggregation methods for remote sensing variables and meteorological elements. For example, pixel-level vegetation index data for the same crop type are aggregated into county-level time series by calculating statistical measures such as mean or quantiles. However, such aggregation methods ignore the significant phenological differences that exist between pixels of the same crop within the same county. Influenced by factors such as differences in sowing dates and inconsistent agricultural management practices, even spatially adjacent fields may exhibit significant differences in crop growth stages. Therefore, the phenological characteristics reflected in the vegetation index time series obtained by county-level aggregation often deviate significantly from the average phenological characteristics at the pixel level, mainly due to the scale effect of phenology. When pixels or farmlands with the same growing season length but different phenological changes within a region are aggregated using vegetation indices, the apparent growing season of that region is prolonged, thus distorting the true expression of crop phenological characteristics. This highlights the importance of explicitly considering intra-regional phenological differences in crop yield prediction models. Therefore, how to reasonably consider the scale effect in the process of multi-source time-series predictor modeling and use regional internal characteristics to characterize the impact of phenological changes is an urgent problem to be solved in the field of crop yield prediction. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a crop yield prediction method based on feature clustering and convolutional neural networks. This invention improves the accuracy and stability of crop yield prediction by performing feature clustering on multi-source time-series prediction factors and employing convolutional neural networks for feature extraction, thereby incorporating the influence of regional phenological changes.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A crop yield prediction method based on feature clustering and convolutional neural networks includes the following steps:
[0008] S1. Obtain multi-source data related to crop growth and historical crop yields within the study area;
[0009] The multi-source data includes vegetation index data, meteorological data, and soil moisture data;
[0010] S2. Standardize and preprocess the multi-source data to construct a unified time-scale pixel-level vegetation index time series, an administrative-level meteorological time series, and an administrative-level soil moisture time series;
[0011] S3. To address the phenological differences in crop growth within the same administrative region, phenological feature clustering is performed on the vegetation index time series characteristic sequences within that administrative region to divide them into multiple vegetation index time series subsets, thereby achieving the separation of different phenological features.
[0012] The average vegetation index time series of all vegetation index time series in the same vegetation index time series subset is processed to obtain the category average vegetation index time series of the subset, and the weight of the category average vegetation index time series is set.
[0013] S4. Construct a convolutional neural network model as a crop yield prediction model;
[0014] During model training, the time series of average vegetation index of all categories, the time series of meteorological data at the administrative level, and the time series of soil moisture at the administrative level are input into the model. The predicted yields obtained by forward propagation are weighted and aggregated to the administrative level to obtain the predicted yields within the administrative region. Then, the error between the predicted yields and the actual yields within the administrative region is used as the loss function to backpropagate and update the model parameters.
[0015] S5. Obtain multi-source data for the year to be predicted and perform standardized preprocessing to obtain pixel-level vegetation index time series feature sequences, administrative-level meteorological time series feature sequences, and administrative-level soil moisture time series feature sequences.
[0016] Step S3 is used to perform phenological feature clustering on the time series feature sequence of pixel-level vegetation index for the year to be predicted, and to obtain the time series of average vegetation index for each category and its corresponding weight.
[0017] The obtained time series of average vegetation index of the year to be tested and its corresponding weights, the time series of meteorological features at the administrative level, and the time series of soil moisture features at the administrative level are input into the trained crop yield prediction model, and the predicted crop yield in the study area is output.
[0018] Preferably, in step S3, the ratio of the number of pixels corresponding to all vegetation index time series belonging to the same vegetation index time series subset to the total number of pixels participating in clustering is used as the weight of the category average vegetation index time series of that subset.
[0019] Preferably, the vegetation index data is extracted from multi-band remote sensing reflectance images; the meteorological data is extracted from meteorological raster remote sensing images; and the soil moisture data is extracted from meteorological raster remote sensing images.
[0020] Preferably, the method for standardizing and preprocessing the multi-source data includes:
[0021] (1) Extract pixel-level vegetation index features from crop planting areas, and perform masking on the vegetation index image based on the administrative division layer to extract the pixel-level vegetation index time series within the administrative region.
