Method for identifying color-changing standing trees based on phenology and multi-temporal remote sensing images

Abstract

The present application relates to a method for identifying color-changing standing trees based on phenology and multi-temporal remote sensing images, belonging to the field of remote sensing image processing and forestry information technology, and solving the deficiencies of existing technologies in terms of color-changing standing tree identification accuracy, efficiency, robustness, etc. The method comprises: classifying multi-temporal remote sensing images into different phenological subsets according to phenological periods; constructing a standardized sample library; training a window-level classification model; performing coarse screening with a large step size and then performing fine determination on the target remote sensing image to be detected; calculating the disease probability accumulation value of each pixel point to generate a disease mask matrix; and sequentially performing connected domain analysis and morphological operation on the disease mask matrix to obtain the final disease distribution mask matrix and output. The present application uses window-level samples for training and reasoning, and realizes an automatic identification process from large image discrimination to high-resolution mask generation through multi-step sliding windows and probability accumulation strategies, achieving efficient, accurate and automatic color-changing standing tree identification.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and forestry information technology, and in particular to a method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images. This method can be applied to fields such as forest pest and disease monitoring, remote sensing image analysis of vegetation phenology, ecosystem protection, and intelligent forestry management. Background Technology

[0002] Forest resources are a vital ecological barrier for the nation, playing an irreplaceable role in maintaining ecological security, addressing climate change, and ensuring the sustainable development of the forestry economy. However, in recent years, major forest pests such as pine wilt disease have continued to spread, causing large areas of pine forests to wither and die, seriously threatening the stability and health of forest ecosystems. How to conduct efficient and accurate large-scale pest and disease monitoring has become a core issue that forestry departments urgently need to address.

[0003] Remote sensing technology, with its advantages of wide coverage, high timeliness, and non-contact acquisition, has become an important means of monitoring forest pests and diseases. Identifying discoloration in pine trees through remote sensing imagery can reveal potentially affected areas on a large scale, enabling early warning and control. However, existing methods still have the following significant shortcomings:

[0004] 1. Pixel-level annotation is extremely labor-intensive and has low feasibility:

[0005] Current deep learning-based identification methods mostly rely on large-scale, pixel-level fine-grained mask samples for training. On high-resolution remote sensing images with a resolution better than 1m, labeling the insect-damaged color-changing areas pixel by pixel in a 10km×10km area, especially repeatedly creating samples under different seasons and different ground cover backgrounds, is extremely labor-intensive and complex in practice, making it difficult to apply on a large scale.

[0006] 2. Macroscale analysis lacks precision:

[0007] Existing methods often rely on macroscopic analysis of whole-view images or processing based on large-scale windows, which can easily mask the subtle spectral and textural features of individual diseased pine trees, resulting in a high rate of missed detection in early disease identification.

[0008] 3. Phenological changes and pathological changes are easily confused:

[0009] Pine trees exhibit color changes during different phenological stages, such as turning green in spring and yellowing in autumn. These natural changes are similar to pathological discoloration in their spectral characteristics, making it difficult for existing methods to effectively distinguish them, resulting in a high misdiagnosis rate.

[0010] 4. Efficiency and accuracy are difficult to balance:

[0011] Faced with massive amounts of high-resolution remote sensing data, pixel-level fine segmentation would require an enormous amount of computation and result in low processing efficiency; while rapid detection methods would struggle to guarantee the accuracy and boundary recognition of small targets. There is a lack of a solution that balances efficiency and accuracy.

[0012] In summary, existing technologies for remote sensing monitoring of pests and diseases generally suffer from bottlenecks such as high annotation costs, insufficient detection accuracy, high false positive rates, and difficulty in balancing efficiency. Therefore, there is an urgent need for a new identification method that can achieve efficient and accurate identification of pine wilt disease-affected trees without large-scale, pixel-level manual annotation. Summary of the Invention

[0013] This invention aims to provide a method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images, in order to overcome the shortcomings of existing technologies in terms of sample preparation, accuracy, efficiency, and robustness of discolored standing tree identification, and to achieve high-precision and automated monitoring of pine wilt disease over a wide range of areas.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] A method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images includes the following steps:

[0016] Step 1: Acquire multi-temporal remote sensing images covering the target forest area. After preprocessing the multi-temporal remote sensing images, classify each multi-temporal remote sensing image into the corresponding phenological period subset according to the acquisition time and center point latitude of the multi-temporal remote sensing images, forming an image database classified according to the time dimension.

