Iterative optimization method for single-tree crown segmentation and related device
By using an iterative optimization method based on confidence decoupling and prior information, and by fusing neural networks and prior information, the problem of insufficient boundary accuracy in complex environments in single tree crown segmentation is solved, and higher accuracy crown segmentation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU UNIVERSITY
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-02
Smart Images
Figure CN122135038A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to an iterative optimization method and related apparatus for single-tree crown segmentation based on confidence decoupling and prior-driven approach. Background Technology
[0002] Biodiversity supports the critical supply and regulation services needed by human society by maintaining ecosystem functions; its loss threatens the very foundation of human well-being. Forest ecosystems play a vital role in protecting global biodiversity and provide humanity with a wealth of valuable ecosystem services. However, in recent years, the world's forests have been gradually decreasing, facing the risk of continuous loss and degradation, and their capacity to produce and provide essential services, as well as their ability to maintain biodiversity, is gradually declining. To better safeguard biodiversity and achieve sustainable human development, it is necessary to conduct systematic statistics and management of existing forest resources.
[0003] Individual tree canopy segmentation (ITCD) is a key technology for forest resource surveys, biomass estimation, and tree health monitoring. With the development of UAV remote sensing technology, its advantages, such as wide data acquisition range, low cost, and the ability to carry a variety of sensors, enable the acquisition of high spatiotemporal resolution, feature-rich images at low cost, making it the primary data source for current ITCD tasks. Traditional ITCD methods rely on manual field surveys or visual interpretation of aerial images, which suffers from limitations such as high manpower and material consumption and low efficiency. The development of high-resolution remote sensing satellite technology and computer vision technology has driven the transformation of ITCD towards automation, progressing from traditional image processing and traditional machine learning to deep learning-based methods. Furthermore, deep learning-based semantic segmentation methods, due to their lack of manual feature design and strong adaptability, have become the mainstream approach in current ITCD. These methods input images into neural networks, extract high-dimensional features, and decode and map them back to the original spatial dimensions to obtain the canopy mask. However, in practical applications, ITCD tasks often face challenges such as different tree canopy sizes and similar tree canopy textures, which limits the final recognition effect and leads to problems such as target misclassification and poor canopy edge performance in downstream applications, thus restricting generalization.
[0004] In deep learning-based single-tree crown segmentation, data is input into a deep learning model to extract high-dimensional features from low-dimensional features. These high-dimensional features are then decoded and mapped back to their original spatial dimensions to obtain the single-tree crown segmentation mask. Therefore, extracting effective and robust crown feature information from limited data is the core problem of single-tree crown segmentation. Based on the approach to crown feature mining, most work can be categorized into two main approaches: improving the model structure to enhance its feature mining capabilities and optimizing data input to improve feature discriminability.
[0005] In the study of improving model structure, the following approaches can be used: (1) Improve the model’s feature extraction and boundary decoupling capabilities by improving the model’s backbone; (2) Design modules to enhance the model’s context awareness capabilities; (3) Optimize the decoder to enhance the decoding effect; (4) Optimize the loss function to suppress noise.
[0006] In the study of optimizing data input, common approaches include: (1) Multi-source heterogeneous information fusion strategy: such as fusing RGB images with vegetation index, integrating UAV and satellite images for multi-scale fusion, and fusing hyperspectral data with point cloud data, thereby improving the semantic separability of tree canopy boundaries in complex backgrounds; (2) Data acquisition strategy optimization: optimizing data quality by using phenological features to amplify the visual contrast between target trees and background; (3) Data quality enhancement: using generative adversarial networks to enhance data and obtain training samples, and using geometric transformation and spectral transformation for data enhancement.
[0007] In summary, existing research on single-tree crown segmentation has explored various aspects of crown segmentation in complex environments, improving segmentation accuracy to some extent. However, it still faces bottlenecks in the accuracy of single-tree crown boundaries. The reason for this is that existing methods are essentially end-to-end single-step prediction paradigms. This paradigm couples and attempts to solve the problems of global structure establishment and local boundary refinement in the single-tree segmentation process simultaneously, leading to difficulties in model optimization. Moreover, any cognitive errors in the model discrimination process are accumulated and cannot be optimized. To overcome the bottlenecks faced by the single-step prediction paradigm, this invention aims to construct a step-by-step iterative optimization paradigm based on confidence decoupling and prior-driven approaches. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides an iterative optimization method and related device for single tree crown segmentation based on confidence decoupling and prior driving, which significantly improves the boundary fineness and regional consistency of remote sensing image segmentation in complex forestry scenarios.
[0009] To address the aforementioned technical problems, embodiments of the present invention provide an iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approaches, the method comprising: Remote sensing image data is obtained and input into a single tree crown segmentation model. After being processed by a neural network encoder and a neural network decoder, the coarse segmentation result of the single tree crown is output. The confidence assessment of the coarse segmentation results of the single tree crown is performed based on the statistical calculation method of pixel-level probability distribution map, thereby forming the confidence assessment result corresponding to the coarse segmentation results of the single tree crown. The confidence evaluation result is used to generate prior information through a prior encoder, and the prior information is fused with the original high-dimensional features output by the neural network encoder to generate fused features. The fused features are input into the neural network decoder for the next optimized segmentation process, and the next optimized segmentation result of the single tree crown is output. The process is then returned to the confidence evaluation step for continuous optimization and iterative segmentation until the number of optimization iterations reaches the preset number, at which point the final optimized segmentation result of the single tree crown is output.
