Method, system and equipment for analyzing stand uniformity of man-made forest under different forest age conditions and medium

Abstract

The invention discloses a stand uniformity analysis method, system and device for a man-made forest under different forest age conditions and a medium, and the method comprises the steps: obtaining a single-tree segmentation model data set, constructing an initial single-tree segmentation model based on a Mask R-CNN algorithm, improving a backbone network, a feature pyramid network and a detection head of the initial single-tree segmentation model, and calculating the stand uniformity of the man-made forest under different forest age conditions. Training the improved individual tree segmentation model to obtain a trained individual tree segmentation model; and performing individual tree segmentation and individual tree age estimation of the to-be-researched artificial forest image based on the trained individual tree segmentation model, calculating the forest stand age by using an SAHI algorithm, extracting the individual tree height and the individual tree crown breadth, and calculating the uniformity of the artificial forest stand under different forest age conditions based on the extracted individual tree height and individual tree crown breadth. According to the technical scheme, accurate single-tree segmentation of the eucalyptus artificial forest and accurate estimation of the forest stand and the forest age at different stages can be achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of forestry resource management, and in particular relates to a method, system, equipment and medium for analyzing uniformity of artificial forest stands under different forest age conditions. Background Art

[0002] Forest age, as an important forest parameter, is the most commonly used and effective auxiliary variable for the inversion of eucalyptus plantation productivity. How to use modern remote sensing technology to accurately extract and timely update the forest age information of eucalyptus plantations is the basis for achieving high-precision inversion of eucalyptus plantation productivity and dynamic change monitoring.

[0003] Deep learning has been widely used in forest tree segmentation. The U-net model is used to segment individual trees on satellite and aerial images. The accuracy of satellite image segmentation is 46.3%, and the recall rate is 63.7%. The accuracy of aerial image segmentation is 52%, and the recall rate is 66.2%. Traditional crown tree segmentation models usually use RGB channels as input for individual tree segmentation. RGB images contain rich texture features, and the task of individual tree segmentation can be completed based on RGB images.

[0004] As a representative algorithm in deep learning, Mask R-CNN has been widely used in research on forest tree segmentation and tree height extraction, and has achieved good results.

[0005] Although Mask R-CNN has been widely used in forest tree segmentation and has achieved good results, the previous segmentation objects were mostly slow-growing trees such as Ginkgo and Pinus sylvestris. Eucalyptus, as a fast-growing plantation, grows fast. When it is young, it grows faster, growing 3cm a day, 1m a month, and 10m a year. Although the growth of eucalyptus trees slows down with the increase of forest age, its growth rate is still faster than other tree species. The rapid growth and diversity of eucalyptus plantations pose greater challenges to the Mask R-CNN single tree segmentation algorithm, which cannot be applied to the segmentation of single trees in eucalyptus plantations at different stages. Summary of the invention

[0006] The purpose of the present invention is to provide a method, system, equipment and medium for analyzing uniformity of artificial forest stands under different forest age conditions, so as to solve the problems existing in the above-mentioned prior art.

[0007] To achieve the above object, the present invention provides a method for analyzing uniformity of plantation stands under different forest age conditions, comprising:

[0008] Acquire a single tree segmentation model dataset, wherein the single tree segmentation model dataset includes artificial forest image data and corresponding tree height information;

[0009] An initial single tree segmentation model is constructed based on the Mask R-CNN algorithm, and a backbone network, a feature pyramid network, and a detection head of the initial single tree segmentation model are improved to obtain an improved single tree segmentation model;

[0010] Training the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model;

[0011] Based on the trained single tree segmentation model, single tree segmentation of the artificial forest image to be studied is performed to obtain the single tree segmentation results and estimate the single tree age. The stand age is calculated using the SAHI algorithm. The single tree height and single tree crown width are extracted based on the single tree segmentation results. Based on the extracted single tree height and single tree crown width, the uniformity of the artificial forest stand under different forest age conditions is calculated.

[0012] Optionally, the step of obtaining a single tree segmentation model dataset specifically includes:

[0013] Use drones to collect stereo images of artificial forests and obtain image data of artificial forests;

[0014] Based on Pix4D software, digital orthophotos, digital surface model data and digital elevation model data in artificial forest image data are obtained;

[0015] Subtracting the digital surface model data from the digital elevation model data to obtain canopy height model data;

[0016] Segmenting the digital orthophoto to obtain a plurality of segmented images, and marking a tree crown label for each of the segmented images;

[0017] The canopy height model data is mapped to each tree crown label to obtain tree height information corresponding to each single tree.

[0018] Optionally, the improvement of the backbone network, feature pyramid network and detection head of the initial single tree segmentation model specifically includes:

[0019] Firstly, the Transnext network is used as the backbone network, and then the feature pyramid network is replaced by the CARAFEFPN method. Finally, SABLHead with Seesawloss as the loss function is used as the detection head to obtain the improved single tree segmentation model.

[0020] Optionally, training the improved single tree segmentation model based on the single tree segmentation model dataset specifically includes:

[0021] The single tree segmentation model data set is divided into a training set and a validation set based on preset rules, the improved single tree segmentation model is trained based on the training set, the trained single tree segmentation model is evaluated based on the validation set, the single tree segmentation model is optimized based on the evaluation result, the iterative training and optimization process is repeated until the preset number of iterations is reached, and the trained single tree segmentation model is output.