[0022] (2) Perform time resampling or time alignment on multi-source data with different time resolutions to unify the time scale of different data;
[0023] (3) Perform interpolation, imputation, or removal processing on multi-source data that has missing or abnormal data;
[0024] (4) Perform Min-Max standardization on multi-source data with different dimensions to eliminate scale differences between different features.
[0025] Preferably, the convolutional neural network model includes at least one one-dimensional convolutional layer (1D-CNN); the one-dimensional convolutional layer is used to extract local temporal variation features from the time series.
[0026] Preferably, the phenological characteristics include at least one of temporal morphology, change trend and key growth stage change characteristics, thereby dividing pixel-level vegetation index temporal characteristic sequences with similar temporal change characteristics into the same subset.
[0027] Preferably, the phenological feature clustering adopts K-means clustering algorithm, ISODATA clustering algorithm (Iterative Self-Organizing Data Analysis Techniques Algorithm), etc.
[0028] Compared with the prior art, the present invention has at least the following beneficial effects:
[0029] (1) This invention models crop growth characteristics at the pixel level and obtains administrative-level prediction results through spatial aggregation, which effectively reduces the distortion of crop growth time sequence characteristics during spatial aggregation.
[0030] (2) By clustering vegetation index time series based on crop phenological differences, samples with similar growth dynamics are processed independently during the modeling process, reducing mutual interference between different growth patterns.
[0031] (3) Using convolutional neural networks to extract temporal feature sequences can effectively capture local temporal change features in the crop growth process and improve the ability to represent crop growth dynamics.
[0032] (4) During the model training process, the pixel-level prediction results are weighted and aggregated to the administrative level to participate in parameter updates, so that the model training process is consistent with the final prediction scale, which improves the rationality of the model structure.
[0033] (5) The method of the present invention is applicable to multi-source remote sensing data, meteorological data and soil moisture data, and has good versatility and scalability. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the overall process of a crop yield prediction method based on feature clustering and convolutional neural networks.
[0035] Figure 2 This is a schematic diagram of the ISODATA clustering results in an embodiment of the present invention (Calhoun County, Illinois, 2008).
[0036] Figure 3 This is a scatter plot of the soybean yield forecast results for 2022-2023 according to an embodiment of the present invention.
[0037] Figure 4 This is a schematic diagram showing the distribution of soybean yield forecast results for 2022-2023 according to an embodiment of the present invention. Among them, (A) is the actual predicted yield and the distribution of predicted yield in 2022 and 2023, and (B) is the distribution of absolute error in 2022 and 2023.
[0038] Figure 5 This is a schematic diagram showing the distribution of soybean yield prediction results under extreme drought conditions (2012) according to an embodiment of the present invention.
[0039] Figure 6 This is a line graph showing the predicted soybean production season for 2022-2023, based on an embodiment of the present invention. Detailed Implementation
[0040] The following detailed embodiments further illustrate the crop yield prediction method based on feature clustering and convolutional neural networks described in this invention, but the scope of protection of this invention is not limited thereto.
[0041] Taking soybeans from 13 states in the US Corn Belt region as an example, in this embodiment, the crop yield prediction process based on clustering features and convolutional neural networks is as follows: Figure 1 As shown, it includes the following steps:
[0042] S1. Obtain multi-source data related to crop growth and historical soybean yield within the study area.
[0043] The multi-source data includes vegetation index data, meteorological data, and soil moisture data. The vegetation index data is extracted from multi-band remote sensing reflectance images and is used to reflect the canopy change characteristics of soybeans during their growth process. The meteorological data includes daily minimum temperature (tmin), maximum temperature (tmax), and precipitation (precip) data, extracted from meteorological raster remote sensing images, and is used to describe the impact of environmental conditions such as temperature and precipitation on crop growth. The soil moisture data is extracted from meteorological raster remote sensing images and is used to describe the direct impact of agricultural irrigation, precipitation, evapotranspiration, etc., on crop growth.
