A mechanism knowledge and multi-source data fusion driven rice above-ground biomass estimation method

Abstract

The application discloses a kind of mechanism knowledge and multi-source data fusion driven rice above-ground biomass estimation method.It includes: using rice growth model, according to the multi-source environmental data of the region to be estimated to generate potential biomass time series information of each growth stage;Multi-source environmental data and rice potential biomass simulation data are input into the trained rice above-ground biomass estimation model to obtain the rice above-ground biomass estimation of each growth stage;The rice above-ground biomass estimation model includes potential biomass module and actual environmental restriction module, potential biomass module and actual environmental restriction module learn potential biomass characteristics and environmental restriction characteristics respectively, and carry out growth stage alignment feature interaction, output the rice above-ground biomass estimation of each growth stage.The application realizes high-precision biomass estimation without a large number of measured data, and performs excellently in interannual generalization, small sample learning and the like, and can be used for rice growth monitoring, yield prediction and agricultural management.

Description

Technical Field

[0001] This invention belongs to the field of agricultural information technology, specifically relating to a method for estimating aboveground biomass of rice based on mechanistic knowledge and multi-source data fusion. Background Technology

[0002] Aboveground biomass of rice is a core indicator reflecting crop growth and yield formation, playing a crucial role in growth monitoring, yield prediction, disaster assessment, and agricultural management decision-making. In recent years, with the development of remote sensing and digital agriculture technologies, crop biomass estimation using multi-source data has become a hot topic in smart agriculture research. However, existing technologies still face challenges in terms of accuracy and generalization ability under complex environments.

[0003] Traditional crop mechanistic models (such as RiceGrow, APSIM, and CERES-Rice) describe physiological processes such as photosynthetic accumulation and dry matter distribution using mathematical equations, and can be used to simulate biomass changes under different conditions. Although mechanistic models have advantages in terms of scientific validity and interpretability, they are highly dependent on field input data and suffer from problems such as complex parameters and weak cross-temporal and spatial transfer, which restricts their widespread application in large-scale and complex environments.

[0004] Deep learning methods can automatically extract features and fit nonlinear relationships from multi-source data such as remote sensing and meteorology, and have been used to predict crop growth status and yield. However, rice biomass accumulation exhibits significant stage-dependent and process-dependent characteristics. Purely data-driven models lack a structured representation of the key mechanism of "potential growth-environmental constraints," making it difficult to maintain stable spatiotemporal generalization. Furthermore, due to the limited number of high-quality biomass observation samples, deep learning models are prone to overfitting under small sample conditions. While existing methods for integrating mechanistic models with deep learning have been explored, they still generally suffer from the following shortcomings: (1) The action pathways of potential growth and environmental limiting factors were not explicitly reflected in the model structure, resulting in insufficient utilization of mechanistic information; (2) The lack of a system pre-training-fine-tuning strategy based on mechanism simulation data limits the effective transfer of mechanism knowledge to model parameters.

[0005] Therefore, it is necessary to develop a method for estimating aboveground rice biomass that can integrate mechanistic knowledge with the advantages of multi-source data, while taking into account process rationality and estimation generalization performance. Summary of the Invention

[0006] To address the problems and needs existing in the background technology, this invention proposes a method for estimating aboveground biomass of rice driven by the fusion of mechanistic knowledge and multi-source data. This invention achieves high-precision biomass estimation without requiring a large amount of measured data, and exhibits excellent performance in interannual generalization and small-sample learning. It can be used for rice growth monitoring, yield prediction, and agricultural management.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: I. A method for estimating aboveground biomass of rice driven by the fusion of mechanistic knowledge and multi-source data.

[0008] Step 1: Using a rice growth model, generate time-series simulation data of potential rice biomass at each growth stage based on multi-source environmental data of the area to be estimated. Step 2: Input the multi-source environmental data and rice potential biomass time series simulation data into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at at least one growth stage. The multi-source environmental data includes remote sensing data, meteorological data, soil data, and geographic location data; The growth stages include the jointing stage, the flowering stage, and the maturity stage; The rice aboveground biomass estimation model includes a potential biomass module and an actual environmental constraint module. The potential biomass module and the actual environmental constraint module learn potential biomass characteristics and environmental constraint characteristics, respectively, and carry out feature interaction aligned with growth stages to generate and output the estimated aboveground biomass of rice at each growth stage.

[0009] Specifically, in the potential biomass module, the daily time-series simulation data of rice potential biomass at each growth stage is sent to the first long short-term memory network unit, and the hidden layer features of the first long short-term memory network unit at each growth stage time step are extracted as the potential biomass features of the corresponding stage.

[0010] Specifically, for the actual environmental constraint module, the remote sensing data and meteorological data from the multi-source environmental data are constructed as time-series inputs according to three growth stages. The actual environmental constraint module receives the daily time-series inputs for each growth stage and inputs them into a second long short-term memory network unit. The hidden layer features of the second long short-term memory network unit at each time step of the growth stage are extracted as the dynamic environmental constraint features for the corresponding stage. The actual environmental constraint module receives the soil data and geographic location data and encodes them through a fully connected layer to obtain static environmental constraint features. The dynamic environmental constraint features and static environmental constraint features for each growth stage are time-aligned and concatenated to obtain the environmental constraint features for that growth stage.

