Forest leaf area index estimation method based on three-dimensional radiation transmission model and deep learning

Abstract

The invention discloses a forest leaf area index estimation method based on a three-dimensional radiation transmission model and deep learning, and belongs to the technical field of ecological remote sensing. The method comprises the steps of 1, obtaining a modeling data set based on a DART model, and preprocessing the modeling data set to obtain a modeling training set; 2, constructing a pre-trained 1D-CNN forest LAI inversion model, and carrying out the training optimization of the pre-trained 1D-CNN forest LAI inversion model based on the modeling training set, and obtaining an optimization model; 3, constructing a migration data set based on Landsat satellite image data and actually measured LAI data, and performing parameter optimization on the optimization model based on the migration data set to obtain a migration learning forest LAI inversion model; and step 4, processing the Landsat image data needing to be predicted based on the transfer learning forest LAI inversion model to obtain an estimation result. According to the invention, a forest LAI estimation result with higher precision can be provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of ecological remote sensing, and particularly relates to a method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning. Background Art

[0002] Vegetation is an important part of the Earth system, and realizes the exchange of matter and energy between the soil-vegetation-atmosphere through biophysical processes such as photosynthesis, respiration and transpiration. Forests are the main body of the terrestrial ecosystem, covering about 30% of the land area. Compared with other ecosystems, they are the ecosystems with the largest surface area, the highest productivity and the richest material resources on the land surface, and are also important carbon sinks on the earth's surface. With the in-depth study of global change and the establishment of models such as forest carbon cycle, the forest leaf area index (LAI) plays a key role in global climate and carbon cycle research as an important input factor. The main way to measure LAI on the ground is by means of optical instruments, which are limited to small-area applications and are easily affected by the environment. Remote sensing technology has the ability of large-scale synchronous continuous observation and has now become the main way to quickly obtain forest LAI.

[0003] From the perspective of data sources, remote sensing inversion of forest LAI can be divided into three categories: Light Detection and Ranging (LiDAR), Synthetic Aperture Radar (SAR), and optical remote sensing. Among them, the LiDAR method obtains the three-dimensional structural information of vegetation by emitting laser pulses and receiving the echo signals returned from the ground objects, and calculates the gap fraction through the Beer-Lambert law to further obtain the forest LAI. However, the data processing process of this technology is complex and is easily affected by the heterogeneity of forest structure; the SAR method for inverting forest LAI analyzes the indirect relationship between the radar backscattering signals at different polarizations and bands and the forest structure, and with the help of empirical models, physical scattering models (such as the water cloud model) or machine learning algorithms, converts the influence of the forest canopy on the radar wave into an estimated value of LAI. However, the SAR signal is not highly sensitive to LAI, is easily affected by multiple factors such as canopy structure and canopy moisture, and the established fitting relationship often depends on samples in specific regions; optical remote sensing methods mainly include data-driven empirical methods, model-driven physical methods, and hybrid methods that combine the two. Among them, the empirical inversion method realizes efficient estimation by constructing the regression relationship between vegetation indices or spectral reflectance and forest LAI, and has the advantages of simple implementation and rapid response. However, it lacks the modeling support for remote sensing physical processes and is difficult to meet the general requirements under different ecosystems; in contrast, the physical model method theoretically has stronger regional applicability. Usually, it generates simulated data with the help of remote sensing radiation transfer models, and then constructs a forest LAI estimation model through iterative optimization or look-up table technology. Although the method has physical consistency, there are still obvious bottlenecks in terms of computational efficiency and estimation accuracy; the hybrid inversion strategy combines the advantages of the high efficiency of the empirical method, introduces the theoretical support of the physical model, and can integrate the advantages of artificial intelligence algorithms in feature extraction and nonlinear modeling, showing better stability, accuracy, and adaptability, and becoming an effective path to improve the accuracy and efficiency of forest LAI inversion.