[0044] In addition to the crop growth-related data mentioned above, historical soybean yield data at the county level corresponding to the study area were also obtained as reference data for model training and prediction.
[0045] S2. Standardize and preprocess the multi-source data to construct a unified time-scale pixel-level vegetation index time-series feature sequence, county-level meteorological time series, and county-level soil moisture time series.
[0046] Standardized preprocessing mainly includes: quality control, filtering, interpolation, and CDL masking of soybean regions in remote sensing images. The EVI2 (Enhanced Vegetation Index 2) of the soybean regions is calculated, and masking is performed according to county boundaries to obtain the time series of vegetation indices at the pixel level for all soybeans within the county. For meteorological and soil moisture data, the time scale is first unified with the remote sensing data. CDL masking is used to select soybean regions, and masking is performed according to county boundaries. The corresponding pixel values of soybean pixels are averaged and aggregated to the county level before interpolation to fill in missing values, resulting in meteorological time series (precipitation time series, maximum temperature time series, and minimum temperature time series) and soil moisture time series.
[0047] S3. To address the phenological differences in crop growth within the same county, the ISODATA clustering method is used to cluster all pixel-level vegetation index time series based on phenological characteristics, dividing them into multiple subsets of vegetation index time series to achieve the separation of different phenological characteristics.
[0048] After dividing the pixel-level vegetation index time series into subsets (categories), the pixel-level vegetation index time series of the same subset are averaged to obtain the category-averaged vegetation index series. The ratio of the number of pixels in that category to the total number of pixels participating in the clustering within the county is then used as the weight for that category in subsequent modeling. An example of the processing results is shown below. Figure 2 As shown, this is the clustering result of Calhoun County in 2008. The three curves represent the average vegetation index sequence of three categories with different phenological characteristics within the county.
[0049] S4. Construct a convolutional neural network model, Group-CNN, as a yield prediction model. This model can effectively capture the local temporal change features during crop growth and improve the ability to represent crop growth dynamics.
[0050] The convolutional neural network model structure in this embodiment mainly includes three sequentially connected convolutional modules: the first convolutional module consists of 64 convolutional kernels with a stride of 1 in its first layer and 32 convolutional kernels with a stride of 2 in its second layer; the second convolutional module consists of 16 convolutional kernels with a stride of 3 and a pooling layer with a window size of 2; the third convolutional module consists of 8 convolutional kernels with a stride of 3 and a pooling layer with a window size of 2. After passing through the three convolutional modules, the features are flattened and reduced in dimensionality by a flattening layer, then further extracted by a fully connected layer with 64 hidden units, and finally the predicted output is obtained through a regression output layer.
[0051] To match the time series dimensions after ISODATA clustering, this invention designs a loss function, GroupMSE, suitable for clustered time series. The principle of GroupMSE is that the average vegetation index time series of each category within a county, along with its corresponding meteorological and soil moisture time series, yields a predicted yield during forward propagation. However, since only county-level soybean yield records are available, the predicted yield is weighted and summed to the county level using the corresponding category weights in S3. Then, the county-level MSE is calculated and backpropagated to update the network parameters. This achieves the goal of predicting county-level soybean yield while considering phenological differences within the county but lacking county-level soybean pixel yield records.
[0052] S5. Divide the data into training, validation, and test sets according to the year; input the obtained time series data into the trained convolutional neural network model, output the predicted soybean yield for the test set, and use scatter plots and error distribution plots to quantitatively evaluate and visualize the prediction results.
[0053] The following three experiments are set up to evaluate the effectiveness of the yield prediction method based on clustering features and convolutional neural networks of the present invention.
[0054] Quantitative evaluation employed RMSE and R... 2 The model is measured using three metrics: Root Mean Square Error (RMSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). RMSE measures the deviation between predicted and actual values, and its value is a non-negative real number; a smaller value indicates higher model accuracy. 2 The value of 0 is used to measure how well the model fits the data. It ranges from 0 to 1, with a value closer to 1 indicating a better fit. The mean absolute percentage error (MAPE) measures the relative error of the prediction. It is a non-negative real number, with a smaller value indicating a more accurate prediction.