[0010] Optionally, the step of inputting the remote sensing image data into the single-tree crown segmentation model and processing it sequentially through a neural network encoder and a neural network decoder to output a coarse segmentation result for the single-tree crown includes: The remote sensing image data is input into the single tree crown segmentation model, and after passing through the neural network encoder, the original high-dimensional features containing crown information are output. The original high-dimensional features are input into the neural network decoder, which outputs the coarse segmentation result of the single tree crown. The coarse segmentation result of the single tree crown includes... The initial coarse segmentation result of the dimension and the corresponding pixel-level probability distribution map, where Original image size It is a set of classification categories.
[0011] Optionally, the statistical calculation method based on pixel-level probability distribution maps is used to perform confidence assessment processing on the coarse segmentation results of the single tree crown, forming a confidence assessment result corresponding to the coarse segmentation results of the single tree crown, including: Obtain the pixel-level probability distribution map corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result, and use a preset statistical calculation formula to perform confidence evaluation processing on each initial coarse segmentation result to form the confidence evaluation result corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result. The preset statistical calculation formula is as follows: ; in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
[0012] Optionally, generating prior information from the confidence assessment result using a prior encoder includes: A proportional threshold that evolves with each iteration is constructed using a cosine annealing function. The confidence evaluation results corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result are sorted in ascending order, and a binary control mask is generated based on the ratio threshold using the ascending sorting results. The binary control mask is used to filter the coarse segmentation result of the single tree crown, resulting in a filtered coarse segmentation result of the single tree crown. The filtered coarse segmentation result of the single tree crown is input into the prior encoder, which outputs prior information aligned with the original high-dimensional feature dimension. The prior encoder is a shallow convolutional network.
[0013] Optionally, the formula for generating the binary control mask based on the ascending sorting result using the ratio threshold is as follows: ; ; in, This is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
[0014] Optionally, the step of fusing the prior information with the original high-dimensional features output by the neural network encoder to generate fused features includes: The prior information is added to and fused with the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ; in, Features of fusion; These are the original high-dimensional features; This is prior information.
[0015] Optionally, the step of inputting the fused features into the neural network decoder for the next optimized segmentation process and outputting the next optimized segmentation result for the single tree crown includes: The fused features are input into the neural network decoder, and the neural network decoder outputs a new coarse segmentation result of the single tree crown; The next optimization segmentation process is performed using a binary control mask, the new coarse segmentation result of the single tree crown, and the previous coarse segmentation result of the single tree crown, and the next optimized segmentation result of the single tree crown is output. The formula for the next optimized segmentation process is as follows: ; in, To optimize the segmentation of the individual tree crown for the next time; This is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
[0016] In addition, embodiments of the present invention also provide an iterative optimization single-tree crown segmentation device based on confidence decoupling and prior-driven iteration, the device comprising: Tree crown segmentation module: used to obtain remote sensing image data, and input the remote sensing image data into the single tree crown segmentation model. After being processed by the neural network encoder and the neural network decoder, the module outputs the coarse segmentation result of the single tree crown. Confidence assessment module: used to perform confidence assessment on the coarse segmentation results of the single tree crown based on the statistical calculation method of pixel-level probability distribution map, and form the confidence assessment result corresponding to the coarse segmentation results of the single tree crown; Feature fusion module: used to generate prior information from the confidence evaluation result through the prior encoder, and to fuse the prior information with the original high-dimensional features output by the neural network encoder to generate fused features; The optimization iteration module is used to input the fused features into the neural network decoder for the next optimization segmentation process, output the next single tree crown optimization segmentation result, and return to the confidence evaluation processing step to perform continuous optimization iteration segmentation processing until the number of optimization iterations reaches the preset number, and then output the final single tree crown optimization segmentation result.
[0017] In addition, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the processor runs a computer program or code stored in the memory to implement the iterative optimization single-tree crown segmentation method as described in any of the above.
[0018] In addition, embodiments of the present invention also provide a computer-readable storage medium for storing a computer program or code, which, when executed by a processor, implements the iterative optimization single-tree crown segmentation method as described above.
[0019] In this embodiment of the invention, the boundary precision and regional consistency of remote sensing image segmentation in complex forestry scenarios are significantly improved. The dynamic region separation strategy based on prediction confidence effectively solves the problem that existing single-step prediction methods use the same processing method for pixels of varying difficulty, leading to difficulty in accurately segmenting contiguous boundaries. Through a priori-driven conditional inference mechanism, historical prediction information is used to guide the current segmentation process. The high-reliability prediction results from the previous round are mapped to feature vectors and fused with current visual features, enabling the model to reference the identified canopy location information during inference. A selective update rule is adopted, accepting new prediction values only in low-reliability regions while maintaining the original results in high-reliability regions. This allows the generated segmentation results to infer the attribution of fuzzy boundaries based on the determined canopy subject, effectively suppressing background interference and ensuring the integrity of individual tree canopy morphology. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a flowchart illustrating the iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approach in an embodiment of the present invention. Figure 2 This is a flowchart illustrating an iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approach, according to another embodiment of the present invention. Figure 3 This is a schematic diagram of the structural composition of the iterative optimization single-tree crown segmentation device based on confidence decoupling and prior driving in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structural composition of the electronic device in an embodiment of the present invention; Figure 5 This is a performance graph showing the change of mIoU with the number of iterations in an embodiment of the present invention; Figure 6 This is a performance variation graph of BF1 as the number of iterations in an embodiment of the present invention; Figure 7 This is a graph showing the variation in model inference time for different iteration numbers in this embodiment of the invention; Figure 8 These are the generation results of various methods in the embodiments of the present invention on the test set. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Example 1, please refer to Figure 1 , Figure 1 This is a flowchart illustrating the iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approaches in an embodiment of the present invention.