[0017] Step 2: Crop the images in the image database according to a fixed window size of 256×256 to construct a standardized sample library containing positive and negative samples. The positive samples are image blocks containing confirmed pine wood nematode diseased wood, and the negative samples are image blocks not containing pine wood nematode diseased wood.

[0018] Step 3: Use a convolutional neural network to train a standardized sample library to extract multi-dimensional spectral and spatial features of image patches. After training, a window-level classification model is obtained. The input of this model is an image patch of size 256×256, and the output is the disease probability.

[0019] Step 4: On the remote sensing image of the target to be detected, create a sliding window of fixed size 256×256 with the first step size as the sliding step size. Each time the sliding window slides, the image patch corresponding to the sliding window is captured and input into the window-level classification model to obtain the disease probability value corresponding to the image patch. It is determined in real time whether the obtained disease probability value is greater than or equal to the first threshold. If it is, the fine judgment process in Step 5 is executed; if the disease probability values ​​corresponding to all captured image patches are less than the first threshold, the entire target remote sensing image is determined to be disease-free, and the disease-free detection result of the entire image is directly output.

[0020] Step 5: Using a second step size smaller than the first step, perform dense sliding window inference on the entire target remote sensing image, calculate the cumulative disease probability value of each pixel, construct a disease probability accumulation matrix, and generate a disease mask matrix based on the disease probability accumulation matrix and the second threshold.

[0021] Step 6: Perform connected component analysis and morphological operations on the disease mask matrix generated in Step 5 to remove noise and smooth regions, and output the final disease distribution mask matrix.

[0022] This invention provides a method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images. This method abandons the traditional approach that requires numerous pixel-level labeled masks, instead relying entirely on window-level samples for training and inference. Through multi-step sliding windows and a probability accumulation strategy, it achieves an automated identification process from large-image discrimination to high-resolution mask generation. It can perform temporal classification by combining phenological information, starting from standardized small-scale units, and utilizing deep learning and multi-model collaboration to achieve efficient, accurate, and automated identification of discolored standing trees affected by pine wilt disease. Compared with existing technologies, this invention has the following advantages:

[0023] (1) No pixel-level manual annotation required: This method is based on window-level sample training, avoiding the high cost of mask production;

[0024] (2) Balancing efficiency and accuracy: By adopting a two-stage strategy of "coarse screening + fine judgment", the amount of computation is greatly reduced while maintaining high accuracy;

[0025] (3) Flexible adaptation to multiple steps: coarse screening is used for rapid detection, and fine screening is used for local refinement, adapting to different application needs;

[0026] (4) Strong robustness: Combining phenological period classification reduces the interference of seasonal color changes on disease identification;

[0027] (5) Good scalability: This method can be extended to other forestry pest and disease monitoring scenarios to achieve efficient automatic identification of cross-regional and multi-temporal remote sensing. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the color-changing tree identification method described in an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of a sliding window on a portion of a remote sensing image of a target, as described in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the disease mask matrix in an embodiment of the present invention. Detailed Implementation

[0031] The specific implementation methods of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0032] like Figure 1 As shown, this embodiment provides a method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images. The method includes the following steps:

[0033] Step 1: Acquisition and classification of multi-temporal remote sensing images;

[0034] Step 2: Construct a standardized sample library;

[0035] Step 3: Feature extraction and window-level classification model training;

[0036] Step 4: Coarse screening of target remote sensing images;

[0037] Step 5: Multi-step probability accumulation mask generation;

[0038] Step Six: Post-processing and Result Output.

[0039] Specifically, step one: acquire multi-temporal remote sensing images covering the target forest area, preprocess the multi-temporal remote sensing images, and classify each multi-temporal remote sensing image into the corresponding phenological period subset according to the acquisition time and center point latitude of the multi-temporal remote sensing images, forming an image database classified according to the time dimension.

[0040] In this step, after acquiring multi-temporal remote sensing images covering the target forest area, the multi-temporal remote sensing images first undergo preprocessing such as radiometric correction, atmospheric correction, and geometric registration. In this embodiment, the multi-temporal remote sensing images are multispectral remote sensing images, which include at least four spectra: blue light band, green light band, red light band, and near-infrared band.

[0041] After preprocessing, based on the acquisition time and center point latitude of the multi-temporal remote sensing images, the images are classified into corresponding phenological period subsets according to the pre-divided phenological periods, forming an image database classified according to the time dimension.