[0055] The design and implementation results of each experiment are as follows:
[0056] Experiment 1: To verify the yield prediction capability of the method of this invention under normal year conditions. 2022 and 2023 were selected as test years. The soybean yield of each county in the study area was predicted and compared with the actual statistical yield.
[0057] In this experiment, the proposed method (Group-CNN) was compared with several contrasting models, including Random Forest (RF), Support Vector Regression (SVR), PB-CNN (Phenology-guided Bayesian-CNN), and a "w / o intra-county" ablation model that does not consider intra-county phenological differences. All models were trained on the same training data and used to make predictions on data from 2022 and 2023.
[0058] Experimental results are as follows Figure 3 , Figure 4 As shown. Among them, Figure 3 A scatter plot of all models' forecasts for 2022-2023 is provided to illustrate the correspondence between predicted and actual outputs. Figure 4 (A) is a distribution map of actual and predicted output in 2022 and 2023, used to visually demonstrate the prediction effect of each model in different output level ranges; Figure 4 (B) is a distribution map of absolute errors in 2022 and 2023, which is used to show the spatial distribution of prediction errors of each model.
[0059] Experimental results show that the Group-CNN method proposed in this invention exhibits high prediction accuracy on two years of data. Its RMSE and MAPE indices are lower than those of the comparative model, while its R² value is higher, indicating that the model can fit the actual yield distribution well. Especially in medium-to-high yield areas, the prediction results of the method in this invention have better linear consistency with the actual yield, indicating that by introducing pixel growth difference information within counties, the regional-scale yield prediction accuracy can be effectively improved.
[0060] Experiment 2: Verify the applicability of the method of the present invention under extreme climatic conditions.
[0061] 2012 was selected as the test year for the experiment. 2012 was a typical drought year in the study area, and soybean yields in that year were severely affected by abnormal weather, thus making it highly representative.
[0062] In this experiment, the same comparative models from Experiment 1 were used to predict yield, and the predicted results were compared with the actual statistical yields. The experimental results are as follows: Figure 5 As shown, this scatter plot is used to illustrate the correspondence between predicted output and actual output.
[0063] Experimental results show that the Group-CNN method proposed in this invention can still maintain high prediction stability under extreme drought conditions. Compared with traditional machine learning models and models that do not consider intra-county differences, the prediction error of the method in this invention is significantly reduced, and the scatter distribution is closer to the 1:1 reference line, indicating that the model still has good generalization ability under extreme climatic conditions. This shows that by explicitly introducing intra-county crop growth difference information into the model, the model's adaptability to abnormal growth years can be enhanced, thereby improving the robustness of yield prediction.
[0064] Experiment 3: Evaluate the ability of the method of the present invention to predict yield in the middle of the growing season.
[0065] In agricultural production management, it is usually desirable to obtain crop yield prediction results as early as possible. Therefore, it is of great significance to evaluate the predictive performance of the model at different stages of the growing season.
[0066] In this experiment, based on year-round time series data, forecasts were made by gradually increasing the available time steps starting from the beginning of the year. Specifically, starting from DOY=201 (approximately the end of July), yield forecasts were made by adding one time step every 8 days until the end of the year. Each forecast used only the vegetation index time series, meteorological time series, and soil moisture time series available up to that date. The experimental results are as follows: Figure 6 As shown in the figure, this graph shows the variation curves of prediction accuracy of different models at different time points in the growing season, where the horizontal axis represents the prediction time point (DOY) and the vertical axis represents the model prediction accuracy index.
[0067] Experimental results show that the prediction accuracy of each model gradually improves with the increase of available time series length. The Group-CNN method proposed in this invention exhibits relatively stable prediction performance from the middle of the growing season and maintains high accuracy in the later stages. Traditional models (RF and SVR) outperform deep learning methods before mid-August, which may be attributed to their stronger robustness to noise in incomplete vegetation index time series. In contrast, Group-CNN shows stable and consistently superior performance from mid-August onwards.