[0004] With the rapid development of computer technology and artificial intelligence methods, the use of hybrid methods to carry out remote sensing inversion of forest LAI has become increasingly widespread, but existing research still has certain limitations. On the one hand, physical models usually use one-dimensional models based on the assumption of turbid media, geometric optical models based on simple geometric bodies (such as assuming that the forest canopy is a cylinder or cone), and computer simulation models that can characterize the three-dimensional real canopy structure of the forest. Among them, the one-dimensional model assumes that the leaves are ideal Lambertian surfaces, and the canopy is a randomly distributed, horizontally uniform, and vertically layered turbid medium, which is mainly suitable for homogeneous vegetation; the geometric optical model takes into account the geometric characteristics of the forest canopy and the bidirectional reflectance of the canopy, calculates the proportion of illuminated canopy, shaded canopy, illuminated soil and shaded soil, and then obtains the simulated canopy reflectance; the three-dimensional computer model accurately describes the three-dimensional canopy structure and spatial distribution information of the forest, more realistically reflects the radiation transmission process, and obtains more accurate simulation data. Existing hybrid inversion studies of forest LAI usually use one-dimensional radiation transfer models or geometric optical models to obtain simulation data, which makes it difficult for the simplified vegetation canopy structure assumption in the model to fully describe the actual complex structural characteristics of the forest, resulting in errors in simulated reflectivity, which further affects the inversion of forest LAI; on the other hand, existing studies usually use machine learning such as support vector machines, artificial neural networks, and random forests as inversion methods, and the constructed inversion relationship is completely dependent on the radiation transfer model simulation data set, lacking the guidance and optimization of remote sensing prior knowledge, resulting in the limited accuracy of the constructed inversion model and prone to prediction results that do not conform to physical common sense. Compared with commonly used machine learning algorithms, deep learning algorithms have a stronger ability to construct complex nonlinear regression relationships, and the field of deep learning has a variety of auxiliary means to facilitate the introduction of prior knowledge. Summary of the invention

[0005] In view of the problems in the prior art that the expression of the complex three-dimensional spatial structure of forests is insufficient, the regression relationship process fitting is inaccurate, and the training process lacks the guidance of prior knowledge, a new solution is proposed. The three-dimensional canopy structure of the forest is constructed through the three-dimensional radiative transfer model DART, and the one-dimensional convolutional neural network (1D-CNN) deep learning algorithm with physical constraints is used to construct the forest LAI inversion model, and the observational knowledge guidance is introduced through transfer learning. In addition, the algorithm design is carried out for Landsat remote sensing data, so that the model can generate long-term forest LAI time series products with a spatial resolution of 30 m, significantly improving the spatio-temporal scale applicability of the forest LAI inversion algorithm. A key breakthrough is achieved in the modeling of the complex three-dimensional structure of the forest canopy and the construction of the LAI inversion model on the basis of the existing methods, which can provide more accurate forest LAI estimation results for the monitoring and modeling of key processes such as the dynamic evolution of forest ecosystems and terrestrial carbon cycles, and has an important promoting effect on improving the accuracy, reliability and practical application level of forest LAI remote sensing inversion.

[0006] To achieve the above object, the present invention provides the following solution: A method for estimating forest leaf area index based on a three-dimensional radiative transfer model and deep learning, comprising the following steps:

[0007] Step 1: Obtain a modeling data set based on the DART model, and preprocess the modeling data set to obtain a modeling training set;

[0008] Step 2: Construct a pre-trained 1D-CNN forest LAI inversion model, and train and optimize the pre-trained 1D-CNN forest LAI inversion model based on the modeling training set to obtain an optimized model;

[0009] Step 3: Construct a transfer data set based on Landsat satellite image data and measured LAI data, and optimize the parameters of the optimized model based on the transfer data set to obtain a transfer learning forest LAI inversion model;

[0010] Step 4: Process the Landsat image data to be predicted based on the transfer learning forest LAI inversion model to obtain an estimation result.

[0011] Further preferably, Step 1 includes:

[0012] S11. Construct a three-dimensional model of a single tree and a three-dimensional model of understory vegetation of a single plant based on measured data, and construct a three-dimensional forest scene based on the three-dimensional model of the single tree and the three-dimensional model of the understory vegetation of the single plant to obtain the modeling data set;

[0013] S12. Analyze the geometric features of the leaf elements in the three-dimensional forest scene to obtain the scene LAI of the three-dimensional forest scene;

[0014] S13. Obtain the spectral components, radiance information, and forest canopy reflectance of the scene elements in the three-dimensional forest scene to obtain the modeling training set.

