A forest resource growth trend dynamic monitoring method based on a canopy height model

Abstract

This invention discloses a method for dynamic monitoring of forest resource growth status based on a canopy height model, belonging to the field of forest resource monitoring and ecological remote sensing technology. The method includes: acquiring multi-temporal UAV image data to construct a canopy height model and extracting vegetation indices from the UAV image data; extracting canopy vertical structure complexity features and phenological response features, and concatenating them with pixel-level features to form an input feature vector; constructing a temporal convolutional feature extraction network based on dilated causal convolution to extract temporal feature tensors; fusing the temporal feature tensors with static site environment features to construct an encoder-predictor two-stage prediction model, outputting predicted canopy height values, predicted growth rate values, and anomaly warning labels; and combining historical monitoring data with statistical confidence intervals to classify growth status levels. This invention achieves dynamic tracking and trend prediction of forest canopy growth.

Description

Technical Field

[0001] This invention belongs to the field of forest resource monitoring and ecological remote sensing technology, specifically involving a dynamic monitoring method for forest resource growth status based on the canopy height model (CHM), which is applicable to the monitoring and trend prediction of growth status in nature reserves, forest parks and rare plant populations. Background Technology

[0002] Traditional forest resource growth status monitoring mainly relies on manual plot surveys, which suffers from low efficiency, long cycles, and insufficient spatial coverage. With the development of UAV remote sensing technology, multispectral / hyperspectral imagery is gradually being applied to vegetation monitoring. However, existing methods mostly focus on extracting vegetation indices at a single time point, lacking full utilization of information on the vertical structure of the forest canopy, making it difficult to achieve dynamic tracking and trend prediction of growth status.

[0003] Although some studies have attempted to use time-series remote sensing data for vegetation change monitoring, existing methods have the following shortcomings when processing canopy height time-series data: (1) They have a weak ability to capture long-term time-series dependencies, making it difficult to effectively model the cumulative effect and seasonal rhythm of canopy growth; (2) They lack a systematic extraction and fusion framework for the multi-scale time-series characteristics of canopy vertical structure; (3) The prediction models are mostly static regression or simple time-series extrapolation, lacking the ability to dynamically perceive the trend of growth rate changes and abnormal fluctuations; (4) Existing deep learning methods directly apply the general time-series prediction architecture without specifically designing for the special characteristics of canopy height data such as sparse sampling, significant phenological periodicity, and strong spatial heterogeneity, making it difficult to achieve high-precision growth prediction under limited time-series observations. Summary of the Invention

[0004] This invention addresses the problems of insufficient utilization of canopy vertical structure information, lack of long-term time-series dependency modeling ability, and low accuracy of dynamic tracking and trend prediction of growth status in existing forest resource monitoring methods. It provides a dynamic monitoring method for forest resource growth status based on a canopy height model, which enables efficient and quantitative monitoring and trend prediction of the growth status of target tree species.

[0005] To achieve the above objectives, the following technical solution is adopted:

[0006] This invention provides a method for dynamic monitoring of forest resource growth status based on a canopy height model, comprising the following steps:

[0007] Acquire multi-temporal UAV image data of the target forest area for multiple growing seasons, construct a canopy height model for each temporal node based on the UAV image data, and extract vegetation indices from the UAV image data;

[0008] Based on the canopy height model, the vertical structure complexity feature of the canopy is extracted, and the phenological response feature is extracted based on the UAV image data. The vertical structure complexity feature of the canopy and the phenological response feature are used as plot-level context features and concatenated with the canopy height model height value of each pixel and the vegetation index to form the input feature vector of each pixel.

[0009] A temporal convolutional feature extraction network based on dilated causal convolution is constructed. The pixel is used as the smallest processing unit. The input feature vector of each pixel is input into the temporal convolutional feature extraction network to extract a temporal feature tensor that integrates the temporal evolution law of canopy height and the fluctuation of vegetation index.

[0010] The temporal feature tensor is fused with static site environment features to construct an encoder-predictor two-stage prediction model, which outputs predicted values ​​of canopy height, canopy growth rate, and growth anomaly warning labels for multiple future time nodes.

[0011] Based on the predicted canopy growth rate and the abnormal growth warning label, combined with the statistical confidence interval of historical monitoring data, the monitoring unit is divided into multiple growth status levels.

[0012] Furthermore, the canopy height model for each time-series node is constructed in the following manner:

[0013] The UAV image data at each time series node are independently encrypted by aerial triangulation, and digital surface model and digital elevation model are constructed for the corresponding time series node. The canopy height model for the corresponding time series node is obtained by calculating the difference between the digital surface model and the digital elevation model.

[0014] Only spatial registration is performed between canopy height models at different time points; there is no process of using subsequent temporal data to participate in the construction of earlier canopy height models.

[0015] The UAV image data includes multispectral image data or hyperspectral image data; the vegetation index includes at least one of normalized vegetation index, ratio vegetation index and enhanced vegetation index.

[0016] Furthermore, the canopy vertical structure complexity features include the Shannon diversity index and the vertical stratification index;

[0017] The formula for calculating the Shannon diversity index is as follows:

[0018]

[0019] In the formula, The Shannon Diversity Index, For the height value in the canopy height model falling within the first... The percentage of pixels within each vertical stratification interval This represents the total number of vertical layers; It is the natural logarithm function;

[0020] The vertical stratification index is defined as the proportion of vertical strata with valid canopy return signals to the total number of strata.

[0021] Furthermore, the phenological response feature is the flowering area index, which is defined as the proportion of the area of ​​pixels in the image that meet the preset flowering period spectral threshold to the total canopy area; the preset flowering period spectral threshold is obtained by automatic thresholding, statistical mean of historical samples, or manual calibration.

[0022] Furthermore, the temporal convolutional feature extraction network stacks multiple residual blocks, each residual block containing two layers of dilated causal convolution, followed by a weight normalization layer, a ReLU activation function layer and a random deactivation layer, and introduces residual connections.

