A Time-Series Remote Sensing Monitoring Method and System for Mangroves Based on Adaptive Denoising and Automatic Sampling

Abstract

This invention discloses a method and system for time-series remote sensing monitoring of mangroves based on adaptive denoising and automatic sampling. The method includes: acquiring time-series remote sensing images of the study area and constructing an annual image sequence; performing adaptive filtering to generate annual low tide composite images and annual median composite images, and extracting multi-dimensional remote sensing features; automatically constructing a candidate sample pool of mangroves and non-mangroves based on a global mangrove distribution dataset, dividing the study area into spatial grid units, optimizing training samples through an iterative adaptive sampling strategy, and constructing a classification model to generate an annual distribution map of mangroves. Post-processing of the classification results yields the monitoring results. This method achieves efficient time-series monitoring of mangroves through adaptive denoising and automatic sampling, and has advantages such as high automation and strong adaptability. It can be widely applied to scenarios such as coastal ecological restoration and blue carbon assessment, representing an innovative application of remote sensing information technology in the field of coastal ecological monitoring.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology application and coastal wetland monitoring, and relates to a time-series remote sensing monitoring method and system for mangroves based on adaptive denoising and automatic sampling. Background Technology

[0002] Mangroves, mainly distributed in the intertidal zone of tropical and subtropical coasts, are a type of coastal wetland ecosystem with important ecological functions, playing a crucial role in maintaining coastal stability, protecting biodiversity, and regulating carbon sinks. With the intensification of global climate change and increased human interference, large-scale, high-frequency dynamic monitoring of mangroves has become an important task for ecological protection and resource management. Remote sensing technology, due to its advantages of wide coverage, short revisit time, and efficient information acquisition, has been widely applied to mangrove monitoring research. However, in tropical island regions, remote sensing imagery still faces many challenges in practical applications. The persistent high-frequency cloud cover in tropical regions, along with tidal periodic changes that cause continuous variations in the exposure of the intertidal surface, makes it difficult for a single temporal image to fully reflect the true spatial distribution pattern of mangroves. To alleviate these problems, existing studies often employ annual image construction methods based on maximum spectral index synthesis. However, under tropical environmental conditions, this synthesis strategy is still easily affected by residual clouds, thin clouds, and aerosols, leading to spectral anomalies in the synthesized images and reducing the accuracy of mangrove information extraction.

[0003] Furthermore, existing remote sensing classification methods for mangroves typically rely on field survey samples or manually interpreted samples to construct training data. However, in island regions where mangroves are scattered and transportation is limited, field surveys are difficult and costly, resulting in low efficiency in sample acquisition and making it difficult to meet the needs of large-scale, long-term dynamic monitoring. Meanwhile, while existing global mangrove distribution data can provide reference information on the spatial distribution of mangroves, it still has limitations in terms of temporal updates and automated monitoring, making it difficult to directly construct high-quality temporal training samples.

[0004] Therefore, there is currently a lack of a time-series remote sensing monitoring method for mangroves that simultaneously achieves adaptive image denoising and automatic training sample construction. There is an urgent need to propose a remote sensing identification method for mangroves that can effectively suppress interference from spectral outliers in remote sensing images, automatically construct training samples, and is applicable to long-term series monitoring, in order to improve the automation and identification accuracy of mangrove monitoring in complex island environments. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a time-series remote sensing monitoring method for mangroves based on adaptive filtering denoising and automatic sampling. By introducing adaptive filtering spectral denoising, automatic iterative selection of sample points, random forest classification, and post-processing mechanisms, it achieves stable extraction and long-term dynamic monitoring of mangrove distribution in complex island environments.