[0024] like Figure 1 As shown, an iterative optimization method for single-tree canopy segmentation based on confidence decoupling and prior-driven approach is presented. The step-by-step iterative canopy segmentation framework based on prediction confidence decoupling and prior-driven approach mainly consists of two parts: the first part is dynamic region separation based on prediction confidence, aiming to divide the entire prediction result into high-reliability reliable regions and low-reliability regions to be optimized; the second part is prior-driven conditional reasoning, which, based on the region separation in the first part, differentiates between reliable regions and regions to be optimized. The reliable regions are encoded to obtain prior information that guides subsequent iterative optimization. Then, the prior information is fused with the original image features output by the backbone network and fed into the decoder to obtain new reasoning results. This allows for a re-examination of the entire image classification based on known reliable information, realizing an iterative optimization reasoning paradigm that continuously incorporates reliable prior information.
[0025] Specifically, the acquired remote sensing image data is input into a neural network encoder to output high-dimensional features. The high-dimensional features are then input into a neural network decoder to output a coarse segmentation result. Continuous optimization is then performed to evaluate the confidence level of the coarse segmentation result. Low-reliability results are discarded, and high-reliability results are input into a prior encoder to output prior information. The prior information is then fused with the high-dimensional features to form a fused feature, which is input into the neural network decoder to output a coarse segmentation result again. Optimization operations continue until the preset number of continuous optimizations is reached. After the preset number of continuous optimizations is reached, the optimized segmentation result is directly output.
[0026] Example 2, please refer to Figure 2 , Figure 2 This is a flowchart illustrating an iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approach, according to another embodiment of the present invention.
[0027] like Figure 2 As shown, an iterative optimization method for single-tree crown segmentation based on confidence decoupling and prior-driven approaches is presented. The method includes: S201: Obtain remote sensing image data, and input the remote sensing image data into the single tree crown segmentation model. After being processed by the neural network encoder and the neural network decoder, the coarse segmentation result of the single tree crown is output. In a specific implementation of this invention, the step of inputting the remote sensing image data into a single-tree crown segmentation model, processing it sequentially through a neural network encoder and a neural network decoder, and then outputting a coarse segmentation result for the single-tree crown includes: inputting the remote sensing image data into the single-tree crown segmentation model, processing it through the neural network encoder, and outputting original high-dimensional features containing crown information; inputting the original high-dimensional features into the neural network decoder, and outputting the coarse segmentation result for the single-tree crown, wherein the coarse segmentation result for the single-tree crown includes... The initial coarse segmentation result of the dimension and the corresponding pixel-level probability distribution map, where Original image size It is a set of classification categories.
[0028] Specifically, after obtaining remote sensing image data, the data is input into a single tree canopy segmentation model, where it is processed sequentially by a neural network encoder and a neural network decoder. After passing through the neural network encoder, the output is a high-dimensional feature containing canopy information. Then the high-dimensional features The input is then fed into the neural network decoder to obtain... Initial coarse segmentation results of the dimension and the pixel-level probability distribution map corresponding to the initial coarse segmentation result. ;in Original image size It is a set of classification categories.
[0029] S202: The confidence assessment of the coarse segmentation results of the single tree crown is performed based on the statistical calculation method of the pixel-level probability distribution map to form the confidence assessment result corresponding to the coarse segmentation results of the single tree crown. In the specific implementation of this invention, the statistical calculation method based on pixel-level probability distribution maps is used to perform confidence assessment processing on the coarse segmentation results of the single tree crown, forming a confidence assessment result corresponding to the coarse segmentation results of the single tree crown, including: Obtain the pixel-level probability distribution map corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result, and use a preset statistical calculation formula to perform confidence evaluation processing on each initial coarse segmentation result to form the confidence evaluation result corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result. The preset statistical calculation formula is as follows: ; in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
[0030] Specifically, pixel-level confidence quantification aims to solve the problem of mathematically representing the difficulty levels of different regions in an image. Generally, in areas with clear textures, such as the center of the tree canopy, the features extracted by the model have significant discriminative power, and their classification probability distribution exhibits a peak-like shape. However, in areas such as the edges of the tree canopy, areas where features are fuzzy or semantically ambiguous, the probability distribution predicted by the model is often relatively flat. Therefore, the maximum classification probability can directly reflect the model's "certainty" regarding the current pixel. By establishing a pixel-level confidence index, the image space can be mapped from a single pixel grid to a "confidence potential surface" with high and low undulations, thus providing a quantitative benchmark for identifying difficult examples. The specific implementation involves statistically analyzing the initial segmentation probability map output by the model. For the Round of iteration output Dimensional pixel-level probability distribution map Define pixels confidence level The maximum predicted probability for this pixel across all categories is given by the following formula: ; in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
[0031] S203: The confidence evaluation result is used to generate prior information through a prior encoder, and the prior information is fused with the original high-dimensional features output by the neural network encoder to generate fused features; In the specific implementation of this invention, the step of generating prior information from the confidence evaluation result through a prior encoder includes: constructing a proportional threshold that evolves with each iteration using a cosine annealing function; sorting the confidence evaluation results corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result in ascending order, and generating a binary control mask based on the proportional threshold using the ascending sorting result; using the binary control mask to filter the single tree crown coarse segmentation result to form a filtered single tree crown coarse segmentation result; inputting the filtered single tree crown coarse segmentation result into the prior encoder, and outputting prior information aligned with the original high-dimensional feature dimension, wherein the prior encoder is a shallow convolutional network.
[0032] Furthermore, the formula for generating a binary control mask based on the ascending sorting result using the ratio threshold is as follows: ; ; in, This is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
[0033] Furthermore, the step of fusing the prior information with the original high-dimensional features output by the neural network encoder to generate fused features includes: performing an additive fusion process on the prior information and the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ; in, Features of fusion; These are the original high-dimensional features; This is prior information.
