A crop yield prediction method based on modeling of reference year time series differences

Abstract

The application discloses a county scale soybean yield prediction method based on reference year time series difference modeling, and relates to the technical field of agricultural remote sensing and intelligent information processing. The method first acquires multi-source crop growth data and administrative level soybean yield data of previous years in a research area, and constructs a spatial scale multi-source time sequence characteristic sequence of a unified space-time scale after preprocessing; then, a reference year is selected to construct a sample pair, and a multi-source time series difference sequence and a yield difference value are calculated; subsequently, a deep learning model is constructed, the difference sequence is taken as input, and the yield difference value is taken as a label to train the model; finally, the difference sequence of a to-be-measured sample is input to obtain a predicted yield difference value, and the yield of the to-be-measured year is obtained by combining the yield of the reference year. The application realizes explicit depiction of interannual differences, effectively expands training samples, reduces extreme value interference, improves prediction stability and precision, adapts to various deep learning models, and is easy to popularize in large-scale areas.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and intelligent information processing technology, and in particular to a method for predicting crop yield based on time-series differences modeled from a reference year. Background Technology

[0002] Against the backdrop of continuous global population growth, intensified climate change, and increasingly scarce land resources, agriculture faces unprecedented pressure. Soybeans, as one of the world's most important food and oilseed crops, play a crucial role in ensuring food security, stabilizing vegetable oil supply, and promoting livestock development. In recent years, with population growth and changes in consumption patterns, global demand for soybeans has continued to increase, making the stability and sustainability of soybean production a key concern in international agriculture. Soybeans are widely used in human food, animal feed, and industrial processing, and the stability and fluctuations in their yield directly affect global food security and agricultural economic development. Therefore, conducting research on soybean yield forecasting and obtaining accurate and timely crop yield predictions is increasingly important. This type of information is a key foundation for achieving efficient agricultural management and scientific decision-making, and is of great significance for improving agricultural production management, optimizing agricultural product market regulation, and serving national food security strategies.

[0003] Traditional yield prediction methods mainly include statistical regression models, crop growth mechanism models, and machine learning-based prediction models. Statistical methods rely on the empirical relationship between meteorological factors and yield, making it difficult to characterize complex nonlinear processes; crop growth models require a large number of parameters and field trial data, making it difficult to generalize to large-scale regions; although deep learning methods that have emerged in recent years can extract dynamic characteristics of crop growth using remote sensing time series data, most models use "year-county" as independent samples, directly learning the mapping relationship between the current year's growth curve and the current year's yield.

[0004] Most current studies use county-level or regional yield as the target and average time-series data as model input, neglecting the spatial heterogeneity of crop growth processes within a region. Furthermore, differences in climate conditions and management practices between years lead to significant variations in crop growth curves, and modeling based solely on a single year's sample is insufficient to fully learn the impact of interannual differences on yield. Within the same county, similar soil backgrounds, planting systems, and management levels exist across different years, with yield variations primarily driven by climate fluctuations, extreme weather, and differences in growth processes. Existing methods ignore this "interannual variation information within the same spatial unit," preventing models from explicitly characterizing the contribution of growth changes to yield changes. Moreover, because samples are typically measured in "county-year" units, the sample size is limited, making deep learning models prone to overfitting and insufficient generalization. Therefore, how to fully utilize growth difference information between different years within the same county under limited sample conditions to construct a model capable of directly predicting yield changes is a critical technical problem that urgently needs to be solved in the field of county-scale crop yield prediction. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a crop yield prediction method based on time-series difference modeling of reference years. This invention establishes a mapping relationship between growth differences and yield differences by constructing multi-source time-series difference features of "the year to be predicted – the reference year." Furthermore, it utilizes a sample expansion mechanism built with multiple reference years to increase the number of training samples for the model, thereby effectively improving the stability and accuracy of crop yield prediction at the administrative level.

[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 time-series differences modeling of reference years includes the following steps:

[0008] S1. Data Acquisition;

[0009] We obtained multi-source data related to crop growth over the years within the study area, as well as historical soybean yield data for administrative regions within the study area.

[0010] The multi-source data includes vegetation index data, meteorological data, and soil moisture data.

[0011] S2. Data preprocessing;

[0012] The multi-source data is standardized and preprocessed to construct an administrative-level multi-source time series feature sequence with a unified time and spatial scale.

[0013] The multi-source time-series feature sequences include vegetation index time-series feature sequences, meteorological time-series feature sequences, and soil moisture time-series feature sequences.

[0014] S3. Reference year selection and sample pairing;

[0015] Select at least one reference year within the same administrative region to construct a "target year - reference year" sample pair.