[0006] The technical solution adopted in this invention is as follows: A time-series remote sensing monitoring method for mangroves based on adaptive denoising and automatic sampling includes the following steps: (1) Acquire multi-temporal remote sensing images of the study area and construct an annual image sequence; (2) Adaptive filtering operations are performed on the images of each year based on the pixel-by-pixel time series statistical characteristics to eliminate spectral anomalies; (3) Generate annual low tide composite images and annual median composite images based on the filtered images, and extract multidimensional remote sensing features, including multispectral bands, mangrove sensitive spectral index and elevation data; (4) Automatically construct a mangrove candidate sample pool and a non-mangrove candidate sample pool based on the global mangrove distribution dataset; (5) Divide the study area into spatial grids, take the grid with the most mangrove candidate samples and non-mangrove candidate samples, use its samples and corresponding multidimensional remote sensing features as the initial training set, construct a random forest classification model, and perform sample adaptive optimization by automatically iterating and selecting sample points. (6) The final classification model is obtained by training the adaptively optimized samples and corresponding multidimensional remote sensing features and the annual distribution map of mangroves is generated. The classification results are then subjected to spatial filtering and temporal smoothing to obtain stable mangrove time-series monitoring results.

[0007] In the above technical solution, further, in step (2), for each pixel in the annual image data, the mean μ and standard deviation σ of the red band reflectance time series and the near-infrared band reflectance time series are calculated respectively. If the red band reflectance or near-infrared band reflectance on a certain observation date exceeds the corresponding interval μ±2σ, the observation is determined to be an outlier, and all band data of the pixel on that date are adaptively removed from the synthetic dataset. After completing the adaptive removal of spectral outliers pixel by pixel year by year, subsequent operations are performed on the remaining image series.

[0008] Further, in step (3), the remaining image sequence after step (2) is processed by performing maximum spectral index synthesis to generate annual low tide composite images and annual median composite images. Based on the annual low tide composite images and annual median composite images, multidimensional remote sensing features are extracted, including multispectral bands, spectral indices, and elevation. The multispectral bands include blue light, green light, red light, near-infrared, shortwave infrared 1, and shortwave infrared 2. The spectral indices include normalized vegetation index (NDVI), enhanced vegetation index (EVI), normalized water index (NDWI), improved normalized water index (MNDWI), land surface water index (LSWI), modular mangrove identification index (MMRI), normalized mangrove index (NDMI), mangrove vegetation index (MVI), and enhanced mangrove vegetation index (EMVI).

[0009] Furthermore, in step (4), the method for constructing the mangrove candidate sample pool is to overlay global mangrove distribution datasets from multiple adjacent years and extract the areas marked as mangroves in each year; perform morphological erosion operations using 3×3 pixel rectangular structural elements, iterating 2 to 5 times; select the centroid from the eroded mangrove patches as the initial candidate sample points; calculate the MVI, EMVI, NDMI, and MMRI indices for the initial candidate sample points, and calculate the mean μ and standard deviation σ for each index; then, for each index, remove abnormal samples whose index is outside the interval μ±σ formed by the corresponding mean μ and standard deviation σ, thereby obtaining the mangrove candidate sample pool.

[0010] Furthermore, in step (4), the method for generating the non-mangrove candidate sample pool is as follows: by superimposing global mangrove distribution datasets from multiple years, regions marked as non-mangroves in each year are extracted, and morphological erosion operations are performed on the patches using 3×3 pixel rectangular structural elements, iterating 2 to 5 times; candidate sample points are randomly generated according to an oversampling ratio of no less than 3 times the number of mangrove candidate samples; MVI, EMVI, NDMI, and MMRI indices are calculated for the candidate samples, and the mean μ and standard deviation σ are calculated for each index. Then, for each index, abnormal samples whose index is outside the interval μ±σ formed by the corresponding mean μ and standard deviation σ are removed; finally, the number of non-mangrove candidate samples after screening is controlled so that the number of samples is kept in a 1:1 ratio with the number of mangrove candidate samples.

[0011] Furthermore, in step (5), the study area is divided into spatial grid units; the number of mangrove candidate samples and non-mangrove candidate samples in each grid is counted, and the sample of the grid with the largest total number of samples is selected as the initial training set to construct a random forest classification model; the model is applied to the entire region, and the F1 value of each spatial grid unit is calculated; in each iteration, the F1 values ​​are sorted from high to low, and the grids with the lowest F1 values ​​in all spatial grid units are selected as candidate regions. Each grid sample in the candidate region is added to the training set and the model is retrained. At the same time, the updated average F1 value of the entire region is calculated, and the grid sample that can maximize the improvement of the average F1 value of the entire region is selected to be formally added to the training set. The training set and model are updated and iterated repeatedly until the change in the average F1 value of two consecutive rounds is less than the preset threshold and the model accuracy is stable.

