A forest growth parameter prediction method based on a cascaded recurrent network

Abstract

The application discloses a forest growth parameter prediction method based on a cascade recurrent network, and comprises the following steps: collecting artificial measurement growth parameters of trees in a forest sample plot at multiple time nodes; acquiring the growth parameters of the trees through collected laser point cloud data; sorting the growth parameters according to data collection time to obtain a time sequence set of single-tree growth parameters; training a forest growth parameter prediction deep learning network by using the time sequence set of single-tree growth parameters; and predicting single-tree growth parameters corresponding to future time nodes by using the trained forest growth parameter prediction deep learning network. The application combines laser radar scanning and artificial sample plot measurement data to construct a training set, and compared with all artificial manual measurement data, the cost is lower and the efficiency is higher. Meanwhile, the future growth parameters of the trees can be effectively predicted, and the prediction accuracy is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest growth parameter research, and particularly relates to a forest growth parameter prediction method based on a cascade recurrent network. BACKGROUND

[0002] Forests, as an important terrestrial ecosystem in China, play an important role in carbon sequestration and oxygen release, soil and water conservation, air environment purification, nutrient accumulation, and biodiversity protection. Accurate extraction and prediction of forest growth parameters, such as diameter at breast height, tree height, and crown width, are important basic work in forest resource cultivation, management, and estimation of regional forest stock volume, carbon storage, growth and development evaluation of forest trees, afforestation strategy, and site quality evaluation.

[0003] At present, there are a large number of studies on prediction models for forest growth parameters. According to the methods used, the prediction models can be mainly divided into two categories: empirical and regression-based prediction models and machine learning-based prediction models. Empirical and regression-based prediction models usually extract single-tree growth parameters such as diameter at breast height, tree height, and crown width from forest land by field measurement, unmanned aerial forest land image, and laser scanning, and use data within a certain time range to construct a regression equation to build a prediction model. Specifically, the Verhaulst prediction model is used to predict the tree height, ground diameter, and crown diameter of a young Pinus elliottii tree, or a competition factor is introduced into a mixed effect model to predict the bark diameter of Larix gmelinii, which has higher prediction accuracy than traditional models, or terrain moisture index and leaf area index are extracted from laser radar data and introduced into a linear mixed effect model to improve the growth prediction of forest trees in southeastern Finland. However, the analysis of complex forest land under different seasonal phases and different environmental conditions by relying on artificial forest investigation and various remote sensing methods under different scales still has problems such as uncertain accuracy, dependence on artificial research and judgment, and weak robustness of the algorithm; secondly, the growth prediction and harvest estimation of forestry mainly rely on the theoretical growth regression model summarized by predecessors in fixed forest areas, which lacks universality and self-adjusting ability of model parameters. Machine learning-based prediction models can achieve high-precision prediction of forest growth parameters by introducing collected forest-related data into machine learning models and self-adaptively adjusting internal parameters. The construction methods of some current prediction models are as follows: a random forest model is used to predict the diameter at breast height of Cunninghamia lanceolata in North China based on 25 years of continuous climate and stand data, a deep learning algorithm is combined with an optimized radial basis neural network to process airborne laser radar data to estimate forest structure parameters of artificial forests in southern China, and a spectral vegetation index obtained by a hyperspectral camera and several machine learning techniques are combined to predict the diameter at breast height and tree height of Eucalyptus. Machine learning-based prediction models can consider more related factors such as climate, competition, and precipitation, thereby providing further more accurate prediction of forest growth parameters.

[0004] At present, the breakthrough of new generation artificial intelligence technology application and the further development of laser scanning technology greatly reduce the cost of obtaining forest growth parameters, and deep learning technology has been widely used in the prediction of forest parameters. For example, using recurrent neural network (RNN) and the climate conditions of the forests in Argentina Patagonia to predict the basal area of 61 trees of two special tree species, or combining recurrent neural network (RNN) and time convolution network (TCN) to predict the yield of greenhouse crops based on historical yield and greenhouse environmental parameters, or combining past plant growth images with convolutional long short-term memory network (LSTM) to predict plant growth, and evaluating the results by using the weighted coverage score of each leaf as quantitative evaluation and predicting appearance as qualitative evaluation, or combining wavelet transform with LSTM with attention mechanism to predict the stem diameter change of plants. However, the application of deep learning technology in plant growth prediction still has the following problems: first, the acquisition of plant growth parameters often relies on manual measurement, and large-scale forest data collection will consume a lot of manpower and material resources and is low in efficiency; second, plant growth requires a long time, and for many trees, the time span can reach 10 to 20 years, which requires regular sample plot measurement by artificial, so regular data collection will cost high cost; finally, most of the plant growth parameters are nonlinear data, and the prediction accuracy of the traditional prediction model is difficult to satisfy. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a forest growth parameter prediction method based on a cascade recurrent network to solve the problems of the prior art. The forest growth parameter prediction method based on the cascade recurrent network combines laser radar scanning and artificial sample plot measurement data to construct a deep prediction network training set, which is low in cost and high in efficiency compared with manual measurement data. The improved recurrent neural network is used to predict the future trend of forest growth parameters, and the long-term dependence problem of the traditional recurrent neural network is solved by introducing an attention module, which can effectively predict the future growth parameters of trees and has high prediction accuracy.

[0006] To achieve the above technical purposes, the technical scheme adopted by the present application is as follows:

[0007] A forest growth parameter prediction method based on a cascade recurrent network, comprising:

[0008] Step 1, collecting growth parameters obtained by manually measuring trees in a forest sample plot at multiple time nodes;

[0009] Step 2, collecting laser point cloud data of the forest sample plot at multiple time nodes, and obtaining growth parameters of the trees through the laser point cloud data of the forest sample plot;

[0010] Step 3: Sort the growth parameters from Step 1 and Step 2 according to the data collection time to obtain a time series set of growth parameters for individual trees;

[0011] Step 4: Construct a deep learning network for predicting tree growth parameters. Use the tree growth parameters corresponding to the most recent time node A in the time series set of tree growth parameters as the output of the deep learning network for predicting tree growth parameters, and use the tree growth parameters corresponding to multiple time nodes before the most recent time node A in the time series set of tree growth parameters as the input of the deep learning network for predicting tree growth parameters, thereby training the deep learning network for predicting tree growth parameters.

[0012] Step 5: Obtain the time series set of individual tree growth parameters before the time node to be tested by following the methods in Step 1, Step 2 and Step 3 in sequence. Input the time series of individual tree growth parameters in the set into the trained deep learning network for forest growth parameter prediction to obtain the individual tree growth parameters corresponding to the time node to be tested.

[0013] As a further improvement of the present invention, the growth parameters include diameter at breast height (DBH), tree height, and crown width.

[0014] As a further improvement to the present invention, step 2 involves obtaining tree growth parameters using laser point cloud data from forest plots, specifically as follows:

[0015] The collected forest laser point cloud data was denoised using Gaussian filtering. The denoised point cloud was then divided into ground points and ground points using cloth simulation filtering, and elevation normalization was performed.

[0016] The processed forest point cloud data is subjected to crown width extraction and individual tree segmentation. The diameter at breast height (DBH) and tree height are then extracted from the segmented individual tree point cloud data.

[0017] As a further improved technical solution of the present invention, the deep learning network for predicting forest growth parameters includes a two-layer Bi-GRU module, an attention module, and a fully connected neural network module.

[0018] As a further improvement of the present invention, the time nodes of the single tree growth parameters input to the deep learning network for predicting forest growth parameters are seven.