[0023] The dilated causal convolution at temporal positions Convolution output for:

[0024]

[0025] in, For the first The convolution kernel of the layer residual block, For the first The convolution kernel of the residual block in the first layer Weight values ​​at each index position The kernel size is the convolution kernel size. For the first The expansion factor of the layer This represents the index position of the convolution kernel. The output features are the residual blocks of the previous layer; the dilated causal convolution satisfies the causal constraint: if Enable causal zero-fill, making ;

[0026] The expansion factor increases exponentially with the number of layers. , ,in This represents the total number of layers in the residual blocks.

[0027] Furthermore, the encoder-predictor two-stage prediction model includes:

[0028] Encoding stage: Perform temporal self-attention operation on the temporal feature tensor to output self-attention features with the same temporal length and channel dimension as the temporal feature tensor; generate a channel-wise scaling factor from the static site environment features through learnable projection; use the channel-wise scaling factor to perform Hadamard element-wise multiplication modulation on the temporal features after self-attention operation to obtain fused features; the static site environment features include altitude, aspect, slope, and slope position;

[0029] Prediction phase: Extract the feature vector of the last time position from the fused features, and feed it into the height prediction branch, rate prediction branch and early warning branch respectively. Output the predicted values ​​of canopy height, canopy growth rate and growth abnormality early warning labels for multiple future time nodes in parallel.

[0030] Furthermore, the height prediction branch outputs the predicted canopy height for the next τ time nodes. Where τ is a preset positive integer representing the prediction step size; the rate prediction branch outputs the predicted canopy growth rate value. The warning branch first outputs the original logit value, followed by the warning probability value after Sigmoid activation. ,when Output a binary warning label when the value is greater than or equal to 0.5. Triggering an abnormal growth warning when Output when less than 0.5 This indicates normal growth.

[0031] Furthermore, the encoder-predictor two-stage prediction model adopts a multi-task joint training strategy, and the total loss function for joint training is:

[0032]

[0033] in, This is the mean squared error loss term for canopy height prediction. This is the mean absolute error loss term for canopy growth rate prediction. For anomaly warning, the binary cross-entropy classification loss term is used. This refers to the multi-task loss weighting coefficient; The predicted canopy height is... This corresponds to the actual value of the canopy height; The predicted value of the canopy growth rate is... This corresponds to the actual value of the canopy growth rate; The predicted value of the growth abnormality warning label. This represents the actual value of the corresponding growth abnormality warning label.

[0034] Furthermore, based on the mean of historical canopy growth rate sequences and standard deviation The growth status levels include:

[0035] Vigorous growth, the criterion is that the predicted growth rate is greater than ;

[0036] Normal growth is defined as the predicted growth rate falling within [ Within the range;

[0037] Growth slowdown is determined when the predicted growth rate is between ( ). Between ), or a single abnormal growth warning occurs but the criteria for degradation are not met;

[0038] Growth degradation is determined when three or more consecutive time-series nodes in the prediction sequence show abnormal growth warnings, or when the prediction step size is less than three, all prediction periods show absolute negative growth and the year-on-year decline exceeds a preset threshold; wherein, the preset threshold is determined based on the interannual growth variation coefficient of the target tree species, or is set using the statistical distribution percentile of the growth rate in the same period of history.

[0039] Furthermore, it also includes a step for extracting the correlation features between crown width and biomass at the individual tree scale:

[0040] Based on the canopy height model, the treetop is located by local maxima, and the canopy region is segmented by the marked watershed algorithm to calculate the canopy width parameters of a single tree.

[0041] A nonlinear empirical model is constructed based on canopy height, the canopy width parameters of individual trees, and measured biomass data to realize the quantitative conversion of canopy height model parameters into biomass indicators;

[0042] The formula for calculating the crown width parameter is as follows:

[0043]

[0044] in, The total pixel area of ​​the single-tree canopy segmentation region. For the crown width;

[0045] The nonlinear empirical model is as follows:

[0046]

[0047] in, To estimate biomass, The canopy height extracted from the canopy height model. The parameters are those of the model to be fitted. This is the random error term.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) This invention designs a dilated causal convolutional network, Canopy-TCN, adapted to canopy height temporal data. When the temporal data is sufficient, it relies on the exponential expansion of the receptive field through the dilated structure to capture long-term growth patterns; in the case of sparse temporal data, it relies on multi-task optimization, domain feature engineering, and environmental modulation to improve accuracy. The example shows a 20.6% reduction in RMSE compared to the traditional TCN baseline. Compared to recurrent neural networks, it avoids the gradient accumulation problem of temporal backpropagation, supports temporal parallel computation, and has a faster training speed; compared to ordinary convolutional networks, the dilated causal architecture strictly constrains causality and prevents future information leakage.

[0050] (2) To address the pain point of sparse canopy time-series sampling, a dual-stage encoder-predictor architecture with a multi-task joint loss function is designed to simultaneously output canopy height numerical prediction, growth rate trend, and growth anomaly early warning, forming a complete monitoring closed loop of prediction-grading-early warning.

[0051] (3) A two-stage fusion coding mechanism of temporal self-attention + environment modulation is proposed: self-attention mines temporal node dependence and completely preserves temporal structure; static environmental features are modulated by channel scaling factor through learnable projection to overcome the defects of traditional cross-modal fusion static features smoothing out the temporal dynamic information, so that the terrain factor regulates the growth representation in the form of weight, and significantly improves the generalization ability of complex mountain forest area models.

[0052] (4) Fully explore the vertical structure information of CHM to make up for the shortcomings of traditional vegetation index that only represents horizontal greenness; rely on CHM extreme value detection + watershed segmentation to extract the crown width of individual trees, and couple the ground measurement to construct a nonlinear inversion model of height-crown width-biomass, so as to realize the quantitative transformation of remote sensing vertical structure parameters into population growth indicators.

[0053] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0054] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0055] Figure 1 This is a flowchart illustrating the steps of a method for dynamically monitoring the growth status of forest resources based on a canopy height model, according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating the overall technical process of the present invention.