[0012] Furthermore, the spatial grid unit is a 1°×1° spatial grid.

[0013] Furthermore, in step (6), the final classification model is trained using the adaptively optimized samples and corresponding multidimensional remote sensing features to classify the study area at the pixel level year by year; the classification results are spatially smoothed by applying 3×3 neighborhood majority filtering to eliminate salt-and-pepper noise and improve classification consistency; then a three-year sliding window is used to smooth and correct the time series classification results to correct the pseudo-changes caused by isolated years; thus, the final mangrove time series monitoring results are obtained.

[0014] The present invention also provides a time-series remote sensing monitoring system for mangroves based on adaptive denoising and automatic sampling, for implementing the method described in any of the preceding claims.

[0015] The present invention also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described in any of the preceding claims.

[0016] The beneficial effects of this invention are: This invention proposes a pixel-by-pixel adaptive filtering image synthesis method, which can effectively suppress cloud shadows and atmospheric interference, retain low tide characteristics while eliminating spectral outliers. Through automatic sample construction and iterative optimization, it achieves automatic generation and quality control of training samples. Combined with spatial filtering and temporal smoothing, it improves the stability and reliability of mangrove time-series monitoring results. This method achieves efficient mangrove time-series monitoring through adaptive denoising and automatic sampling, possessing advantages such as no need for manual selection of training samples, high degree of automation, and strong adaptability. It can be widely applied to scenarios such as coastal blue carbon monitoring, and has significant practical value, representing an innovative application of remote sensing information technology in the field of coastal ecological monitoring. Attached Figure Description

[0017] Figure 1This is a schematic diagram of a method flow according to a specific embodiment of the present invention; Figure 2 This is a schematic diagram comparing the low tide composite image (right) generated by adaptive denoising of the present invention with the low tide composite image (left) without denoising. Figure 3 This is a schematic diagram of model performance changes during sample iterative optimization in a specific embodiment of the present invention (the numbers in the diagram indicate the number of sample points). Figure 4 This is a specific embodiment of the invention showing the change of newly added samples in the feature space as the number of iterations increases.

[0018] Figure 5 This is a specific embodiment of the invention showing the spatial changes in mangrove mapping results as the number of iterations increases.

[0019] Figure 6 This is a comparison of the annual mangrove area monitoring results in the study area and the results from GMW v3.0 in a specific embodiment of the present invention.

[0020] Figure 7 This is a comparative diagram of the local spatiotemporal mapping results for regions A, B, and C (2015–2020). Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0022] According to the present invention, a time-series remote sensing monitoring method for mangroves based on adaptive denoising and automatic sampling is described in the following technical roadmap: Figure 1 As shown, according to a specific embodiment of the present invention, the method includes the following steps: (1) Acquire multi-temporal remote sensing images of the study area and construct an annual image sequence. In this case study area, Landsa-8 OLI multispectral data was used as the data source. Based on Google Earth Engine, all Landsat 8 OLI Collection 2 surface reflectance Tier 1 data covering the study area from 2013 to 2024 were acquired, totaling 57,606 images with a spatial resolution of 30 meters.

[0023] (2) For each year and each pixel, the mean μ and standard deviation σ of the red band reflectance time series and the near-infrared band reflectance time series are calculated respectively. If the red band reflectance or near-infrared band reflectance on a certain observation date exceeds the corresponding interval μ±2σ, the observation is determined to be an outlier, and all band data of that pixel on that date are adaptively removed from the synthetic dataset. Since the red and near-infrared bands are extremely sensitive to clouds (high reflectance) and shadows (low reflectance), this statistically based adaptive removal aims to eliminate abnormal spectral fluctuations caused by residual clouds, cloud shadows, sensor artifacts, or extreme atmospheric scattering. It can effectively identify and remove outliers, ensuring that the data sources participating in the synthesis have high spectral consistency and temporal stability. After completing the pixel-by-pixel and year-by-year adaptive removal of spectral outliers, the maximum spectral index synthesis function ee.ImageCollection.qualityMosaic('NDVI') provided by Google Earth Engine is executed on the remaining image sequences to generate year-by-year low tide synthetic images. In addition, the Google EarthEngine's ee.ImageCollection.median() function was executed on the image sequence to synthesize annual median images.

