A method for retrieving above-ground biomass of a forest

Abstract

A forest aboveground biomass prediction method, comprising the following steps: 1) optical data acquisition and preprocessing; 2) preprocessing of target SAR data; 3) feature extraction of optical and SAR images; 4) selecting two SAR variables with the highest AGB correlation coefficients and combining the vegetation index and biophysical parameters with the absolute value greater than 0.1 to obtain COSI; 5) using PIO iteration process to optimize COSI1 parameter and COSI2 parameter of BiLSTM neural network model; 6) linking the number of hidden layer nodes to the BiLSTM layer, and inputting the optimal parameters of initial learning rate and L2 regularization coefficient into the BiLSTM neural network model; the prediction value of AGB is obtained. The method has more advantages in AGB inversion. Compared with other three models, the relative improvement rate of RMSE is as high as 43.6%.

Description

Technical Field

[0001] This invention relates to a method for obtaining forest aboveground biomass, and more particularly to a method for forest aboveground biomass inversion that integrates swarm intelligence optimization and deep learning. Background Technology

[0002] Forest biomass is closely linked to global carbon cycling and climate change, with aboveground biomass (AGB) comprising 70-90%. Traditional field measurements of AGB are costly and destructive to forest ecosystems, particularly failing to capture large-scale, comprehensive changes. In contrast, remote sensing technologies can rapidly cover vast areas at a lower cost, reaching remote regions inaccessible to human personnel. Among a range of available data sources, Sentinel-1 and Sentinel-2 are suitable choices for large-scale and cost-effective AGB mapping, primarily due to their global coverage and free accessibility. Since spectral information in the infrared and near-infrared bands is generally highly sensitive to AGB, many studies employ vegetation indices and biophysical parameters to estimate AGB. Furthermore, synthetic aperture radar (SAR) can acquire dielectric properties and vertical structure through more appropriate polarization. Previous studies have demonstrated the advantages of synergistic AGB estimation using SAR and optical remote sensing.

[0003] Sentinel-1 and Sentinel-2 images can be used to conveniently extract a large number of feature variables for forest AGB inversion. However, too many feature variables can lead to information redundancy, resulting in significant errors between model building and prediction. Therefore, when combining different data sources, optimization is needed to increase and enhance features relevant to AGB. Most previous research has focused on identifying appropriate features. Generally, correlation calculations can measure the sensitivity of feature variables to forest AGB. These studies often combine techniques such as random forest (RF) algorithms or stepwise linear regression to screen features with robust linear relationships to AGB from a large number of features. While importance-based feature selection undoubtedly produces a more concise feature set, this approach faces significant limitations in obtaining the optimal feature set. When using RF or other algorithms to rank the importance of feature variables, experiments need to ensure that the number of feature variables is greater than the number of sample images. Whether this method is suitable for obtaining the optimal feature set for AGB inversion when the number of sample images is much greater than the number of feature variables warrants careful consideration. Further research is needed to create more reliable methods for obtaining suitable feature sets for AGB inversion.

[0004] Furthermore, constructing an appropriate inversion model is another crucial step in estimating AGB. Parametric models, such as Multiple Linear Regression (MLR) and Generalized Linear Models (GLM), typically require making assumptions. These assumptions need to be based on the fundamental distribution of the data to establish the association between the predictor variables and AGB. In contrast, nonparametric models often infer these relationships by creating flexible mappings, such as Decision Trees (DT), Randomized Randomized Regression (RF), Support Vector Machines (SVM), and Deep Learning (DL). Given the inherent complexity of the relationship between AGB and remote sensing data, parametric models often exhibit limited accuracy. Nonparametric models, on the other hand, have a clear advantage in handling complex data dimensions. Bidirectional Long Short-Term Memory (BiLSTM) neural networks in DL are capable of incorporating bidirectional information from the past and future. This enables BiLSTM to address challenges such as vanishing and exploding gradients in data modeling. However, BiLSTM suffers from parameter selection difficulties and overfitting problems, which are also common in DL models. When model parameters cannot be fully determined, the results are unreliable. In fact, intelligent optimization algorithms can be used to find optimal parameters, improving the robustness of DL models, and have been widely applied in image classification and object detection. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for forest aboveground biomass inversion based on integrated swarm intelligence optimization and deep learning.

[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for predicting forest aboveground biomass, comprising the following steps;

[0007] 1) Acquisition and preprocessing of optical data; acquiring Sentinel-2 optical images of the target location via satellite, and then correcting and unifying the resolution;

[0008] 2) Preprocess the Sentinel-1 SAR data of the target location, including orbit correction, radiometric calibration, multi-look, filtering, terrain correction and geocoding, thermal noise removal and decibel reduction; and obtain polarization parameters through polarization decomposition and interferometric parameters through InSAR technology.

[0009] 3) Feature extraction from optical and SAR images; extracting vegetation indices and biophysical parameters from preprocessed Sentinel-2 optical images;

[0010] 4) Optimization of COSI features: Select the two SAR variables with the highest AGB correlation coefficient and combine them with vegetation indices with an absolute correlation value greater than 0.1 and biophysical parameters to obtain COSI; which includes two types of indicators, namely COSI1 parameters and COSI2 parameters.