[0057] Figure 3 This is a schematic diagram of the Canopy-TCN network architecture according to an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the encoder-predictor two-stage prediction model architecture according to an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0061] Example 1:

[0062] Figure 1 This is a flowchart illustrating the steps of a method for dynamically monitoring the growth status of forest resources based on a canopy height model, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the overall technical process of the present invention. Figure 2 This demonstrates the complete chain from data acquisition, CHM construction and multi-source feature extraction, Canopy-TCN temporal feature extraction, prediction model training to growth status grading assessment, and embeds the internal execution logic of each submodule. Specifically, a method for dynamic monitoring of forest resource growth status based on a canopy height model includes the following steps:

[0063] S1: Acquire multi-temporal UAV image data of the target forest area for multiple growing seasons, construct a canopy height model for each temporal node based on the UAV image data, and extract vegetation index from the UAV image data;

[0064] S1.1, Multi-time series UAV multispectral / hyperspectral image data acquisition:

[0065] Within the fixed sample plots of the target forest area, UAVs equipped with multispectral or hyperspectral cameras were used to acquire multi-temporal, high-resolution UAV imagery data of the target forest area during multiple growing seasons (such as flowering, fruiting, and vigorous growth periods). RTK spatial positioning information was recorded simultaneously during image acquisition to ensure spatial registration accuracy across multiple image periods. Simultaneously, ground surveys were conducted within the sample plots to record auxiliary data such as population characteristics (plant height, diameter at breast height, crown width, root width, biomass, etc.), flowering characteristics (flowering period, number of flowers, inflorescence size, etc.), leaf functional traits (leaf length, width, thickness, dry matter content, specific leaf area, etc.), and site conditions (altitude, aspect, slope, slope position, etc.). This data provides training and validation samples for subsequent model construction.

[0066] S1.2, Construction of the Canopy Height Model (CHM):

[0067] The canopy height model for each time-series node is constructed as follows: Aerial triangulation is performed independently on the UAV imagery data for each time-series node to encrypt the data, and a digital surface model and digital elevation model (DEM) are constructed for each corresponding time-series node. The canopy height model for the corresponding time-series node is obtained by calculating the difference between the DEM and the DEM. Only spatial registration is performed between canopy height models from different time-series nodes; there is no process of using subsequent temporal data to participate in the construction of earlier canopy height models. Specifically:

[0068] The acquired multi-temporal image data is preprocessed (including radiometric correction, geometric correction, atmospheric correction, etc.). Each UAV image independently completes aerial triangulation densification, digital surface model (DSM) construction, and digital elevation model (DEM) construction, and then calculates the corresponding CHM for each period. Only spatial registration is performed between CHMs of different periods. There is no process of using subsequent temporal data to participate in the construction of the earlier CHM.

[0069] Let the first The CHM of each time-series node is represented as a matrix. ,in For the real number field, These represent the number of pixels in the row and column directions, respectively. Simultaneously, multiple vegetation indices for the target tree species are extracted from multispectral / hyperspectral images, including Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Enhanced Vegetation Index (EVI).

[0070] S2: Extract the vertical structure complexity feature of the canopy based on the canopy height model, extract the phenological response feature based on the UAV image data, and use the vertical structure complexity feature of the canopy and the phenological response feature as plot-level context features. These features are then concatenated with the canopy height model height value of each pixel and the vegetation index to form the input feature vector of each pixel.

[0071] Step S2 is used to perform multi-source feature extraction.

[0072] Based on the canopy height model (CHM) and vegetation indices, the following plot-level contextual characteristics were further calculated:

[0073] (1) Characteristics of the vertical structure complexity of the canopy:

[0074] Vertical structure complexity features of the canopy include: Shannon diversity index based on the canopy height model (CHM). And the Vertical Hierarchy Index (VSI).

[0075] Among them, the Shannon Diversity Index definition:

[0076]

[0077] In the formula, The medium height value falls in the first The percentage of pixels within each vertical stratification interval; The total number of vertical layers, as determined in this invention. Divide the canopy at equal intervals up to the maximum canopy height; It is the natural logarithm function.

[0078] The Vertical Segmentation Index (VSI) is defined as the percentage of vertical segments with valid canopy return signals out of the total number of segments. The proportion.

[0079] (2) Phenological response characteristics:

[0080] The phenological response characteristic is the Flowering Area Index (FAI). The FAI is defined as the area of ​​light in an image that meets a preset spectral threshold for the flowering period (increase in reflectance in the red-edge band). The proportion of the pixel area to the total canopy area.

[0081] The above Shannon diversity index Vertical Stratification Index (VSI) and Flowering Area Index (FAI) are plot-level contextual features, broadcast to all pixels within the corresponding plot. They are then concatenated with the CHM height value and vegetation index of each pixel to form a complete input feature vector for each individual pixel. Threshold The threshold can be obtained using Otsu's automatic thresholding method, historical sample statistical mean, or manual calibration, as described in the specific embodiments of this invention. .

[0082] Specifically, threshold The selection method is determined based on image quality and phenological period: when the image signal-to-noise ratio is high and the difference between the flowering period and the background spectrum is significant, the Otsu automatic thresholding method is preferred; when there are sufficient historical samples, the statistical mean of historical samples is used; in scenarios where the data quality is unstable or historical data is lacking, it is manually calibrated by professional technicians based on the actual situation.

[0083] S3: Construct a temporal convolutional feature extraction network based on dilated causal convolution. Using the pixel as the smallest processing unit, input the input feature vector of each pixel into the temporal convolutional feature extraction network to extract a temporal feature tensor that integrates the temporal evolution law of canopy height and the fluctuation of vegetation index.

