A method for inverting total primary productivity based on multi-level deep learning

Abstract

This invention discloses a method for inverting gross primary productivity (GPP) based on multi-level deep learning. The method comprises obtaining spectral, meteorological, and flux information from remote sensing images; generating a lookup table using the PROSAIL model to establish correspondences between spectral information and leaf structure, leaf biochemical, and soil information; constructing a GPP inversion model using multi-level deep learning; inverting GPP using the model and assessing the inversion accuracy. The method utilizes the model to convert spectral information into information such as leaf structure, expanding the spectral information dimension; designing four types of limiting factors to enhance model interpretability; constructing a multi-level deep learning network, utilizing it to automatically optimize the expression of limiting factors, and employing a parallel design to clarify the effects of these factors, thereby improving model accuracy.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing data and deep learning technology, and specifically to a gross primary productivity inversion method based on multi-level deep learning. Background Art

[0002] Gross primary productivity (GPP) is a key link in the carbon cycle of terrestrial ecosystems. It represents the amount of organic carbon fixed by vegetation through photosynthesis, directly determines the initial material and energy input of the ecosystem, and is of great significance for studying the global carbon balance and the evolution of ecosystem functions.

[0003] At present, the calculation methods of gross primary productivity mainly include the following: (1) box measurement method, which indirectly calculates GPP by measuring the net ecosystem exchange (NEE) and ecosystem respiration (ER), but due to the limitations of its measurement method, continuous measurement cannot be achieved; (2) carbon dioxide flux observation method based on eddy covariance method, which uses flux towers to obtain NEE and further decomposes it into ER and GPP. Although continuous monitoring is achieved and it provides a key data source for model verification, the construction cost of flux stations is high and the number is scarce, which makes it difficult to meet the needs of large-scale measurements; (3) using remote sensing models or process models to predict GPP, among which representative remote sensing models such as Carnegie-Ames- The Stanford Approach Model (CASA), the Global Production Efficiency Model (GLO-PEM), the Vegetation Photosynthesis Model (VPM), and process models such as the General Land Surface Process Model (CLM), the Ecosystem Carbon-Water Dynamics Model (ORCHIDEE), and the Plant-Soil-Atmosphere System Carbon Exchange Model (CEVSA), although optimized with the help of flux data, have different calculation formulas, making it difficult to unify the calculation results, affecting the accuracy and versatility of the measurements.

[0004] Furthermore, spectral information contains a wealth of information about vegetation growth conditions. However, in existing remote sensing or process models, spectral information is only calculated in simple forms such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI). A large amount of hidden spectral information is not effectively exploited, limiting the models' ability to accurately reflect the physiological state of vegetation. Furthermore, the GPP of different ecosystems worldwide varies significantly due to the combined influence of multiple factors, including climate, vegetation, soil, and disturbance conditions. Existing remote sensing or process models struggle to fully account for the limiting effects of these complex factors on photosynthesis, resulting in insufficient prediction accuracy across diverse ecosystems. Therefore, there is an urgent need to develop a new measurement method that fully utilizes spectral information and does not rely on traditional fixed formulas to improve the accuracy and universality of gross primary productivity retrieval. Summary of the Invention

[0005] Based on this, the purpose of the present invention is to provide a total primary productivity inversion method based on multi-level deep learning, which can effectively solve the problems existing in the above-mentioned prior art.

[0006] The present invention adopts the following technical solution: a method for inverting total primary productivity based on multi-level deep learning, comprising the following steps:

[0007] (1) Obtaining remote sensing image spectral information, meteorological information and flux information;

[0008] (2) Generate a lookup table using the PROSAIL model to establish the correspondence between spectral information and leaf structure information, leaf biochemical information, and soil information;

[0009] (3) Construct a multi-level deep learning model for total primary productivity inversion;

[0010] (4) Use the model to invert the total primary productivity and evaluate the inversion accuracy.

[0011] Preferably, the spectral information of the remote sensing image in step (1) needs to be preprocessed, and the preprocessing includes radiation correction, atmospheric correction and geometric correction. The meteorological information includes total solar radiation SOL, precipitation W, temperature T, evaporation ET, and saturated water vapor pressure difference VPD. The flux information includes gross primary productivity GPP.