[0024] (3) Based on the annual low tide composite image and the annual median composite image, multidimensional remote sensing features were extracted, including six reflectance bands (Blue, Green, Red, NIR, SWIR1, and SWIR2) and nine spectral indices from the median composite image and the low tide composite image. The spectral indices include NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Difference Vegetation Index), NDWI (Normalized Difference Water Index), MNDWI (Modified Normalized Difference Water Index), LSWI (Land Surface Water Index), MMRI (Modified Mangrove Recognition Index), NDMI (Normalized Difference Mangrove Index), MVI (Mandatory Mangrove Vegetation Index), and EMVI (Enhanced Mangrove Vegetation Index). The calculation methods for these spectral indices are shown in Table 1. In addition, given that mangroves are mainly distributed in the low-altitude coastal zone and have significant topographic differences from inland vegetation, elevation was extracted based on FABDEM (Forest And Buildings removed Copernicus DEM) data and also used as a remote sensing classification feature.

[0025] Table 1. Spectral indices used and their calculation formulas

[0026] (4) Automatically construct mangrove and non-mangrove candidate sample pools based on the global mangrove distribution dataset. The method for constructing the mangrove candidate sample pool is to overlay the global mangrove distribution dataset (Global MangroveWatch, GMW) from multiple adjacent years and extract the areas marked as mangroves in each year; perform morphological erosion operation using 3×3 pixel rectangular structuring elements, iterating 2 to 5 times; select the centroid from the eroded mangrove patches as the initial candidate sample points; calculate the MVI, EMVI, NDMI and MMRI indices for the initial candidate sample points, and remove abnormal samples whose indices are outside the corresponding intervals μ±σ formed by the mean μ and standard deviation σ of each index, thereby obtaining the mangrove candidate sample pool.

[0027] Method for generating non-mangrove candidate samples: By overlaying GMW data layers from multiple years, regions marked as non-mangrove in each year are extracted, and morphological erosion operations are performed on the patches using 3×3 pixel rectangular structuring elements, iterating 2-5 times. Candidate sample points are randomly generated according to an oversampling ratio of no less than 3 times the number of mangrove samples. MVI, EMVI, NDMI, and MMRI indices are calculated for the candidate samples, and abnormal samples whose indices fall outside the interval μ±σ formed by the mean μ and standard deviation σ of each index are eliminated. Finally, the number of non-mangrove samples after screening is controlled to maintain a 1:1 ratio with the number of mangrove candidate samples.

[0028] (5) Divide the study area into spatial grids and perform adaptive optimization of samples by automatically iterating and selecting sample points. Specifically: Divide the study area into multiple spatial grid units (usually 1°×1°). First, count the number of mangrove candidate sample points and the number of non-mangrove candidate samples in all grid units, and select the sample from the grid unit with the largest total number of candidate samples as the initial training sample. Use this sample set and its corresponding multidimensional remote sensing features to train the random forest classification model.

[0029] Subsequently, the trained classification model was applied to classify the entire study area, and the classification accuracy index F1 was calculated within each grid cell. The formula for calculating the F1 value is: F1 = 2 × (PA × UA) / (PA + UA) Where PA represents producer precision and UA represents user precision.

[0030] In each iteration, the grid cells with lower classification accuracy are first identified based on their F1 scores, and these F1 scores are sorted from highest to lowest. The regions with F1 scores in the bottom 50% of all grid cells are selected as candidate regions. Subsequently, samples from these candidate regions are added to the current training set, and the random forest model is retrained, while the updated average F1 score for the entire region is calculated. The grid cell that maximizes the average F1 score across the entire study area is selected, and its samples are formally added to the training set. This iterative process is repeated until the change in the average F1 score is less than 0.01 over two consecutive iterations. At this point, the model accuracy is considered to have reached a stable state.