[0042] Define the phenological window function Subset:

[0043] Subset = f(φ,t)

[0044] Where φ represents the latitude of the center point of the multi-temporal remote sensing image; t represents the acquisition time of the multi-temporal remote sensing image; f() represents the mapping relationship from the image acquisition time t and the center point latitude φ to the phenological period subset (dormancy period, early growth period, etc.). The phenological period window function Subset is used to determine the phenological period to which each multi-temporal remote sensing image belongs.

[0045] This embodiment combines phenological classification, effectively reducing the interference of seasonal color changes on disease identification, and can further improve the accuracy of identifying pine wilt disease-affected trees with discoloration.

[0046] Based on the phenological period division table, this invention divides the annual growth cycle of pine forests into four main phenological periods: Dormancy, Early Growth, Peak Growth, and Senescence. Multi-temporal remote sensing images are automatically categorized into these four subsets: Dormancy, Early Growth, Peak Growth, and Senescence.

[0047] This classification method is simple and effective, reflecting the main physiological changes of pine trees in different seasons. Since latitudinal differences significantly affect plant growth rhythms, this invention further dynamically adjusts the time range of phenological periods based on latitudinal zones.

[0048] The phenological periods in high-latitude regions (≥40°N, such as Liaoning and Jilin) ​​are as follows:

[0049] Dormant period: Late November to mid-March of the following year;

[0050] Early growth period: Late March to late May;

[0051] Peak growing season: Early June to late August;

[0052] Recession period: early September to mid-November.

[0053] The corresponding phenological periods in mid-latitude regions (30°–40°N, such as Hebei and northern Sichuan) are:

[0054] Dormancy period: Mid-December to late February of the following year;

[0055] Early growth period: early March to early May;

[0056] Peak growing season: Mid-May to early September;

[0057] Recession period: Mid-September to early December.

[0058] The corresponding phenological periods in low-latitude regions (≤30°N, such as Fujian and Jiangxi) are:

[0059] Dormancy period: early January to mid-February;

[0060] Early growth period: Late February to early April;

[0061] Peak growing season: Mid-April to early October;

[0062] Recession period: Mid-October to late December.

[0063] It should be noted that the time range of the above-mentioned phenological period will fluctuate with the seasonal changes of pine trees.

[0064] Major physiological changes in different seasons

[0065] Step 2: Crop the images in the image database using a fixed window size of 256×256 to construct a standardized sample library containing positive and negative samples.

[0066] For the categorized images in the image database obtained in step one, a standardized sample library containing positive and negative samples is constructed by cropping the images using a fixed preset window size. The preset window size is 256×256 pixels. The definitions of positive and negative samples are as follows:

[0067] Positive samples: Image blocks containing pine wood nematode lesions confirmed by field surveys, expert interpretation, or existing data.

[0068] Negative samples: Image patches that do not contain pine wood nematode-infected trees, specifically including the following types:

[0069] Healthy pine trees;

[0070] Other types of vegetation;

[0071] Non-vegetated ground features (such as bare land, water bodies, roads, buildings, etc.);

[0072] Non-disease-related discoloration of standing trees caused by phenological or environmental factors.

[0073] Image blocks including healthy pine trees, other vegetation, non-vegetation features, and non-pathological discolored wood.

[0074] The formal representation of a standardized sample library is as follows:

[0075] S={(x i ,y i )∣x i ∈R 256×256×c ,yi ∈{0,1}}

[0076] Where S represents the standardized sample library; x i y is the i-th sample in the standardized sample library; i Let R be the sample label for the i-th sample, where 0 represents a negative sample and 1 represents a positive sample; 256×256×c This represents all three-dimensional tensors composed of real numbers, with a size of 256×256×c, where c represents the number of spectra (or bands), and its value is 4, specifically the four spectra: blue light band, green light band, red light band, and near-infrared band.

[0077] Step 3: Use a convolutional neural network to train a standardized sample library to extract multi-dimensional spectral and spatial features of image patches, and then build a window-level classification model. The input of this model is an image patch of size 256×256, and the output is the disease probability.

[0078] When a convolutional neural network extracts multidimensional spectral and spatial features from an image patch, the convolutional features are expressed as:

[0079]

[0080] Among them, b j As a bias, σ is the activation function, and here a band dimension k is introduced:

[0081] x k (p+q) is the pixel value of the k-th band at position p+q;

[0082] w j,k (q) represents the weight of the convolution kernel on band k;

[0083] F j (p) represents the feature response value extracted by the j-th convolution kernel at position p, which is the output on the feature map after the convolution operation;

[0084] q represents the relative position index of the convolution kernel in the receptive field Ω. For example, in a 3×3 convolution, q traverses these 9 positions in turn.