[0019] As a further improvement of the present invention, the two layers of Bi-GRU modules are respectively referred to as the first layer Bi-GRU module and the second layer Bi-GRU module. Both the first layer Bi-GRU module and the second layer Bi-GRU module include a forward propagation sequence and a backward propagation sequence. The output of each GRU unit in the backward propagation sequence of the first layer Bi-GRU module is denoted as... Let the output of each GRU unit in the forward propagation sequence of the first-layer Bi-GRU module be denoted as... The output vector corresponding to the time node is obtained by vertically concatenating the outputs of the two GRU units at each time node. At the same time, the output of the first layer Bi-GRU module will also be used as the input data of the second layer Bi-GRU module.

[0020] The input vector IN of each GRU unit in the Bi-GRU module at time t (t≤7) t From vector X t sum vector H t-1 It is formed by vertical splicing, X t H is the input vector of the Bi-GRU module at time t, that is, the growth parameters of a certain tree at time t. t-1 IN is the output vector of the GRU unit at time t-1. t =[X t H t-1 ];

[0021] The GRU unit includes update gates and reset gates. The calculation formula for the update gate is: Z t =σ(W z *IN t +b z The formula for calculating the door reset is: R t =σ(W R *IN t +b R );W z To update the gate weight matrix, b z To update the gate bias vector, W R To reset the gate weight matrix, b R To reset the gate bias vector, σ is the sigmoid function;

[0022] After receiving the corresponding gating signal, first use the reset gate R. t Information H from the previous time step t-1 To process, that is:

[0023] H t-1,' Compared with the current input X t After concatenating the two parts, input them into the tanh neural network layer to obtain the current state. Right now: W h W is the weight matrix of the tanh neural network layer. h Size and W z W R The same size, b h This refers to the bias vector of the tanh neural network layer.

[0024] Finally, use the update gate Z. t The state at the current moment Information H from the previous time step t-1 Continuing the processing, we obtain the output vector at the current time point, denoted as H. t ,Right now: Where H t As input to the attention module;

[0025] In the above formula, The symbol represents element-wise multiplication, and * represents matrix multiplication.

[0026] As a further improvement of the present invention, the input of the attention module is the output H = [H] of the second-layer Bi-GRU module. 1 H 2 ,...,H 7 ];

[0027] First, let Q = [q1,q2,...,q7], K = [k1,k2,...,k7], and V = [v1,v2,...,v7]. Calculate matrices Q, K, and V respectively, i.e.:

[0028] q m =W q *H m ;k n =W k *H n ;v r =W f *H r ;

[0029] Q = W q *H;K=W k *H;V=W f *H;

[0030] Among them, H m H n Let H be any two vectors in H. r For the input vector, W q W k and W f Here is the weight matrix; q m Let k be a vector in matrix Q. n Let v be a vector in matrix K. r A vector in matrix V;

[0031] Calculate vector H m and H n Attention score α between m,n That is: α m,n =q m ·kn ;

[0032] The attention score matrix a between all input vectors is calculated, namely:

[0033]

[0034] Wherein K is the number of input vectors. T K is the transpose matrix of matrix K.

[0035] The weight coefficient a' corresponding to each attention score a is calculated by a softmax function m,n m,n Namely:

[0036]

[0037] The matrix V is multiplied by a' to obtain the output matrix A of the attention module, namely:

[0038]

[0039] A = V * a';

[0040] Wherein a is a vector in matrix A. m

[0041] In the above formula, The corresponding elements are multiplied, * is matrix multiplication, and · is vector inner product, is vector multiplication.

[0042] After obtaining the output vector sequence of the attention module, it is summed and brought into the fully connected neural network module, so as to obtain the final output vector, namely the growth parameter of the 8th time node of a single tree, so as to realize the prediction of the growth parameter of the tree.

[0043] As a further improved technical solution of the application, the loss function of the forest growth parameter prediction deep learning network is:

[0044]

[0045] In the formula, N represents the number of samples in training, cosh represents the hyperbolic cosine function, represents the predicted value of the i-th sample under the current network parameter, and y i represents the true value corresponding to the i-th sample.

[0046] The beneficial effects of the application are:

[0047] ​​The present application proposes an accurate extraction method of forest growth parameters for airborne laser radar data, aiming at automatically realizing the extraction of tree diameter at breast height, tree height and crown width at single tree scale in a large area of forest region; laser radar scanning and artificial sample plot measurement data are combined to obtain tree diameter at breast height, tree height and crown width, a time series of single tree growth parameters is constructed and used as a training set of a deep prediction network, compared with all artificial periodic manual measurement data as a training set, the cost is lower and the efficiency is higher; secondly, aiming at the extracted forest growth parameter time series, an improved recurrent neural network is proposed to predict the future trend of forest growth parameters, and the long-term dependence problem of the traditional recurrent neural network is solved by introducing an attention mechanism module. 2 All of the determination coefficients R 2 of the diameter at breast height, tree height and crown width of each tree species and the true value are greater than 0.81, the root mean square error RMSE of the predicted value and the measured value of the diameter at breast height of the four tree species is less than 2.5 cm, and the root mean square error RMSE of the predicted value and the measured value of the tree height and crown width is less than 0.5 m. The method proposed in the present application can automatically extract corresponding diameter at breast height, tree height, crown width and other parameters from forest laser point cloud data, and is suitable for predicting the growth parameters of different tree species, which is one of the application points of the landing development of artificial intelligence technology in the field of smart forestry. The training set of the deep prediction network of the present application can also extract corresponding diameter at breast height, tree height, crown width and other parameters from forest laser point cloud data. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The flowchart of the present application is shown.

[0049] Figure 2 The schematic diagram of the research area of the present application is shown.

[0050] Figure 3 Fig. (a1) is a point cloud data diagram of the research area in 2016 of the Metasequoia sample plot laser point cloud data.

[0051] Figure 3 Fig. (a2) is a point cloud data diagram of the research area in 2018 of the Metasequoia sample plot laser point cloud data.

[0052] Figure 3 Fig. (a3) is a point cloud data diagram of the research area in 2020 of the Metasequoia sample plot laser point cloud data.

[0053] Figure 3 Fig. (b1) is a point cloud data diagram of the research area in 2016 of the cherry sample plot laser point cloud data.

[0054] Figure 3(b2) is a point cloud data map of the study area in 2018, based on laser point cloud data from the cherry blossom plot.

[0055] Figure 3 (b3) is a point cloud data map of the study area in 2020 from the laser point cloud data of the cherry blossom plot.

[0056] Figure 3 (c1) is a point cloud data map of the study area in 2016 from the laser point cloud data of the tulip tree sample plot.

[0057] Figure 3 (c2) is a point cloud data map of the study area in 2018 from the laser point cloud data of the tulip tree sample plot.

[0058] Figure 3 (c3) is a point cloud data map of the study area in 2020 from the laser point cloud data of the tulip tree sample plot.

[0059] Figure 3 The middle (d1) is the point cloud data map of the study area in 2016 from the laser point cloud data of the Ginkgo sample plot.

[0060] Figure 3 The middle (d2) is the point cloud data map of the study area in 2018 from the laser point cloud data of the Ginkgo sample plot.

[0061] Figure 3 The middle (d3) is the point cloud data map of the study area in 2020 from the laser point cloud data of the Ginkgo sample plot.

[0062] Figure 4 The middle (a) plot shows the point cloud data of the dawn redwood trees scanned in 2016, 2018 and 2020 in each coordinate axis region.

[0063] Figure 4 The middle (b) plot shows the point cloud data of the dawn redwood trees scanned in 2016, 2018 and 2020 in each coordinate axis region.

[0064] Figure 4 The middle (c) plot shows the point cloud data of cherry trees scanned in 2016, 2018 and 2020, displayed in each coordinate axis region.