[0031] (6) A classification model is constructed using the automatically optimized training samples, and an annual distribution map of mangroves is generated. Spatial filtering and temporal smoothing are applied to the classification results to obtain stable mangrove time-series monitoring results. The final random forest classification model is trained using the optimized training samples and corresponding multi-dimensional remote sensing features, and the study area is classified at the pixel level year by year. In this embodiment, the number of decision trees in the random forest model is set to 500. Based on the model, remote sensing images from 2013 to 2024 are classified to generate an annual distribution map of mangroves. Subsequently, the classification results are post-processed. First, 3×3 neighborhood majority filtering is used to spatially smooth the classification results to eliminate salt-and-pepper noise and improve the spatial consistency of the classification map; then, a three-year sliding window is used to logically smooth the time-series classification results to correct spurious changes caused by isolated years. Through the above spatial filtering and temporal smoothing processes, the final mangrove time-series monitoring results are obtained.

[0032] Example

[0033] This embodiment takes a specific study area as an example and applies a time-series remote sensing monitoring method for mangroves based on adaptive denoising and automatic sampling. The method includes the following steps: The first step was to acquire Landsat imagery data for the study area. This example uses Google Earth Engine to acquire all Landsat 8 OLI Collection 2 Tier 1 surface reflectance data covering the study area from 2013 to 2024, totaling 57,606 images with a spatial resolution of 30 meters.

[0034] The second step involves calculating the mean μ and standard deviation σ of the red and near-infrared reflectance time series for each year and each pixel. If the red or near-infrared reflectance on a given observation date exceeds the interval μ ± 2σ, the observation is considered an outlier, and all band data for that pixel on that date are adaptively removed from the composite dataset. After completing the pixel-by-pixel, year-by-year adaptive removal of spectral outliers, the Google Earth Engine's maximum spectral index synthesis function ee.ImageCollection.qualityMosaic('NDVI') is applied to the remaining image sequences to generate annual low tide composite images for 2013-2024. Figure 2 As shown, after processing by the method in this embodiment, cloud shadows, thin clouds, and high reflectance anomaly areas in the image are effectively suppressed, patchy noise and discontinuities in the original image are significantly reduced, and the intertidal boundary and mangrove spatial structure are more continuous and clear, thereby improving the spatial consistency and interpretation reliability of the synthetic image. At the same time, the ee.ImageCollection.median() function provided by Google Earth Engine was also executed on the image sequence to synthesize median images from 2013 to 2024.

[0035] The third step involves extracting multidimensional remote sensing features based on annual low tide composite images and annual median composite images. These features include six reflectance bands from the annual median and low tide composite images: Blue, Green, Red, NIR, SWIR1, and SWIR2, as well as nine spectral index features. These spectral indices include NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index), NDWI (Normalized Difference Water Index), MNDWI (Modified Normalized Difference Water Index), LSWI (Land Surface Water Index), MMRI (Modified Mangrove Recognition Index), NDMI (Normalized Mangrove Index), MVI (Mandatory Mangrove Vegetation Index), and EMVI (Enhanced Mangrove Vegetation Index). Furthermore, elevation data extracted from FABDEM data was also used as a remote sensing classification feature.

[0036] The fourth step is to automatically construct a candidate sample pool for both mangroves and non-mangroves based on the Global Mangrove Distribution Dataset (GMW). The method for constructing the mangrove candidate sample pool is as follows: using 2015 as a baseline, overlaying multiple adjacent years' GMW data, and extracting areas marked as mangroves in each year; performing morphological erosion operations using 3×3 pixel rectangular structuring elements, iterating twice; selecting centroids from the eroded mangrove patches as initial candidate sample points; calculating MVI, EMVI, NDMI, and MMRI indices for the candidate sample points, and removing outlier samples based on the interval μ±σ formed by the mean μ and standard deviation σ (this method can remove as many impure samples as possible, retaining only the very pure samples for subsequent iterative selection), thus obtaining the mangrove candidate sample pool, totaling 6,604 mangrove sample points.