[0084] Step S3: Construct a canopy height temporal feature extraction network based on dilated causal convolution (Canopy-TCN):

[0085] This invention proposes a temporal convolutional feature extraction network, Canopy-TCN, for canopy height time-series data. Using dilated causal convolution as the core operator, it captures long-term temporal dependencies in canopy height growth by exponentially expanding the receptive field under the causal constraint of "using only historical information." This invention does not impose any specific limitations on the data sampling frequency; the temporal length... When the network receptive field can be covered, the advantages of long-range modeling of the bloated architecture are fully utilized; When the size is small, dilated causal convolution is equivalent to global convolution and can still extract features normally. The accuracy gain comes from the synergistic effect of pixel-by-pixel temporal sample construction, multi-source feature fusion, multi-task joint optimization and environmental modulation.

[0086] The temporal convolutional feature extraction network stacks multiple residual blocks. Each residual block contains two layers of dilated causal convolution, followed by a weight normalization layer, a ReLU activation function layer, and a random deactivation (Dropout) layer, and residual connections are introduced. Figure 3 This is a schematic diagram of the Canopy-TCN network architecture of the present invention. The backbone is composed of four layers of residual blocks with increasing dilation factors. Each residual block contains two layers of dilated causal convolution, weight normalization, ReLU activation, and random deactivation (Dropout) layers, which, together with residual connections, achieve stable propagation of deep gradients.

[0087] S3.1, Pixel-level temporal sample construction:

[0088] Canopy-TCN uses pixels as the smallest processing unit and independently performs one-dimensional temporal convolution on the multi-temporal feature vectors of each spatial pixel within a single plot.

[0089] Let the total number of pixels in the sample plot be... (That is, the total number of pixels in the row direction × the total number of pixels in the column direction within the sample plot). The first... Each pixel in Feature sequences at each time node:

[0090]

[0091] In the formula: For the first The pixel in the first The dataset consists of concatenated vectors of CHM height, vegetation index, and plot-level contextual features at each time-series node. All pixel time-series vectors constitute the dataset, and individual pixel time-series samples independently participate in training and prediction. This is a spatial location index used to traverse all cells within the sample plot. .

[0092] S3.2, Dilated Causal Convolution Operation:

[0093] Input time series feature sequence Canopy-TCN performs one-dimensional dilated causal convolution pixel-by-pixel in the temporal dimension. Let the... Layer residual block convolution kernel is Core size Inflation factor The dilated causal convolution at the temporal position Convolution output:

[0094]

[0095] in, For the first The convolution kernel of the layer residual block, For the first The convolution kernel of the residual block in the first layer Weight values ​​at each index position This represents the index position of the convolution kernel; : No. The layer expansion factor increases exponentially with the number of layers. Total number of residual block layers; : Kernel size; : The original input sequence of the network; : No. Layer residual block output features; For the next level (i.e., the first) (Layer) residual block output characteristics;

[0096] Dilated causal convolution satisfies the causal constraint: if Enable causal zero-fill, making This ensures that only current and historical information is used, satisfies causal constraints, and enables normal operation at the beginning of the sequence.

[0097] S3.3, Residual Block Structure:

[0098] A single residual block contains two layers of dilated causal convolutions, followed by weight normalization, ReLU activation, and a Dropout layer; residual connections are introduced to ensure effective propagation of deep gradients.

[0099]

[0100] In the formula: Representing the Overall operation of dilated causal convolution and nonlinear transformation within residual blocks.

[0101] If the first Number of layer output channels Number of channels in the previous layer Inconsistent, adopt Convolution for residual branch dimension alignment:

[0102]

[0103] In the formula: It is a learnable linear projection matrix.

[0104] Stacking Formula for calculating the total receptive field of a network when the time sequence length is sufficient for layer residual blocks:

[0105]

[0106] in, This represents the range of historical information that the Canopy-TCN network can cover in the time dimension, and its unit is the time step.

[0107] Instance configuration: Substituting into One timing step size; when the actual timing length If the receptive field is smaller than the theoretical receptive field, the effective receptive field of the network is constrained by the sequence length, degenerating into a global equivalent convolution that covers all historical moments.

[0108] S3.4 Network Configuration and Feature Extraction:

[0109] Specific configuration of this invention: Canopy-TCN stack Each residual block has an expansion factor sequentially. kernel size Unified number of channels Dropout ratio 0.2, output feature dimension Dilated causal convolution does not change the sequence length, but the temporal output length remains the same. The time series information is fully preserved for subsequent prediction modules.

[0110] Through multi-layer residual block nonlinear mapping, Canopy-TCN encodes the original multi-source temporal features into a high-dimensional temporal feature tensor:

[0111]

[0112] This tensor integrates the temporal evolution of canopy height, phenological fluctuations of vegetation indices, and the long-term effects of environmental factors, providing a complete temporal characterization for growth status prediction.

[0113] The Canopy-TCN of this invention is not a simple module stacking; specifically:

[0114] 1. Constructing pixel-by-pixel temporal samples changes the network input granularity and training paradigm, improving spatial generalization ability;

[0115] 2. The pixel-level CHM / vegetation index + plot ecological statistics fusion method is adapted to the characteristics of canopy height data, which is different from the input structure of general time series models;

[0116] 3. The two-stage encoding of temporal self-attention and environment modulation solves the traditional fusion defect of smoothing out the time sequence information of static features, which is a specific design for the canopy monitoring scenario;

[0117] Together, these three constitute a complete methodological system for adapting to sparse canopy time-series data.

[0118] S4: The temporal feature tensor is fused with the static site environment features to construct an encoder-predictor two-stage prediction model, which outputs the predicted values ​​of canopy height, canopy growth rate and growth anomaly warning labels for multiple future time nodes.