[0085] Ω represents the receptive field of the convolution kernel (i.e., the neighborhood range covered by the convolution kernel). For example, for a 3×3 convolution kernel, Ω represents the set of pixels in that 3×3.

[0086] After training the convolutional neural network on a standardized sample library, a window-level classification model f is finally obtained. θ The model takes a 256×256 pixel image patch as input and outputs the disease probability, expressed by the following formula:

[0087] P(W i)=f θ (W i ),P(W i )∈[0,1]

[0088] Among them, P(W i ) represents the probability of disease, W i The input is an image patch of size 256×256.

[0089] This embodiment eliminates the need for pixel-level manual annotation, instead relying on window-level sample training, thus avoiding the high cost of mask creation, resulting in higher efficiency and lower cost.

[0090] Through the above operations, the convolutional neural network can simultaneously extract multi-band spectral and spatial structural features from standardized image patches and convert them into high-level feature representations. Using these features, this invention trains on a large number of labeled positive and negative samples, optimizes the convolutional kernel weights and biases, and finally obtains a window-level classification model. This model can output the probability value of an image patch being a diseased region when the input is a fixed-size image patch, thereby achieving window-level automated discrimination.

[0091] In this step, the convolutional neural network training does not use random remote sensing images, but standardized image patches of 256×256 size. The training result is not a fine pixel grid, but a judgment of whether the 256×256 image patch has insect pest characteristics: "yes" (1) or "no" (0).

[0092] Step 4: On the remote sensing image I to be detected, create a sliding window of fixed size 256×256 with the first step size as the sliding step size. Each time the sliding window slides, the image patch corresponding to the sliding window is captured and input into the window-level classification model to obtain the disease probability value corresponding to the image patch. In real time, it is determined whether the obtained disease probability value is greater than or equal to the first threshold. If so, the fine judgment process in Step 5 is executed; if the disease probability values ​​corresponding to all captured image patches are less than the first threshold, the entire target remote sensing image is determined to be disease-free, and the disease-free detection result of the entire image is directly output.

[0093] This step performs a coarse screening of the remote sensing image I of the target to be detected.

[0094] by Figure 2Taking the remote sensing image I of the target shown as an example, on the remote sensing image I of the target to be detected, a sliding window of a preset size (256×256 pixels) is created with a sliding step size of the first step length S1 (e.g., 128 pixels). The size of the sliding window is the same as the window size in step two. It is traversed sequentially starting from the beginning position (or the end position) of the image. Each time the sliding window slides according to the first step length S1, the image patch corresponding to the sliding window is captured. The image patch is input into the window-level classification model trained in step three. The window-level classification model outputs the disease probability value P(W) corresponding to the image patch. i );

[0095] During the traversal, the probability value P(W) of the disease corresponding to the obtained sliding window is determined in real time. i Is it less than a pre-set first threshold τ?

[0096] If P(W) of all sliding windows i If ) < τ, then all sliding windows are determined to be "insect-free", that is, the entire target remote sensing image is free of defects, and the detection result of no defects in the entire image is directly output;

[0097] If there exists at least one sliding window P(W) i If )≥τ, then it is determined that there is at least one sliding window that is "with insects", that is, the target remote sensing image is a suspected disease image, and at this time the fine judgment process in step five is executed.

[0098] Step 5: Using a second step size smaller than the first step, perform dense sliding window inference on the entire target remote sensing image I, calculate the cumulative disease probability value of each pixel, construct the disease probability accumulation matrix, and generate a disease mask matrix based on the disease probability accumulation matrix and the second threshold.

[0099] If the target remote sensing image I is initially determined to be "infested with insects," a second step size S2 (S2∈2,4,8,…32 pixels) smaller than the first step size S1 is used. A suitable second step size S2 is selected based on the required accuracy. Dense sliding window inference is then performed on the target remote sensing image I to calculate the cumulative disease probability A(p) for each pixel. Finally, a disease probability accumulation matrix is ​​constructed based on the cumulative disease probability values ​​of each pixel. The formula for calculating the cumulative disease probability A(p) for pixel p is as follows:

[0100]

[0101] in:

[0102] p represents a pixel;

[0103] N(p) is the number of windows covering pixel p;

[0104] P(W i) is the window W i The probability value of disease;

[0105] i:p∈W i This means that during the summation process, only those windows W that cover pixel p are summed. i Accumulate. i:p∈W i This represents a constraint that means traversing all window indices i, but only selecting those windows containing pixel p to participate in the summation.