[0012] Preferably, step (2) comprises the following steps:

[0013] (2.1) The PROSAIL model formula is:

[0014] ;

[0015] in, is the canopy reflectance, represents the PROSAIL model operator, is the blade structural parameter, is the leaf area index, is the average leaf inclination angle, is the chlorophyll content, is the carotenoid content, is the equivalent water thickness, is the dry matter content, is the soil wetness coefficient, is the soil hardness coefficient, is the hotspot parameter, To observe the zenith angle, is the solar zenith angle, is the relative azimuth between the sun and the observation;

[0016] (2.2) Use the normal random distribution function to , , , , , , , , , ; , , Determined by the sensor of the remote sensing image, the parameters are brought into the PROSAIL model to obtain the spectral reflectance curve;

[0017] (2.3) Select the band response function of the corresponding remote sensing image and resample the spectral reflectance curve to the band of the corresponding remote sensing image. This completes the correspondence between spectral information and leaf structure information, leaf biochemical information, and soil information. The band response function is:

[0018]

[0019] in, For satellite The equivalent remote sensing reflectivity of the band, and is the band range of the band, for The value of the spectral reflectance curve at wavelength, for Spectral responsivity at wavelength;

[0020] (2.4) The spectral information of the remote sensing image is entered into the lookup table to obtain the corresponding leaf structure information, leaf biochemical information, and soil information.

[0021] Preferably, in step (3), the calculation model of total primary productivity data based on multi-level deep learning is:

[0022]

[0023]

[0024] Among them, GPP stands for gross primary productivity; SOL stands for total solar radiation; is the blade structure limiting factor, It is a biochemical limiting factor in leaves. is the soil limiting factor, It is the environmental limiting factor, and full consideration should be given to the limiting effect of different limiting factors on the total initial productivity.

[0025] Furthermore, the network construction method of multi-level deep learning is:

[0026] Four network models based on the Transformer architecture were built. Each network model represents a corresponding layer, including the encoder-decoder and attention mechanism. The encoder is composed of six stacked encoders, and the decoder is also composed of six stacked encoders. The different layers are connected in parallel, that is, the product of the output results of the four layers is the final output of the network.

[0027] The input data of the first level is , the output is the blade structure constraint factor; the second level input data is , the output is the leaf biochemical limiting factor; the input data of the third level is , the output is the soil limiting factor; the input data of the fourth level is , the output is the environmental limiting factor.

[0028] Furthermore, the training method of multi-level deep learning is:

[0029] The total primary productivity of the flux site is used as a label, and the leaf structure information, leaf biochemical information, soil information and meteorological information are used as features. They are sent to the network model for training to obtain the predicted total primary productivity. The loss function is constructed with the label. The mean square error loss function is selected as the loss function, and the Adam optimizer is selected as the optimizer.

[0030] Preferably, in step (4), the method for inversion accuracy is:

[0031] The model prediction value is compared with the true value to evaluate the inversion accuracy; the root mean square error RMSE and the determination coefficient R are used to 2 As an accuracy evaluation indicator, the formula is as follows:

[0032] Root mean square error RMSE:

[0033] ;

[0034] Coefficient of determination R 2 :

[0035] ;

[0036] Where n is the total number of samples, is the model prediction value, is the true value, is the average value of the true value; the range of the root mean square error is , when the predicted value is completely consistent with the true value, it is equal to 0, that is, perfect prediction; the worse the prediction effect, the larger the root mean square error value; the range of the coefficient of determination is , when the predicted value is completely consistent with the true value, it is equal to 1, that is, a perfect prediction; the worse the prediction effect, the smaller the determination coefficient.

[0037] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0038] 1. In this invention, the PROSAIL model is integrated into the calculation of gross primary productivity. By leveraging its physical mechanism, spectral information is converted into leaf structure information, leaf biochemical information, and soil information relevant to solving gross primary productivity, thus expanding the information dimension of spectral information.

[0039] 2. The gross primary productivity calculation formula designed by this invention expands the traditional limiting factors. Traditional limiting factors only consider the influence of some environmental factors and fail to fully explore the impact of leaf structure and leaf biochemical information on photosynthesis. This invention designs four types of limiting factors to enhance the interpretability of the model;

[0040] 3. The present invention designs a multi-level deep learning network model, which uses deep learning to automatically optimize the functional expression of the limiting factors. The parallel design of the network can clarify the role of different limiting factors, better utilize the loss function for optimization, and improve the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0042] In the attached figure:

[0043] Figure 1 It is a schematic diagram of the process of the present invention;

[0044] Figure 2 Schematic diagram of the algorithm of the present invention;

[0045] Figure 3 This is a network architecture diagram of the present invention;

[0046] Figure 4 This is the inversion result diagram obtained in this embodiment. DETAILED DESCRIPTION

[0047] To make the objects and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is merely used to describe a method for inverting total primary productivity based on multi-level deep learning or several specific implementations of the present invention, and does not strictly limit the scope of protection specifically claimed in the present invention.

