A method for retrieving the ratio of photosynthetically active radiation absorption with high spatio-temporal resolution by coupling a mechanism model and machine learning

Abstract

The present invention discloses a method for retrieving the fraction of photosynthetically active radiation absorbed (FAPAR) with high spatio-temporal resolution by coupling a mechanism model and machine learning, including: simulating the top-of-atmosphere reflectance of FY3B MERSI, retrieving clear-sky FAPAR and the surface reflectance of the corresponding Landsat bands from the top-of-atmosphere reflectance of FY3B MERSI; establishing a first LightGBM regression model to retrieve the 30-meter clear-sky FAPAR value from the true Landsat surface reflectance; establishing a second LightGBM regression model to reconstruct the missing Landsat 30-meter FAPAR value under clouds; applying the first LightGBM regression model and the second LightGBM regression model successively at the regional scale to estimate the fraction of photosynthetically active radiation absorbed with high spatio-temporal resolution; the present invention can accurately estimate the FAPAR time series at a resolution of 30 meters.

Description

Technical Field

[0001] The present invention relates to the technical field of quantitative remote sensing parameter inversion of vegetation, and in particular to a method for inverting the fraction of absorbed photosynthetically active radiation with high spatio-temporal resolution by coupling a mechanism model and machine learning. Background Technique

[0002] The fraction of absorbed photosynthetically active radiation (FAPAR) is a key biophysical parameter characterizing the light energy utilization efficiency of vegetation, which reflects the absorption degree of the solar radiation in the photosynthesis available band (400 - 700 nm) by the vegetation canopy. FAPAR is an important input parameter for estimating the net primary productivity (NPP) of ecosystems and is also an important indicator for evaluating the impacts of global carbon cycling and climate change. Accurately obtaining FAPAR with high spatio-temporal resolution is of great significance for understanding the light energy utilization mechanism of terrestrial ecosystems, simulating vegetation productivity, and evaluating ecosystem service functions. Currently, there are mainly three types of methods for inverting FAPAR based on satellite remote sensing data: 1) Physical methods based on radiative transfer models, such as the inversion method based on the PROSPECT-SAIL model. Such methods have clear physical mechanisms, but are computationally complex, inefficient, and difficult to handle inversion problems under complex terrain conditions; 2) Empirical methods based on vegetation indices, such as the inversion models established based on indices such as NDVI and EVI. Such methods are simple and easy to implement, but lack the support of physical mechanisms and have limitations in applicability under different surface conditions; 3) Methods based on machine learning, such as neural networks and random forests. Such methods have strong non-linear fitting capabilities, but often rely on a large number of training samples and are prone to overfitting problems.

[0003] Most of the existing FAPAR products are generated based on medium and low spatial resolution satellite data (such as MODIS, MERIS, etc.), and the spatial resolution is between 250 meters and 1 kilometer. Such a rough spatial resolution is difficult to meet the needs of refined ecosystem research. Especially in areas with complex terrain and diverse vegetation types, it is impossible to accurately depict the impact of surface heterogeneity on FAPAR. Although there have been studies attempting to invert FAPAR using high-resolution data such as Landsat, the following challenges still remain: 1) Due to the influence of factors such as atmosphere, clouds, and shadows, the effective observation of high-resolution satellite data is severely restricted, resulting in poor temporal continuity of the inversion results; 2) Most of the existing methods directly establish empirical relationships between remote sensing observations and FAPAR, lacking consideration of the physical process of radiative transfer, and it is difficult to guarantee the inversion accuracy in complex environments; 3) There is a scale effect between FAPAR products with different spatial resolutions, and how to effectively use multi-source remote sensing data for high-precision inversion remains a challenge.