[0119] Step S4 is used to construct a dynamic prediction model of growth status based on temporal feature fusion:

[0120] The temporal feature tensor output by Canopy-TCN By integrating static site environmental characteristics such as altitude, aspect, slope gradient, and slope position, a dynamic prediction model for growth status is constructed, employing a two-stage encoder-predictor architecture. Specifically, such as... Figure 4The diagram shown illustrates the encoder-predictor two-stage prediction model architecture of an embodiment of the present invention. Temporal features are first encoded using self-attention, and then fused with static environment feature channels using modulation. The fused temporal features are truncated into terminal vectors and fed into three independent fully connected branches, outputting three types of prediction results in parallel: canopy height, growth rate, and anomaly warning. Furthermore, the encoder-predictor two-stage prediction model includes an encoding stage and a prediction stage:

[0121] (a) Encoding stage: Temporal self-attention + environmental modulation

[0122] Encoding stage: A temporal self-attention operation is performed on the temporal feature tensor, outputting self-attention features with the same temporal length and channel dimension as the temporal feature tensor; the static site environment features are then projected using a learnable method to generate a channel-wise scaling factor, which is used to modulate the self-attention-operated temporal features using Hadamard element-wise multiplication to obtain fused features; the static site environment features include four dimensions: elevation, aspect, slope, and slope position. The details are as follows:

[0123] To preserve the complete temporal structure, the temporal feature tensor is first processed. Perform time-series self-attention operations and set the query / key dimensions of the self-attention mechanism. Define the query, key, and value matrix:

[0124]

[0125] in For the learnable projection matrix, the self-attention output is:

[0126]

[0127] Output The timing length and channel dimension remain unchanged.

[0128] Subsequently, static environmental characteristics ( (For the environmental feature dimension) a channel-wise scaling factor is generated through learnable projection, and self-attention features are adaptively modulated:

[0129]

[0130] in, : Feature projection matrix; Bias vector; : Sigmoid activation function; Hadamard element-wise multiplication; channel weights are copied along the time dimension and then multiplied. Element-wise operations; this mechanism, while fully preserving the temporal structure, enables differentiated weight control of feature channels under different site conditions, adapting to the modeling of heterogeneous forest areas.

[0131] (II) Forecasting Phase:

[0132] Prediction Phase: The feature vector of the last temporal position is extracted from the fused features and fed into the height prediction branch, rate prediction branch, and early warning branch, respectively. The predicted canopy height, canopy growth rate, and growth anomaly warning labels for multiple future temporal nodes are output in parallel. Details are as follows:

[0133] Before performing multi-step prediction, start with fused features Extract the feature vector of the last temporal position The forecast head is fed into the forecast head; the forecast head uses a direct multi-step forecasting strategy to output the future forecast in one go. Prediction results for each time series node A pre-defined positive integer represents the prediction step size; three independent output branches are configured for parallel computation:

[0134] 1. Highly predictive branch: Outputs the future Time series nodes (i.e.) Predicted canopy height (period) ;

[0135] 2. Rate Prediction Branch: Outputs predicted canopy growth rate values. ;

[0136] 3. Early Warning Branch: First, output the raw logit value, then activate it with a Sigmoid function to obtain the early warning probability value. ,Will Binary labels are obtained by comparing with a threshold of 0.5. , Triggering a growth abnormality warning, This indicates normal growth. Specifically, when... Output a binary warning label when the value is greater than or equal to 0.5. Triggering an abnormal growth warning when Output when less than 0.5 This indicates normal growth.

[0137] Each branch is a two-layer fully connected network (hidden layer dimension 128, ReLU activation, Dropout) The weights between branches are not shared, and the output dimensions are as follows: .

[0138] Warning Truth Value Label Construction rule: Statistically analyze the historical canopy height growth rate series of single pixels and solve for the historical mean. Standard deviation ;like ,make ,on the contrary Steps S4 and S5 require at least three complete years of data from the same period (≥9 time series nodes) to ensure the reliability of the statistical benchmark; when the data is ≥5 years, a longer historical window should be used to improve stability.

[0139] The encoder-predictor two-stage prediction model employs a multi-task joint training strategy, and the total loss function for joint training is:

[0140]

[0141] in, This is the mean squared error loss term for canopy height prediction. This is the mean absolute error loss term for canopy growth rate prediction. For anomaly warning, the binary cross-entropy classification loss term is used. This refers to the multi-task loss weighting coefficient; The predicted canopy height is... This corresponds to the actual value of the canopy height; The predicted value of the canopy growth rate is... This corresponds to the actual value of the canopy growth rate; The predicted value of the growth abnormality warning label. This represents the actual value of the corresponding growth abnormality warning label.

[0142] Implementation Example: Initial Weight Configuration Alternatively, GradNorm gradient normalization can be introduced for adaptive parameter tuning to balance the gradient norm across multiple tasks.

[0143] S5: Based on the predicted canopy growth rate and the abnormal growth warning label, and combined with the statistical confidence interval of historical monitoring data, the monitoring unit is divided into multiple growth status levels.

[0144] Step S5 is used to achieve a graded assessment of growth status.

[0145] A hierarchical system is established by combining predicted output, historical monitoring data, and statistical confidence intervals; historical growth mean. Standard deviation Based on at least 3 full years ( The solution is obtained from the observation data of the same time series nodes, and the long-term drift is eliminated by linear detrending; when historical data is insufficient, it is replaced with the benchmark data of neighboring sample plots with the same tree species and similar site conditions.

[0146] It should be noted that step S4 involves label statistics ( ) and step S5 hierarchical statistics ( Same-source calculation ensures consistent logic for tag generation and classification.

[0147] Based on the future Step-by-step prediction and early warning sequence growth rate series The monitoring units are divided into four growth status levels:

[0148] Based on the mean of historical canopy growth rate sequences and standard deviation ,include:

[0149] 1. Vigorous growth: Forecasted growth rate > ;

[0150] 2. Normal growth: The predicted growth rate falls within [ Within the range;

[0151] 3. Slower growth: The predicted growth rate is between ( ), or a single warning However, it does not meet the criteria for degradation.

[0152] 4. Growth degradation:

[0153] If predicting step size Warning: Predicted sequence shows growth anomalies at 3 or more consecutive time nodes. Determine if degradation warning is issued;

[0154] If predicting step size Exemption from continuous early warning criteria, meeting the requirements for all consecutive forecast periods ( A period showing absolute negative growth or a year-on-year decline exceeding a preset threshold (e.g., 10%) is considered a degradation; a single warning or fewer than three consecutive warnings are also considered degradations. The scenario is classified as growth slowdown, filtering out temporary abnormal fluctuations caused by short-term environmental stress.