[0034] Specifically, generating prior information requires dynamic separation and mask generation, with the core challenge being resource allocation and convergence control during iterative optimization. While confidence quantification identifies challenging areas, focusing on only a few boundaries in the early stages of iteration may prevent the model from correcting widespread contour errors; conversely, continuous large-scale updates not only waste computational resources but can also lead to model oscillations. Therefore, this invention introduces a "coarse-to-fine" dynamic separation mechanism that aligns with human cognitive patterns. In the early stages of iteration, a wider range of corrections is allowed to adjust the topology; as iteration progresses, the separation boundaries gradually shrink, forcing the model to focus only on pixel-level boundaries that are extremely difficult to distinguish. This dynamic separation ensures the stability and efficiency of the optimization process.
[0035] The implementation process mainly includes two steps: timing scheduling and space separation. The first step is time-series scheduling: a time-series scheduling mechanism is constructed based on the iteration rounds. Evolutionary Proportion Threshold The cosine annealing function is preferred for nonlinear scheduling, as shown in the following formula: ; in This represents the total number of iterations; this function ensures that the proportion to be optimized smoothly decreases from 1 to 0.
[0036] The second step is spatial separation: sorting all pixels by confidence level. Sort in ascending order and select the first few. A proportional number of pixels are used as the "region to be optimized" to generate a binary control mask. The formula is expressed as: ; in, This is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
[0037] The image is separated into regions that need to be updated and reliable regions that remain unchanged.
[0038] By fusing prior information with the original high-dimensional features, this approach aims to address the lack of contextual reference in models when handling blurred boundaries. When segmenting canopy adhesion boundaries, relying solely on current local visual features often makes it difficult for models to determine boundary affiliation. However, the main canopy area (high-reliability region) identified in the previous iteration contains spatial location and semantic category information; fully leveraging this information can aid in inferring the categories of adjacent blurred pixels. Therefore, by mapping discrete labels to continuous high-dimensional features and explicitly injecting them into the visual features, the model can more accurately repair edges given the knowledge that the surrounding area is a canopy.
[0039] First, there is a priori encoding, using the resulting mask. Regarding the results of the previous round of predictions Filter and keep only the locked area. The filtered results are then fed into a lightweight prior encoder, designed as a shallow convolutional network, to map the filtered segmentation results into high-dimensional features aligned with the visual feature dimensions. .
[0040] Then, the prior information is added to and fused with the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ; in, Features of fusion; These are the original high-dimensional features; This is prior information.
[0041] S204: Input the fused features into the neural network decoder for the next optimized segmentation process, and output the next optimized segmentation result of the single tree crown; In the specific implementation of this invention, the step of inputting the fused features into the neural network decoder for the next optimized segmentation process and outputting the next optimized segmentation result of the single tree crown includes: inputting the fused features into the neural network decoder, the neural network decoder outputting a new coarse segmentation result of the single tree crown; using a binary control mask, the new coarse segmentation result of the single tree crown and the previous coarse segmentation result of the single tree crown for the next optimized segmentation process, and outputting the next optimized segmentation result of the single tree crown. The formula for the next optimized segmentation process is as follows: ; in, To optimize the segmentation of the individual tree crown for the next time; This is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
[0042] Specifically, the fused features are input into the neural network decoder, which then outputs a new coarse segmentation result for the single tree crown. The core principle is to ensure the monotonicity and stability of the iterative process. To avoid compromising the original correct results while correcting errors, this invention employs a strict gating update logic. Based on the previously separated "reliable regions," during state updates, the prediction results are updated only within the identified "regions to be optimized," preserving the results for the reliable regions. This strategy effectively avoids erroneous changes to the inference results, ensuring that the segmentation accuracy steadily improves with each iteration.
[0043] Then, the binary control mask, the new coarse segmentation result of the single tree crown, and the previous coarse segmentation result of the single tree crown are used to perform the next optimization segmentation process, and the next optimized segmentation result of the single tree crown is output. The formula for the next optimized segmentation process is as follows: ; in, To optimize the segmentation of the individual tree crown for the next time; This is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
[0044] This processing ensures that only low-reliability areas indicated by the mask are addressed. When updating the state, results from low-reliability regions are discarded, while results from high-reliability regions are updated. Maintain the previous round of forecast values. constant.
[0045] S205: Determine whether the current optimization iteration segmentation count has reached the preset count. If not, return to S202; otherwise, proceed to S206.
[0046] S206: Output the final optimized segmentation result of a single tree crown when the number of optimization iterations reaches the preset number.
[0047] In the verification and result analysis of this embodiment, the experimental setup, to verify the effectiveness of the present invention in the single-tree crown segmentation task, firstly selected the classic but widely used random forest model method as the baseline comparison method, and then selected advanced single-tree crown segmentation methods from the past three years to conduct comparative experiments on the OAM-TCD dataset, where the training set, validation set, and test set were set in a 3:1:1 ratio. This dataset is sampled from around the world, with rich tree species and diverse crown types, enabling the framework to learn richer single-tree crown feature information; the experimental evaluation metrics used were overall accuracy (OA), region consistency (mIoU), and boundary accuracy (BF1).
[0048] Quantitative Evaluation: The significant performance advantages and quantitative evaluation results of all comparative methods are shown in Table 1, while the optimization effect of the embodiments of the present invention is as follows. Figure 5 , Figure 6 , Figure 7 As shown. Analysis of these results leads to the following conclusions: Table 1. Quantitative evaluation results of various methods on the test set.
