High-resolution photosynthetically active radiation absorption ratio time sequence generation method based on random forest iterative interpolation

Abstract

The invention discloses a high-resolution photosynthetically active radiation absorption ratio time sequence generation method based on random forest iterative interpolation, and the method comprises the steps: building a Gaussian process regression model with Sentinel-2 surface reflectance as input and GLASS FAPAR as output, carrying out the downscaling of the GLASS FAPAR, and obtaining a FAPAR result of a 10-meter clear sky pixel; carrying out pre-filling on the missing value by utilizing GLASS FAPAR, establishing a random forest model, and carrying out second filling on the missing FAPAR; updating the pre-filled FAPAR value by using the result of the second filling, re-establishing the random forest model, and performing third filling on the missing FAPAR; and applying the final model to the tinel-2 reflection data of the target area until the precision reaches a target threshold, and generating a high-resolution space-time complete FAPAR product of the target area. According to the method, the FAPAR time sequence with the resolution ratio of 10 meters can be accurately generated, and important data support is provided for applications such as vegetation monitoring, ecological system evaluation and carbon cycle research.

Description

Technical Field

[0001] This invention relates to the field of algorithms for inverting quantitative remote sensing parameters of vegetation, and in particular to a method for generating a high-resolution time series of photosynthetically active radiation absorptivity based on random forest iterative interpolation. Background Technology

[0002] The fraction of absorbed photosynthetically active radiation (FAPAR), a key biophysical parameter describing the efficiency of vegetation light energy use, plays an important role in carbon cycle research, agricultural monitoring, and ecological environment assessment. The development of satellite remote sensing technology has made large-scale FAPAR monitoring possible, but existing products still face technical challenges in achieving spatiotemporal resolution coordination.

[0003] Currently, FAPAR remote sensing products can be divided into two main categories: The first category consists of medium-resolution products, such as the 500-meter GLASS FAPAR product and the 1-kilometer MOD15 product built based on MODIS sensors. These products have good temporal coverage and relatively stable observation frequency, but their spatial detail representation is limited, making it difficult to capture local changes in surface vegetation, thus restricting their applicability in small-scale applications. The second category consists of high-resolution products, mainly relying on 10-30 meter observation data from sensors such as the Landsat series and Sentinel-2. Although these products perform excellently in spatial detail characterization, they are severely affected by atmospheric conditions, have insufficient effective observation frequency, and poor temporal series integrity.

[0004] To address these issues, scholars have explored various technical approaches: In terms of spatiotemporal data fusion technology, methods such as STARFM and ESTARFM integrate the advantages of different sensors to construct spatiotemporally coordinated products, but they require high-quality reference images and are unstable in rapidly changing areas; In terms of time series reconstruction technology, mathematical methods such as polynomial fitting and cubic splines are used to fill time gaps, but these methods are usually based on simple mathematical assumptions and ignore the biological laws and spatial correlations of vegetation growth; In terms of intelligent algorithm applications, machine learning methods such as support vector regression and artificial neural networks have shown good nonlinear modeling capabilities, but they require a large number of training samples and it is difficult to ensure the temporal consistency of the generated sequences.

[0005] Random forests, as a representative algorithm of ensemble learning, perform exceptionally well in handling high-dimensional data and missing values. Their built-in missing value handling mechanism provides a new approach for remote sensing time series reconstruction. However, traditional applications often employ a one-time imputation strategy, failing to leverage the potential of iterative optimization, and lack specific optimization designs tailored to the characteristics of FAPAR parameters.

[0006] In conclusion, there is an urgent need to develop new technical solutions, effectively integrate the advantages of multi-source remote sensing information, fully explore the iterative optimization capabilities of the random forest algorithm in time series interpolation, and build high-quality, high-resolution, and time-continuous FAPAR products to meet the actual needs of precise vegetation monitoring. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a high-resolution FAPAR time series generation method based on random forest iterative interpolation. This invention can accurately generate FAPAR time series with a resolution of 10 meters, providing important data support for applications such as vegetation monitoring, ecosystem assessment, and carbon cycle research.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation, comprising: S1: Establish a Gaussian process regression model with Sentinel-2 surface reflectance as input and GLASS FAPAR as output. Downscale GLASS FAPAR to obtain FAPAR results for 10-meter clear sky pixels. S2: Use GLASS FAPAR to pre-fill missing values ​​in the 10-meter FAPAR time series. Build a random forest model with the target clear sky 10-meter FAPAR as output and its historical 3-year pre-filled 10-meter FAPAR time series as input to fill the missing FAPAR a second time. S3: Update the pre-filled FAPAR value with the result of the second filling, rebuild the random forest model with the target clear sky 10-meter FAPAR as the output and its historical 3-year 10-meter FAPAR time series after the second filling as the input, and fill the missing FAPAR for the third time. S4: Repeat step S3 until the accuracy reaches the target threshold. Apply the final model to the tinel-2 reflection data of the target region to generate a high-resolution spatiotemporally complete FAPAR product for the target region.