[0065] Figure 4 The middle (d) plot shows the point cloud data of cherry trees scanned in 2016, 2018 and 2020, displayed in each coordinate axis region.

[0066] Figure 4 The middle (e) plot shows the point cloud data of ginkgo trees scanned in 2016, 2018 and 2020 in each coordinate axis region.

[0067] Figure 4Fig. 6 is a diagram showing point cloud data scanned in 2016, 2018 and 2020 for a ginkgo tree in each coordinate axis region.

[0068] Figure 4 Fig. 7 is a diagram showing point cloud data scanned in 2016, 2018 and 2020 for a liriodendron tree in each coordinate axis region.

[0069] Figure 4 Fig. 8 is a diagram showing point cloud data scanned in 2016, 2018 and 2020 for a liriodendron tree in each coordinate axis region.

[0070] Figure 4 Fig. 9 is a diagram showing the results of fitting the diameter at breast height for four cherry trees.

[0071] Figure 4 Fig. 10 is a diagram showing the results of fitting the diameter at breast height for four cherry trees.

[0072] Figure 4 Fig. 11 is a diagram showing the results of fitting the diameter at breast height for four liriodendron trees.

[0073] Figure 4 Fig. 12 is a diagram showing the results of fitting the diameter at breast height for four liriodendron trees.

[0074] Figure 5 Fig. 13 is a diagram showing the deep learning network constructed in the present invention for predicting tree growth parameters.

[0075] Figure 6 Fig. 14 is a diagram showing the loss function curve for the deep learning network constructed in the present invention for predicting cherry tree growth parameters.

[0076] Figure 6 Fig. 15 is a diagram showing the loss function curve for the deep learning network constructed in the present invention for predicting ginkgo tree growth parameters.

[0077] Figure 6 Fig. 16 is a diagram showing the loss function curve for the deep learning network constructed in the present invention for predicting liriodendron tree growth parameters.

[0078] Figure 6 Fig. 17 is a diagram showing the loss function curve for the deep learning network constructed in the present invention for predicting liriodendron tree growth parameters.

[0079] Figure 7 Fig. 18 is a diagram showing the comparison results of the predicted values and the measured values of the diameter at breast height for the four trees in the test data.

[0080] Figure 7 Fig. 19 is a diagram showing the comparison results of the predicted values and the measured values of the tree height for the four trees in the test data.

[0081] Figure 7 Figure C is a comparison chart of the predicted values and the measured values of the four tree crown widths in the test data. DETAILED DESCRIPTION

[0082] The specific embodiments of the present application are further described below with reference to the accompanying drawings:

[0083] Tree growth parameters are of great significance in forest character assessment, forest carbon sink measurement, and optimization of forest management strategies. Today, accurately extracting forest growth parameters from LiDAR data and predicting future tree growth parameters will provide technical support for the digital development of forestry. This study proposes a method for extracting and predicting forest growth parameters based on artificial intelligence, which takes four tree species in Nanjing Forestry University, including cherry, ginkgo, tulip tree, and metasequoia, as experimental objects. First, airborne LiDAR is used to obtain point cloud data of the four tree species plots, and single tree point cloud is extracted through single tree segmentation algorithm. Second, based on the circle fitting and Gaussian filtering method, the diameter at breast height, tree height, crown width, and other growth parameters are automatically extracted from the single tree point cloud in 2016, 2018, and 2020, supplemented by plot survey data (2015, 2017, 2019, 2021, and 2022), to construct the single tree growth parameter time series of different tree species as the training sample set of the deep learning network. Finally, a deep learning network for predicting forest growth parameters is constructed by two layers of GRU (Gated Recurrent Unit), and the attention module is introduced to make up for the deficiency of traditional RNN (Recurrent Neural Network) in capturing long-term dependencies. The network takes the single tree growth parameter time series from 2015 to 2021 as input, relies on the training data and stochastic gradient descent algorithm to make the network parameters approximate the real tree growth, and is used to predict the growth parameters of single trees in 2022. Specifically, tree growth parameters are measured in July-August of each year from 2015 to 2022, and a total of 318 trees are constructed into growth parameter time series, each of which contains 8 sets of growth parameters of the same tree from 2015 to 2022. At the same time, an improved recurrent neural network combined with attention mechanism is designed to predict the growth parameters, taking the data of the first 7 time nodes (2015-2021) in the time series as input to predict the tree growth parameters in 2022, and comparing the prediction results with the field measurement data to provide technical support for the growth trend analysis, garden management, and management of different tree species. The hybrid deep learning network in this paper can effectively predict the future growth of trees and improve the prediction accuracy of forest growth parameters. The introduction of GRU and attention mechanism has a certain robustness in the time series prediction of forest parameters, providing a new idea for the intelligent management and visualization analysis of forests. The flowchart of this study is as follows: Figure 1As shown.

[0084] 1. Materials and Methods.

[0085] 1.1 Study Area and Data Collection:

[0086] The study area is located at the Xinzhuang main campus of Nanjing Forestry University in Nanjing, Jiangsu Province, China (32°4'49"N", 118°48'45"E). Figure 2 As shown, Figure 2 The left image is a satellite image of Nanjing, with a yellow rectangle marking the study area located at Nanjing Forestry University, Nanjing City, Jiangsu Province, China. The middle image is a map of the Nanjing Forestry University campus obtained from Google Earth. Each colored rectangle corresponds to the location of the selected study area. The four images on the right are marked with rectangles of corresponding colors, corresponding to real-world photographs of their respective tree species. The study area, covering 1257 acres, is nestled against Zijin Mountain in Nanjing and adjacent to Xuanwu Lake. The region has a subtropical monsoon climate, characterized by mild temperatures and distinct seasons. The average annual temperature is approximately 15.4℃, with significant monsoons, abundant sunshine and heat, and ample rainfall, averaging about 1106 mm per year. These inherent climatic advantages allow the trees to thrive in this region. Various tree species are distributed in different locations on campus, including cherry blossom (Prunus subg. Cerasus sp.; Sakura), dawn redwood (Metasequoia glyptostroboides; Chinese fir), ginkgo (Ginkgo Biloba L.; Ginkgo), tulip tree (Liriodendron chinense (Hemsl.) Sarg.; Liriodendron), crape myrtle (Lagerstroemia indica L.), weeping willow (Salix babylonica L.), and sycamore (Platanus Orientalis Linn.). This study selected four plants—dawn redwood, cherry blossom, ginkgo, and tulip tree—as the main research subjects. Dawn redwood has a tall and straight trunk, flat branches, a conical crown, drooping twigs, and spirally arranged leaves with few overlapping branches. Cherry blossom has elliptical or obovate leaves that are relatively dense, and a larger crown, but its height is shorter than that of dawn redwood. Young and mature ginkgo trees have conical crowns, while older trees have broad oval crowns. The leaves are fan-shaped and relatively sparse, making their branches more prominent than those of dawn redwood and cherry blossom. Liriodendron tulipifera is a national second-class protected plant. It has an upright shape, shawl-like leaves with smooth or slightly curved sides, and a relatively moderate crown width.

[0087] The cherry blossoms are mainly distributed on both sides of the sports field road and are arranged in a row. The distance between each tree is 3 meters, and the height of each tree is about 5-8 meters. The length of the road is about 200 meters. The cherry blossoms on the side close to the sports field receive more light, so they grow better. The metasequoia is mainly distributed on the side of No. 9 teaching building, and the overall distribution is rectangular. The distance between each tree is 2 meters, and the height of each tree is about 26 meters. It is densely distributed, and there is mutual shielding between the tree crowns. The total area of the forest is about 500 square meters. The ginkgo trees are distributed on both sides of the road in the residential area and on the east side of the dormitory. The distance between each tree is about 4 meters, and the height of each tree is about 13 meters. The total length of the two roads is about 1000 meters. Because the crown width of ginkgo is relatively small compared to the planting distance, there is basically no shading between the trees. The Liriodendron tulipifera is mainly distributed on both sides of the road behind the library, and is symmetrically distributed. The distance between each tree is about 5 meters, the height of each tree is about 16 meters, and the length of the whole road is about 500 meters. Generally, various tree species are distributed relatively concentratedly, in a linear or rectangular shape, and grow very well, which is conducive to subsequent prediction of their growth.