[0015] Further preferably, the spectral components of the scene elements include: leaf spectral characteristics and soil spectral characteristics;

[0016] The method for obtaining the leaf spectral characteristics includes:

[0017] ,

[0018] The method for obtaining the soil spectral characteristics includes:

[0019] ,

[0020] where R(λ) and T(λ) respectively represent the reflectance and transmittance of the leaf at wavelength λ simulated by the PROSPECT-D model; N, C ab , C ar , C w , C m , C brown and λ respectively represent the leaf structure parameters, chlorophyll content, carotenoid content, leaf equivalent water thickness, dry matter content, brown matter content, and wavelength input into the PROSPECT-D model; R soil (λ) represents the soil reflectance; , , are three dry soil components; c sm represents the wet soil component.

[0021] Further preferably, the method for obtaining the forest canopy reflectance includes:

[0022] ,

[0023] where, ,

[0024] where R canopy (band i ) represents the reflectance value corresponding to the i-th band of the Landsat sensor after conversion by the spectral response function; R canopy (λ j ) represents the forest canopy reflectance value simulated by the DART model at wavelength λ j ; R canopy represents the forest canopy reflectance value simulated by the DART model; SRF(λj ) represents the spectral response function value of the Landsat sensor at wavelength λ j ; b represents the total number of wavelength sampling points; λ represents wavelength; θ s and are the solar zenith angle and azimuth angle respectively; θ v and are the sensor's observation zenith angle and azimuth angle respectively; veg represents the vegetation type; LAI is the forest leaf area index extracted based on the 3D scene; R is the leaf reflectance; T is the leaf transmittance; R soil is the soil reflectance; R other represents the reflectance of branches and understory vegetation.

[0025] Further preferably, the expression of the pre-trained 1D-CNN forest LAI inversion model includes:

[0026] LAI out = bias + weight ★ input,

[0027] In the formula, LAI out represents the LAI result calculated by the 1D-CNN network for the input feature values; input is the input feature value; weight represents the convolution kernel for performing cross-correlation operation with the input feature value; bias represents the bias term added after the convolution operation; ★ represents the valid cross-correlation operator.

[0028] Further preferably, the loss function of the transfer learning forest LAI inversion model includes:

[0029] ,

[0030] where,

[0031] ,

[0032] ,

[0033] ,

[0034] ,

[0035] In the formula, Loss w represents the loss value of the transfer learning forest LAI inversion model; α is the proportion of the MSE error in the quantization of the loss value, and (1 - α) represents the proportion of the KL error; β represents the coefficient of additional penalty for the out-of-bounds situation where the model predicts LAI less than 0 and greater than 10; B low and B uprespectively represent the penalty losses calculated by the boundary limit where the prediction result is less than 0 and the boundary limit where the prediction result is greater than 10; n represents the total number of data applied to model verification; MSE w represents the MSE loss weight after weighted calculation; KL represents the relative entropy loss calculated based on the reference LAI and the output LAI; m p in m represents the amount of data in the current interval among the statistical forest LAI; p and q are both parameter coefficients set to strengthen the quantitative characteristics; y i represents the reference LAI value of the i-th training data; represents the model output LAI value of the i-th training data; represents the model output LAI value; Relu represents a non-linear function in deep learning; LAI min and LAI max respectively represent the minimum value of LAI and the maximum value of LAI, with values of 0 and 10 respectively.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] The present invention uses the three-dimensional radiative transfer model DART to generate simulated datasets. Compared with the one-dimensional radiative transfer model and the geometric optical model, it describes the three-dimensional structure of the forest canopy more accurately and is more suitable for forest scenarios with strong heterogeneity; compared with the traditional machine learning inversion scheme, the forest LAI inversion model construction scheme trained by the 1D-CNN algorithm combined with transfer learning used in the present invention can introduce observational knowledge to assist the modeling process, enhance the model's fitting ability for complex non-linear regression relationships, and help improve the model accuracy; aiming at satellite image data with high spatial resolution as the application target, the generated LAI products also have high spatial resolution. Compared with the coarse spatial resolution algorithm, it reduces the uncertainty caused by the mixed pixel problem in the algorithm, greatly guarantees the accuracy of LAI estimation, and at the same time, due to the time span of the Landsat series data, it has the potential to produce long-term forest LAI products. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0039] Figure 1 FIG. is the technical roadmap of the forest leaf area index estimation method based on the three-dimensional radiative transfer model and deep learning in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0041] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Embodiment 1:

[0043] As Figure 1 shown, this embodiment provides a method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning, including the following steps:

[0044] Step 1: Obtain a modeling data set based on the DART model and preprocess the modeling data set to obtain a modeling training set.