[0022] Optionally, the calculation of the uniformity of the plantation stands under different forest age conditions based on the extracted single tree height and single tree crown width specifically includes:

[0023] The age of individual trees is determined based on the results of individual tree segmentation, the age of the stand is determined using the SAHI algorithm, and the height and crown width of individual trees are extracted based on the results of individual tree segmentation;

[0024] Based on the stand age, tree height and tree crown width, PH 3 The Pvar50 index and Pvar50 index were used to calculate the uniformity of plantation stands under different stand ages.

[0025] A system for analyzing uniformity of plantation stands under different forest age conditions, comprising:

[0026] A data acquisition module is used to obtain a single tree segmentation model data set, wherein the single tree segmentation model data set includes artificial forest image data and corresponding tree height information;

[0027] A model training module is used to construct an initial single tree segmentation model according to the Mask R-CNN algorithm, improve the backbone network, feature pyramid network and detection head of the initial single tree segmentation model to obtain an improved single tree segmentation model; train the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model;

[0028] The single tree segmentation module is used to perform single tree segmentation of the artificial forest image to be studied according to the trained single tree segmentation model, obtain the single tree segmentation results and estimate the single tree age, calculate the stand age using the SAHI algorithm, extract the single tree height and single tree crown width based on the single tree segmentation results, and calculate the uniformity of the artificial forest stand under different forest age conditions based on the extracted single tree height and single tree crown width.

[0029] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for analyzing uniformity of artificial forest stands under different forest age conditions.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for analyzing uniformity of artificial forest stands under conditions of different forest ages.

[0031] The technical effects of the present invention are:

[0032] The improved Mask R-CNN model provided by the present invention can achieve accurate single tree segmentation of eucalyptus plantations at different stages, can achieve accurate estimation of the stand age of eucalyptus plantations, and can calculate the uniformity of plantations under different stand ages based on the single tree segmentation results and stand age. The present invention can not only provide accurate single tree crown width and stand age for subsequent eucalyptus plantations yield estimation and management, but also provide methods and technical guidance for single tree segmentation and stand age judgment of other forest types. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 A schematic diagram of the location of the research area in an embodiment of the present invention;

[0036] Figure 2 This is a box diagram of the results of extracting the height of individual eucalyptus trees of different forest ages in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the age distribution results of eucalyptus plantations based on CHM extraction and tree height determination in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the Mask R-CNN structure;

[0039] Figure 5 Schematic diagram of the improved Mask R-CNN structure in an embodiment of the present invention;

[0040] Figure 6 Schematic diagram of the GLU structure in an embodiment of the present invention;

[0041] Figure 7 CARAFEFPN is a schematic diagram of the structure of an embodiment of the present invention;

[0042] Figure 8 is a flow chart of the SAHI algorithm in an embodiment of the present invention;

[0043] Fig. 9is the PR curve of the ablation experiment model in the embodiment of the present invention; wherein, Fig. 9 (a) is the PR curve of model 1 in the embodiment of the present invention; Fig. 9 (b) is the PR curve of model 2 in the embodiment of the present invention; Fig. 9 (c) is the PR curve of model 3 in the embodiment of the present invention; Fig. 9 (d) is the PR curve of model 4 in the embodiment of the present invention;

[0044] Fig.10 Schematic diagram of the results of single tree segmentation in eucalyptus plantations at different forest ages in an embodiment of the present invention; wherein, Fig.10 (a) is the real label in the embodiment of the present invention; Fig.10 (b) is the single tree segmentation result of the Mask R-CNN model; Fig.10 (c) is the single tree segmentation result of the improved Mask R-CNN model in an embodiment of the present invention;

[0045] Fig.11 is the confusion matrix in the embodiment of the present invention;

[0046] Fig.12 The CHM data, single tree segmentation and stand age prediction results of the regional eucalyptus plantation in the embodiment of the present invention are shown in FIG. Fig.12 (a) is the CHM data of the regional eucalyptus plantation in the embodiment of the present invention; Fig.12 (b) shows the individual tree segmentation and stand age prediction results of regional eucalyptus plantations;

[0047] Fig.13 Schematic diagram of DOM data, CHM data and prediction results of region A in an embodiment of the present invention; wherein, Fig.13 (a) is the DOM data of area A in the embodiment of the present invention; Fig.13 (b) is the CHM data of region A in an embodiment of the present invention; Fig.13 (c) is the prediction result of region A in the embodiment of the present invention;

[0048] Fig.14 It is an implementation flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0050] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0051] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art generally understood by the present invention. Although the present invention describes only preferred methods, any method similar or equivalent to that described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0052] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0053] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0054] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] Embodiment 1

[0056] like Figure 1 - Fig.14As shown, this embodiment provides a method for analyzing uniformity of artificial forest stands under different forest age conditions, including: obtaining a single tree segmentation model data set, the single tree segmentation model data set including artificial forest image data and corresponding tree height information; constructing an initial single tree segmentation model based on the Mask R-CNN algorithm, improving the backbone network, feature pyramid network and detection head of the initial single tree segmentation model to obtain an improved single tree segmentation model; training the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model; performing single tree segmentation of the artificial forest image to be studied based on the trained single tree segmentation model to obtain single tree segmentation results and single tree forest ages, using the SAHI algorithm to infer forest age, extracting single tree height and single tree crown width based on the single tree segmentation results, and calculating the uniformity of artificial forest stands under different forest age conditions based on the extracted single tree height and single tree crown width.