[0009] As a further improvement of the present invention, step S1 also includes: The blue, green, red, red-edge, near-infrared, and short-wave infrared bands of Sentinel-2 and the solar zenith angle were selected as input features. The radial basis function was used as the kernel function of the Gaussian process regression model, and the optimal combination of hyperparameters was determined through cross-validation.

[0010] As a further improvement of the present invention, step S2 also includes: The random forest model is set to have 100 decision trees, a maximum tree depth of 15, and a minimum number of leaf node samples of 5. The out-of-bag error is used as the model performance evaluation metric.

[0011] As a further improvement of the present invention, step S3 also includes: Using the same random forest model parameter settings, the number of decision trees is kept at 100, the maximum tree depth is 15, and the minimum number of leaf node samples is 5; The random forest model was retrained using the updated FAPAR time series as new training data to improve the model's accuracy in imputing missing values.

[0012] As a further improvement of the present invention, step S4 also includes: The iteration convergence condition is set as follows: the root mean square error change between two consecutive iterations is less than 0.001, or the number of iterations reaches the maximum value of 6. A block processing strategy was adopted, dividing the target area into 1000×1000 pixels and reconstructing the FAPAR time series separately. Establish a vegetation mask based on 10-meter surface classification products to eliminate interference from vegetation areas of water bodies and building sites.

[0013] The beneficial effects of this invention are: 1. This invention integrates multi-scale data fusion and random forest iterative interpolation technology, making full use of the advantages of multi-source remote sensing data, and effectively solving the problems of low accuracy and poor temporal continuity in generating high-resolution FAPAR products by traditional methods.

[0014] 2. This invention achieves accurate downscaling through Gaussian process regression, and combined with iterative optimization strategies, effectively improves the accuracy of FAPAR estimation and the completeness of the time series.

[0015] 3. The present invention uses the random forest iterative imputation method, which can effectively capture the complex time dependencies of FAPAR time series and significantly improve the accuracy and reliability of missing value imputation.

[0016] 4. This invention can efficiently generate long-term FAPAR products with a resolution of 10 meters, providing important data support for applications such as vegetation dynamics monitoring, ecosystem assessment, and carbon cycle research. Attached Figure Description

[0017] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Example

[0019] A high-resolution photosynthetically active radiation absorptivity (RAPA) time series generation method integrating multi-scale data fusion and random forest iterative interpolation techniques is proposed, specifically including: S1, establishing a Gaussian process regression model with Sentinel-2 surface reflectance as input and GLASS FAPAR as output, downscaling GLASS FAPAR to obtain FAPAR results for 10-meter clear sky pixels; S2, pre-filling missing values ​​in the 10-meter FAPAR time series using GLASS FAPAR, establishing a random forest model with the target clear sky 10-meter FAPAR as output and its historical 3-year pre-filled 10-meter FAPAR time series as input, and performing a second filling of missing FAPAR; S3, updating the pre-filled FAPAR values ​​with the results of the second filling, re-establishing the random forest model, and performing a third filling of missing FAPAR; S4, repeating step S3 until the accuracy reaches the target threshold, applying the final model to Sentinel-2 reflectance data of Sichuan Province, and generating a high-resolution spatiotemporally complete FAPAR product for Sichuan Province. This embodiment combines the advantages of multi-scale data fusion and random forest iterative interpolation technology to solve the problems of low accuracy and poor temporal continuity in generating high-resolution FAPAR products by traditional methods. It generates high-precision, long-term 10-meter resolution FAPAR products, providing data support for vegetation dynamic monitoring.