[0045] Further implementation lies in that Step 1 includes:

[0046] S11. Construct a three-dimensional model of a single tree and a three-dimensional model of understory vegetation of a single plant based on measured data, and construct a three-dimensional forest scene based on the three-dimensional model of the single tree and the three-dimensional model of the understory vegetation of the single plant to obtain the modeling data set.

[0047] Specifically, this embodiment generates a modeling data set through the DART model, mainly including three steps: scene construction, spectral setting and LAI extraction of each component of the scene elements, and simulation of the radiation transfer process and output of reflectance results.

[0048] Among them, the construction of the three-dimensional forest scene based on the real structure is mainly based on single trees, taking into account the expression of the understory vegetation structure information, and generating scenes for three types of coniferous forests, broad-leaved forests, and coniferous and broad-leaved mixed forests through different combinations and spatial distributions of single trees and understory vegetation.

[0049] Generation of 3D models of individual trees: First, use OnyxTREE software to select the corresponding tree type model libraries (such as coniferous trees and broad-leaved trees). Then, according to different tree types, conduct statistical analysis on all the individual tree structures measured on the ground (such as tree height, height to the lowest live branch, crown width, and diameter at breast height), foundation, and forest structures extracted by UAV LiDAR (such as tree height and crown width), etc., so as to adjust the default individual tree structure parameters in the OnyxTREE software model library and generate individual trees of different forest types that conform to the actual situation. Finally, based on the statistical results of the vertical distribution profile of the forest leaf area extracted by the foundation and UAV LiDAR, control the leaf area density of the individual trees to obtain 3D models of individual trees with different leaf areas.

[0050] Generation of 3D models of individual understory vegetation: Understory vegetation mainly consists of shrubs and herbs. Select the corresponding vegetation type model library in OnyxTREE software. Then, according to the information such as the height, diameter at breast height, and leaf area of the understory vegetation surveyed on the ground and detected by the ground-based LiDAR, adjust the default vegetation structure parameters in the model library to generate 3D models of individual shrubs or herbaceous plants in the forest that conform to the actual situation.

[0051] Construction of 3D forest scenes: First, based on the typical 3D models of individual trees, generate combinations of individual trees that meet the actual forest growth conditions in the study area by means of equal-proportion scaling, or scaling that conforms to a certain probability distribution (such as Gaussian distribution), or changing the proportion of leaf area, wood area, tree height, diameter at breast height, and height to the lowest live branch, or rotation angle, etc. Then, based on the 3D models of individual shrubs or herbs, generate combinations of individual vegetation that meet the actual growth conditions of the understory vegetation by means of equal-proportion scaling, or scaling that conforms to a certain probability distribution (such as Gaussian distribution), or changing the leaf area and vegetation height, or rotation angle, etc. Finally, based on the spatial position information of individual trees and understory vegetation measured on the ground and extracted by LiDAR, construct various combined distribution methods that take into account different tree types and understory vegetation types on the DART model platform, thereby generating 3D forest scenes that meet the actual spatial distribution.

[0052] S12. Analyze the geometric characteristics of the leaf elements in the 3D forest scene to obtain the scene LAI of the 3D forest scene.

[0053] The LAI of the 3D forest scene is jointly determined by the leaf area, number of plants, and spatial distribution of each plant in the scene. On the basis of conducting ground surveys and LiDAR detections, determine the spatial distribution and number of plants of each plant, and then adjust the 3D structure and leaf area of each plant in the scene by means of equal-proportion scaling, or scaling that conforms to a certain probability distribution (such as Gaussian distribution), or changing the leaf area volume density and plant height, or rotation angle, etc., so as to change the size of the scene LAI. After completing the scene construction, count the triangular patch areas of all the leaves in the scene to obtain the scene LAI.