[0057] Eucalyptus, as one of the three major fast-growing tree species in the world, not only provides timber for a variety of industries, but also makes significant contributions to carbon sequestration and climate change mitigation. However, due to the rapid growth and diverse growth changes of eucalyptus, it brings certain difficulties to the accurate segmentation of individual trees. At the same time, how to achieve the extraction of the age of eucalyptus plantations while segmenting individual trees also poses a greater challenge to the individual tree segmentation algorithm. Based on this, this embodiment, based on the Mask R-CNN algorithm framework, firstly, uses Transnext to improve the backbone's extraction performance of image texture information and spatial information. Secondly, the original FPN structure is replaced by CARAFEFPN to increase the receptive field, thereby improving the detection effect of the model in detecting dense eucalyptus plantations. Afterwards, SABLHead with See esawloss as the loss function is used as the detection head to solve the problem of sample imbalance and improve the robustness of the model. Finally, the improved Mask R-CNN algorithm is used to perform individual tree segmentation and forest age extraction of eucalyptus plantations. Results The improved Mask R-CNN model can achieve accurate individual tree segmentation in eucalyptus plantations at different stages. When the model parameters are only 33.85M, MAP50 can reach 80.6%, MAP75 can reach 54.2%, and MAP0.5:0.95 can reach 48.9%, all of which are better than the Mask R-CNN model. Based on the improved Mask R-CNN model, the SHAI algorithm can be used to accurately estimate the stand age of eucalyptus plantations. The accuracy of stand age estimation is 100%, which is better than the single tree age estimation result. The stand uniformity results calculated based on the crown width and tree height of individual trees are relatively different, and the stand uniformity calculated based on the height of individual trees is better than the stand uniformity calculated based on the crown width of individual trees. Among the five regions A, B, C, D, and E, the stands in regions B, C, D, and E are relatively uniform, while the stand in region A is relatively uneven. The research results show that the improved Mask R-CNN algorithm can not only realize the segmentation of individual trees in eucalyptus plantations at different stages, but also accurately estimate the stand age of eucalyptus plantations, and can provide data support for the evaluation of uniformity and productivity of eucalyptus plantations.

[0058] Study area: The study area is located in Luzhai County, Liuzhou City, Guangxi Zhuang Autonomous Region, China (109°30'0"E-110°15'0"E, 24°15'0"N-25°0'0"N) (e.g. Figure 1 As shown in the figure, (a) is the CHM image and (b) is the DOM image. The study area is located in the central part of Guangxi Zhuang Autonomous Region, China. The terrain is mainly mountainous, hilly and basin-like. The climate is mild and the rainfall is abundant. It is an important eucalyptus planting area in China. There are 1.4 million hectares of eucalyptus plantations and the annual logging volume reaches 1.21 million cubic meters.

[0059] Acquisition and processing of UAV stereo images: The DJI Phantom 4RTK drone was used to collect UAV stereo images of eucalyptus plantations in July and August 2022 and March 2023. The weather was windless or light windy. The five-way flight mode was adopted, the flight altitude was 70 meters, the flight speed was 5.7m / s, the heading overlap rate was set to 80%, the side overlap rate was set to 70%, and the coordinate system was WGS84. The acquired UAV stereo images were processed using Pix4D software to obtain DOM, DSM, DEM and other data. Afterwards, the DSM was subtracted from the DEM to obtain the corresponding CHM.

[0060] Single tree segmentation model dataset preparation: In order to facilitate deep learning model training, the DOM was segmented into 640×640 pixel tif files. The eucalyptus crown labels were marked on the segmented DOM images using ArcGIS, and the "spatial Analyst Tools-Regional Analysis-Display Zoning Statistics in Table" tool in ArcGIS was used to map the CHM information to each crown label to obtain the height information corresponding to each eucalyptus tree.

[0061] The average height of one-month-old eucalyptus trees in the study area of ​​this embodiment is 1.96m, the average height of one-year-old eucalyptus trees is 10.7m, the average height of two-year-old eucalyptus trees is 13.8m, and the average height of three-year-old eucalyptus trees is 16.7m. The statistical results of the tree heights of eucalyptus plantations of different ages in the study area are as follows: Figure 2 Based on Figure 2 The results shown classify the height of eucalyptus trees in the eucalyptus plantations in the study area into four categories: eucalyptus trees less than one year old (<10m), one year old eucalyptus trees (10m-13m), two years old eucalyptus trees (13m-15.6m) and three years old and above eucalyptus trees (>15.6m).

[0062] Based on CHM, the height of individual trees was extracted and the corresponding forest age was determined. 386 images were annotated, with a total of 10,055 eucalyptus labels. The distribution of forest age categories is as follows: Figure 3 As shown in Figure 2, 338 images are used as training set and 48 images are used as validation set.

[0063] Eucalyptus single tree segmentation model: Improved Mask R-CNN model: In agriculture and forestry management, the Mask R-CNN model is often used to identify and detect fruits and trees, as well as to identify and analyze urban tree distribution. The Mask R-CNN model is composed of backbone, FPN, RPN, ROI Align and head5. The specific structure is as follows Figure 4As shown in FIG, a mask prediction branch is added to Faster R-CNN to make it suitable for instance segmentation tasks. In order to achieve accurate segmentation of individual eucalyptus trees under different forest ages, this embodiment proposes an improved Mask R-CNN model to achieve eucalyptus tree segmentation and forest age estimation. The improvements include using Transnext_tiny as the backbone to enhance the ability to extract texture information and spatial information, using CA RAFE FPN instead of FPN module as the neck part, and using SABLHead with Seaseewloss classification loss function as the detection head to improve the Mask R-CNN model. The specific structure of the improved model is shown in FIG. Figure 5 shown.