[0020] To make the objectives, technical solutions, and advantages of this embodiment clearer, the technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This embodiment provides a method for generating a high-resolution time series of photosynthetically active radiation absorptivity based on random forest iterative interpolation, including: S1: Establish a Gaussian process regression model with Sentinel-2 surface reflectance as input and GLASS FAPAR as output. Downscale GLASS FAPAR to obtain FAPAR results for 10-meter clear sky pixels. S2: Use GLASS FAPAR to pre-fill missing values ​​in the 10-meter FAPAR time series. Build a random forest model with the target clear sky 10-meter FAPAR as output and its historical 3-year pre-filled 10-meter FAPAR time series as input to fill the missing FAPAR a second time. S3: Update the pre-filled FAPAR value with the result of the second filling, rebuild the random forest model with the target clear sky 10-meter FAPAR as the output and its historical 3-year 10-meter FAPAR time series after the second filling as the input, and fill the missing FAPAR for the third time. S4: Repeat step S3 until the accuracy reaches the target threshold. Apply the final model to the Sentinel-2 reflection data of Sichuan Province to generate a high-resolution spatiotemporally complete FAPAR product for Sichuan Province.

[0022] In one implementation, step S1 further includes: Taking Sichuan Province as an example, this region has complex and diverse topography and rich vegetation types, including subtropical evergreen broad-leaved forests, temperate deciduous broad-leaved forests, alpine meadows, and farmland. The blue, green, red, red-edge, near-infrared, and short-wave infrared bands of Sentinel-2, along with the solar zenith angle, were selected as input feature variables. A radial basis function was used as the kernel function for the Gaussian process regression model. The optimal hyperparameter combination was determined using a 5-fold cross-validation method, with the signal variance set to 1.0, the length scale parameter set to 0.5, and the noise variance set to 0.01. Stratified sampling was conducted within Sichuan Province according to topography and vegetation type, selecting 100,000 representative sample points for model training, achieving accurate downscaling of GLASS FAPAR from 500 meters to 10 meters.

[0023] In one implementation, step S2 includes: For the generated 10-meter FAPAR time series, a large number of missing values ​​exist due to factors such as cloud cover. First, pre-filling is performed using the GLASS FAPAR product. For each missing 10-meter pixel, its corresponding 500-meter GLASS pixel is found, and the GLASS FAPAR value is directly assigned to that missing pixel as the initial fill value. A random forest model is then built for a second fill, with 100 decision trees, a maximum tree depth of 15, and a minimum number of leaf node samples of 5. For each pixel at the target time, the FAPAR time series within its historical three-year time window is extracted as input features, and out-of-bag error is used as the model performance evaluation metric.

[0024] In one implementation, step S3 includes: The initial pre-filled values ​​are updated using the results of the second imputation, forming a new complete time series. Maintaining the same random forest model parameter settings as in step S2, the training dataset is reconstructed, using the updated FAPAR time series as new input features. Due to the improved completeness of the time series, the model is able to learn more accurate temporal correlation patterns. A time sliding window strategy is employed to ensure the model fully utilizes the continuity information of the time series; in addition to using the historical three-year time series, spatial neighborhood information of adjacent time points is also considered, enhancing the model's spatiotemporal modeling capabilities. The random forest model is retrained, and the remaining missing values ​​are imputed a third time.

[0025] In one implementation, step S4 includes: A convergence criterion was set: the model was considered convergent when the change in root mean square error (RMSE) between two consecutive iterations was less than 0.001. A maximum of 6 iterations was set to prevent excessive iteration and wasted computational resources. During each iteration, the RMSE change of all missing pixels was calculated. The iteration process stopped when the RMSE change of more than 95% of the pixels was less than the threshold or the maximum number of iterations was reached. Considering the large area of ​​Sichuan Province, a block-based processing strategy was adopted, dividing the entire study area into 1000×1000 pixels. Each block underwent independent FAPAR time-series reconstruction, and finally, seamless stitching was performed. A vegetation mask based on a 10-meter land classification product was established. The ESA WorldCover product was used to select vegetation types such as forest, shrub, grassland, and farmland, excluding interference from non-vegetated areas such as water bodies and built-up areas. A multi-level quality assessment mechanism was established, including pixel-level quality control, time-series quality control, and spatial consistency checks, to ensure the reliability of the generated results.