[0054] S13. Obtain the spectral, radiance information of the scene element components of the three-dimensional forest scene and the forest canopy reflectance to obtain the modeling training set. In this embodiment, the spectral of the scene element components includes: leaf spectral characteristics and soil spectral characteristics.

[0055] First, define the input parameters of the leaf radiative transfer model PROSPECT-D, and obtain the simulated leaf spectral characteristics through Equation (1); determine the input parameters of the soil model GSV, and obtain the simulated soil spectral characteristics through Equation (2); the spectral information (R other ) of branches, understory vegetation, etc. in the three-dimensional forest scene can be measured on-site by a spectrometer or obtained for free from a public database.

[0056] The method for obtaining the leaf spectral characteristics includes:

[0057] . (1)

[0058] The method for obtaining the soil spectral characteristics includes:

[0059] , (2)

[0060] In the formula, R(λ) and T(λ) respectively represent the reflectance and transmittance of the leaf simulated by the PROSPECT-D model at wavelength λ; N, C ab 、C ar 、C w 、C m 、C brown and λ respectively represent the leaf structure parameters, chlorophyll content, carotenoid content, leaf equivalent water thickness, dry matter content, brown matter content and wavelength input into the PROSPECT-D model; R soil (λ) represents the soil reflectance; 、 、 are three dry soil components; c sm represents the wet soil component.

[0061] The method for obtaining the radiance information includes: on the premise of a given solar incident angle (solar zenith angle θ s and solar azimuth angle φ s ), adopt the ray forward-tracing mode, emit photons from the light source (such as direct sunlight, sky scattering), simulate the transmission process of photons in the three-dimensional forest scene, and calculate the radiation energy reaching the ground from direct sunlight, sky scattering and adjacent terrain reflection. Given the sensor observation angle (observation zenith angle θ v and observation azimuth angle φ v), adopting the backward light tracing mode, emitting light from the sensor, simulating the single and multiple scattering of photons in the complex forest scene, and calculating the radiance information received by the sensor.

[0062] The method for obtaining the forest canopy reflectance includes:

[0063] Using the DART model to output the surface reflectance image with a spatial resolution of 30 m, a spectral range of 400 nm - 2500 nm, and a spectral resolution of 1 nm. The conceptual expression of the DART model is described by Equation (3). Then, using the spectral response functions of the Landsat series sensors (i.e., TM, ETM+, OLI), the DART-simulated reflectance image is converted into the corresponding multispectral reflectance, and the fitting formula is expressed as Equation (4). In addition, in order to describe the difference between the simulated band reflectance and the actual remote sensing observation, Gaussian noise is added to it.

[0064] , (3)

[0065] , (4)

[0066] In the formula, R canopy (band i ) represents the reflectance value corresponding to the i-th band of the Landsat sensor after conversion by the spectral response function; R canopy (λ j ) represents the forest canopy reflectance value simulated by the DART model at the wavelength λ j ; R canopy represents the three-dimensional forest canopy directional reflectance value simulated by the DART model; SRF(λ j ) represents the spectral response function value of the Landsat sensor at the wavelength λ j ; b represents the total number of wavelength sampling points; λ represents the wavelength; θ s and are the solar zenith angle and azimuth angle respectively; θ v and are the sensor observation zenith angle and azimuth angle respectively; veg represents the vegetation type; LAI is the forest leaf area index extracted based on the three-dimensional scene; R is the leaf reflectance; T is the leaf transmittance; R soil is the soil reflectance; R other represents the reflectance of branches and understory vegetation.

[0067] Taking the cosine values of green, red, near-infrared, shortwave infrared-1, shortwave infrared-2, and the solar zenith angle as feature values, and the LAI of the three-dimensional forest scene as the target value, the data processed through the above steps is sorted to generate a modeling training set.

[0068] Step 2: Construct a pre-trained 1D-CNN forest LAI inversion model, and train and optimize the pre-trained 1D-CNN forest LAI inversion model based on the modeling training set to obtain an optimized model.