[0088] 1.2, Data acquisition and preprocessing:

[0089] In this study, some growth parameters of four tree species in the study area from 2015 to 2022 were collected, including diameter at breast height, tree height, crown width, etc. Data collection was conducted once a year from July to August. The growth parameters in 2016, 2018 and 2020 were extracted from laser point cloud data, and the growth parameters of the remaining years were obtained by artificial sample plot measurement. When measuring, the tree height was measured using Vertex IV height gauge, the diameter at breast height was measured at 1.3 m outside the tree bark using a diameter gauge, and the width of the tree crown in the north-south direction and the east-west direction was measured and the average value was taken as the crown width.

[0090] The growth parameters of trees in 2016, 2018 and 2020 were extracted from laser point cloud data of the study area, which was obtained by Velodyne HDL-32E laser scanner (see Table 1 for specific parameters). The Velodyne HDL-32E laser scanner was carried by a drone and flew at a vertical height of 166 meters above ground at a speed of 25 kilometers per hour. The flight line side degree was 55%. The scanner repeatedly emitted pulses with a wavelength of 903 nm at a frequency of 15.7 kHz. The vertical field of view angle was +10.67° to -30.67°, and the angular resolution was 1.33°. The horizontal view angle was 360°, and the angular resolution was 0.1° to 0.4°. During data acquisition, the average error of 3 years was about 0.007 m through 13 ground control point measurements.

[0091] Table 1 is information about Velodyne HDL-32E:

[0092]

[0093] The original point cloud data obtained by airborne laser radar includes ground points, object points (vegetation), and noise points. The noise of the data mainly includes high-order gross errors and low-order gross errors. The collected original point cloud is denoised by using a Gaussian filtering method, and the denoised point cloud is divided into above-ground points and ground points by using a cloth simulation filtering (CSF) method. In order to eliminate the influence of terrain undulation in subsequent research as much as possible, the original point cloud is subjected to elevation normalization processing.

[0094] 1.3, single tree segmentation and extraction of single tree growth parameters:

[0095] After completing the preprocessing of the original point cloud data, the crown width extraction and single tree segmentation of the obtained forest point cloud are needed. Here, we first use a method combining a double Gaussian filter and a tree tip screening strategy to extract the crown width of the trees in the study area, Figure 3 The tree top detection and crown width extraction results of four tree species, Metasequoia glyptostroboides, Prunus serrulata, Ginkgo biloba and Talus podocarpus, from airborne laser point cloud data in 2016, 2018 and 2020 are shown. Figure 3 The tree top detection and crown segmentation of some sample plots are shown. Figure 3 (a), (b), (c), (d) represent Metasequoia glyptostroboides, Prunus serrulata, Talus podocarpus and Ginkgo biloba sample plots respectively. (1), (2), (3) correspond to three stages of laser point cloud data (i.e. point cloud data of the study area in 2016, 2018 and 2020), magenta circles and black dividing lines are the detected tree tops and the crown boundaries of each tree respectively. Among the four trees, the height of Prunus serrulata is approximately 5-8m, so the elevation color display and change are not obvious under the unified elevation colorbar.

[0096] Then, based on the watershed algorithm for extracting the crown center, single tree segmentation is performed on the above-ground point cloud by combining manual visual inspection and result correction. Figure 4 The point clouds of the same tree after single tree segmentation of airborne laser scanning data in 2016, 2018 and 2020 are shown. Figure 3 and Figure 4 It can be seen from the above that after two years of growth, the diameter at breast height, tree height and crown width of the same tree have increased to different extents. Figure 4 The single tree separation and diameter fitting schematic diagram is shown, which shows the complete point cloud of 18 trees and the diameter fitting results of four trees for each tree species, i.e. Figure 4 (a), (b), (A1-A4) in Metasequoia glyptostroboides, Figure 4 (c), (d), (B1-B4) in Prunus serrulata, Figure 4 (e), (f), (C1-C4) in Ginkgo biloba, Figure 4Middle (g), (h), (D1-D4): Liriodendron chinense. The point cloud data of the same tree scanned in 2016, 2018 and 2020 is displayed in each coordinate axis area. Among them, the pink, green and blue point clouds respectively represent the point cloud data in 2016, 2018 and 2020. The black represents the ground point cloud. The right upper corner of each coordinate axis is marked from top to bottom with the tree height measured from the laser point cloud in 2020, 2018 and 2016, respectively, in meters (m).

[0097] Next, it is necessary to extract the remaining growth parameters from the single tree point cloud data obtained by segmentation, including diameter at breast height and tree height. We combined the least squares fitting method and the random Hough transform method to extract the diameter at breast height of the tree trunk section 1.3 meters above the ground, wherein the least squares fitting method is used to preliminarily obtain the diameter at breast height and the random Hough transform method is used to verify the preliminarily obtained diameter at breast height. At the same time, Figure 3 In the middle, there are a few single tree point clouds that cannot collect branches due to shielding problems. For data that cannot extract the diameter at breast height due to this problem, we use manual re-measurement to ensure the integrity of the data set.

[0098] The extraction of the height of the single tree is based on the ground point cloud data, and the difference between the ground height and the highest point of the pre-processed single tree point cloud data is taken as the height of the single tree. Table 3 shows the different growth parameters of four kinds of trees obtained from laser point cloud data and sample plot measurements from 2015 to 2022. According to the different data collection times, the obtained forest growth parameters are sorted according to the data collection time to obtain a time series set of the growth parameters of different tree species of single trees. And divide 60% of this time series set into a training set and 40% into a test set, which are used for the training and verification of the cascading recurrent neural network described in Section 1.4.

[0099] Table 2 is the growth parameters of four kinds of trees in different years obtained from laser point cloud and sample plot measurements:

[0100]

[0101]

[0102] Note: CW s-n : Average crown width in north-south direction; CW e-w : Average crown width in east-west direction; H: Average tree height; DBH: Average diameter at breast height.

[0103] 1.4, Improved cascading recurrent neural network:

[0104] With the continuous application breakthroughs of new generation artificial intelligence technology, every industry is facing a revolutionary change. In the field of forestry, the application of artificial intelligence is not only an important measure for the development of smart forestry, but also provides new opportunities for innovation and development and technological power for precise management, ecological assessment, carbon sink measurement, and forest right transaction. The typical neural network currently applied in the prediction field is the long short-term memory (LSTM) network and its variants. LSTM is a special recurrent neural network (RNN). Due to the defects of RNN in processing long-term memory, LSTM emerged to solve the problem of gradient explosion in long sequence training. Compared with simple neural network layers, LSTM has a unique gate mechanism, and there is a special interaction between these gates. A common LSTM consists of a cell, an input gate, an output gate, and a forget gate. The first step of LSTM is to decide which information we want to discard from the cell state, the next step is to decide which new information to store in the cell state, then update the previous cell state value, and finally determine the value we want to output. At the same time, LSTM also has many variants, one of which is the gated recurrent unit (GRU). The essence of LSTM and GRU is to introduce the concept of cell. The cell remembers the value at any time interval, and multiple gates regulate the information flow inside and outside the cell to determine which states should be retained and which states should be forgotten.