[0048] Example: Figure 1As shown in FIG, a method for inverting total primary productivity based on multi-level deep learning includes the following steps:

[0049] S1. Obtain remote sensing image spectral information, meteorological information and flux information;

[0050] In this embodiment, the remote sensing image selected is the MODIS satellite. The MODIS surface reflectance data MODIS09 can be downloaded from the data distribution website of the NASA Goddard Space Flight Center at https: / / ladsweb.modaps.eosdis.nasa.gov / search. It contains spectral information of seven bands, as shown in Table 1. The MODIS09 spectral information has been preprocessed, including radiation correction, atmospheric correction, and geometric correction.

[0051] Table 1 MODIS09 spectral information

[0052] Band wavelength spectrum bands 1 620-670 nm Red bands 2 841-876 nm Near infrared (NIR) bands 3 459-479 nm Blue bands 4 545-565 nm Green bands 5 1230-1250 nm Shortwave Infrared 1 (Swir1) bands 6 1628-1652 nm Shortwave Infrared 2 (Swir2) bands 7 2105-2155 nm Shortwave Infrared 3 (Swir3)

[0053] Meteorological information includes total solar radiation (SOL), precipitation (W), temperature (T), evaporation (ET), and saturated vapor pressure difference (VPD). The total solar radiation (SOL), precipitation (W), temperature (T), and evaporation (ET) can be downloaded from the European Centre for Medium-Range Weather Forecasts at https: / / cds.climate.copernicus.eu. The saturated vapor pressure difference (VPD) can be expressed as the difference between the saturated vapor pressure (SVP) and the actual vapor pressure (AVP), and is calculated using the following formula:

[0054] ;

[0055] Among them, VPD stands for saturated vapor pressure difference, SVP stands for saturated vapor pressure, AVP stands for actual vapor pressure, T stands for air temperature, and RH stands for relative humidity;

[0056] Flux information, including gross primary productivity (GPP), can be downloaded from the Flux2015 Consortium at https: / / fluxnet.org / data / fluxnet2015-dataset / .

[0057] S2. Generate a lookup table using the PROSAIL model to establish the correspondence between spectral information and leaf structure information, leaf biochemical information, and soil information;

[0058] Step S2 includes the following steps:

[0059] The PROSAIL model is a canopy reflectance model. By assigning values ​​to the model parameters, the corresponding relationship between each parameter and spectral information can be obtained. The PROSAIL model formula is:

[0060] ;

[0061] in, is the canopy reflectance, represents the PROSAIL model operator, is the blade structural parameter, is the leaf area index, is the average leaf inclination angle, is the chlorophyll content, is the carotenoid content, is the equivalent water thickness, is the dry matter content, is the soil wetness coefficient, is the soil hardness coefficient, is the hotspot parameter, To observe the zenith angle, is the solar zenith angle, is the relative azimuth between the sun and the observation;

[0062] Use the normal random distribution function to , , , , , , , , , ; , , Determined by the sensor of the remote sensing image, the parameters are substituted into the PROSAIL model to obtain the spectral reflectance curve; the parameter range is shown in Table 2.

[0063] Table 2 Parameter range and parameter step size of input variable parameters

[0064] parameter unit Minimum Maximum average value Standard deviation #timg# / 0.5 2 1.5 0.15 #timg# / 3 7 5 0.5 #timg# ° 30 80 50 5 #timg# <![CDATA[μg / cm 2 ]]> 0 80 4 4 #timg# <![CDATA[μg / cm 2 ]]> 0 20 10 1 #timg# cm 0 100 50 5 #timg# <![CDATA[μg / cm 2 ]]> 0 50 25 2.5 #timg# / 0 1 0.5 0.05 #timg# / 0 1 0.5 0.05 #timg# / 0 1 0.5 0.05

[0065] The band response function of the corresponding remote sensing image MODIS09 is selected, and the spectral reflectance curve is resampled to the 7 bands of the corresponding remote sensing image MODIS09. The correspondence between the spectral information and the leaf structure information, leaf biochemical information, and soil information is completed. The band response function is:

[0066]

[0067] in, For satellite The equivalent remote sensing reflectivity of the band, and is the band range of the band, for The value of the spectral reflectance curve at wavelength, for Spectral responsivity at wavelength.