[0049] First, the method in this embodiment significantly outperforms the comparative methods in both key metrics: region consistency (mIoU) and boundary accuracy (BF1). The mIoU of this embodiment reaches 74.03, a significant improvement of 6.42 compared to the second-best U2-Net method (67.61). This significant improvement in region consistency directly demonstrates that the canopy segmentation mask generated by this embodiment is more morphologically complete and semantically accurate, effectively overcoming the problem of false detections or breaks in complex backgrounds caused by the lack of strong semantic prior guidance in existing technologies. Simultaneously, the BF1 score of this embodiment reaches 36.32, an improvement of 5.74 compared to the second-best method (U2-Net). These results strongly demonstrate that the confidence-based dynamic region separation strategy and prior-driven conditional reasoning mechanism proposed in this embodiment can effectively separate and optimize easy and difficult samples step by step, and effectively utilize high-reliability regions as prior information to guide the iterative generation process of the model. This enables the model to generate single-tree crown segmentation results with clear boundaries and conforming to real phenological characteristics when dealing with complex forest areas.
[0050] Secondly, the method in this embodiment demonstrates significant iterative effectiveness in achieving progressive segmentation optimization from coarse to fine. Experimental data shows that as the number of iterations increases, the model's performance indicators exhibit a clear and continuous upward trend. Specifically, BF1 shows a significant leap starting from the second iteration and maintains a continuous growth trend until reaching 36.32% in the sixth iteration. This continuous performance gain obtained with each iteration directly proves that the segmentation results generated by the method in this embodiment can continuously approach the true boundary through multiple rounds of "reflection-correction" processes, effectively overcoming the limitation of existing end-to-end single-step prediction methods that cannot correct errors once output. These results strongly demonstrate that the dynamic region separation based on confidence decoupling and the prior-driven conditional reasoning mechanism proposed in this embodiment can successfully establish information associations in the time dimension and effectively utilize the reliable prediction results of the previous round as strong semantic priors to guide the resolution of local ambiguities in the current round, thereby ensuring that the model can gradually repair canopy adhesion and edge blurring, achieving a cumulative improvement in segmentation accuracy.
[0051] In summary, the quantitative experimental results confirm that the technical solution proposed in this embodiment establishes a novel and high-performance technical framework for high-resolution single-tree crown fine segmentation in complex forestry scenarios.
[0052] Qualitative assessment results, such as Figure 8 As shown in the visualization, the results intuitively demonstrate the actual effect of the method of the present invention in achieving single-tree crown segmentation, and further confirm that it has the following key advantages compared with existing methods: In the results of the first row, the method of this embodiment not only accurately identified the prominent single tree crown on the left, but also successfully detected the weaker target in the upper right corner that was partially obscured, while effectively avoiding missegmentation of the dense tree cluster area below. In contrast, the comparative method showed multiple missed detections and blurred boundaries in the same area.
[0053] In the results of the second row, the method of this embodiment demonstrates the ability to accurately identify the crowns of small-scale individual trees. The segmentation results have clear boundaries and complete shapes, and no significant background false detections are generated, reflecting the method's sensitivity and discrimination ability to detailed features.
[0054] In the third row of results, the method of this embodiment accurately separates the canopy of a single tree in the grassy background below, overcoming the interference of similar textures; at the same time, in the densely built-up lower and middle areas, the method of this embodiment can effectively identify the continuous canopy structure distributed laterally along the gaps between buildings, maintaining a high degree of target integrity, while the comparative methods generally show breakage or missing phenomena in this area.
[0055] Example 3, please refer to Figure 3 , Figure 3This is a schematic diagram of the structural composition of the iterative optimization single-tree crown segmentation device based on confidence decoupling and prior-driven iterative optimization in an embodiment of the present invention.
[0056] like Figure 3 As shown, an iterative optimization single-tree crown segmentation device based on confidence decoupling and prior-driven iteration is disclosed. The device includes: Tree crown segmentation module 301: used to obtain remote sensing image data, and input the remote sensing image data into the single tree crown segmentation model, which is processed by the neural network encoder and the neural network decoder in sequence, and outputs the coarse segmentation result of the single tree crown; In a specific implementation of this invention, the step of inputting the remote sensing image data into a single-tree crown segmentation model, processing it sequentially through a neural network encoder and a neural network decoder, and then outputting a coarse segmentation result for the single-tree crown includes: inputting the remote sensing image data into the single-tree crown segmentation model, processing it through the neural network encoder, and outputting original high-dimensional features containing crown information; inputting the original high-dimensional features into the neural network decoder, and outputting the coarse segmentation result for the single-tree crown, wherein the coarse segmentation result for the single-tree crown includes... The initial coarse segmentation result of the dimension and the corresponding pixel-level probability distribution map, where Original image size It is a set of classification categories.
[0057] Specifically, after obtaining remote sensing image data, the data is input into a single tree canopy segmentation model, where it is processed sequentially by a neural network encoder and a neural network decoder. After passing through the neural network encoder, the output is a high-dimensional feature containing canopy information. Then the high-dimensional features The input is then fed into the neural network decoder to obtain... Initial coarse segmentation results of the dimension and the pixel-level probability distribution map corresponding to the initial coarse segmentation result. ;in Original image size It is a set of classification categories.