[0105] The overall framework of the forest growth parameter prediction model proposed in this paper is shown in Figure 5 Figure 5 The deep learning network for tree growth parameter prediction constructed for this study consists of two layers of bidirectional gated recurrent unit (Bi-GRU) network modules, an attention module, and a fully connected neural network module. Among them, The input parameters of the first layer Bi-GRU network module (referred to as the first layer Bi-GRU module) are the data of the first 7 time nodes in the time series of a single tree (i.e., the growth parameters of a single tree from 2015 to 2021), each node's data includes the tree's diameter at breast height, tree height, and crown width, so the size of each input parameter is 3x1. Each input parameter of the Bi-GRU network module will have a corresponding output, The corresponding output is As shown in Figure 5 The output of the first layer Bi-GRU network module is transmitted to the second layer Bi-GRU network module (referred to as the second layer Bi-GRU module), i.e. The input parameters of the second layer Bi-GRU network module are the outputs H 1 ,H 2 ,…,H 7 ​At the same time, the incoming attention module is also introduced to make up for the problem of insufficient capture of long-term dependence of Bi-GRU. In the attention module, the attention score a between each pair of inputs is calculated by the dot product attention mechanism u,v (1≤u,v≤7), and is normalized by the softmax function. Finally, the sum of the product of the attention score and the extracted input parameter feature is taken as the output of the attention mechanism module, and is sent to the fully connected neural network module to output the prediction value of the tree growth parameter in 2022. The network is trained by the corresponding time series of 190 trees in the training set, including 52 cherry trees, 50 ginkgo trees, 45 metasequoia trees, and 43 ginkgo trees. The training process is as follows: the data of the first 7 time nodes of each time series of a certain tree species is taken as the input, the tree parameter growth feature of the tree species is extracted step by step in the GRU module and the attention module, and the feature vector set is mapped to the prediction value of the 8th time node of the corresponding sequence by the fully connected neural network. The real measurement value in 2022 and the loss function are combined to update the network parameters by gradient descent algorithm, and the iteration is repeated until the loss in the training set reaches a low level. Finally, the growth parameter prediction model of cherry, ginkgo, metasequoia and ginkgo is obtained.

[0106] 1.4.1, bidirectional gated recurrent unit in deep learning network:

[0107] The introduction of GRU effectively solves the problem of gradient explosion in the training of recurrent neural network. The constructed deep learning network includes two layers of Bi-GRU modules. The hidden size of the first layer Bi-GRU module (i.e. Figure 5 Bi-GRU module 1 shown by light blue rectangle) is set to 30, and the hidden size of the second layer Bi-GRU module (i.e. Figure 5 Bi-GRU module 2 shown by light green rectangle) is set to 50. The two layers of Bi-GRU modules have similar structures. Taking Bi-GRU module 1 as an example, the module includes two sequences in opposite directions (i.e. forward propagation sequence and backward propagation sequence). The output of each GRU unit in the backward propagation sequence is denoted as The output of each GRU unit in the forward propagation sequence is denoted as The output vectors corresponding to each time node are obtained by vertically splicing the outputs of the two GRU units at each time node. At the same time, the output of the module will also be taken as the input data of Bi-GRU module 2 (as shown in the area surrounded by the red dashed box in Figure 5 ). The input vector IN t of each GRU unit of Bi-GRU module at time t (t≤7) is composed of two vectors X t , Ht-1 It is formed by vertical splicing, X t This is the input vector of the Bi-GRU module at time t, i.e., the three growth parameters of a certain tree at time t, with a size of 3×1, H. t-1 It is the output vector of the GRU unit at time t-1, and H in Bi-GRU module 1. t-1 The size is 30×1, and the H in Bi-GRU module 2 t-1 The size is 60×1, and the time step length of the Bi-GRU module is the same as the length of the input time series (7). Therefore, IN t =[X t H t-1 ], Figure 5 Input vectors of Bi-GRU module 1 and Bi-GRU module 2 The subscripts 1 and 2 indicate that it is the input vector of the first or second layer Bi-GRU module. GRU simplifies the gate mechanism based on LSTM; the input gate and forget gate in LSTM are combined into one, called the update gate Z. t This is used to determine how much information from the previous time step and the current time step should be passed to the next time step. Another gate is called the reset gate R. t The formulas (1) and (2) are used to determine how much information from the previous time step should be forgotten. They show the calculation process of the update gate and the reset gate.

[0108] Z t =σ(W z *IN t +b z (1);

[0109] R t =σ(W R *IN t +b R (2);

[0110] In the formula W z W R The weight matrix has sizes of 30×33 and 50×80 in the first-layer Bi-GRU module and the second-layer Bi-GRU module, respectively. z b R The bias vector has sizes of 30×1 in the first-layer Bi-GRU module and 33×1 in the second-layer Bi-GRU module. σ is the sigmoid function, which transforms the data into values ​​within the range of 0-1, thus serving as a gating signal. In the formula... For element-wise multiplication, * represents matrix multiplication, and · represents vector dot product. This refers to vector scalar multiplication, and the same applies below.

[0111] After obtaining the corresponding gating signal, first use the reset gate R t information H t-1 "forget", that is, determine which information should be discarded, H t-1 The result after discarding the "forget" data is denoted as H t-1,' , as shown in equation (3).

[0112]

[0113] At this time, H t-1,' is concatenated with the input X t at the current time to input the tanh neural network layer to obtain the state of the current time after "forgetting" , as shown in equation (4).

[0114]

[0115] where W h is the weight matrix, which has the same size as W z , W R , and b h is the bias vector. Finally, the update gate Z t decides how much data in the current state can be "remembered" and how much data in the previous time step should be "forgotten", obtaining the output vector at the current time point, denoted as H t , which is used as the input of the attention module, as shown in equation (5).

[0116]

[0117] Figure 5 The light orange rectangle in the Bi-GRU module 1 shows the calculation process of the last GRU unit in the forward propagation sequence.

[0118] 1.4.2, Attention module in deep learning network:

[0119] To further improve the prediction efficiency of the model, we also introduce an attention mechanism in the network (shown in the light yellow rectangular frame in Figure 5 ). The attention mechanism (referred to as attention module) can focus on the information that is considered more critical to the process when a large amount of information is input, while reducing the attention of other unimportant information, making the information more cleverly encoded and connected, and improving the efficiency of the processing process. Therefore, we introduce the attention mechanism in the model to improve the prediction efficiency and reduce the attention of some unreasonable tree growth parameters caused by errors in data collection, and filter out more valuable information.

[0120] The input of the attention module is the output of the second layer Bi-GRU network H = [H1 H 2 ,...,H 7 Since the output vector dimension of the Bi-GRU network at each time step is twice the size of the hidden layer, the size of matrix H is 100×7. The attention module first considers the correlation between the input data, given two input data H. m and H n Attention score α between m,n (1≤m,n≤7; m,n∈N), H m and H n Let H be any two vectors in H, with a size of 100×1. Generally, dot product attention and additive attention are two methods for calculating the attention score between input data. Dot product attention is more commonly used because it has lower time and space complexity and the results are similar to those of additive attention. Any input vector H in the dot product attention mechanism... m First, with the weight matrix W q W k and W f Multiplying (weight matrix size is 60×100), the q in the result m and k m It is used to calculate the attention score of two input vectors, vector v m Represents the input vector H m Features, q m k m and v m All sizes are 60×1 (in Figure 5 The black dashed box and the corresponding cyan rectangle in the image represent the 7th input vector H. 7 (And the corresponding calculation methods for q7, k7, v7). From a matrix perspective, the matrix Q = [q1, q2, ..., q7] composed of the vectors q corresponding to each input vector can be derived from the input matrix H = [H 1 H 2 ,...,H 7 ] and matrix W q Multiplying them yields the following: Similarly, we can obtain matrices K = [k1, k2, ..., k7] and V = [v1, v2, ..., v7]. The size of matrices Q, K, and V is 60 × 7. Then q... m k n v r The specific calculation process is shown in formulas (6)-(8).