[0016] S4. Construction of temporal difference features;

[0017] The difference between the multi-source time-series feature sequences of the “target year-reference year” sample pair is calculated to obtain the multi-source time-series difference sequence reflecting the interannual growth difference, and the yield difference of the sample pair is calculated at the same time.

[0018] The multi-source time-series difference sequences include vegetation index time-series difference sequences, meteorological time-series difference sequences, and soil moisture time-series difference sequences.

[0019] S5. Deep learning-based differential mapping modeling;

[0020] Construct a deep learning model; use the multi-source time-series difference sequence as input data for the deep learning model, and the yield difference of the sample pair as the label, and train the model to learn the mapping relationship of "growth difference - yield difference";

[0021] S6. Production back-calculation and result output;

[0022] Construct a "test year - reference year" sample pair and calculate the corresponding multi-source time series difference sequence; input the obtained multi-source time series difference sequence into the trained deep learning model and output the predicted yield difference.

[0023] The predicted production difference is superimposed with the production of the reference year to calculate the production forecast for the year to be tested.

[0024] Preferably, in step S6, if there are multiple reference years, the average of the obtained production forecast results is taken as the final forecast result.

[0025] Preferably, the deep learning model is a one-dimensional convolutional neural network model, CNN, LSTM, or LSTM-TA (LSTM with temporal attention).

[0026] 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.

[0027] Preferably, the method for standardizing and preprocessing the multi-source data includes:

[0028] (1) SG filtering and linear interpolation are used to process the vegetation index data to ensure that the obtained vegetation index time series feature sequence is complete and reliable.

[0029] (2) Perform temporal and spatial resampling and alignment on multi-source data with different spatiotemporal resolutions to unify the temporal and spatial scales of different data.

[0030] (3) Extract multi-source temporal feature sequences at the administrative level by performing spatial aggregation masking based on the administrative region vector boundary.

[0031] Compared with the prior art, the present invention has at least the following beneficial effects:

[0032] (1) Explicitly characterize interannual differences, transforming the traditional yield prediction problem into a difference prediction problem, enabling the model to directly learn the impact mechanism of crop growth changes on yield changes, and improving the pertinence and interpretability of modeling.

[0033] (2) It has strong sample expansion capability. By matching multiple historical reference years for the same year to be tested to construct differential sample pairs, the training data scale can be effectively expanded without introducing additional original data, which significantly alleviates the overfitting problem caused by the limited sample size of deep learning models and improves the training stability of the model.

[0034] (3) The constructed time series difference sequence has the characteristics of near zero mean and small numerical fluctuation range, which can effectively reduce the interference of extreme values ​​on model training and further improve the stability and reliability of yield prediction results.

[0035] (4) It is suitable for large-scale regional promotion and application. This method is based on administrative-level regional statistical data and remote sensing time series data. It does not rely on field test parameters, gets rid of the limitation of field test, and is easy to apply in a large area.

[0036] (5) The model has good adaptability and versatility. The core method of differential modeling does not depend on a specific network structure and can be flexibly combined with various temporal deep learning models such as convolutional neural networks, recurrent neural networks, and Transformers to adapt to different modeling needs and application scenarios. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall process of a crop yield prediction method based on time-series differences in reference years.

[0038] Figure 2 This is a scatter plot of the prediction results for 2022-2023 based on an embodiment of the present invention.

[0039] Figure 3 This is a spatial distribution map of the actual and predicted soybean yields in 2022-2023, based on an embodiment of the present invention.

[0040] Figure 4This is a spatial distribution map of the absolute error in soybean yield prediction for 2022-2023 according to an embodiment of the present invention.

[0041] Figure 5 This is a scatter plot showing the soybean yield prediction results under extreme drought conditions (2012) according to an embodiment of the present invention. Detailed Implementation

[0042] The following detailed description of a crop yield prediction method based on time-series differences of reference years, according to specific embodiments, further illustrates this invention, but the scope of protection of this invention is not limited thereto.

[0043] This study selected 13 major agricultural states in the United States as the research area, including Illinois, Indiana, Iowa, Kansas, Kentucky, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, and Wisconsin. These regions are important soybean producing areas in the United States and are highly representative of agricultural production.

[0044] In this embodiment, the crop yield prediction process based on time-series differences of the reference year is as follows: Figure 1 As shown, it includes the following steps:

[0045] S1, Data Acquisition;

[0046] We obtained multi-source data related to soybean growth over the years within the study area, as well as historical soybean yield data for administrative regions within the study area.