[0011] Specifically, in the trained rice aboveground biomass estimation model, the potential biomass characteristics and environmental constraint characteristics of each growth stage are temporally aligned and fused to obtain the fused characteristics of that growth stage. The fused characteristics of the three growth stages constitute the temporal information of rice growth characteristics. The temporal information of rice growth characteristics is sequentially input into the third long short-term memory network unit and the fully connected layer for processing, and the estimated value of rice aboveground biomass for the corresponding growth stage is output.

[0012] Specifically, the training process of the rice aboveground biomass estimation model includes the following steps: Step D1: Obtain multiple sets of historical multi-source environmental data and the actual aboveground biomass of mature rice corresponding to each set of historical multi-source environmental data. Use the rice growth model to generate simulated potential and actual biomass data of rice corresponding to each set of historical multi-source environmental data. The simulated potential biomass data refers to the daily simulated time series of biomass at each growth stage by the rice growth model without considering actual environmental stress, and is used to characterize the simulated potential biomass information. The simulated actual biomass data refers to the daily simulated time series of biomass at each growth stage by the rice growth model considering actual environmental stress, and is used to characterize the simulated actual biomass information.

[0013] Step D2: Construct a pre-training model for estimating rice aboveground biomass. Take meteorological data, soil data, and geographical location data from each set of historical multi-source environmental data, as well as the corresponding daily simulated time-series data of potential rice biomass, as inputs, and simulated data of actual rice biomass as labels. Pre-train the rice aboveground biomass estimation model so that it learns the biomass accumulation mechanism process, and obtain the pre-trained rice aboveground biomass estimation model.

[0014] Step D3: Fine-tuning the parameters of the rice aboveground biomass estimation pre-training model constructed in step D2: Using the historical multi-source environmental data and the corresponding daily simulated time-series data of potential rice biomass as input, and the actual aboveground biomass of rice at the target growth stage corresponding to the multi-source environmental data as labels, fine-tuning the pre-trained rice aboveground biomass estimation model obtained in step D2 to obtain a trained rice aboveground biomass estimation model.

[0015] Specifically, in step D1, the aboveground biomass data of rice at the actual maturity stage is obtained by converting county-level statistical yield and harvest coefficient.

[0016] Specifically, in step D3, the target growth stage includes at least the mature stage. In practice, the target growth stage is determined by the growth stages covered by the true biomass labels in the dataset. Specifically: if the dataset only contains true biomass labels for the mature stage, then the target growth stage is the mature stage; if it also contains true biomass labels for other growth stages, then the target stage can be expanded accordingly to achieve the estimation of time-series biomass across multiple growth stages.

[0017] Specifically, the remote sensing data includes the enhanced vegetation index and the normalized yellowness index.

[0018] Specifically, the meteorological data includes high-temperature stress factors and daily precipitation.

[0019] Specifically, the soil data includes pH, organic matter content, total nitrogen content, bulk density, and clay content.

[0020] Specifically, the geographic location data includes latitude and longitude.

[0021] II. A rice aboveground biomass estimation system applied to a rice aboveground biomass estimation method driven by the fusion of the aforementioned mechanistic knowledge and multi-source data.

[0022] The rice aboveground biomass estimation system includes: The data acquisition unit is used to acquire multi-source environmental data of the area to be estimated. The mechanism knowledge unit is used to receive multi-source environmental data of the area to be estimated and input it into the rice growth model to obtain time-series simulation data of potential biomass of rice at each growth stage. The rice aboveground biomass estimation unit is used to receive multi-source environmental data of the area to be estimated and time-series simulation data of potential rice biomass at various growth stages, and input them into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at at least one growth stage.

[0023] The beneficial effects of this invention are: 1. The rice aboveground biomass estimation method proposed in this invention, driven by mechanistic knowledge and multi-source data fusion, constructs a dual-branch neural network structure containing a potential growth module and an environmental constraint module. Combined with a pre-training strategy based on RiceGrow simulated samples and a fine-tuning strategy using real samples, it achieves a structured representation of the rice biomass accumulation process. This method maintains stable estimation performance even under conditions of limited measured samples, complex environmental conditions, and significant regional differences.

[0024] 2. Compared to traditional methods that rely solely on mechanistic models, this invention effectively compensates for the sensitivity of mechanistic models to differences in field environments and management uncertainties by introducing multi-source data such as remote sensing indices, meteorological stress, soil properties, and geographical location. This improves the systematic biases that easily occur in complex environments. Compared to purely data-driven models, this invention uses mechanistic simulation pre-training to enable the neural network to learn the phased and procedural priors of biomass accumulation. This allows the model to maintain strong generalization ability under small sample conditions and significantly reduces its dependence on large-scale observational data.

[0025] 3. By extracting and fusing phased features from the potential growth module and the environmental constraint module, this invention can capture the contributions of ideal growth potential and actual environmental constraints to biomass formation, respectively. Ablation results show that the potential growth module performs better in high biomass scenarios, while the environmental constraint module has higher estimation stability under extreme stress conditions. The combination of the two significantly improves the model's adaptability and estimation accuracy under different growth scenarios.

[0026] In summary, this invention overcomes the problems of weak transferability of mechanistic models and lack of mechanistic constraints in data-driven models, and realizes a method for estimating aboveground biomass of rice that combines mechanistic rationality with robustness in complex environments. It is applicable to regional growth monitoring, yield assessment, agricultural management, and carbon sequestration. Attached Figure Description

[0027] Figure 1 A roadmap for the overall technology of rice aboveground biomass estimation driven by the fusion of mechanistic knowledge and multi-source data.