[0011] 5) Construction of the inversion model; Optimize the COSI1 and COSI2 parameters of the BiLSTM neural network model using the PIO iteration process; Obtain the optimal solutions for the initial learning rate, L2 regularization coefficient, and number of hidden layer nodes;

[0012] 6) The optimal parameters of the number of hidden layer nodes, initial learning rate, and L2 regularization coefficient are input into the BiLSTM neural network model; the input variables are connected to the regression layer after passing through several network layers of BiLSTM to obtain the AGB prediction value.

[0013] In the above-mentioned method for predicting forest aboveground biomass, preferably, the correction of the optical image Sentinel-2 includes orthorectification, geometric correction, and atmospheric correction.

[0014] In the above-mentioned method for predicting forest aboveground biomass, preferably, the vegetation indices include ARVI, EVI, MSAVI, MSR, NDVI, OSAVI, PVI, RVI, SAVI, and VARI.

[0015] The above-mentioned method for predicting forest aboveground biomass, preferably, involves the following steps in obtaining interferometric parameters using InSAR technology:

[0016] A) Select the image and perform registration;

[0017] B) Eliminate the leveling effect;

[0018] C) The maximum likelihood method is selected for coherence coefficient calculation, and Goldstein Phase Filtering is used for filtering;

[0019] D) Phase unwrapping: Adding the blurred phase to the principal phase value to obtain the true phase;

[0020] E) Track refinement and re-leveling, recalculating track parameters to prevent the leveling effect from recurring;

[0021] F) Convert the obtained phase to elevation, and finally perform geocoding to convert the slant coordinate system to the geographic coordinate system.

[0022] Compared with existing technologies, the advantages of this invention are as follows: The forest aboveground biomass prediction method of this invention uses Sentinel-1 and Sentinel-2 data to invert forest AGB, and experiments were conducted on two types of tree species. Furthermore, to further verify the performance of the method of this invention, support vector regression, random forest, and BiLSTM models were selected as comparative experiments. The results show that the method proposed in this invention is more advantageous in AGB inversion; the R² of the model in the two types of tree species is 0.8055, RMSEr = 18.1715% and R² = 0.7956, RMSEr = 9.3000%, respectively. Compared with the other three models, the relative improvement rate of RMSE is as high as 43.6%. Attached Figure Description

[0023] Figure 1 Example 1 shows the geographical location and remote sensing images of the study area.

[0024] Figure 2 This is a flowchart for estimating forest AGB in Example 1.

[0025] Figure 3 This is a flowchart of Sentinel-2 data preprocessing and feature extraction.

[0026] Figure 4 This is a flowchart of Sentinel-1 data preprocessing and feature extraction.

[0027] Figure 5 This is a diagram of the BiLSTM neural network structure.

[0028] Figure 6 This describes the map and compass operation modes for PIO, specifically the homing behavior of pigeon flocks.

[0029] Figure 7 This refers to the map and landmark operation mode of PIO, specifically the homing behavior of pigeons.

[0030] Figure 8 This is a structural diagram of the CIOPB framework.

[0031] Figure 9 A heatmap showing the absolute values ​​of the correlation coefficients between remote sensing variables for forest type category-1 and AGB.

[0032] Figure 10 A heatmap showing the absolute values ​​of the correlation coefficients between remote sensing variables for forest type category-2 and AGB.

[0033] Figure 11 This is a histogram of the absolute values ​​of the correlation coefficients between the remote sensing variables of forest type category-1 and AGB.

[0034] Figure 12 This is a histogram of the absolute values ​​of the correlation coefficients between the remote sensing variables of forest type-2 and AGB.

[0035] Figure 13 For category -1 and σ vh Histogram of the absolute values ​​of the correlation coefficients of the combined indices.

[0036] Figure 14 For category -1 and σ vv Histogram of the absolute values ​​of the correlation coefficients of the combined indices.

[0037] Figure 15 This is a histogram of the absolute values ​​of the correlation coefficients of the indices associated with Alpha in category-2.

[0038] Figure 16 This is a histogram of the absolute values ​​of the correlation coefficients of the indices associated with Anisotropy in category-2.

[0039] Figure 17 This is a scatter plot of the measured AGB and the model-predicted AGB of the CIOPB framework proposed in Example 1. Note: Results for five feature groups in two categories. The figure illustrates the different feature groups in the two categories, including (a, b, c, d, e) category-1; (f, g, h, i, j) category-2.

[0040] Figure 18 This is a spatial distribution map of the CIOPB framework in Example 1 obtained throughout the study area.

[0041] Figure 19 An AGB plot of the CIOPB framework proposed in Example 1 and all COSI1 and COSI2 combined variables.

[0042] Figure 20 Comparison of model evaluation metrics across all feature groups using the SVR, RF, BiLSTM, and CIOPB frameworks for Category-1.

[0043] Figure 21 Comparison of model evaluation metrics for Category-2 across all feature groups using SVR, RF, BiLSTM, and CIOPB frameworks. Detailed Implementation

[0044] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0045] It should be noted that when a component is described as being "fixed to, attached to, connected to or connected to" another component, it can be directly fixed to, attached to, connected to or connected to the other component, or it can be indirectly fixed to, attached to, connected to or connected to the other component through other intermediate connectors.