[0047] The multi-source data includes vegetation index data, meteorological data, and soil moisture data. The vegetation index data uses MODIS satellite product MOD09A1 data, and crop growth status information is obtained by calculating the enhanced vegetation index EVI2; this data has a spatial resolution of 500m and a temporal resolution of 8 days. The meteorological data uses the Daymet Version 4 meteorological dataset, obtaining daily minimum temperature (tmin), maximum temperature (tmax), and precipitation (precip) data, with a spatial resolution of 1km. The soil moisture data uses profile soil moisture data from the NASA GLDAS v2.2 CLSM land surface model product, with a spatial resolution of 0.25°.

[0048] Meanwhile, we obtained historical soybean production data for each county-level administrative region within the study area, spanning from 2008 to 2023.

[0049] S2, Data Preprocessing;

[0050] The multi-source data undergoes standardized preprocessing to construct a multi-source temporal series feature sequence with a unified time and spatial scale. Specifically, this includes:

[0051] First, the EVI2 (vegetation index data) is interpolated and filtered to reduce the impact of remote sensing noise on the time series.

[0052] Secondly, spatial resampling and regional averaging were performed on EVI2, meteorological data, and soil moisture data to ensure consistency with county-level administrative regions (spatial scale).

[0053] Then, the temporal resolution of different data sources is unified to an 8-day time step (time scale), and a year-round time series feature sequence is constructed; in this embodiment, each year contains 46 time steps. The final result is a multi-source time series with a unified structure: vegetation index time series, meteorological time series (precipitation time series, maximum temperature time series, minimum temperature time series), and soil moisture time series. These sequences together constitute the multi-source feature input for each county for each year.

[0054] S3. Reference year selection and sample pairing;

[0055] Within the same administrative region, at least one reference year is selected to construct a "target year - reference year" sample pair. In this embodiment, under the condition that the reference year is less than the target year, every year is used as the reference year. For example, if the target year is 2012, then the reference years can be 2008, 2009, and 2010, thus forming sample pairs 2012-2008, 2012-2009, 2012-2010, and 2012-2011, and so on, which will expand the sample size several times.

[0056] S4. Construction of temporal difference features;

[0057] The difference between the multi-source time-series feature sequences of the “target year-reference year” sample pairs is calculated to obtain the multi-source time-series difference sequence reflecting interannual growth differences. At the same time, the yield difference of the sample pair is calculated to obtain complete sample features and labels.

[0058] The multi-source time-series difference sequences include vegetation index time-series difference sequences, meteorological time-series difference sequences (precipitation time series, maximum temperature time series, minimum temperature time series) and soil moisture time-series difference sequences.

[0059] S5. Deep learning-based differential mapping modeling;

[0060] In this embodiment, deep learning models are constructed, including CNN, LSTM, LSTM-TA (LSTM with added time attention mechanism), and CNN-RNN models, to illustrate the advantages of the differential modeling of the present invention. The multi-source temporal differential sequence is used as the input data of the deep learning model, and the yield difference of the sample pair is used as the label. The model is trained to learn the mapping relationship of "growth difference - yield difference".

[0061] S6. Production back calculation and result output;

[0062] Construct a "test year - reference year" sample pair and calculate the corresponding multi-source time series difference sequence; input the obtained multi-source time series difference sequence into the trained deep learning model and output the predicted yield difference.

[0063] The predicted output difference is superimposed with the output of the reference year to calculate the output forecast for the year to be tested. That is, the predicted output for the year to be tested = the output of the reference year + the predicted output difference.

[0064] In addition, to improve the accuracy of prediction, sample pairs can be constructed between the year to be predicted and multiple reference years to obtain the predicted output. Then, the average of all predicted outputs can be taken to obtain the final prediction result for the county in that year.

[0065] The following two experiments are designed to evaluate the effectiveness of the crop yield prediction method based on time-series differences of reference years in this invention.

[0066] 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.

[0067] The design and implementation results of each experiment are as follows:

[0068] Experiment 1: Used to verify the performance improvement of the reference year difference modeling method proposed in this invention compared with the traditional absolute sequence modeling method.

[0069] In this experiment, 2008-2019 were selected as training years, 2020 and 2021 as validation years, and 2022 and 2023 as test years to predict soybean yields in counties within the study area.

[0070] Yield prediction was performed using both the absolute sequence modeling method and the differential modeling method proposed in this invention, and the prediction results were compared. The results of the absolute sequence modeling are shown in Table 1:

[0071] Table 1. Absolute sequence modeling results

[0072] Evaluation indicators CNN LSTM CNN-RNN LSTM-TA RMSE (bu / ac) 6.17 6.21 5.72 5.89 <![CDATA[R 2 ]]> 0.75 0.74 0.78 0.77 MAPE (%) 12.08 11.79 10.78 10.96

[0073] Figure 2 A scatter plot of the prediction results of the difference modeling method in 2022-2023 is used to show the correspondence between the predicted output and the actual output obtained by different deep learning models in this invention. Figure 3 The images show the actual production distribution for 2022 and 2023, as well as the predicted production distribution using LSTM-TA. Figure 4 The absolute error distribution plots for LSTM-TA in 2022 and 2023 are used to illustrate the error distribution and predicted output of differential modeling.