[0028] Figure 2 A neural network structure diagram for estimating aboveground biomass of rice based on mechanistic knowledge.

[0029] Figure 3 This is a diagram showing the accuracy of the aboveground biomass estimation for rice at maturity in this embodiment.

[0030] Figure 4 The diagram shows the biomass accumulation process learned by the model during pre-training in this embodiment and the changes in model estimation before and after fine-tuning. (a) shows the biomass accumulation process learned during pre-training, and (b) shows the comparison of model error before and after fine-tuning.

[0031] Figure 5 This is a comparison chart of the accuracy of model structure ablation in this embodiment, where (a) is the model accuracy of high biomass samples and (b) is the model accuracy of extreme stress samples.

[0032] Figure 6 This is a comparison chart of model accuracy under different training sample sizes in this embodiment. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] This invention provides a method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion. The method first extracts temporal information of potential rice biomass based on the RiceGrow crop mechanism model and integrates multi-source data such as remote sensing indices, meteorological stress, soil properties, and geographical location to construct an environmental constraint dataset. Then, a rice aboveground biomass estimation model is constructed, including a potential growth module and an environmental constraint module. Phenologically aligned feature fusion is performed at the jointing, flowering, and maturity stages to output the aboveground biomass estimation results. During training, the network is first pre-trained using RiceGrow mechanism simulation samples to learn the biomass accumulation mechanism; then, fine-tuning is performed using real biomass samples to correct mechanistic biases and improve the model's adaptability to complex environments.

[0035] The method of the present invention specifically includes the following steps: Step 1: Using a rice growth model, generate potential biomass time-series information for each growth stage based on multi-source environmental data of the area to be estimated.

[0036] Specifically, the multi-source environmental data includes remote sensing data, meteorological data, soil data, and geographic location data, which are used to characterize the canopy growth limitation, meteorological stress, soil constraints, and regional differences experienced by rice in actual habitats.

[0037] Specifically, the remote sensing data includes the Enhanced Vegetation Index (EVI) and the Normalized Yellowness Index (NDYI); the meteorological data includes the high-temperature stress factor (KDD) and daily precipitation (PRCP); the soil data includes pH, organic matter content, total nitrogen content, bulk density, and clay content; and the geographic location data includes latitude and longitude.

[0038] Specifically, the growth stages include the jointing stage, the flowering stage, and the maturity stage. That is, the three growth stages of transplanting-jointing, jointing-flowering, and flowering-maturity.

[0039] Preferably, the time series information is daily time series information.

[0040] Preferably, the area to be estimated is at the county level, and the historical multi-source environmental data spatially covers a certain county-level area and temporally covers the entire growth cycle of rice from transplanting to maturity.

[0041] Step 2: Input multi-source environmental data and rice potential biomass simulation data into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at each growth stage.

[0042] Specifically, the rice aboveground biomass estimation model includes a potential biomass module and an actual environmental constraint module. The potential biomass module and the actual environmental constraint module learn potential biomass characteristics and environmental constraint characteristics, respectively, and carry out feature interaction aligned with growth stages to output the estimated aboveground biomass of rice at each growth stage.

[0043] Specifically, in the potential biomass module, the daily time-series information of potential biomass in the three growth stages is sequentially fed into the first long short-term memory network (LSTM) unit, and the hidden layer features of the LSTM unit at each growth stage time step are extracted as the potential biomass features of that stage.

[0044] Specifically, in multi-source environmental data, remote sensing data and meteorological data are constructed as time-series inputs according to three growth stages.

[0045] Specifically, in the actual environmental constraint module, the daily time-series inputs of each growth stage are sequentially fed into the second Long Short-Term Memory (LSTM) network unit to extract the dynamic environmental constraint features of the corresponding growth stage; the soil data and geographic location data are encoded by a fully connected layer to obtain the static environmental constraint features; the dynamic environmental constraint features of each growth stage are concatenated with the static environmental constraint features to obtain the environmental constraint features of the corresponding growth stage.

[0046] Specifically, by fusing the potential biomass characteristics and environmental constraint characteristics of each growth stage, rice growth characteristic time series information composed of the fused characteristics of the three growth stages is obtained. This information is then sequentially input into the third LSTM unit and the fully connected layer to obtain the estimated value of rice aboveground biomass for the corresponding stage.

[0047] Specifically, the training process of the rice aboveground biomass estimation model includes the following steps: Step D1: Obtain multiple sets of historical multi-source environmental data and the actual aboveground biomass of mature rice corresponding to each set of historical multi-source environmental data. Use the rice growth model to generate simulated potential and actual biomass data of rice corresponding to each set of historical multi-source environmental data. The simulated potential biomass data refers to the daily simulated time series of biomass at each growth stage by the RiceGrow mechanism without considering actual water and nutrient stress, and is used to characterize the simulated potential biomass information. The simulated actual biomass data refers to the daily simulated time series of biomass at each growth stage by the RiceGrow mechanism model considering actual water and nutrient stress, and is used to characterize the simulated actual biomass information.