[0068] The spectral information of the remote sensing image MODIS09 is entered into the lookup table to obtain the corresponding leaf structure information, leaf biochemical information, and soil information.

[0069] S3. Construct a multi-level deep learning total primary productivity inversion model;

[0070] like Figure 2 As shown, the leaf structure information was obtained from MODIS09 , leaf biochemical information , soil information Then, combine it with weather information Together, different levels of the Transformer model are input respectively. The first level outputs leaf structure limiting factors, the second level outputs leaf biochemical limiting factors, the third level outputs soil limiting factors, and the fourth level outputs environmental limiting factors.

[0071] The network construction method of multi-level deep learning is:

[0072] Establish four network models based on Transformer architecture. Transformer architecture is as follows Figure 3 As shown in the figure, each network model represents a corresponding layer, including the encoder-decoder and attention mechanism. The encoder is composed of 6 stacked encoders, and the decoder is also composed of 6 stacked encoders. Different layers are connected in parallel, that is, the product of the output results of the four layers is the final output of the network.

[0073] The input data of the first level is , the output is the blade structure constraint factor; the second level input data is , the output is the leaf biochemical limiting factor; the input data of the third level is , the output is the soil limiting factor; the input data of the fourth level is , the output is the environmental limiting factor.

[0074] The calculation model of total primary productivity data based on multi-level deep learning is:

[0075]

[0076]

[0077] Among them, GPP stands for gross primary productivity; SOL stands for total solar radiation; is the blade structure limiting factor, It is a biochemical limiting factor in leaves. is the soil limiting factor, It is the environmental limiting factor, and full consideration should be given to the limiting effect of different limiting factors on the total initial productivity.

[0078] The training method of multi-level deep learning is:

[0079] The total primary productivity of the flux site is used as a label, and the leaf structure information, leaf biochemical information, soil information and meteorological information are used as features. They are sent to the network model for training to obtain the predicted total primary productivity. The loss function is constructed with the label. The mean square error loss function is selected as the loss function, and the Adam optimizer is selected as the optimizer.

[0080] The effectiveness of multi-level deep learning lies in:

[0081] If all data were fed into the same Transformer network model, it would be difficult to adjust the correlation between the final limiting factors and the corresponding inputs in a timely manner. For example, when weather information changes, the environmental limiting factors themselves will change, but due to the influence of other input features, the final limiting factors will not change significantly. By distinguishing different levels, the relationship between each limiting factor is independent of each other, which can better simulate the impact of different factors on limiting factors in real scenarios and thus accurately simulate gross primary productivity.

[0082] S4. Use the model to invert the total primary productivity and evaluate the inversion accuracy.

[0083] The inversion accuracy method is to compare the model prediction value with the true value to evaluate the inversion accuracy; the root mean square error RMSE and the determination coefficient R 2 As an accuracy evaluation indicator, the formula is as follows:

[0084] Root mean square error RMSE:

[0085] ;

[0086] Coefficient of determination R 2 :

[0087] ;

[0088] Where n is the total number of samples, is the model prediction value, is the true value, is the average value of the true value; the range of the root mean square error is , when the predicted value is completely consistent with the true value, it is equal to 0, that is, perfect prediction; the worse the prediction effect, the larger the root mean square error value; the range of the coefficient of determination is , when the predicted value is completely consistent with the true value, it is equal to 1, that is, perfect prediction; the worse the prediction effect, the smaller the coefficient of determination. Figure 4 The fitting effect shows that the predicted value is consistent with the true value trend, and the calculated RMSE is lower than 0.91 and R² is greater than 4.67, indicating that the model has high accuracy.