[0069] 1D-CNN automatically extracts the input remote sensing feature values by constructing a deep neural network and learns the non-linear mapping relationship between them and the simulated LAI. When specifically training the 1D-CNN forest LAI inversion model, the cosine values of green, red, near-infrared, short-wave infrared-1, short-wave infrared-2, and solar zenith angle will be used as one-dimensional sequence data, and a one-dimensional convolutional kernel will slide on the sequence to extract the input data features. The theoretical expression of 1D-CNN can be described by Equation (5).

[0070] LAI out =bias + weight ★ input, (5)

[0071] In the formula, LAI out represents the LAI result output after calculating the input feature values through the 1D-CNN network; input is the input feature value; weight represents the convolutional kernel that performs cross-correlation operation with the input feature value; bias represents the bias term added after the convolutional operation; represents the effective cross-correlation operator.

[0072] When the pre-trained 1D-CNN forest LAI inversion model is pre-trained according to three-dimensional forest simulation data, first, the output LAI value is obtained through forward calculation based on the input feature values and the initial network parameters. Then, the difference between the model-output LAI and the reference LAI is quantified, and the network structure parameters are adjusted according to the loss. To add the guidance of remote sensing prior knowledge during the network training process, different from the conventional training method that constructs the loss function with the Mean Squared Error (MSE), two additional optimization terms are added to the loss function in this embodiment. One is to apply the Kullback-Leibler Divergence (KL) to the construction of the loss function. The KL divergence is an asymmetric measure used to measure the difference between two probability distributions. In practical applications, the probability distributions of the reference LAI value and the predicted LAI value are obtained respectively, and then the information loss between the two probability distributions is quantified through the KL divergence. A KL divergence of 0 indicates that the distributions of the predicted LAI value and the reference LAI value are exactly the same, and this measure is always non-negative. The larger the KL divergence, the more significant the distribution difference between the predicted LAI value and the reference LAI value. Measuring the deviation between the model prediction value and the reference value at the probability distribution level through the KL divergence can provide deeper error constraints. The other is that the physical boundary constraint uses the Relu activation function to penalize the predicted values that exceed the reasonable range of LAI, so as to ensure that the output conforms to the actual physical laws. The finally optimized loss function is adjusted by the hyperparameters α and β, where α determines the weight of the MSE loss term, (1-α) is used as the coefficient of the KL loss term, and β is used to constrain the intensity of the physical boundary penalty. The calculation process of the difference between the LAI output by the 1D-CNN forest LAI inversion model and the reference LAI is shown in Equations (6) to (10):

[0073] , (6)

[0074] , (7)

[0075] , (8)

[0076] , (9)

[0077] , (10)

[0078] In the formula, Loss represents the value of the loss function during the training process of the 1D-CNN forest LAI inversion model; α is the proportion of the MSE error in the quantified loss value, and (1-α) represents the proportion of the KL error; β represents the coefficient of the additional penalty for the out-of-bounds cases where the model predicts LAI less than 0 and greater than 10; B low and B uprespectively represent the penalty losses calculated by the boundary limit where the prediction result is less than 0 and the penalty losses calculated by the boundary limit where the prediction result is greater than 10; n represents the total number of data applied to model validation; MSE and KL respectively represent the mean squared error loss and relative entropy loss calculated based on the reference LAI and the output LAI; p is a parameter coefficient set to strengthen the quantitative features; y i represents the reference LAI value of the i-th training data; represents the model output LAI value of the i-th training data; represents the model output LAI value; Relu represents a non-linear function in deep learning; LAI min and LAI max respectively represent the minimum value of LAI and the maximum value of LAI, with values of 0 and 10 respectively.

[0079] The training parameters of the model mainly include the number of iterations, batch size, optimizer, learning rate, etc. Among them, the batch size can be set to four variables: 64, 128, 256, and 512; the number of iterations can be set to variables such as 5, 10, 50, 100, 500, and 1000; the optimizer can be set to variables such as RMSProp, Adam, AdaGrad, and SGD; the learning rate can be set to variables such as 0.00001, 0.00005, 0.0001, and 0.0005. The above variables can be cross-combined to determine the parameter combination with the optimal training effect, and the optimal model is saved to obtain the optimized model.