[0004] Therefore, there is an urgent need to develop new methods that combine the mechanism of physical models with the efficiency of machine learning, make full use of the advantages of multi-source remote sensing data and advanced algorithms, generate FAPAR products with high spatio-temporal resolution and good temporal continuity, and provide more reliable data support for ecosystem research. Summary of the Invention

[0005] To solve the problems existing in the prior art, the object of the present invention is to provide a method for retrieving the fraction of absorbed photosynthetically active radiation (FAPAR) with high spatio-temporal resolution by coupling a mechanism model and machine learning. The present invention can accurately estimate the FAPAR time series at a resolution of 30 meters, and provide important data support for applications such as the study of light energy utilization in terrestrial ecosystems, the simulation of vegetation productivity, and the assessment of the carbon cycle.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a method for retrieving the fraction of absorbed photosynthetically active radiation (FAPAR) with high spatio-temporal resolution by coupling a mechanism model and machine learning, comprising the following steps:

[0007] S1: Use a land-atmosphere coupled radiative transfer model to simulate the top-of-atmosphere reflectance of FY3B MERSI, and retrieve the clear-sky FAPAR and the corresponding Landsat surface reflectance of the corresponding band by means of the SCE optimization algorithm from the top-of-atmosphere reflectance of FY3B MERSI;

[0008] S2: Take the retrieved Landsat surface reflectance of the corresponding band and the solar zenith angle as inputs, and the retrieved clear-sky FAPAR as the output, and establish a first LightGBM regression model to retrieve the 30-meter clear-sky FAPAR value from the true Landsat surface reflectance;

[0009] S3: Take the Landsat 30-meter clear-sky vegetation photosynthetically active radiation absorption coefficient FAPAR as the output, find similar pixels in the area centered on the pixel where the FAPAR value is located, and use the Landsat 30-meter clear-sky FAPAR, the latitude, the day of the year, the elevation, the GLASS FAPAR, the surface type, and the environmental variables corresponding to the target FAPAR pixel of the similar pixels in the same period as the inputs to establish a second LightGBM regression model to reconstruct the missing Landsat 30-meter FAPAR value under the cloud;

[0010] S4: Apply the first LightGBM regression model and the second LightGBM regression model successively at the regional scale to estimate the fraction of absorbed photosynthetically active radiation with high spatio-temporal resolution.

[0011] As a further improvement of the present invention, in step S1, the radiation transfer model of the land-atmosphere coupling that simulates the top-of-atmosphere reflectance of FY3B MERSI includes the Walthall&Price soil model, the PROSPECT-D and PROSPECT-VISIR leaf models, the 4SAIL canopy model, the physically-based terrain correction model, and the LibRadtran 2.0.3 atmospheric model;

[0012] When the model is simulated, globally homogeneous representative training sample points are selected, and the K-means unsupervised classification method is used to cluster the GLASS FAPAR time series. The sample points selected after clustering should follow the following requirements: 1) All the Landsat 30-meter pixels corresponding to a 250-meter pixel belong to the same land use type; 2) The sample points should be evenly selected in each MODIS tile, and the number of samples belonging to each category should be equivalent.

[0013] As a further improvement of the present invention, step S2 is specifically as follows:

[0014] When establishing the first LightGBM regression model, the spatio-temporal cross-validation method is used to determine the optimal model; for spatial cross-validation, the global sample points are divided into 10 groups according to their spatial positions. In each of the ten repeated experiments, 9 groups are selected to train the model, and the other group is used for independent model verification; for temporal cross-validation, the global samples are divided into 10 groups according to months. In each of the ten repeated experiments, 9 groups are selected to train the model, and the other group is used for independent model verification;

[0015] When applying the first LightGBM regression model to retrieve the 30-meter clear-sky FAPAR value from the true Landsat surface reflectance, the Fmask 4.6 algorithm is used to obtain the Landsat cloud mask, and the model is only applied to the Landsat pixels where the cloud mask is labeled as clear sky to retrieve the Landsat 30-meter clear-sky FAPAR.

[0016] As a further improvement of the present invention, in step S3, when looking for similar pixels, a 3 km × 3 km spatial window is defined with the target pixel as the center. The average annual FAPAR curve is obtained by averaging the Landsat 30-meter clear-sky FAPAR of all pixels within the spatial window over a 3-year time window. If the effective value in the FAPAR curve is not higher than 70%, the time window is expanded to five years, and so on; the correlation coefficient between the annual average FAPAR curve of the target pixel and that of other pixels is calculated, and the top 5 pixels with the highest correlation coefficient are selected as similar pixels.