[0074] Experimental results show that, for the same model structure (CNN, LSTM, CNN-RNN, LSTM-TA), compared to directly using absolute sequences for modeling, the method proposed in this invention based on reference year difference modeling can significantly reduce prediction errors and improve the model's ability to characterize yield changes. Among them, the LSTM-TA model exhibits the best prediction performance within the difference modeling framework, with its predicted yield distribution showing high consistency with the actual yield distribution, and its absolute error being more spatially concentrated.

[0075] Experiment 2: Used to verify the applicability of the method of the present invention under extreme climatic conditions.

[0076] 2012 was a typical year of extreme drought in the US Corn Belt, with crop growth significantly impacted by abnormal weather. 2012 was selected as the test year, and the validation set remained 2020-2021, but samples with a reference year of 2012 were explicitly removed. The training sets, 2008-2011 and 2012-2019, also required explicit removal of samples with a reference year of 2012.

[0077] Similarly, CNN, LSTM, CNN-RNN, and LSTM-TA models were used to predict soybean yields in various counties in 2012, and the predicted results were compared with the actual yields. Figure 5 This is a scatter plot of the 2012 prediction results using the difference modeling method of this invention, used to show the correspondence between predicted output and actual output.

[0078] Experimental results show that, under conditions of extreme drought, the reference year time series difference modeling method proposed in this invention can still maintain high prediction stability, indicating that the method can more effectively capture the impact of climate anomalies on crop growth and yield changes.

Claims

1. A crop yield prediction method based on time-series difference modeling of a reference year, characterized in that, Includes the following steps: S1. Data Acquisition; Acquire multi-source data related to crop growth over the years within the study area, as well as historical soybean yield data for administrative regions within the study area; S2. Data preprocessing; The multi-source data is standardized and preprocessed to construct an administrative-level multi-source time-series feature sequence with a unified time and spatial scale; S3. Reference year selection and sample pairing; Within the same administrative region, select at least one reference year to construct a "year to be tested - reference year" sample pair; S4. Construction of temporal difference features; The difference between the multi-source time series characteristic sequences of the "test year - reference year" sample pair is calculated to obtain the multi-source time series difference sequence reflecting the interannual growth difference, and the yield difference of the sample pair is calculated at the same time. S5. Deep learning-based differential mapping modeling; Construct a deep learning model; use the multi-source time-series difference sequence as input data for the deep learning model, and the yield difference of the sample pair as the label, and train the model to learn the mapping relationship of "growth difference - yield difference"; S6. Production back calculation and result output; Construct "test year - reference year" sample pairs and calculate the corresponding multi-source time series difference sequences; input the obtained multi-source time series difference sequences into the trained deep learning model and output the predicted yield difference; The predicted production difference is superimposed with the production of the reference year to calculate the production forecast for the year to be tested.

2. The crop yield prediction method based on time-series difference modeling of a reference year as described in claim 1, characterized in that, The multi-source data includes vegetation index data, meteorological data, and soil moisture data; The multi-source time-series feature sequences include vegetation index time-series feature sequences, meteorological time-series feature sequences, and soil moisture time-series feature sequences; The multi-source time-series difference sequences include vegetation index time-series difference sequences, meteorological time-series difference sequences, and soil moisture time-series difference sequences.

3. A crop yield prediction method based on time-series difference modeling of a reference year as described in claim 1 or 2, characterized in that, In step S6, if there are multiple reference years, the average of the obtained production forecast results is taken as the final forecast result.

4. The crop yield prediction method based on time-series difference modeling of a reference year as described in claim 3, characterized in that, The deep learning model is a one-dimensional convolutional neural network model, CNN, LSTM, or LSTM-TA.

5. The crop yield prediction method based on time-series difference modeling of a reference year 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.

6. The crop yield prediction method based on time-series difference modeling of a reference year as described in claim 5, characterized in that, The methods for standardizing and preprocessing the multi-source data include: (1) SG filtering and linear interpolation are used to process the vegetation index data to ensure that the obtained vegetation index time series feature sequence is complete and reliable; (2) Perform temporal and spatial resampling and alignment on multi-source data with different spatiotemporal resolutions to unify the temporal and spatial scales of different data. (3) Extract multi-source temporal feature sequences at the administrative level by performing spatial aggregation masking based on the administrative region vector boundary.

Images