[0046] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the invention.

[0047] Example 1

[0048] like Figure 1 As shown, the study area of ​​this embodiment is located in the southeast of Yueyang City, Hunan Province, China (113°51′52″~113°58′24″E, 28°31′7″~28°38′N), covering a total area of ​​4762 hectares. Its high elevation forms a natural boundary with surrounding towns. Located in the transition zone between the mid-subtropical and north-subtropical zones, the area has a humid continental climate, characterized by warmth and humidity. The average annual temperature ranges from 15.8 to 9.6°C, and the average annual precipitation is 1968.8 mm. The terrain is relatively flat, with higher elevations in the south and lower elevations in the north. Red soil is distributed below 600 meters above sea level, while mountain yellow soil is distributed between 600 and 1200 meters, and mountain yellow-brown soil is found above 1200 meters. Therefore, the site conditions in this study area are favorable. The experimental plot is dominated by evergreen broad-leaved forests, with Chinese fir and hardwoods being the dominant tree species.

[0049] 1. Field AGB Calculation

[0050] The China Forest Resources Planning and Design Survey, also known as the Second Class Forest Resources Survey, is an operational task primarily implemented by state-owned forest farms and other forest management units or county-level administrative regions. This survey provides the most comprehensive data on forest resources within management units, including forest type, dominant tree species, stock volume, and AGB (Australasian Biomass). Its main purpose is to obtain the quantity and quality of forest resources and related natural geographical and ecological environmental factors within the assigned unit. The data for this example area comes from the 2020 China Forest Resources Planning and Design Survey, which includes 300 forest plots covering 1032.4653 hectares. Data types include plot latitude and longitude, dominant tree species, diameter at breast height (DBH), height, and number of trees per hectare. Based on this data, dominant tree species are divided into two categories: Chinese fir and hardwood species. Trees with a height greater than 1.3m and a DBH greater than 5cm are selected, and their height, DBH, forest type, and number of trees per plot are recorded. According to the allometric growth equations for different tree species groups in Table 1, the AGB for each tree species is calculated accordingly, ultimately determining the forest biomass of all plots. Detailed forest information is shown in Table 2.

[0051] Table 1. Formulas for calculating biomass of different tree species.

[0052]

[0053] Note: This represents the total AGB value. Indicates stem biomass. Indicates branch biomass. The values ​​represent leaf biomass, D represents diameter at breast height (cm), and H represents tree height (m).

[0054] Table 2. Detailed information on the sample plots.

[0055]

[0056] 1.1 Acquisition and Preprocessing of Optical Data

[0057] In this embodiment, the optical imagery was sourced from Open Access Hub (copernicus.eu). The Sentinel-2 L1C-level product used was acquired on May 19, 2020, close to the timeframe of the forest resource planning and design survey, resulting in nearly cloudless Sentinel-2 images. Sentinel-2 comprises two satellites, 2A and 2B. This embodiment utilizes the Sentinel-2A satellite, covering 13 operating bands. The revisit time for a single satellite is 10 days, and each image displays a 290 km swath. Since the Sentinel-2 L1C-level product has already undergone orthorectification and geometric correction, only atmospheric correction was performed using the Sen2cor 2.8.0 plugin in the SNAP software. To facilitate subsequent analysis and unify the resolution of all bands across all images, bilinear interpolation was used to resample the generated L2A-level product, achieving a spatial resolution of 10 m.

[0058] 1.2 Acquisition and Preprocessing of SAR Data

[0059] SAR images were acquired from Open Access Hub (copernicus.eu) and EarthExplorer (usgs.gov) on August 30, 2020, and September 11, 2020 (details are shown in Table 3). The Sentinel-1 is equipped with a C-band dual-polarized SAR sensor, employing both vertical transmit-vertical receive (VV) and vertical transmit-horizontal receive (VH) polarization. In forest biomass research, the IW mode is the primary acquisition mode, using progressive scanning topographic observation to capture three sub-regions: IW1, IW2, and IW3, ensuring uniform image quality across the entire area. This embodiment's preprocessing of the Sentinel-1 includes both SLC and GRD products. SLC product preprocessing includes orbit correction, radiometric calibration, multi-look (4 slant range looks and 1 azimuth look), filtering (Refined Lee filter), topographic correction, and geocoding. GRD product preprocessing, in addition to the above, requires thermal noise removal and decibel reduction. The polarization parameters of Sentinel-1 need to be obtained through polarization decomposition, while the interferometric parameters need to be obtained through InSAR technology. These operations are performed in SNAP software.

[0060] Table 3. Parameter information of SAR images of the study area.

[0061]

[0062] 2 Methods

[0063] Figure 2 shows the flowchart for estimating forest AGB in this embodiment, which includes: (1) extracting feature factors; (2) generating COSI; (3) combining feature scenes; (4) constructing an inversion model; and (5) evaluating the model and comparing the results.