[0026] Finally, steps S1 to S4 were used in Sichuan Province to obtain Sentinel-2 FAPAR products with a spatial resolution of 10 meters and a temporal resolution of 5 days for the period 2018-2023. Comparative analysis with ground-based observation data verified the effectiveness and reliability of this method. The generated FAPAR products significantly outperformed existing products in terms of spatial resolution, temporal continuity, and accuracy.

[0027] This embodiment integrates multi-scale data fusion and random forest iterative interpolation techniques, fully leveraging the advantages of multi-source remote sensing data and advanced algorithms. It considers the complex temporal dependencies and spatial heterogeneity of FAPAR time series to generate a photosynthetically active radiation absorptivity (EPAA) product with high spatiotemporal resolution and good temporal continuity. The implementation process of this embodiment can be found in [link to implementation details]. Figure 1 The specific implementation method is further described below with reference to the embodiments. A method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation specifically includes the following steps: Step 1: Construct a multispectral downscaling model based on Gaussian process regression, using Sentinel-2 multiband surface reflectance data as input features and GLASS FAPAR products as target variables to achieve accurate conversion from 500 meters to 10 meters spatial resolution and obtain high-resolution FAPAR estimation results under clear sky conditions. Taking Sichuan Province in my country as an example, this region, located in southwestern my country, covers an area of ​​approximately 486,000 square kilometers. It features complex and diverse topography, with elevations ranging from 200 meters to over 7,000 meters and rich vegetation cover. Sentinel-2 L2A-level surface reflectance products and GLASS FAPAR products covering Sichuan Province from 2018 to 2023 were collected as the basic data source. Six key spectral bands—blue, green, red, red-edge, near-infrared, and shortwave infrared—were selected from the 13 bands of Sentinel-2, along with solar zenith angle information as auxiliary features. A Gaussian process regression model was constructed using radial basis functions, and the optimal hyperparameters were determined through grid search and cross-validation. A stratified sampling strategy was employed within Sichuan Province, selecting 100,000 representative sample points based on elevation and vegetation type for model training. This ensured sufficient representativeness for various terrain and vegetation conditions, ultimately achieving high-precision downscaling of GLASS FAPAR data.

[0028] Step 2: To address the missing values ​​in the 10-meter resolution FAPAR time series caused by cloud and fog obstruction, we first use the corresponding GLASS FAPAR values ​​for preliminary filling. Then, we construct a random forest regression model based on historical time series features and use the time change pattern of the target pixels over the past 3 years to accurately estimate the missing values. Invalid observations such as clouds, cloud shadows, and snow cover were identified using quality assessment information from Sentinel-2 products. Statistical analysis showed that the average effective observation rate in Sichuan Province was approximately 60%. For each missing 10-meter pixel, its corresponding 500-meter GLASS pixel was first located, and the GLASS FAPAR value was used as the initial imputation value. When constructing the random forest model, the number of decision trees was set to 100, the maximum tree depth was limited to 15 layers, and the minimum number of leaf node samples was 5. For each pixel at the target time t, the FAPAR observation sequence within the same time window of the previous 3 years for that pixel was extracted to form a multi-dimensional temporal feature vector as the model input. A bootstrap sampling strategy was used to construct each decision tree, and the model performance was evaluated by out-of-bag error, achieving a second accurate imputation of missing FAPAR values.

[0029] Step 3: Replace the initial pre-filled values ​​with the improved results obtained from the second filling to form a higher quality complete time series data. On this basis, retrain the random forest model to further improve the accuracy of missing value filling through more accurate learning of time dependencies. The second imputation result was used to update the original coarse pre-imputed values, significantly improving the continuity and data quality of the time series. When reconstructing the training samples, the updated high-quality FAPAR time series was used as input features, enabling the model to capture more realistic vegetation dynamics. The consistency of the random forest model parameters was maintained, using the same number, depth, and node settings for decision trees. A multi-timescale feature extraction strategy was introduced, including the mean and standard deviation of the short-term 7-day window, the changing trend of the medium-term 30-day window, and the climatological characteristics of long-term interannual variations. An incremental learning strategy was used to retrain the model, utilizing previous learning results while quickly adapting to new data features, and a third high-precision imputation was performed on missing values.