[0106] Based on the disease probability accumulation matrix, the disease probability accumulation value A(p) of each pixel is compared with the second threshold τ. A The formula for generating the disease mask matrix is ​​as follows:

[0107]

[0108] Where I represents the remote sensing image of the target to be detected.

[0109] This yields a high-resolution lesion mask matrix M(p), whose resolution corresponds to the selected second step size S2 (2, 4, 8…32 pixels). Figure 3 for Figure 2 The lesion mask matrix M(p) generated corresponding to the remote sensing image of the partial target shown.

[0110] This embodiment employs a two-stage strategy of "coarse screening + fine judgment," significantly reducing computational load while maintaining high accuracy, thus balancing efficiency and precision in identifying discolored trees. Furthermore, coarse screening is used for rapid detection, while fine screening is used for localized refinement, making it more flexible and adaptable to different application needs.

[0111] Step 6: Perform connected component analysis and morphological operations on the disease mask matrix generated in Step 5 in sequence, remove noise and smooth the region to obtain the final disease distribution mask matrix and output it. The region with a pixel value of 1 accurately marks the spatial location of the discolored standing tree.

[0112] This step performs connected component analysis and morphological operations on the disease mask matrix M(p), removing isolated small patches and noisy regions to obtain and output the final disease distribution mask matrix. This disease distribution mask matrix is ​​the identification result of discolored trees in the target remote sensing image of this invention. It marks the spatial location and distribution range of the identified discolored trees in the target remote sensing image with pixel-level precision. The disease distribution mask matrix can not only mark the spatial location of suspected disease areas, but also provide refined disease representation at different resolutions (2 / 4 / 8 / 16 / 32 pixels).

[0113] After generating the initial disease mask matrix, this invention further optimizes the results using connected component analysis and morphological operations, specifically including:

[0114] (1) Connected component analysis includes the following steps:

[0115] The disease mask matrix generated in step five is labeled with connected components, and the area of ​​each connected component is calculated.

[0116] A minimum area threshold is set, and the area of ​​each connected component is compared with the preset minimum area threshold. Connected components with an area smaller than the minimum area threshold are removed, that is, small isolated areas are removed, thereby removing noise interference and avoiding noise points from interfering with the recognition results.

[0117] (2) Morphological operations include the following steps:

[0118] The defect mask matrix after connected component analysis is processed by opening operation (erosion + dilation) to remove isolated small noise.

[0119] The mask matrix after the opening operation is processed by the closing operation (dilation + erosion) to fill the internal voids, smooth the boundaries, and enhance the continuity of the diseased area.

[0120] After the above post-processing, the output disease distribution mask matrix is ​​smoother and more complete, which can accurately mark the spatial location of suspected disease areas and effectively suppress noise interference.

[0121] In the above embodiments, the preset window size is selected as 256×256 pixels. This is the optimal balance scale determined based on experimental results from multiple sets of sub-meter resolution remote sensing images.

[0122] While small windows (such as 32×32 or 64×64) can capture local details, they are easily affected by noise and differences in ground background, leading to unstable recognition results.

[0123] While large windows (such as 512×512 or 1024×1024) can cover a larger spatial range, they introduce too much heterogeneous information, reducing the model's sensitivity to diseases in individual plants or small areas, and increasing the computational load.

[0124] At sub-meter resolution, a 256×256 pixel window ensures sufficient contextual information for pattern discrimination while maintaining reasonable computational overhead. Test results show that this scale achieves an optimal balance between recognition accuracy and processing efficiency.