[0058] Confidence assessment module 302: used to perform confidence assessment processing on the coarse segmentation result of the single tree crown based on the statistical calculation method of the pixel-level probability distribution map, and form the confidence assessment result corresponding to the coarse segmentation result of the single tree crown; In the specific implementation of this invention, the statistical calculation method based on pixel-level probability distribution maps is used to perform confidence assessment processing on the coarse segmentation results of the single tree crown, forming a confidence assessment result corresponding to the coarse segmentation results of the single tree crown, including: Obtain the pixel-level probability distribution map corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result, and use a preset statistical calculation formula to perform confidence evaluation processing on each initial coarse segmentation result to form the confidence evaluation result corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result. The preset statistical calculation formula is as follows: ; in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
[0059] Specifically, pixel-level confidence quantification aims to solve the problem of mathematically representing the difficulty levels of different regions in an image. Generally, in areas with clear textures, such as the center of the tree canopy, the features extracted by the model have significant discriminative power, and their classification probability distribution exhibits a peak-like shape. However, in areas such as the edges of the tree canopy, areas where features are fuzzy or semantically ambiguous, the probability distribution predicted by the model is often relatively flat. Therefore, the maximum classification probability can directly reflect the model's "certainty" regarding the current pixel. By establishing a pixel-level confidence index, the image space can be mapped from a single pixel grid to a "confidence potential surface" with high and low undulations, thus providing a quantitative benchmark for identifying difficult examples. The specific implementation involves statistically analyzing the initial segmentation probability map output by the model. For the Round of iteration output Dimensional pixel-level probability distribution map Define pixels confidence level The maximum predicted probability for this pixel across all categories is given by the following formula: ; in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
[0060] Feature fusion module 303: used to generate prior information from the confidence evaluation result through the prior encoder, and to fuse the prior information with the original high-dimensional features output by the neural network encoder to generate fused features; In the specific implementation of this invention, the step of generating prior information from the confidence evaluation result through a prior encoder includes: constructing a proportional threshold that evolves with each iteration using a cosine annealing function; sorting the confidence evaluation results corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result in ascending order, and generating a binary control mask based on the proportional threshold using the ascending sorting result; using the binary control mask to filter the single tree crown coarse segmentation result to form a filtered single tree crown coarse segmentation result; inputting the filtered single tree crown coarse segmentation result into the prior encoder, and outputting prior information aligned with the original high-dimensional feature dimension, wherein the prior encoder is a shallow convolutional network.
[0061] Furthermore, the formula for generating a binary control mask based on the ascending sorting result using the ratio threshold is as follows: ; ; in, This is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
[0062] Furthermore, the step of fusing the prior information with the original high-dimensional features output by the neural network encoder to generate fused features includes: performing an additive fusion process on the prior information and the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ; in, Features of fusion; These are the original high-dimensional features; This is prior information.
[0063] Specifically, generating prior information requires dynamic separation and mask generation, with the core challenge being resource allocation and convergence control during iterative optimization. While confidence quantification identifies challenging areas, focusing on only a few boundaries in the early stages of iteration may prevent the model from correcting widespread contour errors; conversely, continuous large-scale updates not only waste computational resources but can also lead to model oscillations. Therefore, this invention introduces a "coarse-to-fine" dynamic separation mechanism that aligns with human cognitive patterns. In the early stages of iteration, a wider range of corrections is allowed to adjust the topology; as iteration progresses, the separation boundaries gradually shrink, forcing the model to focus only on pixel-level boundaries that are extremely difficult to distinguish. This dynamic separation ensures the stability and efficiency of the optimization process.
[0064] The implementation process mainly includes two steps: timing scheduling and space separation. The first step is time-series scheduling: a time-series scheduling mechanism is constructed based on the iteration rounds. Evolutionary Proportion Threshold The cosine annealing function is preferred for nonlinear scheduling, as shown in the following formula: ; in This represents the total number of iterations; this function ensures that the proportion to be optimized smoothly decreases from 1 to 0.
[0065] The second step is spatial separation: sorting all pixels by confidence level. Sort in ascending order and select the first few. A proportional number of pixels are used as the "region to be optimized" to generate a binary control mask. The formula is expressed as: ; in, This is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
[0066] The image is separated into regions that need to be updated and reliable regions that remain unchanged.
[0067] By fusing prior information with the original high-dimensional features, this approach aims to address the lack of contextual reference in models when handling blurred boundaries. When segmenting canopy adhesion boundaries, relying solely on current local visual features often makes it difficult for models to determine boundary affiliation. However, the main canopy area (high-reliability region) identified in the previous iteration contains spatial location and semantic category information; fully leveraging this information can aid in inferring the categories of adjacent blurred pixels. Therefore, by mapping discrete labels to continuous high-dimensional features and explicitly injecting them into the visual features, the model can more accurately repair edges given the knowledge that the surrounding area is a canopy.
[0068] First, there is a priori encoding, using the resulting mask. Regarding the results of the previous round of predictions Filter and keep only the locked area. The filtered results are then fed into a lightweight prior encoder, designed as a shallow convolutional network, to map the filtered segmentation results into high-dimensional features aligned with the visual feature dimensions. .
[0069] Then, the prior information is added to and fused with the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ; in, Features of fusion; These are the original high-dimensional features; This is prior information.
[0070] Optimization Iteration Module 304: This module is used to input the fused features into the neural network decoder for the next optimization segmentation process, output the next single tree crown optimization segmentation result, and return to the confidence evaluation process step for continuous optimization iteration segmentation until the number of optimization iterations reaches a preset number, at which point the final single tree crown optimization segmentation result is output.
[0071] In the specific implementation of this invention, the step of inputting the fused features into the neural network decoder for the next optimized segmentation process and outputting the next optimized segmentation result of the single tree crown includes: inputting the fused features into the neural network decoder, the neural network decoder outputting a new coarse segmentation result of the single tree crown; using a binary control mask, the new coarse segmentation result of the single tree crown and the previous coarse segmentation result of the single tree crown for the next optimized segmentation process, and outputting the next optimized segmentation result of the single tree crown. The formula for the next optimized segmentation process is as follows: ; in, To optimize the segmentation of the individual tree crown for the next time; This is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
[0072] Specifically, the fused features are input into the neural network decoder, which then outputs a new coarse segmentation result for the single tree crown. The core principle is to ensure the monotonicity and stability of the iterative process. To avoid compromising the original correct results while correcting errors, this invention employs a strict gating update logic. Based on the previously separated "reliable regions," during state updates, the prediction results are updated only within the identified "regions to be optimized," preserving the results for the reliable regions. This strategy effectively avoids erroneous changes to the inference results, ensuring that the segmentation accuracy steadily improves with each iteration.