[0048] The process of acquiring multiple sets of historical, multi-source environmental data includes: First, non-rice pixels are screened out using rice planting distribution products. Then, based on the MCD43A4 data product from the MODIS satellite, the enhanced vegetation index (EVI) and normalized yellowness index (NDYI) are calculated to depict the changes in the greenness and yellowness of the rice canopy with phenological stages. Finally, the average values ​​of remote sensing indicators at the pixel scale are aggregated to the county-level scale. Secondly, based on meteorological elements such as temperature, precipitation, and sunshine duration observed at meteorological stations, a continuous spatial pattern is generated and aggregated to the county level through spatial interpolation methods to characterize the impact of meteorological stress on the biomass accumulation process. Furthermore, in soil parameter processing, rice planting distribution products were used to eliminate interference from non-rice pixels. Key soil properties such as organic matter content, pH, bulk density, and clay content were extracted from the soil data product SoilGrid v2.0. The mean values ​​were calculated and aggregated to the county scale to reflect the long-term limiting effect of soil on rice growth. Finally, latitude and longitude information for each county-level region was extracted based on the county-level administrative divisions of the National Bureau of Statistics of China to describe regional differences and potential variations in management measures. Through the above process, a multi-source environmental constraint dataset reflecting the actual growth limitations of rice was constructed, laying the foundation for subsequent model training. The acquisition details of various data types are shown in Table 1 below: Table 1. Data Acquisition Status Preferably, in step D1, the aboveground biomass data of rice at the actual maturity stage is obtained by converting the county-level statistical yield and harvest coefficient.

[0049] Step D2: Construct a knowledge-guided rice aboveground biomass estimation model (PL-BioNet). Take meteorological data, soil data, and geographic location data from each set of historical multi-source environmental data, as well as the potential biomass time series information from the rice potential biomass simulation data corresponding to the historical multi-source environmental data, as inputs. Use the actual rice biomass simulation data corresponding to the historical multi-source environmental data as labels to pre-train the rice aboveground biomass estimation model, thereby learning the biomass accumulation mechanism.

[0050] Specifically, this step constructs and pre-trains a knowledge-guided rice aboveground biomass estimation model (PL-BioNet). A potential biomass module and an actual environmental constraint module are designed to learn potential biomass characteristics and environmental constraint characteristics, respectively. The potential biomass module uses the potential biomass time series of the three growth stages simulated by RiceGrow as input to characterize the growth potential under ideal conditions. The environmental constraint module only inputs multi-source features such as meteorological time series, soil properties, and latitude and longitude consistent with the RiceGrow simulation, excluding remote sensing time series features to ensure that the input structure completely corresponds to the mechanism simulation data. Based on the rice phenological process, all time series data are divided into three stages: transplanting-jointing, jointing-flowering, and flowering-maturity. Stage-specific potential biomass characteristics and environmental constraint characteristics are extracted using a Long Short-Term Memory (LSTM) neural network, and feature interactions aligned with the growth stages are performed. Finally, the aboveground biomass of rice at the jointing, flowering, and maturity stages is output, achieving a structured expression of the biomass accumulation mechanism of "potential growth-environmental constraint." The PL-BioNet model is pre-trained based on RiceGrow simulated samples. The aboveground biomass of the three stages of jointing, flowering and maturity is kept consistent with the simulation labels of the RiceGrow mechanism model. Mechanism-guided parameter initialization of the model is performed, which provides a stable and reasonable model parameter basis for the supervised fine-tuning in step 4, which introduces remote sensing time series and adjusts the output layer to adapt to the actual estimation needs (estimating aboveground biomass at maturity in this invention).

[0051] Step D3: Fine-tuning the knowledge-guided rice aboveground biomass estimation model framework using real labels: The potential biomass time series information from historical multi-source environmental data and corresponding simulated rice potential biomass data is used as input, and the actual aboveground biomass of mature rice corresponding to the multi-source environmental data is used as labels. The pre-trained rice aboveground biomass estimation model obtained in Step D2 is fine-tuned to correct the systematic bias of the mechanism model under complex environments and obtain a well-trained rice aboveground biomass estimation model with high accuracy and high generalization.

[0052] Specifically, this step utilizes county-level real biomass labels to fine-tune the pre-trained model, correcting mechanistic simulation biases and enhancing the model's adaptability to actual habitat conditions. First, based on county-level statistical yield data and combined with provincial rice harvest coefficients, yield is converted into aboveground biomass at rice maturity, serving as a model supervision label. Subsequently, based on the pre-trained model, remote sensing temporal features reflecting rice growth changes are introduced into the input to enhance the model's ability to extract comprehensive environmental constraints. Since this stage only has mature-stage biomass labels, the original model's three-stage biomass output layer is replaced with a single mature-stage biomass output structure to meet actual needs. Finally, the model undergoes full-parameter fine-tuning training, further optimizing it based on pre-trained parameters to accurately learn stress effects and regional differences in the real environment, thereby achieving high-precision estimation of aboveground biomass at rice maturity.

[0053] The present invention also provides a rice aboveground biomass estimation system that is applied to the rice aboveground biomass estimation method driven by the fusion of the above-mentioned mechanism knowledge and multi-source data.

[0054] The rice aboveground biomass estimation system of this invention includes: The data acquisition unit is used to acquire multi-source environmental data of the area to be estimated. The mechanism knowledge unit is used to receive multi-source environmental data of the area to be estimated and input it into the rice growth model to obtain the potential biomass time series information of each growth stage. The rice aboveground biomass estimation unit is used to receive multi-source environmental data and potential biomass time-series information at each growth stage of the area to be estimated, and input them into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at each growth stage.

[0055] The present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it is able to implement the steps of the rice aboveground biomass estimation method driven by the fusion of the aforementioned mechanistic knowledge and multi-source data.