[0125] This invention provides a method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images. This method abandons the traditional approach that requires numerous pixel-level labeled masks, instead relying entirely on window-level samples for training and inference. Through multi-step sliding windows and probability accumulation strategies, it achieves an automated identification process from large-image discrimination to high-resolution mask generation. It can perform temporal classification by combining phenological information, starting from standardized small-scale units, and utilizing deep learning and multi-model collaboration to achieve efficient, accurate, and automated identification of discolored standing trees affected by pine wilt disease. This method can be extended to other forestry pest and disease monitoring scenarios, enabling efficient and automatic identification across regions and multiple temporal phases using remote sensing.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for identifying discolored standing trees based on phenological changes and multi-temporal remote sensing images, characterized in that, Includes the following steps: Step 1: Acquire multi-temporal remote sensing images covering the target forest area. After preprocessing the multi-temporal remote sensing images, classify each multi-temporal remote sensing image into the corresponding phenological period subset according to the acquisition time and center point latitude of the multi-temporal remote sensing images, forming an image database classified according to the time dimension. Step 2: Crop the images in the image database according to a fixed window size of 256×256 to construct a standardized sample library containing positive and negative samples. The positive samples are image blocks containing confirmed pine wood nematode diseased wood, and the negative samples are image blocks not containing pine wood nematode diseased wood. Step 3: Use a convolutional neural network to train a standardized sample library to extract multi-dimensional spectral and spatial features of image patches. After training, a window-level classification model is obtained. The input of this model is an image patch of size 256×256, and the output is the disease probability. Step 4: On the remote sensing image of the target to be detected, create a sliding window of fixed size 256×256 with the first step size as the sliding step size. Each time the sliding window slides, the image patch corresponding to the sliding window is captured and input into the window-level classification model to obtain the disease probability value corresponding to the image patch. It is determined in real time whether the obtained disease probability value is greater than or equal to the first threshold. If it is, the fine judgment process in Step 5 is executed; if the disease probability values ​​corresponding to all captured image patches are less than the first threshold, the entire target remote sensing image is determined to be disease-free, and the disease-free detection result of the entire image is directly output. Step 5: Perform dense sliding window inference on the entire target remote sensing image using a second step size that is smaller than the first step size, calculate the cumulative disease probability value of each pixel, construct a disease probability accumulation matrix, and generate a disease mask matrix based on the disease probability accumulation matrix and the second threshold. Step 6: Perform connected component analysis and morphological operations on the disease mask matrix generated in Step 5 to remove noise and smooth regions, and output the final disease distribution mask matrix.

2. The method for identifying discolored standing trees according to claim 1, characterized in that, The formula for calculating the cumulative probability of disease for each pixel is as follows: in: p represents a pixel; A(p) is the cumulative probability of disease at pixel p; N(p) is the number of windows covering pixel p; P(W i ) is the window W i The probability value of disease.

3. The method for identifying discolored standing trees according to claim 2, characterized in that, The formula for the disease masking matrix M(p) is as follows: Where, τ A This is the second threshold.

4. The method for identifying discolored standing trees according to claim 1, characterized in that, Multi-temporal remote sensing images include at least the blue light band, green light band, red light band, and near-infrared band.

5. The method for identifying discolored standing trees according to claim 1, characterized in that, The first step is 128 pixels long.

6. The method for identifying discolored standing trees according to claim 1, characterized in that, The second step size is any one of 2, 4, 8, 16, or 32 pixels.

7. The method for identifying discolored standing trees according to claim 1, characterized in that, The phenological period includes a dormant period, early growth period, peak growth period, and decline period, which are dynamically adjusted according to different latitude zones. The latitude zones include high-latitude regions, mid-latitude regions, and low-latitude regions.

8. The method for identifying discolored standing trees according to claim 7, characterized in that, The phenological periods corresponding to high latitude regions are: Dormancy period: Late November to mid-March of the following year; Early growth period: Late March to late May; Peak growing season: Early June to late August; Recession period: Early September to mid-November; The corresponding phenological periods for mid-latitude regions are: Dormancy period: Mid-December to late February of the following year; Early growth period: early March to early May; Peak growing season: Mid-May to early September; Recession period: Mid-September to early December; The phenological periods corresponding to low latitude regions are: Dormancy period: early January to mid-February; Early growth period: Late February to early April; Peak growing season: Mid-April to early October; Recession period: Mid-October to late December.

9. The method for identifying color-changing standing trees according to claim 1, characterized in that, The connected component analysis in step six includes the following steps: The disease mask matrix is ​​labeled with connected components, and the area of ​​each connected component is calculated. The area of ​​each connected component is compared with a preset minimum area threshold, and connected components with an area smaller than the minimum area threshold are removed to eliminate noise interference.

10. The method for identifying discolored standing trees according to claim 9, characterized in that, The morphological operations in step six include the following steps: The disease mask matrix after connected component analysis is processed by opening operation to remove isolated small noises; The mask matrix after the opening operation is processed by the closing operation to fill the internal holes and smooth the boundaries.

Images