[0073] Then, the binary control mask, the new coarse segmentation result of the single tree crown, and the previous coarse segmentation result of the single tree crown are used to perform the next optimization segmentation process, and the next optimized segmentation result of the single tree crown is output. The formula for the next optimized segmentation process is as follows: ; in, To optimize the segmentation of the individual tree crown for the next time; This is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
[0074] This processing ensures that only low-reliability areas indicated by the mask are addressed. When updating the state, results from low-reliability regions are discarded, while results from high-reliability regions are updated. Maintain the previous round of forecast values. constant.
[0075] The final optimized segmentation result for a single tree crown is output when the preset number of optimization iterations is reached.
[0076] In this embodiment of the invention, the boundary precision and regional consistency of remote sensing image segmentation in complex forestry scenarios are significantly improved. The dynamic region separation strategy based on prediction confidence effectively solves the problem that existing single-step prediction methods use the same processing method for pixels of varying difficulty, leading to difficulty in accurately segmenting contiguous boundaries. Through a priori-driven conditional inference mechanism, historical prediction information is used to guide the current segmentation process. The high-reliability prediction results from the previous round are mapped to feature vectors and fused with current visual features, enabling the model to reference the identified canopy location information during inference. A selective update rule is adopted, accepting new prediction values only in low-reliability regions while maintaining the original results in high-reliability regions. This allows the generated segmentation results to infer the attribution of fuzzy boundaries based on the determined canopy subject, effectively suppressing background interference and ensuring the integrity of individual tree canopy morphology.
[0077] This invention provides a computer-readable storage medium storing a computer program. When executed by a processor, this program implements the iterative optimization method for single-tree crown segmentation as described in any of the above embodiments. The computer-readable storage medium includes, but is not limited to, any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. In other words, the storage device includes any medium that stores or transmits information in a readable form by a device (e.g., a computer, a mobile phone), and can be a read-only memory, a disk, or an optical disk, etc.
[0078] This invention also provides a computer application running on a computer, which is used to execute the iterative optimization single-tree crown segmentation method of any of the above embodiments.
[0079] also, Figure 4 This is a schematic diagram of the structural composition of the electronic device in an embodiment of the present invention.
[0080] This invention also provides an electronic device, such as... Figure 4 As shown. The electronic device includes components such as a processor 402, a memory 403, an input unit 404, and a display unit 405. Those skilled in the art will understand that... Figure 4The structural components of the illustrated electronic device do not constitute a limitation on all devices and may include more or fewer components than illustrated, or combine certain components. Memory 403 can be used to store application program 401 and various functional modules. Processor 402 runs application program 401 stored in memory 403, thereby performing various functional applications and data processing of the device. Memory can be internal memory or external memory, or both. Internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory. External memory may include hard disks, floppy disks, ZIP disks, USB flash drives, magnetic tapes, etc. The memory disclosed in this invention includes, but is not limited to, these types of memory. The memory disclosed in this invention is only an example and not a limitation.
[0081] Input unit 404 is used to receive signal input and user-input keywords. Input unit 404 may include a touch panel and other input devices. The touch panel can collect user touch operations on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel) and drive the corresponding connection device according to a pre-set program; other input devices may include, but are not limited to, one or more of physical keyboards, function keys (such as play control buttons, power buttons, etc.), trackballs, mice, joysticks, etc. Display unit 405 can be used to display user-input information or information provided to the user, as well as various menus of the terminal device. Display unit 405 may be in the form of a liquid crystal display, organic light-emitting diode, etc. Processor 402 is the control center of the terminal device, connecting various parts of the entire device through various interfaces and lines, performing various functions and processing data by running or executing software programs and / or modules stored in memory 403, and calling data stored in memory.
[0082] As one embodiment, the electronic device includes: one or more processors 402, a memory 403, and one or more application programs 401, wherein the one or more application programs 401 are stored in the memory 403 and configured to be executed by the one or more processors 402, and the one or more application programs 401 are configured to execute the iteratively optimized single-tree crown segmentation method corresponding to any of the embodiments described above.
[0083] In this embodiment of the invention, the boundary precision and regional consistency of remote sensing image segmentation in complex forestry scenarios are significantly improved. The dynamic region separation strategy based on prediction confidence effectively solves the problem that existing single-step prediction methods use the same processing method for pixels of varying difficulty, leading to difficulty in accurately segmenting contiguous boundaries. Through a priori-driven conditional inference mechanism, historical prediction information is used to guide the current segmentation process. The high-reliability prediction results from the previous round are mapped to feature vectors and fused with current visual features, enabling the model to reference the identified canopy location information during inference. A selective update rule is adopted, accepting new prediction values only in low-reliability regions while maintaining the original results in high-reliability regions. This allows the generated segmentation results to infer the attribution of fuzzy boundaries based on the determined canopy subject, effectively suppressing background interference and ensuring the integrity of individual tree canopy morphology.
[0084] Furthermore, the above provides a detailed description of the iterative optimization method and related apparatus for single-tree crown segmentation based on confidence decoupling and prior-driven approaches provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A single-tree crown segmentation method based on confidence decoupling and prior-driven iterative optimization, characterized in that, The method includes: Remote sensing image data is obtained and input into a single tree crown segmentation model. After being processed by a neural network encoder and a neural network decoder, the coarse segmentation result of the single tree crown is output. The confidence assessment of the coarse segmentation results of the single tree crown is performed based on the statistical calculation method of pixel-level probability distribution map, thereby forming the confidence assessment result corresponding to the coarse segmentation results of the single tree crown. The confidence evaluation result is used to generate prior information through a prior encoder, and the prior information is fused with the original high-dimensional features output by the neural network encoder to generate fused features. The fused features are input into the neural network decoder for the next optimized segmentation process, and the next optimized segmentation result of the single tree crown is output. The process is then returned to the confidence evaluation step for continuous optimization and iterative segmentation until the number of optimization iterations reaches the preset number, at which point the final optimized segmentation result of the single tree crown is output.