[0121] q m =W q *H m Q = W q *Η (6);

[0122] kn =W k *H n K = W k *Η (7);

[0123] v r =W f *H r V = W f *Η (8);

[0124] Secondly, the two input vectors H m and H n The corresponding vector q m and k n The inner product of the vectors is used as vector H. m and H n Attention score α between m,n α m,n Real numbers. As shown in formula (9).

[0125]

[0126] As can be seen from formula (9), the attention mechanism typically calculates the vector H. m And its own attention score. When a certain vector H m When the attention scores for all other vectors (including itself) have been calculated, we can obtain α. m,1 ,α m,2 ,...,α m,7 A set of 7 ratings can be used to obtain the weighting coefficient α' by substituting the data into the softmax function. m,1 ,α' m,2 ,...,α' m,7 As shown in formula (10).

[0127]

[0128] Finally, relying on the weight coefficients of each input vector to other input vectors, key information is extracted from the input vector set to calculate the output vector, i.e., the input vector H. m and its weight coefficients α' with other input vectors m,1 ,α' m,2 ,...,α' m,7 Perform scalar multiplication on the feature vectors v1, v2, ..., v7 corresponding to the input vectors and sum them. At this point, we can obtain each input vector H. m The corresponding attention module has an output vector a1, a2, ..., a7, each with a size of 60×1, as shown in formula (11).

[0129]

[0130] After obtaining the output vector sequence of the attention mechanism, it is summed and brought into a fully connected neural network layer, which converts the sum vector into the final output vector, i.e., the 3 growth parameters of a single tree at the 8th time point (2022), to realize the prediction of tree growth parameters.

[0131] 1.4.3, Loss function:

[0132] The deep learning network takes Log-Cosh as the loss function of the network, as shown in formula (15). During the training process, the loss function is defined as follows.

[0133]

[0134] In formula (15), N represents the number of samples in training, cosh represents the hyperbolic cosine function, represents the prediction value of the i-th sample under the current network parameters, y i represents the true value corresponding to the i-th sample.

[0135] 2, Results and analysis.

[0136] 2.1, Analysis of growth parameter extraction results:

[0137] We segmented single trees and extracted growth parameters from laser point cloud data in 2016, 2018 and 2020. To evaluate the accuracy of growth parameter extraction, we randomly selected 10 trees in each tree species each year, a total of 120 samples for statistical analysis. For each sample, we compared the growth parameters extracted from point cloud data with those obtained by manual measurement. The accuracy (RMSE) of diameter at breast height was 1.18 cm, and the accuracy (RMSE) of tree height and crown width was 0.48 m and 0.39 m, respectively. Statistical tests confirmed that although the growth parameter extraction method based on point cloud data slightly overestimated and underestimated in single tree measurement, there was no significant difference between the field measurement data and the growth parameters extracted from point cloud data.

[0138] In addition, we also analyzed the growth characteristics of different tree species and compared the advantages and disadvantages of manual measurement method and point cloud data extraction method. We found that one of the main problems of extracting crown width from point cloud is the occlusion of adjacent crowns. For tree species with larger crowns and more branches, traditional measurement methods are easily limited by viewing angle and obstacles, while point cloud data extraction methods can more comprehensively reflect the growth characteristics of trees. The extraction of diameter at breast height is also affected by the shape of the trunk, because some tree trunks are not completely round, and for tree species with thin branches, the accuracy of point cloud data extraction of diameter at breast height needs to be further improved.

[0139] 2.2, Training and testing results of the deep learning network:

[0140] The training and testing of the deep learning network were performed on a server equipped with an Intel(R) Core(TM) i7-10750H CPU@2.80GHz processor (Intel Inc., Santa Clara, CA, USA) and 16 GB-RAM running Windows 10-64-bit. Since the training of the deep learning network requires a large amount of sample data to optimize the network parameters in the forward and backward propagation processes, in order to improve the computational efficiency, a NVIDIA GTX 2080Ti GPU (NVIDIA Inc., Santa Clara, CA, USA) was used instead of the CPU to reduce the training time of the algorithm. During the construction of the deep learning network, we tried various parameter settings, Figure 6 The network loss curves under different batch sizes are shown. During the network training process, different batch sizes will affect the network training results. When the batch size is small, there is a large difference between batches, which will cause the loss curve to oscillate more severely and the convergence speed will be slower. In this experiment, when the batch size is set to 16, the network loss curve has a large amplitude of oscillation. When the batch size is large, the difference between batches is relatively small, the loss curve is relatively smooth, and the convergence speed is relatively fast. However, due to the smooth gradient descent, it is easy to fall into a local minimum. In this experiment, when the batch size is set to 64, the network converges to a local minimum. Finally, after adjusting the parameters of the network, we set the learning rate to 0.0001, the batch size to 32, the training epoch to 300, and use the stochastic gradient descent algorithm to optimize the network.

[0141] With the continuous network training, the training loss of the training sample shows a downward trend, indicating that the training of the deep learning network is a global optimization process, and the network is constantly approaching the true growth curve of the forest growth parameters. The training loss decreases significantly in the first 50 Epochs, indicating that the parameters of each neuron in the deep network are rapidly reaching the requirements for predicting forest growth parameters. Due to the inevitable occlusion phenomenon of point cloud data, the extraction of some tree growth parameters may be offset compared to their true values, which will lead to an increase in the complexity of the time series, weaken the learning efficiency of the deep network during training, and cause fluctuations in the loss function value. However, the overall downward trend of the curve indicates that the deep learning network constructed in this paper has good convergence during the training process, and the loss of the training sample reaches 0.03 after 300 Epochs, indicating that the deep learning network constructed in this paper has the ability to predict forest growth parameters.

[0142] 2.3, Growth parameter prediction result analysis:

[0143] The time series of several trees in the test set are substituted into the trained model, i.e. the growth parameters of a single tree from 2015 to 2021 are brought into the model to calculate the growth parameters of the corresponding tree in 2022. Table 3 gives the actual measured values and deep network prediction values of the growth parameters of cherry, ginkgo, tulip tree and metasequoia test samples, including diameter at breast height, tree height and crown width. The data in Table 3 are the growth parameters of the test set in 2022, the average value of the actual measured value and the predicted value of a growth parameter of a tree species is the average value of the growth parameter in 2022, and the positive and negative error values represent the maximum difference of the growth parameter in 2022 data from the average value. At the same time, through three comparison indexes, the determination coefficient R 2 , root mean square error (Root Mean Square Error, RMSE) and relative root mean square error (Relative Root Mean Square Error, rRMSE) are used to quantitatively analyze the effectiveness of the research method. Figure 7 The specific parameter comparison results are given. Figure 7 The comparison chart of the prediction results of various tree growth parameters in the test data and the field measurement data is given.

[0144] Table 3 compares the deep network predicted forest growth parameters with the field measured values:

[0145]

[0146] Figure 7 Figure (a) shows the comparison results of the prediction values and measured values of the diameter at breast height of the four trees by the research method. The comparison results of the deep network prediction values and field measurements of cherry, tulip tree and metasequoia are (R 2= 0.94, RMSE = 1.38 cm, rRMSE = 7.14%), (R 2 = 0.87, RMSE = 0.97 cm, rRMSE = 2.69%) and (R 2 = 0.88, RMSE = 1.07 cm, rRMSE = 2.53%). Compared with the above three trees, the network performed poorly in predicting the diameter at breast height of ginkgo trees (R 2 = 0.81, RMSE = 2.49 cm, rRMSE = 10.06%). This is mainly because the ginkgo trees have dense branches and leaves during data collection, and the point cloud quality of the crown part obtained by the airborne laser radar is high. However, due to the shielding phenomenon of leaves, the point cloud data at the breast height is missing, which leads to a large deviation between the true value and the part of the time series, and thus affects the prediction result of the breast height of ginkgo trees.