[0064] Transnext: Neural networks have always been the mainstream model for computer vision. Applying Transformer and its multi-head attention mechanism to computer vision tasks can avoid the limitations of traditional convolutional neural networks in dealing with long-distance dependencies. In recent years, the backbone based on Transformer has been widely used in various computer vision tasks and has shown excellent feature extraction capabilities. Aggregate attention and length-scaling pre-attention are used to mitigate the potential depth attenuation of the model, enhance pixel focus attention, and enhance the ability of existing attention mechanisms to expand multi-scale inputs. In addition, a convolutional GLU is proposed, which can use the nearest neighbor image features to achieve channel attention and improve the robustness of the model. The GLU structure is as follows: Figure 6 In this embodiment, using Transnext as bankbone can better extract the features of the eucalyptus image.

[0065] CARAFE FPN: Pyramid features are often used in computer vision tasks to process targets of different scales. Upsampling is a crucial operation, which can change the size of the feature map so that the feature map can better integrate feature information at different levels. Common upsampling methods used in FPN include nearest neighbor, unilinear interpolation, bilinear interpolation and other methods. These upsampling methods can only use adjacent pixels to expand the image, which easily causes image information loss. The Content-Aware ReAssembly of Features (CARAF E) method reassembles features in a predefined area centered on each position through weighted combination. Each position has multiple sets of such upsampling weights. Feature upsampling is completed by regularly rearranging the generated features into spatial blocks. This upsampling method has a larger receptive field and content-aware hand-holding, and the amount of calculation is small. It plays a vital role in processing dense prediction tasks. In this embodiment, the eucalyptus plantations are densely planted, the eucalyptus crown density in the drone image is high, and the crowns overlap with each other. It is difficult to process the mutually adhered crowns using other upsampling methods. CARAFE is used as the upsampling method of FPN, as follows Figure 7 As shown in the figure, the eucalyptus features can be better extracted during the upsampling process, improving the robustness of the model.

[0066] SABL and Seesawloss: In the two-stage target detection network, bounding box regression is mainly used to locate the detection target, but the accuracy of bounding box regression is still not high. Side-Aware Boundary Localization (SABL) obtains four features F from the feature map after ROIAlign through Side-aware Feature Extraction left 、F right 、F top 、F down , the four features are input into Boundary Localization With Bucketing to determine in which bucket the four sides of the bounding box are located, and the offset between the center of the bucket and the position of the bounding box is calculated. Finally, the confidence of the bucket where the four sides are located is averaged through Bucketing-Guided Rescoring, and the result of multiplying it with the category confidence is used in NMS to retain the high-precision Bounding Box to improve the accuracy of bounding box regression.

[0067] In addition, loss functions play a vital role in deep learning. In the original Mask R-CNN model, the author used CrossEntropyLoss as a commonly used classification loss function. Its formula is shown in formula (1), which is suitable for multi-classification tasks. However, when faced with the problem of uneven distribution of categories in the data set, CrossEntropyLoss has poor classification effect on minority categories. Therefore, seesawloss is used to solve this problem.

[0068]

[0069] In the classification task, in order to solve the problem of poor classification results caused by sample imbalance, seeawl oss adjusts the loss function through mitigation factor and compensation factor to reduce the negative signal of the majority class on the gradient of other classes, thereby reducing the suppression of other classes and solving the long tail problem. Figure 2 As shown in the figure, there is an imbalance in the number of samples of the four types of eucalyptus. In order to increase the accuracy of a small number of sample types, the Seesawloss loss function is used to solve this problem. The specific calculation formulas are shown in formulas (2)-(5).

[0070]

[0071] S ij =M ij ·C ij (4)

[0072]

[0073] Where N i is the number of species i, σ i is the probability of being classified as class i, p and q are custom weights, y∈(0,1) is a one-hot true value label, z i is the predicted log of the classifier.

[0074] SAHI: Instance segmentation is one of the most important computer vision applications to date. However, in actual situations, there are still some problems in detecting targets in large images, such as scale mismatch and weak target information extraction, which need to be further improved. SAHI algorithm provides a general slice-assisted reasoning and fine-tuning pipeline for small target detection, which can be applied to any target detector. By dividing the super large image into multiple small slices, detecting the small slices in turn, and then splicing the detection results, the specific process is as follows Figure 8As shown. Predictions are made on multiple small slices, and adjacent slices have partially overlapping areas. This can solve the problem of the same object appearing in two slices, and the predicted results are mapped to the super-large image, thereby achieving the purpose of applying target detection and instance segmentation to super-large images. In this embodiment, 640×640 drone images of eucalyptus are used as training images, and the model weights obtained by training are used to use the SAHI algorithm to identify eucalyptus in super-large images, so as to better determine the number of years the eucalyptus forest has been planted, and predict the age distribution of eucalyptus forests in larger areas.

[0075] Experimental environment and parameters: This example is based on a eucalyptus image dataset in a natural environment to verify the effectiveness of the model instance segmentation. The specific specifications of the hardware platform for the experimental operation are shown in Table 1. The training hyperparameter settings are shown in Table 2.

[0076] Table 1 Experimental environment

[0077]

[0078] Table 2 Hyperparameter settings

[0079]

[0080] Accuracy evaluation of single tree segmentation model: In order to evaluate the accuracy of eucalyptus single tree segmentation results of different models, the evaluation indicators used in the study are mainly precision, recall and average precision (MAP). The specific calculation formulas are shown in formulas (6) to (9).