[0064] 1.3 Feature Extraction from Optical and SAR Images

[0065] Band operations were performed on the preprocessed Sentinel-2 optical images to obtain vegetation indices. Based on previous research, ten vegetation indices were selected in this paper. Furthermore, biophysical parameters can monitor aspects such as canopy structure, light use efficiency, vegetation nutrient status, and vegetation growth environment, and can further reflect the level of AGB (Ambient Light Gauge Index). This paper used the Biophysical Processor S2 toolbox in SNAP to extract five biophysical parameters. Figure 3 The flowchart illustrates the process of optical feature extraction from Sentinel-2. Feature information from Sentinel-2 images is shown in Table 4.

[0066] Table 4 Feature extraction and calculation of Sentinel-2 images.

[0067]

[0068] Previous studies have shown that backscattering coefficients and polarization decomposition characteristics have good potential for AGB inversion. This embodiment uses the H-Alpha Dual Pol method to perform polarization decomposition on SLC products, obtaining three polarization parameters: Entropy (H), Anisotropy (A), and Alpha angle (alpha). The incident angle and two backscattering coefficients σ are then extracted from the GRD product. vh , σ vv In addition, the interferometric parameters in the Sentinel-1 image also have a good effect on AGB inversion. The interferometric parameters are mainly obtained through InSAR technology: (1) Select images and register them. It is worth noting that an excessively large baseline will reduce the quality of the interferometry. Therefore, it is necessary to select images with a relatively close time interval. In this embodiment, the Sentinel-1 image on August 30, 2020 is selected as the main image and the Sentinel-1 image on September 11 is selected as the secondary image. In addition, in order to eliminate the error between the two images and correct the offset between the pixels, it is necessary to register the images. (2) In the process of generating the interferogram, the interferometric fringes may be too dense, especially in flat areas, which is called the flat-ground effect. The flat-ground effect will cause great difficulties for subsequent processing and the influence of this effect needs to be eliminated. (3) Coherence coefficient calculation and filtering. In this experiment, a window size of 3×3 is set, the maximum likelihood method is selected for coherence coefficient calculation, and Goldstein Phase Filtering is used for filtering. (4) Phase unwrapping. The blurred phase is added to the phase principal value to obtain the true phase. (5) Track refinement and re-leveling were performed, and track parameters were recalculated to prevent the leveling effect from recurring. (6) The obtained phase was converted to elevation, and finally geocoding was performed to convert the slant distance coordinate system to the geographic coordinate system. This paper extracted two parameters, phase and coherence coefficient, and obtained a digital elevation model (DEM) for the study area. Figure 4 The processing flow for SAR feature extraction is shown in Table 5. Parameter information from the Sentinel-1 image is shown in Table 5.

[0069] Table 5. Features extracted from Sentinel-1 images.

[0070]

[0071] 1.4 Combining Optical and SAR Indicator System (COSI)

[0072] Optical remote sensing provides rich spectral information through vegetation indices and biophysical variables, while SAR can penetrate the forest canopy to obtain the dielectric properties and vertical structure of the forest. Combining the two has a very positive significance for forest aboveground biomass inversion. Inspired by the combination of optical remote sensing and LiDAR variables, this embodiment improves upon this by proposing a combined optical and SAR index system (COSI). Specifically, two SAR variables with the best performance (the two SAR variables with the highest AGB correlation coefficient) are selected and combined with optical variables with high correlation (i.e., vegetation indices and biophysical parameters with an absolute correlation value greater than 0.1) to obtain COSI, which includes two types of indices (COSI1 and COSI2). This can be written as:

[0073] (1)

[0074] (2)

[0075] in This refers to the two SAR variables with the highest correlation. These are vegetation indices or biophysical parameters with an absolute correlation value greater than 0.1. It is worth noting that before calculating COSI2, it is necessary to... and Normalization is performed to ensure that the two variables can be directly compared on the same scale and range.