[0017] As a further improvement of the present invention, in step S3, the Landsat 30-meter clear-sky FAPAR of the same period as the target pixel in the three years before and after the selected similar pixels is used as the spatio-temporal constraint part in the input features, and the latitude, day of the year, elevation, GLASS FAPAR, land surface type, and environmental variables corresponding to the target FAPAR pixel are used as the input to establish a second LightGBM regression model, where the environmental variables include temperature, vapor pressure deficit, and solar radiation.

[0018] As a further improvement of the present invention, step S4 is specifically as follows:

[0019] For a given study area, the first LightGBM regression model is applied to invert the clear-sky FAPAR from the pixels marked as clear sky by the Fmask4.6 cloud mask, and the second LightGBM regression model is applied to reconstruct the missing FAPAR under the cloud from the pixels marked as cloud or cloud shadow by the Fmask4.6 cloud mask, and finally a spatio-temporally seamless Landsat 30-meter 16-day FAPAR product for the study area is obtained.

[0020] The beneficial effects of the present invention are as follows:

[0021] 1. The present invention innovatively integrates a physical model and a machine learning method, which not only maintains the physical mechanism of the radiation transfer theory but also fully utilizes the high efficiency and non-linear fitting ability of the machine learning algorithm.

[0022] 2. The present invention effectively improves the reliability and accuracy of FAPAR inversion through multi-level data quality control and strict model verification strategies.

[0023] 3. The present invention adopts a spatio-temporal reconstruction method based on similarity, effectively solves the problem of data loss caused by cloud contamination, and realizes the spatio-temporal continuity of high-resolution FAPAR products.

[0024] 4. The 30-meter resolution FAPAR product generated by the present invention can better characterize the surface heterogeneity and provide important data support for the study of ecosystem light energy utilization and carbon cycle assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Embodiment 1

[0028] As Figure 1 shown, a method for inverting the fraction of absorbed photosynthetically active radiation with high spatio-temporal resolution by coupling a mechanism model and machine learning includes:

[0029] S1: Use the land-atmosphere coupled radiative transfer model to simulate the top-of-atmosphere reflectance of FY3B MERSI. With the help of the SCE optimization algorithm, retrieve the clear-sky FAPAR and the surface reflectance of the corresponding Landsat bands from the top-of-atmosphere reflectance of FY3B MERSI;

[0030] S2: Take the retrieved surface reflectance of the corresponding Landsat bands and the solar zenith angle as inputs, and the retrieved clear-sky FAPAR as the output to establish the first LightGBM regression model, and retrieve the 30-meter clear-sky FAPAR value from the true Landsat surface reflectance;

[0031] S3: Take the Landsat 30-meter clear-sky FAPAR as the output, search for similar pixels within a 3 km×3 km area centered on the pixel where the FAPAR value is located. Take the Landsat 30-meter clear-sky FAPAR, the latitude, the day of the year, the elevation, the GLASS FAPAR, the surface type, and the environmental variables corresponding to the target FAPAR pixel of the similar pixels in the same period within 3 years as inputs to establish the second LightGBM regression model to reconstruct the missing Landsat 30-meter FAPAR value under the cloud;

[0032] S4: Apply the first and second LightGBM regression models successively at the regional scale to estimate the photosynthetically active radiation absorption ratio with high spatio-temporal resolution.

[0033] In one implementation, step S1 further includes:

[0034] Taking the Qinghai-Tibet Plateau region as an example, the terrain of this region is complex and the ecosystem types are diverse, with a wide distribution from alpine meadows to subalpine coniferous forests. First, establish a complete land-atmosphere coupled radiative transfer model system, including the Walthall&Price soil model, the PROSPECT-D and PROSPECT-VISIR leaf models, the 4SAIL canopy model, the physics-based terrain correction model, and the LibRadtran 2.0.3 atmospheric model.

[0035] In the training sample selection stage, use the K-means clustering method to analyze the GLASS FAPAR time series from 2000 to 2023. It is verified that setting the number of clusters to 12 can better reflect the characteristics of different vegetation functional types. To ensure the sample quality, two criteria are strictly implemented: one is to ensure that all 30-meter Landsat pixels under the 250-meter pixel belong to the same land class; the other is that the sample points are evenly distributed in each MODIS tile and the quantities of each class are quite equal. Finally, about 500,000 high-quality training samples are selected in the study area.