[0056] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the steps of the rice aboveground biomass estimation method driven by the fusion of the above-mentioned mechanism knowledge and multi-source data.

[0057] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, enables the implementation of the steps of the above-mentioned method for estimating aboveground biomass of rice driven by the fusion of the aforementioned mechanistic knowledge and multi-source data.

[0058] Specific embodiments of the present invention are as follows: Example 1 This embodiment covers the main single-season rice-producing areas in the middle and lower reaches of the Yangtze River in China, including four provinces. The implementation process of this embodiment is as follows: Figure 1 and Figure 2 As shown, the specific steps include: Step S1: Construct a rice potential biomass simulation dataset based on the RiceGrow rice growth mechanism model.

[0059] Step S1 is as follows: This step uses the RiceGrow crop mechanistic model to generate potential rice biomass data. RiceGrow is a mechanistic model specifically designed for rice growth simulation, capable of calculating the rice biomass accumulation process under ideal conditions based on meteorological, varietal, management, and soil information. In this embodiment, the input data for RiceGrow is constructed as follows: First, meteorological data, including daily maximum and minimum temperatures, precipitation, and sunshine duration observed at meteorological stations of the China Meteorological Administration from 2001 to 2020, were used to generate county-level daily meteorological sequences via spatial interpolation to drive the RiceGrow model. Second, to obtain variety parameters applicable to different counties, grid search was used to correct the model's sensitive parameters. Forty sets of sensitive parameter combinations were randomly generated within the variety parameter range (Table 2), while the remaining parameters remained at their default settings. Simulations were performed for each combination using the period from 2001 to 2014 as parameter correction years, and the optimal parameter set was selected for each county based on the root mean square error (RMSE) of aboveground biomass at maturity. For management data, management parameters such as rice planting date were aggregated to the provincial scale and input into the model based on the average values ​​of agricultural meteorological station observations, while other water and fertilizer management measures used the default values ​​of RiceGrow. Soil input used SoilGrids v2.0 data product with a spatial resolution of 250m. Rice planting distribution product was used to exclude interference from non-rice pixels. Soil data of rice pixels were aggregated to the county level by averaging. pH, organic matter content, total nitrogen content, bulk density and clay content of the 0-15 cm soil layer were selected as soil properties.

[0060] After completing the above input preparation and parameter correction, the RiceGrow model was run to simulate the potential biomass formation process of rice without setting water and nutrient limitations. Daily potential biomass time-series data of the three key growth stages—transplanting-jointing, jointing-flowering, and flowering-maturation—were extracted to construct a potential biomass dataset representing the theoretical maximum growth capacity of rice. In addition, the RiceGrow model was set to consider water and nutrient limitations to simulate the actual biomass formation process of rice. Daily actual biomass time-series data of the three key growth stages—transplanting-jointing, jointing-flowering, and flowering-maturation—were also extracted to construct a simulated dataset of actual rice biomass for subsequent model pre-training.

[0061] Table 2. Parameters of sensitive varieties in the RiceGrow mechanism model. Step S2: Construct an environmental constraint dataset that affects actual biomass accumulation based on multi-source data such as remote sensing, meteorology, soil, and latitude and longitude.

[0062] Step 2 is as follows: In this embodiment, the environmental constraint dataset includes four types of data: remote sensing (EVI, NDYI), meteorological (KDD, PRCP), soil (pH, organic matter content, total nitrogen content, bulk density, clay content), and geographical location (latitude and longitude). These data are used to characterize the canopy growth limitation, meteorological stress, soil constraints, and regional differences experienced by rice in actual habitats.

[0063] The specific construction method is as follows: First, to obtain canopy growth characteristics, after filtering out non-rice pixels using rice planting distribution products, the Normalized Difference Vegetation Index (EVI) and Normalized Yellowness Index (NDYI) were calculated from the MCD43A4 reflectance product of the MODIS satellite. These were used to characterize the changes in greenness and yellowness of the rice canopy during the growth process, and the pixel-scale remote sensing indicators were averaged to the county-level scale. The calculation process for EVI and NDYI is as follows: Among them, red, blue, green, and nir represent red light, blue light, green light, and near-infrared bands, respectively.

[0064] Secondly, to obtain meteorological characteristics, the high-temperature stress factor KDD for rice growth was calculated based on the daily maximum temperature observed at meteorological stations: Wherein, is the daily maximum temperature on day d, and is the high temperature threshold, which is taken as 35℃ in this invention.

[0065] By combining daily precipitation meteorological elements, a continuous grid is generated through spatial interpolation, and then aggregated at the county level to reflect the phased impact of meteorological conditions on biomass accumulation.

[0066] Furthermore, in the construction of soil properties, non-rice pixels were excluded based on the rice planting distribution product. Key soil variables such as organic matter content, pH, bulk density, and clay content were extracted from the SoilGrids v2.0 data product and aggregated to the county level to describe the long-term limiting effect of stable soil background on rice growth.

[0067] Finally, the corresponding central latitude and longitude information is obtained based on the county-level administrative boundaries to characterize regional geographical differences and spatial variations in potential management measures.

[0068] Through the above steps, an environmental constraint dataset consisting of remote sensing, meteorological, soil, and geographic information was constructed. Together with the potential biomass dataset constructed in step 1, it provides a complete and structured input foundation for subsequent model training.