[0155] Optionally, in some embodiments of the present invention, the method further includes step S6, extraction of the correlation features between canopy width and biomass at the individual tree scale, including: locating the treetop by local maxima based on the canopy height model, segmenting the canopy region using a marked watershed algorithm, and converting the individual tree canopy width parameters; constructing a nonlinear empirical model based on canopy height, the individual tree canopy width parameters, and measured biomass data to achieve quantitative conversion of canopy height model parameters into biomass indicators. Specifically, as follows:

[0156] Based on sample plot survey data, the canopy height parameters retrieved from the CHM (Conditional Tree Model) are correlated with measured individual tree height, diameter at breast height (DBH), crown width, root width, and biomass. This module is optional. The biomass estimation results output by this optional module can be used as auxiliary validation labels for the prediction model in step S4 to cross-validate the ecological rationality of the canopy height prediction; it can also be applied independently to forest productivity estimation scenarios, providing forest managers with biomass accumulation information in addition to canopy height growth trends. Because this module is independent of the Canopy-TCN time-series prediction main process fitting, and the module parameters are based on fitting of all sample plots, its parameter estimation does not participate in the optimization of the main model parameters, ensuring the decoupling between the main prediction model and the biomass retrieval module, facilitating flexible selection according to monitoring needs in practical applications.

[0157] Crown width parameters Solution: Locate the treetop using the CHM local maxima algorithm, segment the canopy region using the marked watershed algorithm, and convert the canopy diameter to an equivalent circle.

[0158]

[0159] in, The total pixel area of ​​the single-tree canopy segmentation region. For the crown width;

[0160] Construct a nonlinear empirical model of metalayer height-canopy width-biomass:

[0161]

[0162] in, To estimate biomass, The canopy height extracted for the canopy height model (CHM). The parameters are those of the model to be fitted. It is a zero-mean, independent and identically distributed random error term.

[0163] Nonlinear least squares (NLS) is used to fit the data with the objective of minimizing the sum of squared residuals. This enables CHM inversion of individual tree and stand growth parameters, and establishes a quantitative connection between remote sensing monitoring and ground-based measured data.

[0164] Furthermore, this module in step S6 establishes a quantitative connection between remotely sensed canopy vertical structure parameters and ground-measured growth indicators, enabling the method of this invention to not only output time-series predictions of canopy height changes, but also to map the predicted height changes to expected changes in biomass, providing richer decision support information for forest resource monitoring.

[0165] Example 2:

[0166] 1. Monitoring Targets: Rhododendron populations in a provincial nature reserve, with *Rhododendron simsii*, *Rhododendron dewdrop*, and *Rhododendron davidii* selected as target species; three monitoring areas were established in Pudi, Jinpo, and Gamu, totaling nine fixed plots, with each plot having a specific size... ,correspond Pixel (spatial resolution 10cm).

[0167] 2. Data Acquisition: The monitoring period is from March 2021 to May 2025. Three key phenological stages are selected annually: pre-flowering (early March), full bloom (mid-April), and fruiting (early June). A DJI M300 RTK drone equipped with a DJI P4 Multispectral camera is used for aerial surveying at an altitude of 120 m, with a ground resolution of approximately 8 cm, 80% forward overlap, and 70% lateral overlap. RTK positioning is recorded simultaneously, accumulating 15 valid image periods (5 years). (Period). Ground-level data including plant height, diameter at breast height, crown width, and biomass were measured in the field at each sample plot during the same period.

[0168] 3. CHM Construction and Feature Extraction:

[0169] Single-period imagery is independently spatially aligned, generating DSM and DEM respectively. Interpolation is then performed to obtain the corresponding temporal CHM. For multi-period CHMs, only spatial registration is performed; there is no cross-temporal back-firing interferometry. Pix4Dmapper aerial survey processing software is used to process the images, outputting 10cm resolution DSM, DEM, and CHM in single-time-series CHM matrix format. .

[0170] Three vegetation indices, NDVI, RVI, and EVI, were extracted from multispectral images; and calculated on a sample-by-sample basis. Three types of statistical features are broadcast to all pixels across the entire region; the final single pixel has 7-dimensional input features: CHM height + 3 vegetation indices + 3 plot-level ecological features.

[0171] 4. Dataset partitioning and moving sample construction:

[0172] Spatially isolated dataset partitioning: 3 plots in Pudi were used as an independent validation set (only for final accuracy evaluation, not for training or statistical calculation); 6 plots in Jinpo and Jiamu were used as the training set to avoid overfitting caused by the correlation of spatially adjacent samples.

[0173] Nine time-series nodes from 2021 to 2023 were selected as historical benchmarks to solve for each pixel in the training set.

[0174] Sliding window sample construction: Input window length 12 periods, predict the next 2 periods. Altitude, rate, and warning labels; Example window: Input 12 periods from March 2021 to March 2024, predict the results for periods April 2024 and June 2024; Warning labels are generated based on the 2021-2023 benchmark; Settings for the first 150 training rounds. Highly predictive convergence recovery Initiate collaborative training for classification tasks; the validation set is not involved in mean and standard deviation statistics or model parameter updates throughout the process.

[0175] Total sample size: training set One pixel time-series sample; validation set One pixel time series sample; single sample input time series length Feature dimension 7; batch size 16.

[0176] This embodiment The receptive field is smaller than 61 in the network theory, so the advantages of long-range expansion cannot be fully utilized. The model gain comes from feature engineering, multi-task training, and environment modulation. The fixed random seed 42 ensures the reproducibility of the experiment.

[0177] Rolling prediction verification: Using the latest 12 consecutive periods (June 2022 - March 2025) as input, predict the canopy growth status in April 2025 and June 2025. The length of the input time series is consistent with that of the training.