[0147] Figure 7 Fig. 3(b) shows the comparison between the predicted and measured values of the tree height of the four trees by the method of the present study. It can be seen that good accuracy is obtained for the four different trees. Among them, ginkgo (R 2 = 0.95, RMSE = 0.31 m, rRMSE = 2.41%) and Talus (R 2 = 0.91, RMSE = 0.28 m, rRMSE = 1.8%) achieved better results among the four tree species. This is mainly because the planting distance of ginkgo and talus in the selected experimental plot is larger, and the accuracy of tree height measurement is higher. The mutual shielding between trees is not significant when airborne laser scanning is performed, so that the quality of the crown part point cloud obtained is higher, and the error between the extracted tree height from the point cloud data and the true value is smaller, so that the network prediction result has higher accuracy.

[0148] Figure 7 Fig. 3(c) shows the comparison between the predicted and measured values of the crown width of the four trees by the method of the present study. Among them, the prediction accuracy of the crown width of water fir is lower than that of other tree species (R 2 = 0.81, RMSE = 0.15 m, rRMSE = 2.83%), and there is an obvious overestimation trend. This can be explained as water fir is planted in a rectangular forest land with a small planting distance, and the growth rate of the crown width is gradually reduced due to intraspecific competition. Since the time span of the data in the training set is short, the influence between water firs is not significant, thus leading to the overestimation trend of the deep learning network for the crown width of water fir.

[0149] The comparison results of the deep learning network constructed in this paper with the linear regression based on tree age and the forest growth parameter prediction method based on LSTM network are shown in Table 4. The coefficient of determination (R 2), root mean square error (RMSE) two aspects of the performance of the algorithm. Based on the age of the linear regression method of the original data is more sensitive to outliers, a small number of outliers can make the regression accuracy decreased significantly, and the method depends on the artificial design, learning ability is limited, in the prediction of diameter at breast height, tree height and crown width R 2 are not higher than 0.75. In comparison, the LSTM solves the problem of gradient explosion in traditional RNN, and has high precision in the time series prediction of forest growth parameters, but due to the large number of parameters to be trained in the LSTM network, the time complexity of calculation is high, and the training time is long. And because the number of parameters is too large, the network is prone to overfitting, resulting in unsatisfactory training accuracy. The hybrid deep learning framework used in this paper learns the characteristics, trends and development rules of parameter changes in the original forest growth parameter time series, and combines attention mechanism to strengthen the attention to relevant information in the sequence, making up for the short board of GRU in capturing long-term dependencies, so it has higher prediction accuracy. In the prediction of diameter at breast height, tree height and crown width R 2 are not less than 0.86. Table 4 compares the growth parameter prediction accuracy of the linear regression based on tree age, the deep learning network based on LSTM and the method in this paper in the same test data:

[0150]

[0151] 2.4、Discussion:

[0152] In this paper, the growth attribute inversion and growth trend prediction of different trees are combined with GRU module and attention mechanism to construct a cascade recurrent neural network, and a growth parameter prediction model for different tree species is constructed and applied to the prediction of different growth parameters of cherry, ginkgo, tulip tree and metasequoia. It provides a technical solution for forest precision management and biomass measurement analysis. The results show that the network constructed in this study has good performance in the prediction accuracy of forest parameters, and the prediction accuracy of tree height and crown width is better than the prediction model of average height of larch plantation based on BP neural network (R2=0.84, RMSE=1.18m) and the prediction of crown width of Phoebe bournei based on mixed effect model (R2=0.85, RMSE=0.85m). The prediction accuracy of diameter at breast height is also superior to the diameter at breast height prediction model of spruce in Gansu Xishui Forest Farm based on nonlinear equation set (RMSE=1.82cm). This shows that it is feasible to use cascade recurrent neural network to predict growth based on time series data. The network can adaptively adjust the attention to different time steps, so as to better capture the information of different time steps. This research has important significance for forest protection and sustainable utilization, and also provides a reference for using deep learning technology to carry out ecological environment monitoring and prediction research.

[0153] The work in this paper has achieved high accuracy in predicting forest growth parameters, but there are still some uncertainties in some technical details, mainly in the following three aspects: (1) The data time span in this study is 8 years, however, the different stages of forest growth have a significant impact on tree development, the growth cycle of forest can be divided into four stages: juvenile, youth, prime and old age, the growth rate of forest in each stage will change over time, and the growth process of some trees often lasts for nearly ten years or even dozens of years. The short data time span in this study may lead to large errors in predicting the growth parameters of some tree species with long growth cycles. Long-term and stable observation of related forest parameters and accumulation of sequence data will be a way to solve this problem, and the correction coefficient representing the development cycle of forest and the habitat and climate data in different periods can be added to the network to improve the prediction accuracy of the model through more detailed data; (2) The structure parameters in deep learning network have a great influence on the prediction results of this paper, which is manifested as the size of the hidden layer of each module is set too large, which will prolong the training time of the network and lead to overfitting phenomenon of the network, and too small will make the network unable to achieve the expected target. Due to the low interpretability of deep learning model at present, the adjustment of parameters often depends on experience, so there is uncertainty in determining the appropriate size of the hidden layer to control the network training time within a reasonable range and achieve the expected prediction accuracy; (3) Long-term and regular measurement of forest growth parameters and proper preservation will be a necessary condition for growth parameter prediction. Deep learning network needs a large amount of reliable data for model optimization, while the current collection of growth parameters is still time-consuming and laborious, and further exploration of machine learning and deep learning to accurately measure forest growth parameters from multi-source remote sensing data will effectively improve the collection efficiency of training samples.

[0154] In summary, in future scientific research, this work can be improved in the following aspects: (1) The stages of tree growth can be further refined, and the influencing factors of tree growth parameters in each stage can be considered. For example, in the juvenile stage, trees mainly undergo vegetative growth and have strong adaptability to environmental conditions. Therefore, the influence of growth environment, such as soil and nutrients, should be considered when predicting. In the youth stage, the growth rate of trees is fast, and the branching of the crown increases rapidly. Therefore, the branching of trees and their demand for environmental conditions such as light and water should be considered in the prediction. In the mature stage, the growth rate of trees reaches a peak, but the plasticity decreases. Therefore, the influence of habitat conditions, secondary succession of forest, and interspecific and intraspecific competition on tree growth should be considered in the prediction. In the old age stage, the physiological function of trees declines significantly, and the number of new branches decreases significantly. Therefore, the health status of trees and the influence of natural stress should be considered in the prediction; (2) The network structure can be further optimized, such as using the competitive random search method to calculate the attention score to avoid the network falling into local optimization, or introducing a tensor discriminator into the GRU module to adaptively determine whether to update the hidden state information at each time step, and to reduce the influence of extreme fluctuations in the training data on the network as much as possible. In data processing, isolated forest algorithm can also be used for anomaly detection and elimination; (3) This paper has proved that it is effective and reliable to extract tree growth parameters from laser point cloud. However, due to the fact that the laser scanner is mounted on an aircraft, the point cloud of trees near the ground side often has missing data due to shielding, which will cause large errors in the extraction of some growth parameters. In the future, combining ground scanning techniques such as backpack laser radar scanning with airborne laser scanning can obtain more complete tree point cloud information, making the growth parameters extracted from the point cloud more accurate.