[0037] Non-mangrove sample generation method: By overlaying GMW data layers from multiple years, regions marked as non-mangrove in all years are extracted, and morphological erosion operations are performed on the patches using 3×3 pixel rectangular structuring elements, iterated twice. Candidate sample points are randomly generated with an oversampling ratio of no less than three times the number of mangrove samples. MVI, EMVI, NDMI, and MMRI indices are calculated for the candidate samples, and outlier samples are removed based on the interval μ±σ formed by the mean μ and standard deviation σ. Finally, the number of non-mangrove samples after screening is controlled, with 6,604 randomly selected to maintain a 1:1 ratio with the number of mangrove candidate samples.

[0038] The fifth step is to divide the study area into spatial grids and perform adaptive optimization of samples through iteration. In this embodiment, the study area is divided into multiple 1°×1° spatial grid units. First, the number of mangrove candidate sample points and the number of non-mangrove candidate sample points are counted in all grid units, and the samples in the grid unit with the largest total number of candidate samples are selected as the initial training samples. The random forest classification model is trained using this sample set and its corresponding multidimensional remote sensing features. Subsequently, the trained classification model is applied to the entire study area for classification, and the classification accuracy index F1 value is calculated in each grid unit. In each iteration, grid units with low classification accuracy are first identified based on the F1 value of each grid unit, and the regions with F1 values ​​in the bottom 50% of all grid units (ranked from highest to lowest F1 value) are selected as candidate regions. Then, the samples in the candidate regions are added to the current training set in turn, and the random forest model is retrained, while the updated average F1 value of the entire region is calculated. The grid unit that can maximize the improvement of the average F1 value of the entire study area is selected, and its samples are formally added to the training set. Repeat the above iterative process. When the average F1 value change is less than 0.01 in two consecutive iterations, stop the iteration. At this point, the model accuracy is considered to have reached a stable state. In this embodiment, as... Figure 3 As shown, with the iteration process, the grid-averaged F1 score increased from 0.38 in the initial model to 0.71 after the third iteration. Further increasing the number of samples thereafter had limited effect on improving model accuracy, indicating that the model performance had reached a stable state. Therefore, the samples from the third iteration were selected for the final model construction.

[0039] Furthermore, analysis of the high-dimensional feature space of the samples reveals that, with iteration, the newly added samples gradually fill the sparse regions of the original samples in the feature space, thereby improving the model's ability to represent and generalize different land cover types (see...). Figure 4 Regarding the classification results, after iterative optimization, the mangrove mapping results show enhanced patch continuity and clearer boundaries in space, and a significant reduction in confusion with adjacent terrestrial vegetation (see...). Figure 5 Finally, the samples generated in the third iteration and their corresponding multidimensional remote sensing features were selected for model training, including 311 mangrove samples and 311 non-mangrove samples, to build the final classification model.

[0040] The sixth step is to construct a classification model using the automatically optimized training samples and generate an annual distribution map of mangroves. Spatial filtering and temporal smoothing are then applied to the classification results to obtain stable mangrove time-series monitoring results. In this embodiment, a random forest classification model is constructed using the optimized training samples and corresponding multi-dimensional remote sensing features, and the study area is classified at the pixel level year by year. In this embodiment, the number of decision trees in the random forest model is set to 500. Based on the model, remote sensing images from 2013 to 2024 are classified to generate an annual distribution map of mangroves. Subsequently, the classification results are post-processed. First, a 3×3 neighborhood majority filter is used to spatially smooth the classification results to eliminate salt-and-pepper noise and improve the spatial consistency of the classification map. Then, a three-year sliding window is used to logically smooth the time-series classification results to correct spurious changes caused by isolated years.