[0036] Parameter inversion is carried out through the SCE optimization algorithm, with the population size set to 100, the maximum number of iterations set to 1000, and the convergence threshold set to 10 -6 During the optimization process, a multi-initialization strategy is adopted to avoid local optimal solutions, and parallel computing is used to improve the inversion efficiency. The optimized results can simultaneously obtain the clear-sky FAPAR value and the corresponding Landsat band surface reflectance

[0037] In one implementation, step S2 includes:

[0038] To ensure the generalization ability of the model, a strict spatio-temporal cross-validation strategy is adopted. In the spatial dimension, the global sample points are divided into 10 groups according to their spatial positions. Each time, 9 groups are selected for model training, and the remaining 1 group is used for independent verification. In the temporal dimension, the samples are divided into 10 groups according to months, and the leave-one-out method is also used for cross-validation to ensure the applicability of the model in different seasons

[0039] The parameters of the first LightGBM model are determined through grid search: the learning rate is set to 0.01, the maximum tree depth is 8, the minimum number of samples in leaf nodes is 5, and the feature sampling ratio is 0.8. In practical applications, the Fmask4.6 algorithm is first used to detect clouds in Landsat images, and the model is only applied to the pixels marked as clear sky for FAPAR inversion

[0040] In one implementation, step S3 includes:

[0041] The size of the spatial window is set to 3km×3km, which not only ensures that enough potential similar pixels are included but also avoids heterogeneity caused by excessive terrain changes. For all pixels within the spatial window, the clear-sky FAPAR observation values within a 3-year time window are extracted, and the annual average FAPAR curve is calculated. If the effective observation rate of a certain pixel is lower than 70%, the time window is extended to 5 years

[0042] By calculating the correlation coefficient between the target pixel and the annual average FAPAR curves of other pixels within the window, the top 5 pixels with the highest correlation coefficient are selected as similar pixels. At the same time, various types of auxiliary information corresponding to these similar pixels are collected, including variables such as geographical location, terrain, phenology, and environment (temperature, vapor pressure deficit, solar radiation), etc., to jointly construct the input features of the second LightGBM model

[0043] In one implementation, step S4 includes:

[0044] For the study area, first, the first LightGBM model is applied to process the pixels marked as clear sky by Fmask4.6 to obtain the clear sky FAPAR values at a resolution of 30 meters. For the pixels marked as clouds or cloud shadows, the second LightGBM model is applied to reconstruct FAPAR based on the historical observations of similar pixels and various auxiliary information.

[0045] During the model application process, a partitioned parallel processing strategy is adopted to improve the computational efficiency. The prediction results of each pixel are evaluated for quality, including the uncertainty analysis of model prediction and the temporal continuity test. After strict quality control, the Landsat FAPAR products with a spatial resolution of 30 meters and a temporal resolution of 16 days for the Qinghai-Tibet Plateau region from 2000 to 2022 are finally generated.

[0046] The generated FAPAR products are compared and verified with the observed data from ground flux stations and existing medium- and low-resolution FAPAR products. The results show that the high-resolution FAPAR products generated by this method have high spatio-temporal consistency (R 2 > 0.85) and good temporal continuity, and can accurately describe the spatial heterogeneity and temporal dynamic characteristics of vegetation light energy utilization under complex terrain conditions.

[0047] This embodiment can integrate the advantages of mechanism models and machine learning techniques, solve problems such as insufficient physical constraints and significant cloud contamination effects in traditional methods when generating high spatio-temporal resolution FAPAR products, generate high-precision, spatio-temporally continuous FAPAR products with a resolution of 30 meters, and provide important data support for the research on vegetation light energy utilization, vegetation productivity simulation, and carbon cycle assessment in terrestrial ecosystems; integrate mechanism models and machine learning techniques, make full use of the advantages of the physical mechanism of radiative transfer and advanced algorithms, consider the physical constraints of FAPAR and the characteristics of surface heterogeneity, and generate photosynthetically active radiation absorption ratio products with high spatio-temporal resolution and a good physical basis.