[0081]

[0082] Parameters=(K H ×K W ×C IN +1)×C OUT (9)

[0083] In the formula, TP is the number of true positive samples, FP is the number of false positive samples, FN is the number of false negative samples, Precision is the percentage of correctly identified objects in the number of identified items, Recall is the percentage of correctly identified objects in the number of real objects, and C is the number of categories. H is the height of the convolution, K W is the width of the convolution, C IN is the number of channels of the input feature layer, C OUT is the number of channels of the output feature layer.

[0084] Eucalyptus plantation stand age and uniformity estimation: Based on field survey data, the improved Mask R-CNN algorithm is used to segment individual eucalyptus trees and extract individual tree age. Since eucalyptus plantations in the same stand are planted in the same year, the majority of consistent eucalyptus individual tree ages in the same stand are taken as the stand age, and compared with the field survey age data. On the basis of stand age extraction, the tree height and crown width are extracted based on the individual tree segmentation results, and the PH 3 50, Pvar50 index is used to calculate the stand uniformity of the forest stand. The specific calculation formula is shown in formulas (10)-(11).

[0085]

[0086] Where Pvar50 is the cumulative percentage of the crown area of ​​the 50% of trees with the smallest crowns among the planted trees. n is the number of trees planted in ascending order. ji It refers to the crown area of ​​the i-th tree at the j-th age. 3 50 means the cumulative percentage of the smallest 50% of trees in terms of height; H 3 ji Represents the cube of the height of the i-th tree at the j-th age.

[0087] Results and analysis: Single tree segmentation results with different input data: Based on the Mask R-CNN model with Resnet50 as the backbone, RBG (three channels), RGB+CHM (four channels), RGB+CHM×2 (five channels) and RGB+CHM×3 (six channels) are input into the model respectively. The results are shown in Table 3.

[0088] Table 3. Eucalyptus tree segmentation results of Mask R-CNN model with different input data

[0089]

[0090] By analyzing the results shown in Table 3, it can be seen that when only RGB images are used as input features, the single tree segmentation results of the Mask R-CNN model are relatively poor, with MAP50 of only 59%, MAP75 of 43.2%, and MAP of 37%. When RGB and CHM information are used as input features together, the single tree segmentation results of the model are significantly improved, and the single tree segmentation results of the Mask R-CNN model are the best when a layer of CHM information is superimposed, with MAP50 of only 77.3%, MAP75 of 44.7%, and MAP of 43.3%. By stacking more layers of CHM, the single tree segmentation results of the model are not significantly improved. When 3 layers of CHM are stacked, MAP50 is only 78.0%, MAP75 of 45%, and MAP of 43.2%, which is not much different from the single tree segmentation results when one layer of CHM is stacked. Therefore, the subsequent experiments in this embodiment adopt the data input form of RGB_CHM.

[0091] Ablation experiment: Parameters, MAP50, MAP75 and MAP are used to evaluate the accuracy of ablation experiment results of different improved methods of Tansnext, CARAFEFP N and SABLHead. The results are shown in Table 4. The PR curves of each model are shown in Fig. 9 shown.

[0092] Table 4 Ablation experimental results of different improved methods

[0093]

[0094]

[0095] right Fig. 9 From the analysis of the ablation experiment PR curve shown in the figure, it can be seen that the area under the PR curve of the improved Mask R-CNN model is larger than the area under the PR curve of the original Mask R-CNN model. The results show that the improved Mask R-CNN model is better than the original Mask R-CNN model in application.

[0096] From the analysis of the results shown in Table 4, we can see that the parameters of the original Mask R-CNN model reached 63.39M, AP50 reached 77.3%, MAP75 reached 43%, and MAP reached 43.3%. After replacing the backbone of Mask R-CNN with Transnext, the model parameters were reduced by 30.69M, which greatly downgraded the model parameters, and MAP75 increased by 6.7% and MAP increased by 2.2%. On this basis, CARAFE FPN was used as the neck, and the model parameters increased by 0.89M, and the model was improved in terms of MAP50 and MAP. Using SABLHead as the detection head, MAP50 reached 80.6%, MAP75 reached 54.2%, M AP reached 48.9%, and the model only increased by 0.26%. The improved model reduced 29.54MB in parameters compared with the original model, and improved MAP50, MAP75 and MAP by 3.3%, 11.2% and 5.6% respectively. The above research results show that the improved Mask R-CNN model has lower model parameters and higher detection accuracy, and is more suitable for single tree segmentation and forest age identification of eucalyptus plantations under different forest age conditions.

[0097] Eucalyptus single tree segmentation results based on the improved Mask R-CNN model: The Mask R-CNN model and the improved Mask R-CNN model were used to segment single trees in eucalyptus plantations of different forest ages. The results are shown in Figure 2. Fig.10 shown.

[0098] right Fig.10 The results shown in the figure show that both the original Mask R-CNN model and the improved Mask R-CNN model perform well in detecting one-year eucalyptus plantations. When the tree crowns are small and the tree heights are not much different, the models can perform eucalyptus crown segmentation well. Fig.10 As shown in the first column. Fig.10 As shown in the second column, the eucalyptus trees in this area should be one-year-old eucalyptus trees, but the original model mistakenly identified some of the eucalyptus trees in the red box area as two-year-old eucalyptus trees, while the improved model had correct results in crown segmentation and classification. Fig.10 The third column shows that the eucalyptus trees in this area are two-year-old eucalyptus plantations. However, due to various natural conditions, the eucalyptus trees in the red frame area are generally short and do not meet the standard of two-year-old eucalyptus trees. There are a total of 8 trees, only 3 of which were detected in the original model, while the improved model successfully detected 6. Fig.10The area shown in the fourth column is a eucalyptus plantation that is three years old or older. In the area marked by the red frame, the original model failed to accurately extract the tree height information and thus misclassified the eucalyptus trees that were more than three years old as two to three years old. The improved model can better judge that the area shown in the picture is a eucalyptus plantation that is three years old or older. From the above analysis, it can be concluded that the improved model can more accurately extract the height of eucalyptus trees from the input feature map, thereby judging the age of the eucalyptus forest, and judging which eucalyptus trees in this area have their growth inhibited and have not reached the expected growth height. At the same time, the improved model also has the problem of misclassification and omission. Through the confusion matrix of the classification results, we can clearly see the misclassification and omission of different eucalyptus plantation ages. The results are shown in the figure. Fig.11 shown.