[0076] 1.5 Construction of the CIOPB Framework

[0077] The BiLSTM neural network model and the PIO algorithm are important components of the CIOPB framework. Figure 5 The diagram shows the neural network structure of BiLSTM. BiLSTM is an extension of Long Short-Term Memory (LSTM) neural networks, consisting of a forward LSTM and a backward LSTM, and has been widely used in natural language processing, speech recognition, and other fields. However, BiLSTM also suffers from difficulties in parameter selection and is prone to overfitting. This limitation can be addressed using swarm intelligence algorithms. PIO is an emerging intelligent optimization algorithm inspired by the homing behavior of pigeons in nature. Figure 6 This describes the map and compass operation modes for PIO, specifically the homing behavior of pigeon flocks. Figure 7 This refers to the map and landmark operation mode of PIO, specifically the homing behavior of pigeons. Figure 6 The gray pigeon on the far right represents the best flight direction. The thin arrows in the grid represent the previous flight direction of each pigeon, and the thick arrows indicate the adjusted flight direction. Figure 7In the diagram, the central blue circle represents the destination for the flock's iterations. Approximately half of the flock outside the circle will follow the pigeons inside, who will quickly reach the destination. For example... Figure 8 As shown, the CIOPB framework consists of two aspects: COSI feature optimization and the construction of the inversion model. In the design of feature optimization, we combined the optical and SAR variables that met the requirements in the Sentinel image to obtain a new COSI index. Then, we set five different feature combination scenarios for the newly obtained COSI index and the original remote sensing image variables. Table 6 shows the detailed grouping information. Combination 1: All variables extracted from the optical and SAR images; Combination 2: All COSI1 calculated using formula (1); Combination 3: All COSI2 calculated using formula (2); Combination 4: All COSI1 and COSI2 variables; Combination 5: All variables in combination 1 and combination 4. In the construction of the inversion model, we used the PIO iteration process to optimize the parameters of the BiLSTM neural network model. First, our experiment set three parameters to be optimized: initial learning rate, L2 regularization coefficient, and number of hidden layer nodes. In this model, the number of hidden layer nodes is an embedded parameter of the BiLSTM layer, which is the core of the entire neural network framework. All input features are processed by the BiLSTM layer, and the network structure of the BiLSTM layer needs to be optimized by obtaining the optimal solution for the number of hidden layer nodes. After repeated updates and iterations, the optimal solutions for the three parameters were obtained. Then, the number of hidden layer nodes is linked to the BiLSTM layer, and the other optimal parameters are automatically input into the model. Finally, the input variables are connected to the regression layer after passing through several layers of the BiLSTM network to obtain the predicted value of AGB. As can be seen from the above process, the CIOPB framework effectively combines the feature information of optical and SAR images, solving the difficulty in model parameter selection. In addition, the framework can adjust its own parameters according to the characteristics of the input variables, further curbing the occurrence of overfitting. In this embodiment, 300 sample plots are divided into training and test sets in a 3:1 ratio, and the total number of validation samples for the two tree species is 74. In order to further quantitatively analyze the inversion results of the regression model, the coefficient of determination (R2), root mean square error (RMSE), and relative root mean square error (RMSE) are selected. r The accuracy and applicability of the model are evaluated using the following formula:

[0078] (3)

[0079] (4)

[0080] (5)

[0081] in, The measured value representing AGB. The estimated value representing AGB, This represents the average value of the measured AGB. This represents the number of sample plots.

[0082]

[0083] Table 6 shows five different feature combinations used in the AGB model.

[0084] 2 Results

[0085] 2.1 Relationship between feature variables and AGB

[0086] Based on the previous feature optimization analysis, the AGB correlation of all extracted remote sensing features was calculated using the Spearman method to determine the optical and synthetic aperture radar variables used for COSI. The results are shown in Figure 9. Figure 10 , Figure 11 and Figure 12 As shown. In category-1, σ vh σ vv The correlation with the ten vegetation indices was significantly higher than that with other variables. This indicates that in Category 1, the backscattering coefficient and vegetation indices are extremely sensitive to AGB. Furthermore, Alpha, Anisotropy, Entropy, DEM, LAI, and FVC showed relatively high correlations with AGB (absolute values ​​of correlation coefficients greater than 0.1). In Category 2, the correlations with Alpha, Anisotropy, and Entropy increased significantly, while σ... vh and σ vv The correlation between the first and second categories of vegetation indices decreased, while the correlation between the ten vegetation indices remained very high. These results indicate that both types of vegetation indices are highly sensitive to forest AGB. Furthermore, the backscattering coefficient σ in the first category... vh σ vv It is highly correlated with the polarization decomposition parameters in Class 2.

[0087] Based on the correlation analysis above, detailed variables for the two types of COSI were determined. In type 1, σ vh and σ vv These are the two variables with the highest correlation to AGB in synthetic aperture radar. In terms of optical variables, the 10 vegetation indices undoubtedly meet the requirements. Furthermore, given the physical significance of LAI and FVC among biophysical variables, and their relatively high sensitivity to AGB, these two variables are also considered qualified variables. Using formulas (1) and (2), σ... vh and σvv Combining with each optical variable, COSI1 and COSI2 were generated respectively. In the second category, Alpha and Anisotropy from the synthetic aperture radar showed the highest correlation with AGB, while only 10 vegetation indices among the optical variables had high correlation; therefore, Alpha and Anisotropy were combined with each of the 10 vegetation indices. Table 7 shows all the generated new indices. Category 1 generated 48 new indices, and Category 2 generated 40 new indices, which together constitute the COSI system.

[0088] Table 7 Newly generated COSI indicators.

[0089]

[0090] Note: The 12 and 10 in parentheses represent the number of newly generated indicators.

[0091] To conduct a more detailed analysis, a correlation analysis was also performed on the newly generated indices. Figure 13 As shown in Figure 14, the overall AGB correlation of COSI1 and COSI2 remains high in category-1, and COSI2 is more sensitive to AGB than COSI1. Furthermore, the AGB correlation of most of the new indices is improved compared to the original image features. However, as... Figure 15 As shown in Figure 16, different phenomena occurred in category 2. Figure 15 In the new indices combined with alpha, COSI2 achieved higher AGB correlation, while COSI1 remained at a relatively low level. Furthermore, compared to the original image features, COSI2 showed improved correlation, while COSI1 showed decreased correlation. Interestingly, in Figure 16, the new indices combined with anisotropy exhibited opposite performance. In summary, except for a few indices in Class 2, the newly generated indices showed significantly improved AGB correlation compared to the original image features, demonstrating the potential of the COSI system for AGB inversion.