[0068] The results show that the method of the present invention can not only achieve high-precision yield prediction after the end of the growing season, but also provide stable prediction results in the middle of the growing season, which has high practical application value.
Claims
1. A crop yield prediction method based on feature clustering and convolutional neural networks, characterized in that, Includes the following steps: S1. Obtain multi-source data related to crop growth and historical crop yields within the study area; The multi-source data includes vegetation index data, meteorological data, and soil moisture data; S2. Standardize and preprocess the multi-source data to construct a unified time-scale pixel-level vegetation index time series, an administrative-level meteorological time series, and an administrative-level soil moisture time series; S3. To address the phenological differences in crop growth within the same administrative region, phenological feature clustering is performed on the vegetation index time series characteristic sequences within that administrative region to divide them into multiple vegetation index time series subsets, thereby achieving the separation of different phenological features. The average vegetation index time series of all vegetation index time series in the same vegetation index time series subset is processed to obtain the category average vegetation index time series of the subset, and the weight of the category average vegetation index time series is set. S4. Construct a convolutional neural network model as a crop yield prediction model; During model training, the time series of average vegetation index of all categories, the time series of meteorological data at the administrative level, and the time series of soil moisture at the administrative level are input into the model. The predicted yields obtained by forward propagation are weighted and aggregated to the administrative level to obtain the predicted yields within the administrative region. Then, the error between the predicted yields and the actual yields within the administrative region is used as the loss function to backpropagate and update the model parameters. S5. Obtain multi-source data for the year to be predicted and perform standardized preprocessing to obtain pixel-level vegetation index time series feature sequences, administrative-level meteorological time series feature sequences, and administrative-level soil moisture time series feature sequences. Step S3 is used to perform phenological feature clustering on the time series feature sequence of pixel-level vegetation index for the year to be predicted, and to obtain the time series of average vegetation index for each category and its corresponding weight. The obtained time series of average vegetation index of the year to be tested and its corresponding weights, the time series of meteorological features at the administrative level, and the time series of soil moisture features at the administrative level are input into the trained crop yield prediction model, and the predicted crop yield in the study area is output.
2. The crop yield prediction method based on feature clustering and convolutional neural networks as described in claim 1, characterized in that, In step S3, the ratio of the number of pixels corresponding to all vegetation index time series belonging to the same vegetation index time series subset to the total number of pixels participating in clustering is used as the weight of the category average vegetation index time series of that subset.
3. The crop yield prediction method based on feature clustering and convolutional neural networks as described in claim 2, characterized in that, The phenological characteristics include at least one of temporal morphology, change trend, and key growth stage change characteristics.
4. The crop yield prediction method based on feature clustering and convolutional neural networks as described in claim 3, characterized in that, The vegetation index data was extracted from multi-band remote sensing reflectance images; the meteorological data was extracted from meteorological raster remote sensing images; and the soil moisture data was extracted from meteorological raster remote sensing images.
5. The crop yield prediction method based on feature clustering and convolutional neural networks as described in claim 4, characterized in that, The methods for standardizing and preprocessing the multi-source data include: (1) Extract pixel-level vegetation index features from crop planting areas, and perform masking on the vegetation index image based on the administrative division layer to extract the pixel-level vegetation index time series within the administrative region. (2) Perform time resampling or time alignment on multi-source data with different time resolutions to unify the time scale of different data; (3) Perform interpolation, imputation, or removal processing on multi-source data that has missing or abnormal data; (4) Perform Min-Max standardization on multi-source data with different dimensions to eliminate scale differences between different features.
6. A crop yield prediction method based on feature clustering and convolutional neural networks as described in any one of claims 3-5, characterized in that, The convolutional neural network model includes at least one one-dimensional convolutional layer for extracting local temporal variation features from the time series.
7. The crop yield prediction method based on feature clustering and convolutional neural networks as described in claim 6, characterized in that, The phenological features are clustered using either the K-means clustering algorithm or the ISODATA clustering algorithm.