[0080] Step 3: Construct a transfer dataset based on Landsat satellite image data and measured LAI data, and optimize the parameters of the optimized model based on the transfer dataset to obtain a transfer learning forest LAI inversion model.

[0081] Although this embodiment obtains forest scene modeling data based on the DART model and uses Gaussian noise technology to simulate the observation situation of the sensor, its description of the mapping relationship between reflectivity and LAI in the actual situation is not comprehensive enough and still needs to be further optimized. By introducing actual observation knowledge and narrowing the difference between the modeling dataset and the real data, the LAI inversion model can be further optimized. In this embodiment, the transfer learning method developed in the field of deep learning is used to fine-tune the parameters of the 1D-CNN network based on actual observation data to solve the problem of poor actual application effect of the model caused by differences between different data, and an algorithm more suitable for the forest LAI inversion task is constructed. Therefore, by integrating Landsat satellite image data and measured LAI data to construct a transfer dataset, the 1D-CNN pre-trained model is further optimized. The specific implementation steps are as follows:

[0082] 1) Obtain Landsat satellite transit images of the measurement mission, integrate image reflectance data and measured forest LAI data to construct a migration dataset. The cosine values of green, red, near infrared, shortwave infrared-1, shortwave infrared-2 and solar zenith angle are used as feature values, and the actual LAI measured in the forest scene is used as the target value.

[0083] 2) Transfer learning uses the pre-trained 1D-CNN forest LAI inversion model as the source model, so the network structure is consistent with the pre-trained model. The transfer learning process first loads the pre-trained 1D-CNN forest LAI inversion model, then adjusts the loss function of the transfer training according to the migration dataset, and finally adjusts the transfer learning training parameters and obtains the forest LAI inversion model after transfer learning.

[0084] The training process of the transfer learning model is easily affected by the unevenness of the sample data set, which leads to the deviation of the inversion model. Therefore, the transfer learning process needs to adjust and optimize the loss function calculation process to better quantify the difference between the model output and the reference LAI, so as to build an accurate forest LAI inversion model. The calculation process of the transfer learning loss function is shown in equations (11) to (12):

[0085] , (11)

[0086] , (12)

[0087] In the formula, Loss w represents the loss value of the transfer learning forest LAI inversion model; α1 is the proportion of MSE error in the quantified loss value, (1-α1) represents the proportion of KL error; n represents the total number of data used for model verification; MSE w Represents the MSE loss weight after weighted calculation; m p Here, m represents the amount of data in the current interval in the statistical forest LAI; p and q are parameter coefficients set to enhance the quantitative characteristics.

[0088] The transfer training parameters of the model mainly include the number of iterations, batch size, optimizer and learning rate. Among them, the batch size can be set to four variables: 64, 128, 256 and 512; the number of iterations can be set to variables such as 5, 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100; the optimizer can be set to variables such as RMSProp, Adam, AdaGrad and SGD; the learning rate can be set to variables such as 0.00001, 0.00005, 0.0001 and 0.0005. The above variables can be cross-combined to determine the parameter combination with the best transfer learning training effect, and the best transfer learning forest LAI inversion model can be saved.

[0089] Step 4: Process the Landsat image data to be predicted based on the transfer learning forest LAI inversion model to obtain the estimation result.

[0090] Obtain the Landsat image data to be predicted, extract the forest area therein, organize its data into the eigenvalue format consistent with the model training process (cosine values of green, red, near-infrared, short-wave infrared-1, short-wave infrared-2, and solar zenith angle), load the saved transfer learning trained forest LAI inversion model, conduct forest LAI prediction, and save the prediction result in the tif image format.

[0091] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning, characterized in that, It includes the following steps: Step 1: Obtain a modeling dataset based on the DART model, and preprocess the modeling dataset to obtain a modeling training set; Step 2: Construct a pre-trained 1D-CNN forest LAI inversion model, and train and optimize the pre-trained 1D-CNN forest LAI inversion model based on the modeling training set to obtain an optimized model; Step 3: Construct a transfer dataset based on Landsat satellite image data and measured LAI data, and optimize the parameters of the optimized model based on the transfer dataset to obtain a transfer learning forest LAI inversion model; Step 4: Process the Landsat image data to be predicted based on the transfer learning forest LAI inversion model to obtain an estimation result.