[0041] Based on the above technical process, the annual distribution results of mangroves in the study area from 2013 to 2024 were obtained, such as... Figure 6 As shown in the figure. Monitoring results indicate that the total area of ​​mangroves in the study area generally showed an increasing trend during this period, increasing from 590,493 hectares to 643,646 hectares, a net increase of 53,153 hectares, an increase of approximately 9.0%, and an average annual growth rate of approximately 0.8%. Figure 6 The document also presents corresponding data from the existing global mangrove dataset GMW v3.0. It shows that within the overlapping years (2015–2020), the mangrove area estimation results obtained in this embodiment are in good agreement with the GMW dataset in terms of overall magnitude, with the values ​​being generally close. Furthermore, in terms of time series completeness, the GMW dataset only provides a limited time phase from 2015 to 2020, and the overall variation is relatively small; while this embodiment constructs a continuous time series from 2013 to 2024, exhibiting more obvious dynamic variation characteristics, especially showing a continuous growth trend after 2020. This variation characteristic is related to climate driving factors in the study area, and may be particularly affected by events that alter precipitation, sea level, and storm activity frequency, thereby influencing mangrove growth and recovery processes. These results demonstrate that this embodiment has a stronger capability in characterizing long-term time series.

[0042] In terms of spatial detail representation, this embodiment is able to identify fine-scale ecological processes in a variety of coastal environments (see...). Figure 7For example, in region A, this embodiment identified the recovery and succession process of mangroves after external disturbances; in region B, it detected localized mangrove reduction caused by agricultural activities; and in region C, it identified mangrove loss due to port construction. In contrast, these local changes are not clearly represented or are omitted in the GMW dataset. In summary, the GMW dataset has certain limitations in terms of temporal continuity and characterization of local dynamic changes, while the method proposed in this embodiment can achieve long-term continuous monitoring and improve the ability to identify interannual fluctuations and local changes, thus making it more suitable for dynamic monitoring of fragmented mangrove ecosystems within a region.

[0043] In this embodiment, the accuracy of the monitoring results is verified based on higher-resolution remote sensing imagery and manual visual interpretation. A total of 800 verification sample points were selected, including 600 stratified random sample points from 2013 (initial year), 2015 (baseline year), and 2024 (final year), and 200 random sample points located in the change area, to assess the type and timing of mangrove changes. The verification results show that the overall accuracy of mangrove classification for the above years is 94.5%, 96.0%, and 95.5%, respectively, with corresponding Kappa coefficients of 0.89, 0.92, and 0.91, indicating that the method can achieve stable and reliable identification of mangrove spatial distribution. The identification accuracy for mangrove expansion and reduction is 89.5%, with a corresponding Kappa coefficient of 0.81. Furthermore, the detected mangrove forest change years showed high consistency with the results of manual visual interpretation, with an overall accuracy of 97% (Kappa coefficient of 0.96), indicating that the method of the present invention has good stability and reliability in long-term mangrove forest time series monitoring.

[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] The embodiments described above are merely some preferred embodiments of the present invention, and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A time-series remote sensing monitoring method for mangroves based on adaptive denoising and automatic sampling, characterized in that, Includes the following steps: (1) Acquire multi-temporal remote sensing images of the study area and construct an annual image sequence; (2) Adaptive filtering operations are performed on the images of each year based on the pixel-by-pixel time series statistical characteristics to eliminate spectral anomalies; (3) Generate annual low tide composite images and annual median composite images based on the filtered images, and extract multidimensional remote sensing features, including multispectral bands, mangrove sensitive spectral index and elevation data; (4) Automatically construct a mangrove candidate sample pool and a non-mangrove candidate sample pool based on the global mangrove distribution dataset; (5) Divide the study area into spatial grids, take the grid with the most mangrove candidate samples and non-mangrove candidate samples, use its samples and corresponding multidimensional remote sensing features as the initial training set, construct a random forest classification model, and perform sample adaptive optimization by automatically iterating sample points. (6) The final classification model is obtained by training the adaptively optimized samples and corresponding multidimensional remote sensing features and the annual distribution map of mangroves is generated. The classification results are then subjected to spatial filtering and temporal smoothing to obtain stable mangrove time-series monitoring results.

2. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (2), for each pixel in the annual image data, the mean μ and standard deviation σ of the red band reflectance time series and the near-infrared band reflectance time series are calculated respectively. If the red band reflectance or near-infrared band reflectance on a certain observation date exceeds the corresponding interval μ±2σ, the observation is determined to be an outlier, and all band data of the pixel on that date are adaptively removed from the synthetic dataset. After completing the adaptive removal of spectral outliers pixel by pixel year by year, subsequent operations are performed on the remaining image series.

3. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (3), the remaining image sequence after step (2) is processed by maximum spectral index synthesis to generate annual low tide composite images and annual median composite images. Based on the annual low tide composite images and annual median composite images, multidimensional remote sensing features are extracted, including multispectral bands, spectral indices and elevation. The multispectral bands include blue light, green light, red light, near infrared, shortwave infrared 1 and shortwave infrared 2. The spectral indices include normalized vegetation index NDVI, enhanced vegetation index EVI, normalized water index NDWI, improved normalized water index MNDWI, land surface water index LSWI, modular mangrove identification index MMRI, normalized mangrove index NDMI, mangrove vegetation index MVI and enhanced mangrove vegetation index EMVI.

4. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (4), the method for constructing the mangrove candidate sample pool is to overlay global mangrove distribution datasets from multiple adjacent years and extract the areas that are marked as mangroves in each year. Morphological erosion is performed using 3×3 pixel rectangular structuring elements, iterating 2 to 5 times. Centroids are selected from the eroded mangrove patches as initial candidate sample points. MVI, EMVI, NDMI, and MMRI indices are calculated for the initial candidate sample points, and the mean μ and standard deviation σ are calculated for each index. For each index, abnormal samples outside the interval μ±σ formed by the corresponding mean μ and standard deviation σ are removed, thereby obtaining the mangrove candidate sample pool.

5. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (4), the method for generating the non-mangrove candidate sample pool is as follows: by superimposing global mangrove distribution datasets from multiple years, regions marked as non-mangroves in each year are extracted, and morphological erosion operations are performed on the patches using 3×3 pixel rectangular structural elements, iterating 2 to 5 times; candidate sample points are randomly generated according to an oversampling ratio of no less than 3 times the number of mangrove candidate samples; MVI, EMVI, NDMI, and MMRI indices are calculated for the candidate samples, and the mean μ and standard deviation σ are calculated for each index. Then, for each index, abnormal samples whose index is outside the interval μ±σ formed by the corresponding mean μ and standard deviation σ are removed; finally, the number of non-mangrove candidate samples after screening is controlled so that the number of samples is kept in a 1:1 ratio with the number of mangrove candidate samples.

6. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (5), the study area is divided into spatial grid units; the number of mangrove candidate samples and non-mangrove candidate samples in each grid is counted, and the samples of the grid with the largest total number of samples are selected as the initial training set to construct a random forest classification model. The model is applied to the entire region, and the F1 value of each spatial grid cell is calculated. In each iteration, the F1 values ​​are sorted from high to low, and the grids with the lowest F1 values ​​in all spatial grid cells are selected as candidate regions. Each grid sample in the candidate regions is added to the training set and the model is retrained. At the same time, the updated average F1 value of the entire region is calculated. The grid sample that can maximize the improvement of the average F1 value of the entire region is selected and formally added to the training set. The training set and model are updated and iterated repeatedly until the change in the average F1 value of two consecutive rounds is less than the preset threshold, and the model accuracy is stable.

7. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 6, characterized in that, The spatial grid unit is a 1°×1° spatial grid.

8. The mangrove time-series remote sensing monitoring method based on adaptive denoising and automatic sampling according to claim 1, characterized in that, In step (6), the final classification model is trained using the adaptively optimized samples and corresponding multidimensional remote sensing features to classify the study area at the pixel level year by year; the classification results are spatially smoothed by applying 3×3 neighborhood majority filtering to eliminate salt-and-pepper noise and improve classification consistency. Subsequently, a three-year sliding window was used to smooth and correct the time series classification results in order to correct the spurious changes caused by isolated years; This yielded the final time-series monitoring results for the mangroves.

9. A time-series remote sensing monitoring system for mangroves based on adaptive denoising and automatic sampling, characterized in that, Used to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method of any one of claims 1-8.

Images