[0069] Step S3: Construct and pre-train a knowledge-guided rice aboveground biomass estimation model (PL-BioNet). Specifically, design a potential biomass module and an actual environmental constraint module to learn potential biomass characteristics and environmental constraint characteristics, respectively, conduct feature interactions aligned with growth stages, and output the aboveground biomass of rice at the jointing, flowering, and maturity stages. Use the potential biomass simulation dataset and environmental constraint dataset constructed in Steps 1 and 2 to pre-train the rice aboveground biomass estimation model, enabling it to learn the biomass accumulation mechanism.

[0070] Step 3 specifically involves: This step involves designing a knowledge-guided rice aboveground biomass estimation network structure (PL-BioNet), comprising two parallel branches: a potential biomass module and an environmental constraint module. Feature-level fusion is implemented at three key growth stages (jointing, flowering, and maturity) to estimate the actual biomass at each stage and ultimately. The model structure is designed as follows: The potential biomass module takes the daily time series of potential rice biomass simulated by the RiceGrow mechanism model as input. This time series is divided into three growth stages: transplanting-jointing, jointing-flowering, and flowering-maturity. The input sequence of each stage is fed into the first long short-term memory network (LSTM network) unit to extract the potential biomass features of the corresponding stage, which are used to characterize the upper limit of biomass accumulation under ideal conditions, which is jointly determined by light and temperature resources and varietal attributes.

[0071] The environmental constraint module takes as input multiple sources, including remote sensing vegetation indices (EVI, NDYI), meteorological stress factors (high temperature stress KDD, precipitation PRCP), soil properties (pH, organic matter, total nitrogen, bulk density, and clay content in the 0-15cm soil layer), and county-level latitude and longitude. Remote sensing and meteorological variables are constructed as time-series inputs in three stages, sequentially fed into the second LSTM network unit to extract stage-specific environmental constraint features. Static variables such as soil and geographical location are encoded by a fully connected layer and then concatenated with the stage-specific environmental features to form a comprehensive environmental constraint feature representation for rice growth. The model employs a growth stage-aligned feature fusion mechanism, fusing the feature vectors output by the potential growth module and the environmental constraint module at each growth stage and inputting them into the corresponding LSTM output layer. Finally, the model outputs estimated aboveground biomass values ​​for rice at the jointing, flowering, and maturity stages.

[0072] A rice aboveground biomass estimation network structure (PL-BioNet) based on pre-trained knowledge from RiceGrow mechanistic simulation data was developed to fully leverage the process knowledge advantages of the mechanistic model. During the pre-training phase, a large-scale simulated sample set generated by the RiceGrow mechanistic model from 2001 to 2020 under meteorological driving data and county-by-county calibrated variety parameters was used as pre-training data. The pre-training input included potential biomass time series and limiting factors such as meteorological factors, soil properties, and geographical location consistent with RiceGrow simulations, but excluded remote sensing variables to ensure consistency between the network input structure and the mechanistic simulation mechanism. The pre-trained model output consisted of aboveground rice biomass at the jointing, flowering, and maturity stages, with a loss function applied to the simulated values ​​of actual aboveground rice biomass at these stages obtained from the mechanistic simulation. Pre-training ensured a well-initialized model parameter set guided by mechanistic knowledge.

[0073] During the pre-training phase, the mean squared error loss function was chosen to minimize the error between the estimated biomass and the simulated actual biomass. The optimizer used was Adam, with an initial learning rate of 0.01 and a learning rate configuration of ReduceLROnPlateau. The batch size during training was set to 800, and the maximum number of training epochs was set to 200. An early stopping mechanism was also introduced, whereby the error would not decrease for 10 consecutive epochs or the learning rate decay would be less than 1 × 10⁻⁶. -5 If the condition is met, training should be stopped to prevent overfitting.

[0074] Step 4: Fine-tuning the knowledge-guided rice aboveground biomass estimation model framework based on real labels. Building upon the pre-trained model from Step 3, fine-tuning the model using mature rice aboveground biomass samples obtained from county-level statistical yields and harvest coefficients corrects systematic biases in the mechanistic model under complex environments, thereby obtaining a high-precision, highly generalizable rice aboveground biomass estimation model.

[0075] Step 4 specifically involves: In the fine-tuning phase, county-level aboveground biomass samples at maturity (converted from statistical yields and rice harvest coefficients of each province) from 2001 to 2018 were used as supervision labels to fine-tune all parameters of the pre-trained model. During this phase, remote sensing temporal feature inputs were added to the model's environmental constraint module to enhance its ability to represent actual growth and environmental stress. The output layer was also changed from three stages to a single maturity stage output to match the actual biomass labels. Through fine-tuning, the systematic bias of the RiceGrow mechanism model in specific regions can be effectively corrected, improving the model's adaptability in real-world complex environments and achieving high-precision estimation of rice aboveground biomass.

[0076] During the fine-tuning training phase, the mean squared error loss function was selected to minimize the error between the estimated biomass and the simulated actual biomass. The optimizer used was Adam, with an initial learning rate of 0.001 and a learning rate configuration of ReduceLROnPlateau. The batch size for training was set to 400, and the maximum number of training epochs was set to 200. An early stopping mechanism was introduced, whereby the error did not decrease for 10 consecutive epochs or the learning rate decayed to less than 1 × 10⁻⁶. -6 If the condition is met, training should be stopped to prevent overfitting.