2. The method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning according to claim 1, wherein Step 1 includes: S11. Construct a three-dimensional model of a single tree and a three-dimensional model of understory vegetation of a single tree based on measured data, and construct a three-dimensional forest scene based on the three-dimensional model of the single tree and the three-dimensional model of the understory vegetation of the single tree to obtain the modeling dataset; S12. Analyze the geometric characteristics of leaf elements in the three-dimensional forest scene to obtain the scene LAI of the three-dimensional forest scene; S13. Obtain the spectral characteristics of scene elements, radiance information, and forest canopy reflectance of the three-dimensional forest scene to obtain the modeling training set.

3. The method for estimating forest leaf area index based on the three-dimensional radiation transfer model and deep learning according to claim 2, wherein The spectral characteristics of the scene elements include: leaf spectral characteristics and soil spectral characteristics; The method for obtaining the leaf spectral characteristics includes: ， The method for obtaining the soil spectral characteristics includes: ， In the formula, R(λ) and T(λ) respectively represent the reflectance and transmittance of the leaf at wavelength λ simulated by the PROSPECT-D model; N, C ab , C ar , C w , C m , C brown and λ respectively represent the leaf structure parameters, chlorophyll content, carotenoid content, leaf equivalent water thickness, dry matter content, brown matter content and wavelength input by the PROSPECT-D model; R soil (λ) represents the soil reflectance; , , are three dry soil components; c sm represents the wet soil component.

4. The method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning according to claim 2, characterized in that, The method for obtaining the forest canopy reflectance includes: ， Among them, , where R canopy (band i ) represents the reflectance value corresponding to the i-th band of the Landsat sensor after conversion by the spectral response function; R canopy (λ j ) represents the forest canopy reflectance value simulated by the DART model at wavelength λ j ; R canopy represents the forest canopy reflectance value simulated by the DART model; SRF(λ j ) represents the spectral response function value of the Landsat sensor at wavelength λ j ; b represents the total number of wavelength sampling points; λ represents the wavelength; θ s and are the solar zenith angle and azimuth angle respectively; θ v and are the sensor observation zenith angle and azimuth angle respectively; veg represents the vegetation type; LAI is the forest leaf area index extracted based on the three-dimensional scene; R is the leaf reflectance; T is the leaf transmittance; R soil is the soil reflectance; R other represents the reflectance of branches and understory vegetation.

5. The method for estimating forest leaf area index based on three-dimensional radiation transfer model and deep learning according to claim 1, characterized in that, The expression of the pre-trained 1D-CNN forest LAI inversion model includes: LAI out =bias + weight ★ input, where LAI out represents the LAI result output by calculating the input eigenvalue through the 1D-CNN network; input is the input eigenvalue; weight represents the convolution kernel for performing cross-correlation operation with the input eigenvalue; bias represents the bias term added after the convolution operation; ★ represents the effective cross-correlation operator.

6. The method for estimating forest leaf area index based on a three-dimensional radiation transfer model and deep learning according to claim 1, wherein The loss function of the transfer learning forest LAI inversion model includes: ， Among them, ， ， ， ， In the formula, Loss w represents the loss value of the transfer learning forest LAI inversion model; α is the proportion of the MSE error in the quantified loss value, and (1-α) represents the proportion of the KL error; β represents the coefficient of additional penalty for the out-of-bounds cases where the model predicts LAI less than 0 and greater than 10; B low and B up respectively represent the penalty loss calculated from the boundary limit where the prediction result is less than 0 and the penalty loss calculated from the boundary limit where the prediction result is greater than 10; n represents the total number of data applied to model verification; MSE w represents the MSE loss weight after weighted calculation; KL represents the relative entropy loss calculated based on the reference LAI and the output LAI; m p in m represents the amount of data in the current interval among the statistically forest LAI; p and q are both parameter coefficients set to strengthen the quantitative features; y i represents the reference LAI value of the i-th training data; represents the model output LAI value of the i-th training data; represents the model output LAI value; Relu represents a non-linear function in deep learning; LAI min and LAI max respectively represent the minimum value of LAI and the maximum value of LAI, taking the values of 0 and 10 respectively.

Images