[0030] Step 4: Establish a dual termination mechanism based on accuracy convergence and iteration count, and continuously repeat the model update and filling process in step 3 until the algorithm converges. Then, apply the optimized final model to the entire Sichuan Province to generate a spatiotemporally continuous and complete high-resolution FAPAR time series product. A dual convergence judgment mechanism was established, setting the accuracy convergence threshold to a change in RMSE of less than 0.001 between two consecutive iterations, while limiting the maximum number of iterations to 6. Model performance changes were monitored in detail during each iteration, and iteration was stopped when more than 95% of the pixels met the convergence criteria. For the vast area of ​​486,000 square kilometers in Sichuan Province, a spatial block parallel processing strategy was adopted, dividing the area into independent processing units of 1000×1000 pixels. A vegetation mask was constructed using the ESA WorldCover 10-meter global land cover product, screening for vegetation types such as forests, shrubs, grasslands, and crops, while excluding non-vegetated areas such as water bodies, construction land, and bare land. A comprehensive quality control system was established, including multi-level quality assurance measures such as physical rationality verification, time series smoothness checks, and spatial consistency verification. Cross-validation with ground data from forest ecosystem observation stations, MODIS, and VIIRS, among other mature products, ensured the scientific validity and reliability of the generated product.

[0031] Based on the above steps, a Sentinel-2 FAPAR time series product with a spatial resolution of 10 meters and a temporal resolution of 5 days was finally generated for Sichuan Province from 2018 to 2023. This product features high spatiotemporal resolution, good temporal continuity, and excellent accuracy, with a correlation coefficient exceeding 0.85 with ground observation data. It provides crucial foundational data support for regional vegetation dynamics monitoring, ecosystem assessment, and carbon cycle research.

[0032] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for generating a high-resolution time series of photosynthetically active radiation absorptivity (PARA) based on random forest iterative interpolation, characterized in that, include: S1: Establish a Gaussian process regression model with Sentinel-2 surface reflectance as input and GLASS FAPAR as output. Downscale GLASS FAPAR to obtain FAPAR results for 10-meter clear sky pixels. S2: Use GLASS FAPAR to pre-fill missing values ​​in the 10-meter FAPAR time series. Build a random forest model with the target clear sky 10-meter FAPAR as output and its historical 3-year pre-filled 10-meter FAPAR time series as input to fill the missing FAPAR a second time. S3: Update the pre-filled FAPAR value with the result of the second filling, rebuild the random forest model with the target clear sky 10-meter FAPAR as the output and its historical 3-year 10-meter FAPAR time series after the second filling as the input, and fill the missing FAPAR for the third time. S4: Repeat step S3 until the accuracy reaches the target threshold. Apply the final model to the tinel-2 reflection data of the target region to generate a high-resolution spatiotemporally complete FAPAR product for the target region.

2. The method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation according to claim 1, characterized in that, Step S1 also includes: The blue, green, red, red-edge, near-infrared, and short-wave infrared bands of Sentinel-2 and the solar zenith angle were selected as input features. The radial basis function was used as the kernel function of the Gaussian process regression model, and the optimal combination of hyperparameters was determined through cross-validation.

3. The method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation according to claim 1, characterized in that, Step S2 also includes: The random forest model is set to have 100 decision trees, a maximum tree depth of 15, and a minimum number of leaf node samples of 5. The out-of-bag error is used as the model performance evaluation metric.

4. The method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation according to claim 1, characterized in that, Step S3 also includes: Using the same random forest model parameter settings, the number of decision trees is kept at 100, the maximum tree depth is 15, and the minimum number of leaf node samples is 5; The random forest model was retrained using the updated FAPAR time series as new training data to improve the model's accuracy in imputing missing values.

5. The method for generating a high-resolution photosynthetically active radiation absorptivity time series based on random forest iterative interpolation according to claim 1, characterized in that, Step S4 also includes: The iteration convergence condition is set as follows: the root mean square error change between two consecutive iterations is less than 0.001, or the number of iterations reaches the maximum value of 6. A block processing strategy was adopted, dividing the target area into 1000×1000 pixels and reconstructing the FAPAR time series separately. Establish a vegetation mask based on 10-meter surface classification products to eliminate interference from vegetation areas of water bodies and building sites.

Images