2. The iterative optimization method for single-tree crown segmentation according to claim 1, characterized in that, The process of inputting the remote sensing image data into the single-tree crown segmentation model, which is then processed sequentially by a neural network encoder and a neural network decoder, outputs a coarse segmentation result for the single-tree crown, including: The remote sensing image data is input into the single tree crown segmentation model, and after passing through the neural network encoder, the original high-dimensional features containing crown information are output. The original high-dimensional features are input into the neural network decoder, which outputs the coarse segmentation result of the single tree crown. The coarse segmentation result of the single tree crown includes... The initial coarse segmentation result of the dimension and the corresponding pixel-level probability distribution map, where Original image size It is a set of classification categories.
3. The iterative optimization method for single-tree crown segmentation according to claim 1, characterized in that, The statistical calculation method based on pixel-level probability distribution maps is used to perform confidence assessment on the coarse segmentation results of the single tree crown, forming a confidence assessment result corresponding to the coarse segmentation results of the single tree crown, including: Obtain the pixel-level probability distribution map corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result, and use a preset statistical calculation formula to perform confidence evaluation processing on each initial coarse segmentation result to form the confidence evaluation result corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result. The preset statistical calculation formula is as follows: ； in, A collection of categories; For pixels Confidence level at iteration t; for The pixel-level probability distribution of the dimension at iteration round t; The closer the value is to 1, the more it indicates The corresponding initial coarse segmentation result is a high-reliability region; in The closer the value is to 0, the more it indicates The corresponding initial coarse segmentation result is a low-reliability region.
4. The iterative optimization method for single-tree crown segmentation according to claim 1, characterized in that, The step of generating prior information from the confidence assessment result using a prior encoder includes: A proportional threshold that evolves with each iteration is constructed using a cosine annealing function. The confidence evaluation results corresponding to each initial coarse segmentation result in the single tree crown coarse segmentation result are sorted in ascending order, and a binary control mask is generated based on the ratio threshold using the ascending sorting results. The binary control mask is used to filter the coarse segmentation result of the single tree crown, resulting in a filtered coarse segmentation result of the single tree crown. The filtered coarse segmentation result of the single tree crown is input into the prior encoder, which outputs prior information aligned with the original high-dimensional feature dimension. The prior encoder is a shallow convolutional network.
5. The iterative optimization method for single-tree crown segmentation according to claim 4, characterized in that, The formula for generating a binary control mask based on the ascending sorting result using the ratio threshold is as follows: ； ； in, It is a binary control mask; This is the proportional threshold that varies with the iteration round t. This refers to the preset number of optimization iterations; For pixels The confidence level at iteration round t.
6. The iterative optimization method for single-tree crown segmentation according to claim 1, characterized in that, The step of fusing the prior information with the original high-dimensional features output by the neural network encoder to generate fused features includes: The prior information is added to and fused with the original high-dimensional features output by the neural network encoder to generate fused features; The formula for addition fusion is expressed as follows: ； in, Features of fusion; The original high-dimensional features; This is prior information.
7. The iterative optimization method for single-tree crown segmentation according to claim 1, characterized in that, The step of inputting the fused features into the neural network decoder for the next optimized segmentation process and outputting the next optimized segmentation result for the single tree crown includes: The fused features are input into the neural network decoder, and the neural network decoder outputs a new coarse segmentation result of the single tree crown; The next optimization segmentation process is performed using a binary control mask, the new coarse segmentation result of the single tree crown, and the previous coarse segmentation result of the single tree crown, and the next optimized segmentation result of the single tree crown is output. The formula for the next optimized segmentation process is as follows: ； in, To optimize the segmentation of the individual tree crown for the next time; It is a binary control mask; This is the result of a new coarse segmentation of the single tree crown; This is the result of the previous coarse segmentation of the single tree crown; To replace the previous low-confidence assessment results with the new coarse segmentation results of the single tree crown; To preserve the high-confidence evaluation result area from the previous iteration.
8. A single-tree crown segmentation device based on confidence decoupling and prior-driven iterative optimization, characterized in that, The device includes: Tree crown segmentation module: used to obtain remote sensing image data, and input the remote sensing image data into the single tree crown segmentation model. After being processed by the neural network encoder and the neural network decoder, the module outputs the coarse segmentation result of the single tree crown. Confidence assessment module: used to perform confidence assessment on the coarse segmentation results of the single tree crown based on the statistical calculation method of pixel-level probability distribution map, and form the confidence assessment result corresponding to the coarse segmentation results of the single tree crown; Feature fusion module: used to generate prior information from the confidence evaluation result through the prior encoder, and to fuse the prior information with the original high-dimensional features output by the neural network encoder to generate fused features; The optimization iteration module is used to input the fused features into the neural network decoder for the next optimization segmentation process, output the next single tree crown optimization segmentation result, and return to the confidence evaluation processing step to perform continuous optimization iteration segmentation processing until the number of optimization iterations reaches the preset number, and then output the final single tree crown optimization segmentation result.
9. An electronic device comprising a processor and a memory, characterized in that, The processor runs a computer program or code stored in the memory to implement the iterative optimization method for single-tree crown segmentation as described in any one of claims 1 to 7.
10. A computer-readable storage medium for storing computer programs or code, characterized in that, When the computer program or code is executed by a processor, it implements the iterative optimization method for single-tree crown segmentation as described in any one of claims 1 to 7.