[0089] The above describes the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. After knowing the contents described in the present invention, ordinary technicians in this technical field can make several equivalent changes and substitutions without departing from the principles of the present invention. These equivalent changes and substitutions should also be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for inverting total primary productivity based on multi-level deep learning, characterized in that: The following steps are involved: Step (1) obtaining pre-processed remote sensing image spectral information, meteorological information and flux information; Step (2) using the PROSAIL model to generate a lookup table and establish the corresponding relationship between spectral information and leaf structure information, leaf biochemical information, and soil information; Step (3) constructing a total primary productivity inversion model based on a multi-level deep learning parallel Transformer architecture; Step (4) inverting the gross primary productivity using the model, and evaluating the inversion accuracy using the root mean square error and the coefficient of determination; Wherein, the PROSAIL model formula in step (2) is: Among them, ρ c is the canopy reflectance, P represents the PROSAIL model operator, N is the leaf structure parameter, LAI is the leaf area index, ALA is the average leaf inclination angle, C ab is the chlorophyll content, C car is the carotenoid content, C w is the equivalent water thickness, C m is the dry matter content, P soil is the soil dryness coefficient, R soil is the soil hardness coefficient, Hotspot is the hotspot parameter, θ v is the observation zenith angle, θ χ is the solar zenith angle, is the relative azimuth between the sun and the observation; The generation of the lookup table includes: using a normal random distribution function to determine N, LAI, ALA, C ab , C car , C w , C m , P soil , R soil , Hotspot; θ v ,θ χ , Determined by the sensor of the remote sensing image, the parameters are brought into the PROSAIL model to obtain the spectral reflectance curve. Using the band response function, the spectral reflectance curve is resampled to the remote sensing image band, and the correspondence between the spectral information and the leaf structure information, leaf biochemical information, and soil information is established. The band response function is: Among them, ρ(band i ) is the equivalent remote sensing reflectivity of the satellite's i-band, λ1 and λ2 are the band ranges of the band, ρ(λ) is the value of the spectral reflectivity curve at the wavelength of λ, and SRF(λ) is the spectral response rate at the wavelength of λ; The spectral information of remote sensing images is put into the lookup table to obtain the corresponding leaf structure information, leaf biochemical information, and soil information; The total primary productivity calculation model of step (3) is: GPP=SOL*f1ε(N,LAI,ALA)*f2ε(C ab ,C car ,C w ,C m )*f3ε(P soil ,R soil )*f4ε(W,T,ET,VPD) Where GPP represents gross primary productivity; SOL represents total solar radiation; f1ε(N,LAI,ALA) is the leaf structure limiting factor, f2ε(C ab ,C car ,C w ,C m ) is the leaf biochemical limiting factor, f3ε(P soil ,R soil ) is the soil limiting factor, and f4ε(W, T, ET, VPD) is the environmental limiting factor, fully considering the limiting effect of different limiting factors on the total initial productivity.

2. The method for inverting total primary productivity based on multi-level deep learning according to claim 1, characterized in that: In the step (1): The preprocessing of remote sensing image spectral information includes radiation correction, atmospheric correction and geometric correction; Meteorological information includes total solar radiation SOL, precipitation W, temperature T, evaporation ET, and saturated vapor pressure difference VPD; Flux information includes gross primary productivity (GPP).

3. The method for inverting total primary productivity based on multi-level deep learning according to claim 1, characterized in that: The multi-level deep learning model is based on the Transformer architecture and consists of four parallel network layers: The input data of the first level are N, LAI, and ALA, and the output is the blade structure limitation factor; The input data of the second level is C ab ,C car ,C w ,C m , the output is the leaf biochemical limiting factor; The input data of the third level is P soil ,R soil , the output is the soil limiting factor; The input data of the fourth level are W, T, ET, and VPD, and the output is the environmental limiting factor; Each layer of the network contains an encoder-decoder structure and an attention mechanism. The encoder and decoder are composed of 6 encoders stacked together, and the final output is the product of the results of each layer.

4. The method for inverting total primary productivity based on multi-level deep learning according to claim 3, characterized in that: The training method of the multi-level deep learning model is: The GPP of the flux site is used as the label, and the leaf structure information, leaf biochemical information, soil information and meteorological information are used as features. They are sent to the network model for training to obtain the predicted total primary productivity. The loss function is constructed with the label. The mean square error loss function is selected as the loss function, and the Adam optimizer is selected as the optimizer.

5. The method for inverting total primary productivity based on multi-level deep learning according to claim 1, characterized in that: The inversion accuracy assessment method of step (4) is: The model prediction value is compared with the true value to evaluate the inversion accuracy; the root mean square error RMSE and the determination coefficient R are used to 2 As an accuracy evaluation indicator, the formula is as follows: Root mean square error RMSE: Coefficient of determination R 2 : Where n is the total number of samples, is the model prediction value, y={y1,y2,...,y n } is the true value, is the average value of the true value; the range of the root mean square error is [0, +∞), which is equal to 0 when the predicted value is completely consistent with the true value, that is, a perfect prediction; the worse the prediction effect, the larger the root mean square error value; the range of the determination coefficient is [0, 1], which is equal to 1 when the predicted value is completely consistent with the true value, that is, a perfect prediction; the worse the prediction effect, the smaller the determination coefficient.

Images