[0048] Example 2

[0049] A method for retrieving the photosynthetically active radiation absorption ratio with high spatio-temporal resolution by coupling a mechanism model and machine learning specifically includes the following steps:

[0050] Step 1, use a land-atmosphere coupled radiative transfer model to simulate the top-of-atmosphere reflectance of FY3B MERSI, and retrieve the clear sky FAPAR and the surface reflectance of the corresponding Landsat bands from the top-of-atmosphere reflectance of FY3B MERSI with the help of the SCE optimization algorithm;

[0051] Taking the Qinghai-Tibet Plateau region of China as an example, a complete land-atmosphere coupled radiation transfer model system was first constructed. Considering the unique high-altitude environment of the Qinghai-Tibet Plateau, special attention was paid to the spatial variability of parameters such as atmospheric optical thickness and aerosol characteristics during model design. The model system includes the Walthall&Price soil model, the PROSPECT-D and PROSPECT-VISIR leaf models, the 4SAIL canopy model, a physically based terrain correction model, and the LibRadtran 2.0.3 atmospheric model. Each model was optimized in terms of parameters and locally adjusted according to the characteristics of the plateau environment to accurately describe the radiation transfer process under complex terrain and high-altitude conditions.

[0052] Based on the constructed radiation transfer model system, the SCE optimization algorithm was used to invert FAPAR and the corresponding Landsat band surface reflectance from the top-of-atmosphere reflectance of FY3B MERSI. During the optimization process, the population size was set to 100, the maximum number of iterations was 1000, and the convergence threshold was 10 -6 . To improve the inversion accuracy, the spectral matching degree of multiple MERSI bands and the physical rationality constraint of FAPAR were simultaneously considered in the optimization objective function. The multiple initialization strategy was adopted to avoid local optimal solutions, and the MPI parallel computing framework was used to improve the computing efficiency. The inversion results include both FAPAR values consistent with the MERSI spatial resolution and simulated Landsat band surface reflectances.

[0053] During the training sample selection process, a detailed analysis was carried out on the GLASS FAPAR time series from 2000 to 2023. The statistical characteristics (annual mean value, seasonal variation, interannual trend, etc.) and phenological indicators (beginning of the growing season, end of the growing season, length, etc.) of the time series were extracted, and the K-means clustering method was used to divide the vegetation in the study area into 12 functional type categories. To ensure the sample quality, strict screening criteria were established: using Global Land30 data to ensure the surface cover consistency of Landsat pixels within 250-meter pixels; requiring the effective observation ratio of sample points from 2000 to 2023 to be not less than 80%; ensuring the spatial balanced distribution of samples in MODIS tiles; considering the uniform coverage of different altitude, slope, and aspect conditions. Finally, about 500,000 high-quality training sample points were screened out in the Qinghai-Tibet Plateau region.

[0054] Step 2, establish the first LightGBM regression model to simulate the relationship between surface reflectance and high-resolution FAPAR inversion. Using the surface reflectance of the corresponding Landsat bands and the solar zenith angle inverted in Step 1 as input features, and the clear-sky FAPAR value inverted by the physical model as the output label, use machine learning algorithms to mine the non-linear relationship between the input features and FAPAR, and finally achieve a high-precision mapping from real Landsat observation data to the 30-meter resolution clear-sky FAPAR value;

[0055] In the Qinghai-Tibet Plateau region, to ensure the reliability and stability of the LightGBM model, a strict spatio-temporal cross-validation strategy is adopted. In the spatial dimension, the sample points in the study area are divided into 10 sub-regions according to the longitude and latitude grid. Each time, the samples in 9 regions are selected for model training, and the remaining 1 region is used for independent verification. In the time dimension, the samples are divided into 10 groups according to months, and the same cross-validation strategy is adopted to ensure that the model can adapt to the surface and atmospheric conditions in different seasons.

[0056] Determine the optimal parameter combination of the first LightGBM model through grid search: the learning rate is set to 0.01, the maximum tree depth is 8, the minimum number of samples in the leaf nodes is 5, and the feature sampling ratio is 0.8. During the model training process, an early stopping strategy is adopted to prevent overfitting, and L1 and L2 regularization terms are introduced to control the model complexity. Feature importance analysis shows that the reflectance in the near-infrared band and the red band contribute the most to FAPAR estimation, followed by the influence of the solar zenith angle.

[0057] In the actual application stage, first use the Fmask4.6 algorithm to perform cloud detection on Landsat images. The applicability of this algorithm in the plateau region has been verified, and it can effectively identify clouds, cloud shadows, and snow cover. Only apply the trained LightGBM model to the pixels marked as clear sky for FAPAR inversion. To further ensure the reliability of the inversion results, a valid value range constraint based on physical mechanisms is set to screen out possible outliers.