[0178] 5. Canopy-TCN Training and Accuracy Results:

[0179] Network configuration: 4-layer residual blocks, expansion factor convolution kernel 64 channels, Dropout Output dimension 64; Optimizer Adam, initial learning rate batch The total number of iterations is 200, and the early stopping strategy has a patience of 30.

[0180] Validation set accuracy: Canopy height prediction , .

[0181] 6. Optional biomass inversion module verification:

[0182] Based on the measured data from all 9 sample plots, an NLS model was used to fit the biomass model. The fitting parameters for Rhododendron simsii were:

[0183] Coefficient of determination This module performs independent fitting and does not participate in the optimization of the main prediction model parameters; it only serves as evidence of the effectiveness of remote sensing inversion.

[0184] 7. Class Imbalance Handling and Early Warning Indicators:

[0185] Based on the 2021-2023 benchmark, a warning positive and negative sample set was constructed. However, the low proportion of abnormal samples introduced class imbalance. After flattening the temporal features output by Canopy-TCN, SMOTE oversampling was used to augment a small number of abnormal samples. This oversampling was applied only to the training set, and the ratio of positive to negative samples was controlled to not exceed a certain limit. The training strategy is the same as above: freeze the classification loss in the early stage and perform joint optimization in the later stage.

[0186] Validation set alert performance: F1-score .

[0187] The control ablation experiments (mean ± standard deviation of 5 different random seeds) are shown in Table 1.

[0188] Table 1

[0189]

[0190] 8. Experimental Conclusion:

[0191] (1) Compared with LSTM and GRU, Canopy-TCN has a RMSE reduction of 37.2% and 30.8% respectively, and its training efficiency is significantly better than that of the loop structure;

[0192] (2) Timing length Under the given conditions, there is no significant difference in accuracy between incrementally expanding TCN and iso-expanding TCN (Wilcoxon rank-sum test). The long-range receptive field advantage of the inflated structure was not fully realized in this embodiment with a limited time length; the slightly higher time consumption of the standard TCN is due to the overhead of the hierarchical inflated index operation.

[0193] (3) The RMSE of the present invention is reduced by 20.6% compared with the two types of TCN baselines. The accuracy improvement comes from the whole domain adaptation design of pixel-by-pixel sample construction + multi-source ecological feature fusion + multi-task joint training + environmental modulation; the RMSE of steep slope plots with a slope of >15° is reduced by about 18%; the small increase in training time comes from the self-attention matrix operation and the forward operation of the three independent prediction branches.

[0194] On-site verification of growth grading:

[0195] The validation set of the baseline plots independently solved the grading benchmark using its own historical data from 2021 to 2023. , They do not participate in training, ensuring independent verification; Degradation criteria will use short-sequence rules: a year-on-year decline exceeding [a certain threshold] for two consecutive periods. This means determining growth degradation.

[0196] The overall grading results of the 9 sample plots were as follows: 3 plots showed normal growth, 4 plots showed slowed growth, and 2 plots showed degraded growth. Field verification of the causes of degraded plots showed that the canopy height of the No. 1 validation plot decreased by 12% year-on-year due to drought stress in the winter of 2024; the No. 2 training plot was inhibited by competition among surrounding shrubs, resulting in a decrease of 11%, which was consistent with the model prediction.

[0197] This embodiment only verifies the feasibility of the method, and the sample size and prediction step size are limited. Large-scale systematic verification can be carried out by relying on a larger monitoring network.

[0198] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0199] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0200] This invention also provides a dynamic monitoring system for forest resource growth status based on a canopy height model, comprising the following modules:

[0201] The data acquisition and CHM construction module is used to acquire multi-temporal UAV image data of multiple growing seasons in the target forest area, construct canopy height models for each temporal node based on the UAV image data, and extract vegetation indices from the UAV image data.

[0202] The multi-source feature extraction module is used to extract canopy vertical structure complexity features based on the canopy height model and phenological response features based on the UAV image data. The canopy vertical structure complexity features and the phenological response features are used as plot-level context features and concatenated with the canopy height model height value and the vegetation index of each pixel to form the input feature vector of each pixel.

[0203] The temporal feature extraction module is used to construct a temporal convolutional feature extraction network based on dilated causal convolution. Taking the pixel as the smallest processing unit, the multi-temporal input feature vector of each pixel is input into the temporal convolutional feature extraction network to extract a temporal feature tensor that integrates the temporal evolution law of canopy height and the fluctuation of vegetation index.

[0204] The prediction model construction module is used to fuse the temporal feature tensor with static site environment features to construct an encoder-predictor two-stage prediction model, and output the predicted values ​​of canopy height, canopy growth rate and growth anomaly warning labels for multiple future time nodes.

[0205] The growth status grading assessment module is used to divide the monitoring unit into multiple growth status levels based on the predicted canopy growth rate and the abnormal growth warning label, combined with the statistical confidence interval of historical monitoring data.

[0206] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0207] The above description is only a preferred embodiment of the present invention and is used only to explain the present invention. It is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0208] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0209] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for dynamic monitoring of forest resource growth status based on a canopy height model, characterized in that, Includes the following steps: Acquire multi-temporal UAV image data of the target forest area for multiple growing seasons, construct a canopy height model for each temporal node based on the UAV image data, and extract vegetation indices from the UAV image data; Based on the canopy height model, the vertical structure complexity feature of the canopy is extracted, and the phenological response feature is extracted based on the UAV image data. The vertical structure complexity feature of the canopy and the phenological response feature are used as plot-level context features and concatenated with the canopy height model height value of each pixel and the vegetation index to form the input feature vector of each pixel. A temporal convolutional feature extraction network based on dilated causal convolution is constructed. The pixel is used as the smallest processing unit. The input feature vector of each pixel is input into the temporal convolutional feature extraction network to extract a temporal feature tensor that integrates the temporal evolution law of canopy height and the fluctuation of vegetation index. The temporal feature tensor is fused with static site environment features to construct an encoder-predictor two-stage prediction model, which outputs predicted values ​​of canopy height, canopy growth rate, and growth anomaly warning labels for multiple future time nodes. Based on the predicted canopy growth rate and the abnormal growth warning label, combined with the statistical confidence interval of historical monitoring data, the monitoring unit is divided into multiple growth status levels.