[0099] right Fig.11 The results shown in the figure show that the prediction accuracy of eucalyptus trees less than one year old is 80.90%, and 14.60% of the eucalyptus trees are predicted to be one year old. The prediction accuracy of one-year-old eucalyptus trees is 75.43%, and 10.22% and 10.40% are predicted to be less than one year old and two years old, respectively. The prediction accuracy of two-year-old eucalyptus trees is 76.29%, and 12.77% are predicted to be three years old and above. The prediction accuracy of three years and above eucalyptus trees is 70%, and 23.55% are predicted to be two-year-old eucalyptus trees. 52.63% of the background trees are mistakenly classified as eucalyptus trees less than one year old. The reason is that eucalyptus trees less than one year old are of lower height and smaller crown. In actual situations, there are a large number of other plants in the artificial forest. Some other tree species have similar texture features to eucalyptus trees, and the height information is similar, so they are mistakenly judged as eucalyptus trees less than one year old. Analysis of other misclassification cases showed that most of them were in adjacent categories. Misclassification is easy to occur in adjacent categories when the highly similar information is present. However, a few misclassifications are within a reasonable range and do not affect the identification of the age of eucalyptus plantations.

[0100] Results of individual tree segmentation and stand age prediction for regional eucalyptus plantations: From the above analysis, it can be seen that when using 640×640 pixel images for training and verification, the improved Mask R-CNN model obtained good individual tree segmentation and stand age estimation results. In order to obtain individual tree segmentation and stand age extraction for large-area eucalyptus plantations, the study used the SAHI algorithm to apply the improved Mask R-CNN model to images of larger areas, that is, the 4040×3904 pixel image was divided into 640×640 pixels with an overlap ratio of 40%, and then divided into 110 small images for prediction. After that, the prediction results of each small image were spliced, thereby achieving individual tree segmentation and stand age extraction for eucalyptus plantations in larger area images. The results are shown in Figure 2. Fig.12 shown.

[0101] Table 5 Forest age extraction results

[0102]

[0103] right Fig.11 According to the analysis of the results shown in Table 5, this area mainly includes 5 eucalyptus plantations planted at different times. Among them, 95.15% of the single tree age extraction results in area A are less than one year, so the stand age of the eucalyptus plantation in area A is judged to be less than one year; 53.09% of the single tree age extraction results in area B are two years, so the stand age of the eucalyptus plantation in area B is judged to be two years; 100% of the single tree age extraction results in area C are less than one year, so the stand age of the eucalyptus plantation in area C is judged to be less than one year; 63% of the single tree age extraction results in area D are one year, so the stand age of the eucalyptus plantation in area D is judged to be one year; 78.35% of the single tree age extraction results in area E are less than one year, so the stand age of the eucalyptus plantation in area E is judged to be less than one year.

[0104] Results of stand uniformity estimation: The crown width of individual trees and the age of the eucalyptus plantations were extracted based on the improved Mask R-CNN model, and the tree height was extracted based on the CHM. The uniformity of eucalyptus plantations of different ages was calculated based on the crown width and tree height. The results are shown in Table 6.

[0105] Table 6 Estimation results of stand uniformity of eucalyptus plantations of different ages

[0106]

[0107] Analysis of the results shown in Table 6 shows that the results of stand uniformity calculated based on single tree crown width and tree height are relatively different. Among them, for region A, the stand uniformity calculated based on single tree crown width is 35.07%, while the stand uniformity calculated based on single tree height is 20.13%, with a difference of 14.94%. For region D, the stand uniformity calculated based on single tree crown width is 27.95%, while the stand uniformity calculated based on single tree height is 44.04%, with a difference of 16.09%. However, for region C, the stand uniformity calculated based on single tree crown width and tree height is 40.05% and 40.03%, respectively, and the results are basically consistent. The uniformity of uniform eucalyptus plantations is between 37-50%, and eucalyptus plantations tend to be uniform as the forest age increases. Therefore, it is believed that the uniformity results calculated based on single tree height are relatively accurate, while the stand uniformity results calculated based on single tree crown width are relatively poor. Therefore, it is determined that the forest stands in areas B, C, D, and E are uniform, while the forest stand in area A is uneven. Because it is more difficult to obtain the vertical structural parameters of a single tree than the horizontal structural parameters of a single tree, if the vertical structural parameters of a single tree cannot be effectively obtained in the future, the crown width of a single tree can also be used to characterize the uniformity of the forest stand to a certain extent. However, if you want to accurately characterize the uniformity of the forest stand, you need to further develop a forest stand uniformity calculation formula suitable for the crown width of a single tree.