[0058] Step 3, based on the obtained Landsat 30-meter clear-sky FAPAR, identify similar pixels by establishing a search window around the target pixel. At the same time, combine multi-source information such as the historical observations, geographical locations, terrain conditions, phenological characteristics, and environmental variables of the similar pixels to establish a second LightGBM regression model to achieve high-precision reconstruction of the missing FAPAR under clouds;

[0059] Under the complex topographic conditions of the Qinghai-Tibet Plateau, the setting of the spatial window size is particularly important. Through experimental analysis, it is determined that a 3 km × 3 km centered on the target pixel is the most appropriate, which can not only include enough potential similar pixels but also ensure the relative consistency of surface conditions. For all pixels within the spatial window, the clear-sky FAPAR observation sequences within a 3-year time window are extracted, and the annual average FAPAR curve is calculated. For pixels with less than 70% effective observations, more observation samples are obtained by gradually expanding the time window to 5 years.

[0060] The identification of similar pixels adopts a method based on time series similarity. Calculate the correlation coefficient of the annual average FAPAR curves between the target pixel and other pixels within the window, and select the top 5 pixels with the highest correlation coefficient as references. At the same time, collect multi-source auxiliary information of these similar pixels, including environmental variables such as latitude, day of the year, elevation, slope, aspect, GLASS FAPAR, land use type, temperature, vapor pressure deficit, solar radiation, etc., as the input features of the second LightGBM model.

[0061] During the training process of the second LightGBM model, special attention is paid to maintaining time continuity. The model structure design adopts a multi-input and multi-output form, considering both the observation value sequences of similar pixels and the historical observation records of the target pixel, and ensuring the time smoothness of the reconstruction result through the design of the loss function. The model validation adopts the holdout method, and 20% of the samples are used for independent testing.

[0062] Step 4, apply Fmask 4.6 to all Landsat images in the study area for cloud detection. For pixels marked as clear sky, directly use the first LightGBM model for FAPAR inversion; for pixels identified as clouds or cloud shadows, apply the second LightGBM model for FAPAR reconstruction.

[0063] Divide the study area into several sub-regions according to the UTM projection zone, process each sub-region separately, and finally perform seamless mosaicking. To ensure the reliability of the inversion results, a multi-level quality control system is established: first, the valid value range test based on physical mechanisms; second, the time continuity test to eliminate abnormal time jumps; finally, the spatial consistency test to ensure the smoothness of spatial transitions.

[0064] The inversion results have passed strict accuracy verification. By comparing with the FAPAR data of the ground flux observation station, the 30-meter resolution and 16-day time interval FAPAR products generated in the Qinghai-Tibet Plateau region from 2000 to 2022 have achieved good accuracy (R 2 > 0.85, RMSE < 0.1). The products not only maintain good time continuity but also can clearly depict the spatial heterogeneity characteristics of vegetation light energy utilization under the complex topographic conditions of the plateau.

[0065] The above-described embodiments merely represent specific implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several variations and improvements can still be made, and these all fall within the protection scope of the present invention.

Claims

1. A method for retrieving the ratio of photosynthetically active radiation absorption with high spatio-temporal resolution by coupling a mechanism model and machine learning, characterized in that It includes the following steps: S1: Use the land-atmosphere coupled radiative transfer model to simulate the top-of-atmosphere reflectance of FY3B MERSI, and retrieve the clear-sky FAPAR and the surface reflectance of the corresponding Landsat bands from the top-of-atmosphere reflectance of FY3B MERSI with the help of the SCE optimization algorithm; S2: Take the retrieved surface reflectance of the corresponding Landsat bands and the solar zenith angle as inputs, and the retrieved clear-sky FAPAR as the output, and establish the first LightGBM regression model to retrieve the 30-meter clear-sky FAPAR value from the real Landsat surface reflectance; S3: Take the Landsat 30-meter clear-sky fraction of photosynthetically active radiation absorbed by vegetation (FAPAR) as the output, search for similar pixels in the area centered on the pixel where the FAPAR value is located, and take the Landsat 30-meter clear-sky FAPAR, the latitude, the day of the year, the elevation, the GLASS FAPAR, the surface type, and the environmental variables corresponding to the target FAPAR pixel of the similar pixels in the same period as inputs, and establish the second LightGBM regression model to reconstruct the missing Landsat 30-meter FAPAR value under clouds; S4: Apply the first LightGBM regression model and the second LightGBM regression model successively at the regional scale to estimate the fraction of photosynthetically active radiation absorbed with high spatio-temporal resolution.