2. The method according to claim 1, characterized in that, The canopy height model for each time-series node is constructed in the following manner: The UAV image data at each time series node are independently encrypted by aerial triangulation, and digital surface model and digital elevation model are constructed for the corresponding time series node. The canopy height model for the corresponding time series node is obtained by calculating the difference between the digital surface model and the digital elevation model. Only spatial registration is performed between canopy height models at different time points; there is no process of using subsequent temporal data to participate in the construction of earlier canopy height models. The UAV image data includes multispectral image data or hyperspectral image data; the vegetation index includes at least one of normalized vegetation index, ratio vegetation index and enhanced vegetation index.

3. The method according to claim 1, characterized in that, The vertical structural complexity characteristics of the canopy include the Shannon diversity index and the vertical stratification index; The formula for calculating the Shannon diversity index is as follows: In the formula, The Shannon Diversity Index, For the height value in the canopy height model falling within the first... The percentage of pixels within each vertical stratification interval This represents the total number of vertical layers; It is the natural logarithm function; The vertical stratification index is defined as the proportion of vertical strata with valid canopy return signals to the total number of strata.

4. The method according to claim 1, characterized in that, The phenological response feature is the flowering area index, which is defined as the proportion of the area of ​​pixels in the image that meet the preset flowering period spectral threshold to the total canopy area; the preset flowering period spectral threshold is obtained by automatic thresholding, statistical mean of historical samples, or manual calibration.

5. The method according to claim 1, characterized in that, The temporal convolutional feature extraction network stacks multiple residual blocks. Each residual block contains two layers of dilated causal convolution, followed by a weight normalization layer, a ReLU activation function layer, and a random deactivation layer, and residual connections are introduced. The dilated causal convolution at temporal positions Convolution output for: in, For the first The convolution kernel of the layer residual block, For the first The convolution kernel of the residual block in the first layer Weight values ​​at each index position The kernel size is the convolution kernel size. For the first The expansion factor of the layer This represents the index position of the convolution kernel. The output features are the residual blocks of the previous layer; the dilated causal convolution satisfies the causal constraint: if Enable causal zero-fill, making ; The expansion factor increases exponentially with the number of layers. , ,in This represents the total number of layers in the residual blocks.

6. The method according to claim 1, characterized in that, The encoder-predictor two-stage prediction model includes: Encoding stage: Perform temporal self-attention operation on the temporal feature tensor to output self-attention features with the same temporal length and channel dimension as the temporal feature tensor; generate a channel-wise scaling factor from the static site environment features through learnable projection; use the channel-wise scaling factor to perform Hadamard element-wise multiplication modulation on the temporal features after self-attention operation to obtain fused features; the static site environment features include altitude, aspect, slope, and slope position; Prediction phase: Extract the feature vector of the last time position from the fused features, and feed it into the height prediction branch, rate prediction branch and early warning branch respectively. Output the predicted values ​​of canopy height, canopy growth rate and growth abnormality early warning labels for multiple future time nodes in parallel.

7. The method according to claim 6, characterized in that, in: The height prediction branch outputs the predicted canopy height for the next τ time nodes. Where τ is a preset positive integer representing the prediction step size; The rate prediction branch outputs predicted canopy growth rate values. ; The warning branch first outputs the original logit value, followed by the warning probability value after Sigmoid activation. ,when Output a binary warning label when the value is greater than or equal to 0.

5. Triggering an abnormal growth warning when Output when less than 0.5 This indicates normal growth.

8. The method according to claim 1, characterized in that, The encoder-predictor two-stage prediction model adopts a multi-task joint training strategy, and the total loss function for joint training is: in, This is the mean squared error loss term for canopy height prediction. This is the mean absolute error loss term for canopy growth rate prediction. For anomaly warning, the binary cross-entropy classification loss term is used. This refers to the multi-task loss weighting coefficient; The predicted canopy height is... This corresponds to the actual value of the canopy height; The predicted value of the canopy growth rate is... This corresponds to the actual value of the canopy growth rate; The predicted value of the growth abnormality warning label. This represents the actual value of the corresponding growth abnormality warning label.

9. The method according to claim 1, characterized in that, Based on the mean of historical canopy growth rate sequences and standard deviation The growth status levels include: Vigorous growth, the criterion is that the predicted growth rate is greater than ; Normal growth is defined as the predicted growth rate falling within [ ]. Within the range; Growth slowdown is determined when the predicted growth rate is between ( ). Between ), or a single abnormal growth warning occurs but the criteria for degradation are not met; Growth degradation is determined when three or more consecutive time-series nodes in the prediction sequence show abnormal growth warnings, or when the prediction step size is less than three, all prediction periods show absolute negative growth and the year-on-year decline exceeds a preset threshold; wherein, the preset threshold is determined based on the interannual growth variation coefficient of the target tree species, or is set using the statistical distribution percentile of the growth rate in the same period of history.

10. The method according to claim 1, characterized in that, It also includes a step for extracting the correlation features between crown width and biomass at the individual tree scale: Based on the canopy height model, the treetop is located by local maxima, and the canopy region is segmented by the marked watershed algorithm to calculate the canopy width parameters of a single tree. A nonlinear empirical model is constructed based on canopy height, the canopy width parameters of individual trees, and measured biomass data to realize the quantitative conversion of canopy height model parameters into biomass indicators; The formula for calculating the crown width parameter is as follows: in, The total pixel area of ​​the single-tree canopy segmentation region. For the crown width; The nonlinear empirical model is as follows: in, To estimate biomass, The canopy height extracted from the canopy height model. The parameters are those of the model to be fitted. This is the random error term.

Images