[0092] Based on the feature combination settings in Table 6, the generated new index is combined with the initial image features to determine the number of variables for each feature group to be input into the inversion model (Table 8). By setting the combination scheme, the feature groups that contribute the most to the forest AGB inversion can be determined according to the model's performance.

[0093] Table 8. Variables / indices in each feature group.

[0094]

[0095] 2.2 Model Performance of Different Feature Groups under the CIOPB Framework

[0096] To further verify that the COSI index system derived using the CIOPB framework can improve the inversion accuracy of AGB, five feature groups were input into the inversion model within the CIOPB framework (Table 9). Compared with the initial variable (combination 1), the COSI groups in both categories (combination 2, combination 3, combination 4, and combination 5) significantly improved the accuracy of AGB estimation. For the first category, compared with combination 1, combination 4 had a higher R-value. 2 The RMSE decreased by 29.4% (from 17.4063 Mg / ha to 12.2876 Mg / ha) as the concentration increased from 0.6079 to 0.8055. r It decreased by 7.6% (from 25.7413% to 18.1715%). Furthermore, the accuracy of combination 3 was second only to combination 4 (R... 2 =0.8042, RMSE=12.3277 Mg / ha, RMSE r =18.2308%). For the second category, compared with combination 1, the R of combination 5 is... 2 The RMSE decreased by 34.1% (from 3.4727 Mg / ha to 2.2887 Mg / ha) as the concentration increased from 0.5293 to 0.7956. r The accuracy decreased by 4.8% (from 14.1109% to 9.3000%). Therefore, combinations 4 and 5 are the best feature groups for classes 1 and 2, respectively. Furthermore, combinations 2 and 3 also performed well. These results demonstrate that the COSI index system significantly improves the inversion accuracy of AGB, validating the application potential of COSI in AGB inversion. Interestingly, the RMSE and RMSE of class 2... r The results are significantly lower than those of the first category. This phenomenon may be related to the distribution of AGB and forest age classes: the AGB distribution of category-1 has a large range, ranging from 20.57 Mg / ha to 141.56 Mg / ha, indicating a more complex forest structure and age class composition; while the AGB distribution of category-2 has a small range and standard deviation, indicating a more uniform forest structure and age class composition. This shows that the proposed CIOPB framework can achieve good inversion results in two tree species with completely different structures and compositions, further demonstrating the applicability of this method in AGB inversion.

[0097] Table 9 shows the performance of the five feature groups under the proposed CIOPB framework.

[0098]

[0099] To better demonstrate the performance of the CIOPB framework in AGB inversion, such as Figure 17 As shown, this embodiment plots a scatter plot of AGB values. The plot depicts different feature groups in two categories: category-1 (a, b, c, d, e) and category-2 (f, g, h, i, j). The 1:1 line reflects the equivalence between the forest AGB predictions and observations. Clearly, the COSI four groups show significantly better fits than combination 1, with combination 4 showing the best fit for the first category and combination 5 showing the best fit for the second. In their respective scatter plots, the trend lines for these two groups deviate least from the 1:1 line. Furthermore, combination 3 in category 1 and combination 2 in category 2 also show commendable fits. However, combination 1 shows the worst fit in both categories, with large deviations between predictions and observations at most points. This indicates the limitations of using the initial image features of combination 1 as input variables for the inversion model. In conclusion, after feature optimization, the errors between AGB predictions and observations are reduced, further demonstrating that COSI is more effective than the initial image features in AGB inversion.

[0100] 2.3 Comparison of Model Performance of Different Methods in AGB Prediction

[0101] To better verify the performance of the proposed CIOPB framework, comparative experiments were conducted by inputting the optimal feature set into RF, SVR, and BiLSTM models. See Table 10 and... Figure 18 As shown, the COSI framework achieves significantly higher AGB inversion accuracy than the other three models, and the predicted AGB shows greater consistency with the observed AGB. In both categories, compared to the other three models, R... 2 The RMSE increased by 0.1648 to 0.3324, while the RMSE decreased by 26.4% to 43.6%. r The accuracy decreased by 3.5% to 10.3%. Among the basic BiLSTM models (without PIO iterative optimization), it ranked second in accuracy, only behind the CIOPB framework. However, the BiLSTM model encountered challenges in parameter selection and overfitting during training. Using the parameters obtained from the CIOPB framework and fine-tuning them improved the inversion accuracy (Class-1: R). 2 =0.6407, RMSE=16.6998 Mg / ha, RMSE r =24.6965%; Category-2: R 2=0.6154, RMSE=3.1392 Mg / ha, RMSE r =12.7558%. The scatter plot of the BiLSTM model shows that some points in the second class fit well, but the BiLSTM model produces excessive errors in estimating high and low AGB values. These errors weaken the overall performance of the model. Although the CIOPB framework also produces some overestimations and underestimations, the errors are relatively small. In addition, the SVR model performs poorly, while the RF model performs slightly better. The scatter plots of the RF model and the SVR model show similarity, leading to a large bias. Therefore, there are significant differences in the modeling performance of the four methods. The above results demonstrate the advantages of the CIOPB framework and further validate its potential in AGB inversion.