2. The method for inverting the photosynthetically active radiation absorption ratio with high spatio-temporal resolution by coupling a mechanism model and machine learning according to claim 1, characterized in that, In step S1, the land-atmosphere coupled radiative transfer model for simulating the top-of-atmosphere reflectance of FY3B MERSI includes the Walthall&Price soil model, the PROSPECT-D and PROSPECT-VISIR leaf models, the 4SAIL canopy model, the physically based terrain correction model, and the LibRadtran 2.0.3 atmospheric model; When the model is simulated, select globally homogeneous representative training samples, and use the K-means unsupervised classification method to cluster the GLASS FAPAR time series. The samples selected after clustering should meet the following requirements: 1) All the Landsat 30-meter pixels corresponding to a 250-meter pixel belong to the same land type; 2) The samples should be evenly selected in each MODIS tile, and the number of samples belonging to each category should be comparable.

3. The high spatio-temporal resolution photosynthetically active radiation absorption ratio inversion method by coupling a mechanism model with machine learning according to claim 1, characterized in that Step S2 is specifically as follows: When establishing the first LightGBM regression model, use the spatio-temporal cross-validation method to determine the optimal model; for spatial cross-validation, divide the global sample points into 10 groups according to their spatial positions, and select 9 groups to train the model in each of the ten repeated experiments, and use the other group for independent model validation; for temporal cross-validation, divide the global samples into 10 groups according to months, and select 9 groups to train the model in each of the ten repeated experiments, and use the other group for independent model validation; When applying the first LightGBM regression model to retrieve the 30-meter clear-sky FAPAR value from the real Landsat surface reflectance, use the Fmask4.6 algorithm to obtain the Landsat cloud mask, and apply the model to retrieve the Landsat 30-meter clear-sky FAPAR only on the Landsat pixels where the cloud mask is labeled as clear sky.

4. The method for inverting the photosynthetically active radiation absorption ratio with high spatio-temporal resolution by coupling a mechanism model and machine learning according to claim 1, wherein In step S3, when looking for similar pixels, a 3 km × 3 km spatial window is defined with the target pixel as the center. The average annual FAPAR curve is obtained by averaging the Landsat 30-meter clear-sky FAPAR within the 3-year time window for all pixels in the spatial window. If the effective value in the FAPAR curve is not higher than 70%, the time window is extended to five years, and so on. The correlation coefficient between the annual average FAPAR curve of the target pixel and that of other pixels is calculated, and the top 5 pixels with the highest correlation coefficient are selected as similar pixels.

5. The high spatio-temporal resolution photosynthetically active radiation absorption ratio inversion method coupling a mechanism model and machine learning according to claim 4, characterized in that, In step S3, the Landsat 30-meter clear-sky FAPAR of the same period as the target pixel in the three years before and after the selection of similar pixels is used as the spatio-temporal constraint part of the input features. The latitude, day of the year, elevation, GLASS FAPAR, land surface type, and environmental variables corresponding to the target FAPAR pixel are used as the input to establish the second LightGBM regression model. Among them, the environmental variables include temperature, vapor pressure deficit, and solar radiation.

6. The method for inverting the photosynthetically active radiation absorption ratio with high spatio-temporal resolution by coupling the mechanism model and machine learning according to claim 1, characterized in that, Step S4 is specifically as follows: For a given study area, the first LightGBM regression model is used to invert the clear-sky FAPAR from the pixels marked as clear sky in the Fmask4.6 cloud mask, and the second LightGBM regression model is used to reconstruct the missing FAPAR under the cloud from the pixels marked as cloud or cloud shadow in the Fmask4.6 cloud mask, and finally a temporally and spatially seamless Landsat 30-meter 16-day FAPAR product for the study area is obtained.

Images