[0077] The effectiveness of the rice aboveground biomass estimation method driven by mechanistic knowledge and multi-source data fusion proposed in this invention will be evaluated below.

[0078] First, the proposed model (PL-BioNet) is compared with traditional deep learning models (Long Short-Term Memory Neural Network LSTM), machine learning models (Random Forest RF), and the RiceGrow mechanism model. Data from 2001-2014 was selected as the training set (for pre-training, fine-tuning, and variety parameter correction), and data from 2015-2018 was used as an independent test set to ensure that test data did not appear during the training phase, thus objectively evaluating the model's generalization ability. Under the same data conditions and evaluation system, experimental results show that the proposed model achieves optimal performance in estimating aboveground biomass at maturity. Test set results show that the root mean square error (RMSE) of PL-BioNet is 1.61 t / ha, significantly lower than LSTM (1.91 t / ha), RF (1.88 t / ha), and RiceGrow (2.55 t / ha). Compared with the baseline models, PL-BioNet's RMSE is reduced by 14%-37%. Furthermore, the Pearson correlation coefficient (r = 0.76) between the biomass estimate and the label value of the PL-BioNet model was higher than all the comparison methods, indicating that the model is more accurate in capturing the trend of biomass change.

[0079] The results demonstrate that the method of this invention, through the structured fusion of potential growth information and multi-source environmental limiting factors, combined with a mechanism simulation-guided pre-training and real data fine-tuning process, can significantly improve the accuracy of rice biomass estimation and maintain stable and reliable predictive performance in cross-year tests. The accuracy of aboveground biomass estimation at rice maturity using this invention and the baseline method is shown in the following figure. Figure 3 As shown.

[0080] The PL-BioNet model proposed in this invention, after pre-training with RiceGrow mechanism simulation samples, can learn biomass accumulation patterns consistent with the mechanism model. Specifically, the stage-specific biomass distribution output by the model at the three key growth stages—jointing, flowering, and maturity—is highly consistent with the biomass distribution trend simulated by RiceGrow. This result demonstrates that the PL-BioNet model can effectively learn the aboveground biomass accumulation characteristics of rice from vegetative to reproductive growth, laying a solid foundation for parameter initialization and structural priors for subsequent fine-tuning training with real data, and achieving effective embedding of the process mechanism. The biomass accumulation process learned by the PL-BioNet model through pre-training is illustrated in the figure below. Figure 4 As shown in (a), the method of this invention demonstrates significant bias correction capability in the test samples from various provinces. The pre-trained model exhibits a certain degree of systematic underestimation when using RiceGrow to simulate labels, particularly in regions with complex climate and management conditions such as Anhui and Hubei. After fine-tuning with real samples, the estimation errors in all provinces significantly decreased, and the error distribution approached zero, indicating that the model can effectively adapt to the real environmental differences in various regions while maintaining mechanistic consistency. Experimental results are shown in the figure below. Figure 4 As shown in (b).

[0081] To verify the effectiveness of the proposed potential biomass-environmental constraint dual-module structure (PL-BioNet), an ablation experiment was conducted on the model structure. The specific experimental setup included: retaining only the potential growth module, retaining only the environmental constraint module, a conventional LSTM structure, and the complete PL-BioNet model of this invention. Based on the test results of high biomass samples and extreme stress samples, the following performance was obtained: (1) In high biomass samples, such as Figure 5 As shown in (a), the model containing only the potential growth module performs better than other single-module structures, indicating that the potential biomass feature can effectively characterize the upper limit of growth under ideal conditions and provide more effective information for biomass estimation in high-yield environments. (2) In extreme stress samples, such as Figure 5 As shown in (b), the model containing only the environmental constraint module is better than the model containing only the potential growth module. This indicates that features such as meteorological stress, vegetation index, and soil properties, which reflect actual environmental constraints, have a stronger explanatory power for biomass formation under stress conditions and can more accurately reflect the rice growth process under actual environmental constraints. (3) In both types of samples, the PL-BioNet structure proposed in this invention achieved the lowest estimation error, significantly outperforming the traditional LSTM and the two single-module structures. This result indicates that PL-BioNet can simultaneously utilize potential growth capacity and environmental constraint features, and achieve complementarity between the two types of information through a staged fusion mechanism, thereby improving the robustness and adaptability of the model under high biomass and stress conditions. This further verifies that the "growth potential-environmental constraint" dual-module structure proposed in this invention is a key technical means to improve the accuracy of biomass estimation.

[0082] Subsequently, to verify the applicability of the PL-BioNet model based on the mechanism simulation pre-training-fine-tuning strategy proposed in this invention under limited sample conditions, the estimation performance of three training methods under different training sample sizes was compared: the pre-trained-fine-tuned PL-BioNet model proposed in this invention, the directly trained PL-BioNet model, and the traditional random forest model. The experiment used different training year windows (2011-2014, 2007-2014, 2001-2014) as the incrementing setting for the training sample size, and evaluated the root mean square error of the model on the test set. The results are as follows: Figure 6 As shown, with a small training sample size (2011-2014), the directly trained model exhibited a significant decrease in accuracy, indicating that the purely data-driven method is highly dependent on the number of training samples. However, the pre-trained-fine-tuned PL-BioNet model of this invention maintained high prediction accuracy, demonstrating good small-sample learning ability. With an increase in the training sample size (2007-2014, 2001-2014), the accuracy of the directly trained model improved, but the root mean square error of the method in this invention remained the lowest, outperforming the directly trained model. This result demonstrates that pre-training the PL-BioNet model based on simulated sample data provided by the RiceGrow mechanism model can achieve model parameter initialization consistent with the mechanism process, helping to reduce the dependence of deep neural networks on large-scale labeled samples.