[0102] Table 10 shows the performance of SVR, RF, BiLSTM, and CIOPB methods in both categories.

[0103]

[0104] like Figure 18 The figure shows scatter plots of AGB observed using four methods versus AGB predicted by the models. The figure illustrates different inversion models for the two categories, including (a, e) the CIOPB framework; (b, f) BiLSTM; (c, g) RF; and (d, h) SVR. Note: Results for the best feature set in each category.

[0105] 2.4 Comparison of Model Performance of Different Methods in AGB Prediction

[0106] Figure 19 The spatial distribution map of the CIOPB framework across the entire study area is shown.

[0107] The total biomass of the study area was 5.25 × 10⁵ Mg, with an average of 110.35 Mg / ha. AGB values ​​below 0 Mg / ha were considered to underestimate biomass, and these areas were non-forested; therefore, these values ​​were uniformly labeled as 0 Mg / ha during mapping.

[0108] 3. Discussion

[0109] In the RF, SVR, and BiLSTM models, in addition to applying the optimal feature set, other feature sets were also applied to invert AGB. The results of all methods were compared together, and a detailed analysis was conducted from both feature optimization and inversion model perspectives (Table 11).

[0110] Table 11 shows the performance of the four models under the five feature groups.

[0111]

[0112] In feature optimization, all four COSI groups outperformed the initial image features of Combination 1, therefore a quantitative analysis was performed by comparing the enhancement effects of the COSI groups. In the first class, compared with Combination 1, the R of the COSI group was significantly higher. 2 The RMSE increased by 0.118 to 0.196, while the RMSE decreased by 10.4% to 29.4%. r The R2 decreased by 3.3% to 7.6%. In the second category, the R2 in the COSI group increased by 0.106 to 0.266, and the RMSE decreased by 12.1% to 34.1%. r The accuracy decreased by 2.3% to 4.8%. Furthermore, among all methods, the fourth and fifth groups remained the best feature groups for the first and second classes, respectively. This indicates that COSI improves the inversion accuracy of each model, rather than relying solely on a single inversion model, further validating the potential of the proposed COSI index system in AGB inversion. In previous studies, Zhang et al. (Zhang, L., Z. Shao, J. Liu, and Q. Cheng. 2019. Deep Learning Based Retrieval of Forest Aboveground Biomass from Combined LiDAR and Landsat 8 Data. Remote Sensing11, no. 12 (June 20): 1459.) designed a combined optical and LiDAR index, finding that applying the combined index performed better than applying either optical or LiDAR data alone in each combination. Li et al. established a multiplicative combined optical and LiDAR model, successfully inverting the biophysical parameters of maize. However, these studies have rarely focused on the combination of optics and synthetic aperture radar, providing a window for our research. In the inversion model, Figure 20 and Figure 21 The model performance of all methods is shown in Figure 20 and... Figure 21 As can be seen from R 2 The values ​​gradually increase from left to right, RMSE and RMSE. rThe performance gradually decreases. Furthermore, regardless of the feature group, the modeling metrics of the CIOPB framework are significantly superior to other models. This demonstrates the advantages of combining intelligent optimization algorithms with DL models. In fact, many studies have applied DL models to estimate forest AGB. Ghosh and Behera compared the performance of the Interferometric Water Cloud Model (IWCM) and the DL model in mangrove AGB inversion, arguing that estimating mangrove AGB using sentinel data requires a greater emphasis on DL models rather than semi-empirical models. Zhang explored the potential of estimating carbon density based on Convolutional Neural Networks (CNNs) and Keras, finding that CNN models outperform traditional estimation models. However, few studies have focused on the inherent limitations of DL models. In this embodiment, the PIO algorithm, with its global optimization capability and strong adaptability, solves the problems of parameter selection difficulty and overfitting in the BiLSTM model, making the model more robust. The BiLSTM model performs second best. In this embodiment, the DL model outperforms the machine learning model. This difference can be attributed to the inherent multi-layer neural network structure of the DL model, which enables it to systematically learn features and patterns in the data. This model offers significant potential for improving the predictive accuracy of Forest AGB. Among the two machine learning models, the RF model constructs a large number of decision trees by randomly sampling and replacing features, and its satisfactory performance has been confirmed in numerous studies. However, it's important to note that in regression problems, RF algorithms cannot provide continuous outputs, and they may fail to make predictions beyond the training set. Compared to other models, the SVR model performs worse. This may be because the model heavily relies on the choice of kernel function. If the high-dimensional mapping of the kernel function cannot provide sufficient explanatory power, the model's performance may be affected.

[0113] In this embodiment, intelligent optimization algorithms are methods that utilize the concept of collective intelligence in nature to simulate the behavior of natural groups, thereby discovering optimal solutions. These studies indicate that incorporating intelligent algorithms into the construction of AGB inversion models can potentially yield more reliable estimation results. However, in forest AGB inversion, few studies have applied intelligent optimization algorithms to address the inherent problems of deep learning models.