[0108] Comparison of eucalyptus tree segmentation performance before and after the improvement of Mask R-CNN model: In order to achieve accurate segmentation of single tree crowns of eucalyptus plantations under different forest age conditions, this embodiment proposes to improve the Mask R-CNN model to make it applicable to single tree segmentation of eucalyptus plantations at different stages. Even if Transnext is used as the backbone to extract features, the Transnext backbone network has better accuracy than other backbone networks in target detection and semantic segmentation tasks, with a COCO object detection mAP of 57.1% and an ADE20K semantic segmentation mIo U of 54.7%. Transnext can be used as the backbone to extract the texture and height information of eucalyptus. Using CARAFEFPN as the neck, based on the Tood model, using CARAFEFPN as the neck in the task of detecting dense strawberries, the AP reached 72.9%, which can better detect strawberries and judge the maturity of strawberries. For artificial eucalyptus forests, which are also dense tasks, using CARAFEFPN as the neck can better segment dense and overlapping eucalyptus crowns. The SABL head with Seesawloss as the loss function is used as the detection head. Seesawloss is often used to solve the problem of data imbalance. Using Seasawloss as the loss function can effectively improve the robustness of the model. In summary, the improved model has a good single tree segmentation effect in artificial eucalyptus forests of different forest ages.

[0109] Analysis of the results of extracting individual tree age and forest stand age by improving the Mask R-CNN model: Previous studies have shown that the Mask R-CNN model can perform individual tree segmentation, and the commonly used input feature variable is RGB image information. In order to extract the individual tree age while segmenting eucalyptus, this embodiment improves the Mask R-CNN model, introduces CHM as height information based on RGB input, so that it can extract the height of individual trees based on individual tree segmentation, thereby judging the forest age. In the prior art, the YOL Ov8x model using the RGB-DSM combination has achieved excellent performance in segmenting potato leaves and stems, with an F1 score of 86%. The results show that the height mutation of the edge plays a vital role in the segmentation of potato stems. Xie et al. fused RGB images with CHM images to obtain the height information of trees, thereby improving the accuracy of individual tree segmentation. The Mask R-CNN model segmentation with ResNext as the bacbone achieved an AP of 54.946%, confirming the effectiveness of the data fusion strategy for individual tree segmentation. In this embodiment, CHM is used as input, and the extracted crown judges the forest age of eucalyptus according to the height information. The number of trees contained in a single image is limited, and it is impossible to intuitively judge the scope and age of the entire eucalyptus forest. The SHAI algorithm is used to apply model detection to larger images, so that the age and planting range of eucalyptus plantations can be better judged, and the information of eucalyptus plantations can be better obtained, and the age of eucalyptus can be judged based on height information. Although the age of individual trees extracted based on the improved Mask R-CNN model is not very accurate in some forest stands, this is mainly because the height information is too single, and CHM is calculated by the 3D model generated by oblique photogrammetry technology, which has errors, but the forest age extracted based on the age of individual trees is relatively accurate. Further research can combine the near-infrared band to generate more information about height as input, and then combine other vegetation index information to further improve the accuracy of extracting tree height information.

[0110] Analysis of the uniformity estimation results of eucalyptus plantation stands: This example uses Pvar50 and PH 3 Pvar50 evaluated the uniformity of the stand from the two dimensions of crown width and tree height, and found that there were significant differences between the uniformity calculation results based on tree height and that based on individual tree crown width. Specifically, crown width mainly reflects the horizontal distribution of the stand, while tree height mainly reflects the vertical structural distribution of the stand. Through calculation, the stand uniformity obtained by Pvar50 (based on crown width) is lower, while that obtained by PH 3 50 (based on tree height) resulted in a higher uniformity, indicating that the vertical structure of the stand is more sensitive to the uniformity calculation. 3 The 50 index is a variation of the PV50, using the PH 3The 50 index estimates the uniformity of the stand, which is highly correlated with PV50 and is used to evaluate the uniformity of the plantation. In this embodiment, after using the SHAI algorithm to obtain the forest age and range of the eucalyptus plantation, the tree height corresponding to the single tree is obtained according to the drone-derived height, and the stand uniformity PH of the corresponding eucalyptus plantation at the corresponding forest age is estimated. 3 50. In this example, the uniformity of regions B, C, D, and E is between 37% and 50%, and the uniformity increases with the age of the forest. The uniformity of region A is 20.13%, which is much lower than the uniformity of normal eucalyptus plantations and does not meet the uniformity of eucalyptus plantations. Fig.13 As shown in Figure 1, the height of eucalyptus trees in area A is mostly consistent with the height of eucalyptus trees with a one-year age. Fig.13 (b) It can be clearly seen that the eucalyptus trees in area A should be planted in two different batches. The height difference of the eucalyptus trees is large, resulting in a low estimated uniformity.

[0111] This embodiment improves the Mask R-CNN algorithm framework through TranNext, CARAFEFPN, SABLHead and Sawseeloss, so that it can not only achieve the segmentation of individual trees in eucalyptus plantations at different stages, but also achieve accurate estimation of the stand age of eucalyptus plantations. The main conclusions are as follows:

[0112] (1) The improved Mask R-CNN model in this embodiment can achieve accurate single tree segmentation in eucalyptus plantations at different stages. With a model parameter of only 33.85M, MAP50 can reach 80.6%, MAP75 can reach 54.2%, and MAP0.5:0.95 can reach 48.9%, all of which are better than the Mask R-CNN model;

[0113] (2) Based on the improved Mask R-CNN model, the SHAI algorithm can be used to accurately estimate the stand age of eucalyptus plantations. The accuracy of stand age estimation is 100%, which is better than the single tree age estimation result.

[0114] (3) The results of stand uniformity calculated based on tree crown width and tree height are relatively different, and the stand uniformity calculated based on tree height is better than the stand uniformity calculated based on tree crown width. Among the five regions A, B, C, D, and E, the stands in regions B, C, D, and E are relatively uniform, while the stand in region A is relatively uneven.