[0083] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and not to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the disclosure of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for estimating aboveground biomass of rice driven by the fusion of mechanistic knowledge and multi-source data, characterized in that, Includes the following steps: Step 1: Using a rice growth model, generate time-series simulation data of potential rice biomass at each growth stage based on multi-source environmental data of the area to be estimated. Step 2: Input the multi-source environmental data and rice potential biomass time series simulation data into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at at least one growth stage. The multi-source environmental data includes remote sensing data, meteorological data, soil data, and geographic location data; The growth stages include the jointing stage, the flowering stage, and the maturity stage; The rice aboveground biomass estimation model includes a potential biomass module and an actual environmental constraint module. The potential biomass module and the actual environmental constraint module learn potential biomass characteristics and environmental constraint characteristics, respectively, and carry out feature interaction aligned with growth stages to generate and output the estimated aboveground biomass of rice at each growth stage.

2. The method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion according to claim 1, characterized in that, In the potential biomass module, the daily time-series simulation data of rice potential biomass at each growth stage is sent to the first long short-term memory network unit, and the hidden layer features of the first long short-term memory network unit at each growth stage time step are extracted as the potential biomass features of the corresponding stage.

3. The method for estimating aboveground biomass of rice driven by the fusion of mechanistic knowledge and multi-source data according to claim 1, characterized in that, The remote sensing data and meteorological data in the multi-source environmental data are constructed as time-series inputs according to three growth stages; the actual environmental constraint module receives the daily time-series inputs of each growth stage and inputs them into the second long short-term memory network unit, extracts the hidden layer features of the second long short-term memory network unit at each growth stage time step, and uses them as the dynamic environmental constraint features of the corresponding stage. The actual environmental constraint module receives the soil data and geographic location data and processes them through a fully connected layer to obtain static environmental constraint features; the dynamic environmental constraint features of each growth stage are spliced ​​with the static environmental constraint features to obtain the environmental constraint features.

4. The method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion according to claim 1, characterized in that, In the trained rice aboveground biomass estimation model, the potential biomass characteristics and environmental constraint characteristics of each growth stage are fused to obtain rice growth characteristic time series information composed of the fused characteristics of the three growth stages. This information is then sequentially input into the third long short-term memory network unit and the fully connected layer for processing, and outputs the estimated value of rice aboveground biomass for the corresponding growth stage.

5. The method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion according to claim 1, characterized in that, The training process of the rice aboveground biomass estimation model includes the following steps: Step D1: Obtain multiple sets of historical multi-source environmental data and the actual aboveground biomass of mature rice corresponding to each set of historical multi-source environmental data. Use the rice growth model to generate daily simulated time-series data of potential and actual rice biomass corresponding to each set of historical multi-source environmental data. The potential biomass simulation data refers to the daily simulated time-series data of biomass at each growth stage by the rice growth model without considering actual environmental stress. The actual biomass simulation data refers to the daily simulated time-series data of biomass at each growth stage by the rice growth model considering actual environmental stress. Step D2: Construct a pre-trained model for estimating rice aboveground biomass. Take meteorological data, soil data, and geographical location data from each set of historical multi-source environmental data, as well as the corresponding daily simulated time-series data of potential rice biomass, as inputs and the daily simulated time-series data of actual rice biomass as labels. Pre-train the rice aboveground biomass estimation model to obtain the pre-trained rice aboveground biomass estimation model. Step D3: Using the historical multi-source environmental data and the corresponding daily simulated time-series data of potential rice biomass as input, and the actual aboveground biomass of rice at the target growth stage as the label, train the pre-trained rice aboveground biomass estimation model obtained in Step D2 to obtain the trained rice aboveground biomass estimation model.

6. The method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion according to claim 5, characterized in that, In step D1, the aboveground biomass data of rice at the actual maturity stage is obtained by converting county-level statistical yield and harvest coefficient.

7. The method for estimating aboveground biomass of rice driven by mechanistic knowledge and multi-source data fusion according to claim 5, characterized in that, In step D3, the target growth stage includes at least the maturity stage.

8. The method for estimating aboveground biomass of rice driven by the fusion of mechanistic knowledge and multi-source data according to claim 1, characterized in that, The remote sensing data includes enhanced vegetation index and normalized yellowness index; the meteorological data includes high temperature stress factors and daily precipitation; the soil data includes pH, organic matter content, total nitrogen content, bulk density, and clay content; and the geographic location data includes latitude and longitude.

9. A rice aboveground biomass estimation system applied to a rice aboveground biomass estimation method driven by the fusion of mechanistic knowledge and multi-source data as described in any one of claims 1 to 8, characterized in that, include: The data acquisition unit is used to acquire multi-source environmental data of the area to be estimated. The mechanism knowledge unit is used to receive multi-source environmental data of the area to be estimated and input it into the rice growth model to obtain the potential biomass time series information of each growth stage. The rice aboveground biomass estimation unit is used to receive multi-source environmental data and potential biomass time-series information at each growth stage of the area to be estimated, and input them into the trained rice aboveground biomass estimation model to obtain the estimated aboveground biomass of rice at each growth stage.

Images