[0114] Therefore, to estimate the AGB of subtropical forests more accurately and efficiently, we propose an optical and synthetic aperture radar composite index (COSI) optimization inversion framework based on the PIO iterative procedure and a BiLSTM neural network (CIOPB). This framework consists of two core parts: feature optimization and model building. Feature optimization is achieved by developing a novel metric system, COSI. The development process utilizes and improves the combined index method of optical and lidar to be applicable to the combination of optical and synthetic aperture radar. Subsequently, COSI and original image features are divided into five combination schemes. This allows the most suitable feature set to be obtained through the combined index, rather than simply relying on feature importance ranking. When building the model, considering the Pigeon Heuristic Optimization (PIO) algorithm's reputation for its simple and straightforward principles and computational ease, a PIO iterative process is embedded in the BiLSTM model. This enables BiLSTM to obtain the desired optimal parameters through the PIO iterative process, thereby eliminating the difficulties of parameter selection and overfitting. Furthermore, to evaluate the effectiveness of the CIOPB framework, this embodiment also employed Support Vector Regression (SVR), Randomization (RF), and a basic BiLSTM model (unoptimized) for experiments. The above analysis shows that the CIOPB framework cannot achieve good results relying on a single component. Optimizing only the feature variables without processing the model itself results in limited inversion accuracy. Correspondingly, optimizing only the model without processing the input feature variables also fails to yield good inversion results. Combining both parts yields better results. In future research, we will validate with more tree species and consider using intelligent optimization algorithms to build more deep learning models to address the shortcomings of our current work.

Claims

1. A method for predicting forest aboveground biomass, characterized in that: Includes the following steps; 1) Acquisition and preprocessing of optical data; acquiring Sentinel-2 optical images of the target location via satellite, and then correcting and unifying the resolution; 2) Preprocess the Sentinel-1 SAR data of the target location, including orbit correction, radiometric calibration, multi-look, filtering, terrain correction and geocoding, thermal noise removal and decibel reduction; and obtain polarization parameters through polarization decomposition and interferometric parameters through InSAR technology. 3) Feature extraction from optical and SAR images; extract vegetation indices and biophysical parameters from preprocessed Sentinel-2 optical images, and extract SAR variables from preprocessed Sentinel-1 images; the vegetation indices include ARVI, EVI, MSAVI, MSR, NDVI, OSAVI, PVI, RVI, SAVI, and VARI; the biophysical parameters include LAI, FVC, CAB, CW, and FAPAR; the SAR variables include three polarization parameters, incident angle, two backscattering coefficients, phase, coherence coefficient, and digital elevation model; 4) Design an index system COSI that combines optical and SAR, including: selecting the two SAR variables with the highest AGB correlation coefficient, and vegetation indices and biophysical parameters with an absolute correlation value greater than 0.1, and combining them to obtain COSI, which contains two types of indicators, COSI1 and COSI2, as follows: ； in, This refers to the two SAR variables with the highest correlation coefficient. These are vegetation indices or biophysical parameters with an absolute correlation value greater than 0.1; before calculating COSI2, [the following is considered]... and Perform normalization processing; The COSI feature optimization combination includes: combination 1 is all SAR variables, vegetation index and biophysical parameters; combination 2 is the COSI1 index calculated using formula (1); combination 3 is the COSI2 index calculated using formula (2); combination 4 is the COSI1 and COSI2 indices; combination 5 is all variables in combination 1 and combination 4. 5) Construct the CIOPB framework, which consists of COSI feature optimization combination and an inversion model. The COSI feature optimization combination is used as the input to the inversion model. In the construction of the inversion model, a BiLSTM neural network model and the PIO algorithm are adopted. The parameters of the BiLSTM neural network model are optimized by the PIO iteration process. The parameters include the initial learning rate, L2 regularization coefficient, and the number of hidden layer nodes. The number of hidden layer nodes is an embedded parameter of the BiLSTM layer. The BiLSTM layer is the core of the entire neural network framework. All input features will be calculated by the BiLSTM layer. The network structure of the BiLSTM layer needs to be optimized by obtaining the optimal solution of the number of hidden layer nodes. After updating and iterating, the optimal solutions of the three parameters are obtained. The optimal parameters of the number of hidden layer nodes, initial learning rate, and L2 regularization coefficient are input into the BiLSTM neural network model. The input variables are connected to the regression layer after passing through several network layers of BiLSTM to obtain the predicted value of aboveground biomass.

2. The method for predicting forest aboveground biomass according to claim 1, characterized in that: The correction of the optical image Sentinel-2 includes orthorectification, geometric correction, and atmospheric correction.

3. The method for predicting forest aboveground biomass according to claim 1, characterized in that: The InSAR technique for obtaining interferometric parameters includes the following steps: A) Select the image and perform registration; B) Eliminate the leveling effect; C) The maximum likelihood method is selected for coherence coefficient calculation, and Goldstein Phase Filtering is used for filtering; D) Phase unwrapping: Adding the blurred phase to the principal phase value to obtain the true phase; E) Track refinement and re-leveling, recalculating track parameters to prevent the leveling effect from recurring; F) Convert the obtained phase to elevation, and finally perform geocoding to convert the slant coordinate system to the geographic coordinate system.

Images