[0155] 3. Conclusion.

[0156] The accurate extraction and prediction of tree growth parameters play an important role in forest management, forest volume and carbon sink estimation, and forest resource cultivation. First, this study proposes an accurate extraction method for tree growth parameters based on airborne laser radar data, aiming to automatically extract tree diameter at breast height, tree height, and crown width at single tree scale in large forest areas. Second, for the extracted time series of tree growth parameters, an improved recurrent neural network is proposed to predict the future trend of tree growth parameters. The attention mechanism module is introduced to solve the long-term dependence problem of traditional recurrent neural networks. The comparison and analysis of the predicted data and the measured data of four different tree species show that the deep learning method proposed in this paper can effectively predict the future growth of trees, and the determination coefficients R 2The root mean square error (RMSE) of the predicted value and the measured value of the diameter at breast height of the four tree species is less than 2.5 cm, and the root mean square error (RMSE) of the predicted value and the measured value of the tree height and crown width is less than 0.5 m. The method proposed in the present application can automatically extract corresponding parameters such as diameter at breast height, tree height and crown width from forest laser point cloud data, and is suitable for predicting growth parameters of different tree species. It is one of the application points of the development of artificial intelligence technology in the field of smart forestry. The future research direction will focus on further building a lightweight, low sample demand, and deployable deep learning network on mobile devices, and improving the parameter extraction accuracy of different forms of forest laser point cloud data (such as ground and backpack laser radar), and the prediction accuracy of growth parameters of different tree species.

[0157] The protection scope of the present application includes but is not limited to the above embodiments, and the protection scope of the present application is subject to the claims, any replacement, deformation, improvement of the present technology easily thought by the skilled in the art falls within the protection scope of the present application.

Claims

1.A method for predicting a forest growth parameter based on a cascaded recurrent network, the method comprising: The method comprises the following steps: Step 1: collecting growth parameters of trees in a sample forest at multiple time nodes through manual measurement; Step 2: collecting laser point cloud data of the sample forest at multiple time nodes, and obtaining the growth parameters of the trees through the laser point cloud data of the sample forest; Step 3: sorting the growth parameters obtained in steps 1 and 2 according to the data collection time to obtain a time sequence set of the growth parameters of a single tree; Step 4: constructing a deep learning network for predicting forest growth parameters, taking the growth parameters of a single tree corresponding to the latest time node A in the time sequence set of the growth parameters of a single tree as the output of the deep learning network for predicting forest growth parameters, and taking the time sequence of the growth parameters of a single tree corresponding to multiple time nodes before the latest time node A in the time sequence set of the growth parameters of a single tree as the input of the deep learning network for predicting forest growth parameters, so as to train the deep learning network for predicting forest growth parameters; Step 5: sequentially obtaining the time sequence set of the growth parameters of a single tree before the to-be-measured time node according to the methods in steps 1, 2 and 3, inputting the time sequence of the growth parameters of a single tree in the set into the trained deep learning network for predicting forest growth parameters, and obtaining the growth parameters of a single tree corresponding to the to-be-measured time node; The growth parameters include diameter at breast height, tree height and crown width; The deep learning network for predicting forest growth parameters comprises two layers of Bi-GRU modules, an attention module and a fully connected neural network module; The two-layer Bi-GRU module is respectively denoted as a first-layer Bi-GRU module and a second-layer Bi-GRU module, and the first-layer Bi-GRU module and the second-layer Bi-GRU module both include a forward propagation sequence and a backward propagation sequence, and the output of each GRU unit in the backward propagation sequence of the first-layer Bi-GRU module is denoted as The output of each GRU unit in the forward propagation sequence of the first-layer Bi-GRU module is denoted as The output vectors corresponding to the time nodes are obtained by longitudinally splicing the outputs of the two GRU units corresponding to each time node, and the output of the first-layer Bi-GRU module also serves as input data of the second-layer Bi-GRU module. Each GRU unit in the Bi-GRU module takes the input vector IN at time t (t≤7) t from the vector X t and the vector H t-1 t is the input vector of the Bi-GRU module at time t, i.e., the growth parameters of a certain tree at time t, H t-1 is the output vector of the GRU unit at time t-1, IN t = [X t , H t-1 ]​ The GRU unit includes an update gate and a reset gate, a calculation formula of the update gate is: Z t = σ (W z *IN t + b z ) ; a calculation formula of the reset gate is: R t = σ (W R *IN t + b R ) ; W z is an update gate weight matrix, b z is an update gate bias vector, W R is a reset gate weight matrix, b R is a reset gate bias vector, and sigma is a sigmoid function. After obtaining the corresponding gating signal, first use the reset gate R t information H of the previous time step t-1 processing, that is: H t-1,' Compared with the current input X t After concatenating the two parts, input them into the tanh neural network layer to obtain the current state. Right now: W h W is the weight matrix of the tanh neural network layer. h Size and W z W R The same size, b h This refers to the bias vector of the tanh neural network layer. Finally, use the update gate Z. t The state at the current moment Information H from the previous time step t-1 Continuing the processing, we obtain the output vector at the current time point, denoted as H. t ,Right now: Where H t As input to the attention module; in the above formulae, is the multiplication of corresponding elements, and * is the matrix multiplication; The input of the attention module is the output H = [H 1 , H 2 ,...,H 7 ] of the second layer Bi-GRU module; Firstly, Q=[q1, q2,..., q7], K=[k1, k2,..., k7] and V=[v1, v2,..., v7] are calculated, that is: q m = W q * H m ; k n = W k * H n ; v r = W f * H r ; Q = W q *H; K = W k *H; V = W f *H; where H m , H n are any two vectors in H r , H q is an input vector, W k and W f are weight matrices; q m is a vector in matrix Q, k n is a vector in matrix K, and v r is a vector in matrix V. Compute vector H m and attention score a between H n and H m,n , i.e. a m,n = q m · k n ; An attention score matrix α between all input vectors is calculated, that is: where K is the matrix T is the transpose matrix of matrix K; Each attention score a is calculated by a softmax function m,n The corresponding weight coefficient a′ m,n That is: The output matrix A=[a1, a2,..., a7] of the attention module is obtained by multiplying the matrix V and α', that is: A=V*α'; where a m is a vector in matrix A; In the above formulae, is the multiplication of corresponding elements, * is the matrix multiplication, and · is the vector inner product, is the vector number multiplication; After obtaining the output vector sequence of the attention module, the output vector sequence is summed and input into the fully connected neural network module, so as to obtain the final output vector, that is, the growth parameters of a single tree at the eighth time node, so as to realize the prediction of the growth parameters of the tree; The loss function of the deep learning network for predicting forest growth parameters is: where N represents the number of samples in the training, cosh represents the hyperbolic cosine function, represents the predicted value of the i-th sample under the current network parameters, y i represents the true value corresponding to the i-th sample. 2.The forest growth parameter prediction method based on the cascaded recurrent network according to claim 1, characterized in that, In step 2, the growth parameters of the trees are obtained through the laser point cloud data of the sample forest, specifically: Gaussian filtering method is used to denoise the collected forest laser point cloud data, cloth simulation filtering method is used to divide the denoised point cloud into above-ground points and ground points, and height normalization processing is performed; The processed forest point cloud data is subjected to crown width extraction and single tree segmentation, and the diameter at breast height and the tree height are extracted from the single tree point cloud data obtained through the segmentation. 3.The forest growth parameter prediction method based on the cascaded recurrent network according to claim 1, characterized in that, The time nodes of the input growth parameters of a single tree of the deep learning network for predicting forest growth parameters are seven.

Images