[0115] In summary, this embodiment can not only provide accurate individual tree crown width and forest age for subsequent eucalyptus plantation yield estimation and management, but also provide methods and technical guidance for individual tree segmentation and forest age judgment of other forest types.

[0116] A system for analyzing uniformity of plantation stands under different forest age conditions, comprising:

[0117] A data acquisition module is used to obtain a single tree segmentation model data set, wherein the single tree segmentation model data set includes artificial forest image data and corresponding tree height information;

[0118] A model training module is used to construct an initial single tree segmentation model according to the Mask R-CNN algorithm, improve the backbone network, feature pyramid network and detection head of the initial single tree segmentation model to obtain an improved single tree segmentation model; train the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model;

[0119] The single tree segmentation module is used to perform single tree segmentation of the artificial forest image to be studied according to the trained single tree segmentation model, obtain the single tree segmentation results and estimate the single tree age, calculate the forest stand age using the SAHI algorithm, extract the single tree height and single tree crown width based on the single tree segmentation results, and calculate the uniformity of the artificial forest stand under different forest age conditions based on the extracted single tree height and single tree crown width.

[0120] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for analyzing uniformity of artificial forest stands under different forest age conditions.

[0121] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for analyzing uniformity of artificial forest stands under conditions of different forest ages.

[0122] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for analyzing uniformity of plantation stands under different forest age conditions, characterized in that: include: Acquire a single tree segmentation model dataset, wherein the single tree segmentation model dataset includes artificial forest image data and corresponding tree height information; An initial single tree segmentation model is constructed based on the Mask R-CNN algorithm, and a backbone network, a feature pyramid network, and a detection head of the initial single tree segmentation model are improved to obtain an improved single tree segmentation model; Training the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model; Based on the trained single tree segmentation model, single tree segmentation of the artificial forest image to be studied is performed to obtain the single tree segmentation results and estimate the single tree age. The stand age is calculated using the SAHI algorithm. The single tree height and single tree crown width are extracted based on the single tree segmentation results. Based on the extracted single tree height and single tree crown width, the uniformity of the artificial forest stand under different forest age conditions is calculated.

2. The method for analyzing uniformity of artificial forest stands under different forest age conditions according to claim 1, characterized in that: The step of obtaining a single tree segmentation model dataset specifically includes: Use drones to collect stereo images of artificial forests and obtain image data of artificial forests; Based on Pix4D software, digital orthophotos, digital surface model data and digital elevation model data in artificial forest image data are obtained; Subtracting the digital surface model data from the digital elevation model data to obtain canopy height model data; Segmenting the digital orthophoto to obtain a plurality of segmented images, and marking a tree crown label for each of the segmented images; The canopy height model data is mapped to each tree crown label to obtain tree height information corresponding to each single tree.

3. The method for analyzing uniformity of artificial forest stands under different forest age conditions according to claim 1, characterized in that: The improvement of the backbone network, feature pyramid network and detection head of the initial single tree segmentation model specifically includes: Firstly, the Transnext network is used as the backbone network, and then the feature pyramid network is replaced by the CARAFEFPN method. Finally, SABLHead with Seesawloss as the loss function is used as the detection head to obtain the improved single tree segmentation model.

4. The method for analyzing uniformity of artificial forest stands under different forest age conditions according to claim 1, characterized in that: The improved single tree segmentation model is trained based on the single tree segmentation model data set, specifically including: The single tree segmentation model data set is divided into a training set and a validation set based on preset rules, the improved single tree segmentation model is trained based on the training set, the trained single tree segmentation model is evaluated based on the validation set, the single tree segmentation model is optimized based on the evaluation result, the iterative training and optimization process is repeated until the preset number of iterations is reached, and the trained single tree segmentation model is output.

5. The method for analyzing uniformity of artificial forest stands under different forest age conditions according to claim 1, characterized in that: The calculation of the uniformity of the plantation stands under different forest age conditions based on the extracted single tree height and single tree crown width specifically includes: The age of individual trees is determined based on the results of individual tree segmentation, the age of the stand is determined using the SAHI algorithm, and the height and crown width of individual trees are extracted based on the results of individual tree segmentation; Based on the stand age, tree height and tree crown width, PH 3 The Pvar50 index and Pvar50 index were used to calculate the uniformity of plantation stands under different stand ages.

6. A system for analyzing uniformity of plantation stands under different forest age conditions, characterized in that: include: A data acquisition module is used to obtain a single tree segmentation model data set, wherein the single tree segmentation model data set includes artificial forest image data and corresponding tree height information; A model training module is used to construct an initial single tree segmentation model according to the Mask R-CNN algorithm, improve the backbone network, feature pyramid network and detection head of the initial single tree segmentation model to obtain an improved single tree segmentation model; train the improved single tree segmentation model based on the single tree segmentation model data set to obtain a trained single tree segmentation model; The single tree segmentation module is used to perform single tree segmentation of the artificial forest image to be studied according to the trained single tree segmentation model, obtain the single tree segmentation results and the single tree age, use the SAHI algorithm to obtain the forest stand age, extract the single tree height and single tree crown width based on the single tree segmentation results, and calculate the uniformity of the artificial forest stand under different forest age conditions based on the extracted single tree height and single tree crown width.

7. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for analyzing uniformity of artificial forest stands under different forest age conditions according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements a method for analyzing uniformity of artificial forest stands under different forest age conditions as described in any one of claims 1 to 5.

Images