Landsat-based spatio-temporal evolution monitoring method for tidal wetlands

Abstract

The application discloses a Landsat-based tidal wetland spatio-temporal evolution monitoring method, which comprises the following steps: S1, acquiring multi-temporal Landsat remote sensing image data of a target region and performing consistency preprocessing; S2, constructing a spectral index time sequence comprising NDWI, NDVI, PSRI and NIRv based on the preprocessed multi-temporal Landsat remote sensing image data; S3, generating key period images based on the spectral index time sequence; S4, performing initial classification of tidal flats and wetland vegetation based on the key period images to obtain initial classification maps year by year; S5, performing segmentation by using a LandTrendr algorithm based on the spectral index time sequence, and performing time sequence consistency optimization on the initial classification maps according to the obtained segmentation results to obtain optimized classification results; and S6, performing post-processing on the optimized classification results to generate a tidal wetland year-by-year dynamic classification data set. The application can generate a high-quality data set accurately reflecting the year-by-year dynamic evolution of a tidal wetland.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing, and particularly relates to a Landsat-based tidal wetland spatio-temporal evolution monitoring method. BACKGROUND

[0002] Tidal wetland ecosystems, mainly including tidal flats, mangroves and salt marshes, are distributed in the land-sea interface and play an irreplaceable role in wind and wave protection, carbon sequestration and storage, maintenance of biodiversity, and protection of the livelihoods of coastal residents. However, globally, affected by reclamation development, sea level rise and increasing human activities, the area of tidal wetlands continues to shrink, leading to weakening of their carbon sink function, fragmentation of biological habitats, and posing a serious threat to ecological security and regional sustainable development. Therefore, carrying out high-precision, long-term dynamic monitoring and mapping of tidal wetlands is of great scientific significance for revealing their evolution rules and driving mechanisms, supporting international convention implementation, delineating ecological protection red lines, and promoting blue carbon trading policies.

[0003] Traditional ground survey methods are limited by high cost and low efficiency, making it difficult to achieve effective monitoring on a large scale and over a long period of time. Remote sensing technology provides a feasible alternative approach, but existing remote sensing solutions all have their limitations: high spatial resolution images can finely depict surface features, but their high acquisition cost hinders large-scale, repetitive long-term observation applications; low spatial resolution images have high-frequency revisit advantages, but their spatial resolution of hundreds of meters is not sufficient to accurately identify and distinguish complex tidal wetland types. In contrast, Landsat and Sentinel series images, with spatial resolution of about 30 meters, 10 to 60 days of revisit period, and more than 30 years of data accumulation, strike a good balance between spatio-temporal resolution and time span, and can effectively capture vegetation phenology and tidal cycle characteristics, thus becoming an important data source for tidal wetland spatio-temporal evolution monitoring.

[0004] Based on remote sensing technology, tidal wetland dynamic mapping mainly relies on the temporal, spatial and spectral characteristics of images to depict the distribution and changes of tidal flats, mangroves and salt marshes. However, this technology faces many challenges in practical application: (1) spectral fluctuations caused by tidal inundation can lead to underestimation of wetland vegetation at high tide; (2) dynamic changes in tidal cycles increase the difficulty of accurately extracting the range of tidal flats; (3) spectral similarity between wetland vegetation and some inland landscapes can cause classification confusion; (4) frequent cloud cover in coastal areas limits the acquisition of ideal low-tide images, affecting the continuity of monitoring. How to effectively reduce the interference of tides, phenological changes and cloud pollution, and thus improve the accuracy and temporal consistency of monitoring results, is a core problem that needs to be solved in current tidal wetland remote sensing research.

[0005] Currently, the mainstream methods of tidal wetland monitoring using the repeated observation characteristics of remote sensing images mainly include three categories:

[0006] (1) The method based on optimal image selection, that is, to select single image that meets the ideal tidal conditions and has less cloud cover within a certain time window for wetland information extraction. This method is difficult to balance the requirements of range accuracy, type discrimination and cloud interference exclusion at the same time.

[0007] (2) The method based on time series synthesis, which synthesizes the images within a short period to smooth the influence of tidal changes. Although this method can weaken the interference, it loses a lot of original time series information, and when superimposing multi-period classification results, it is easy to produce cumulative errors, leading to inconsistent monitoring results in different periods, reducing the monitoring accuracy of year-by-year dynamic changes.

[0008] (3) The method based on dense time series analysis, which uses change detection model to analyze the trend of pixel-level time series to identify changes. This method can use complete time series information, but it has limitations in representing tidal and extreme state of phenology, which limits its ability to distinguish salt marshes and mangroves and other vegetation types, thereby affecting the accuracy of large-scale multi-type tidal wetland extraction.

[0009] In addition to the limitations of the above methods, they also face a common problem: due to the limitations of early data sources and the influence of coastal weather, the number of early high-quality images available for analysis is limited, resulting in significantly lower classification accuracy in early years than in later years, making it difficult to consistently reflect the dynamic evolution process of tidal wetlands. Although some studies have tried to introduce time consistency testing methods for optimization, this method usually relies on manually set thresholds or parameters, which is highly subjective and has limited universality.

[0010] Based on existing methods, the generated tidal wetland remote sensing products mainly present two forms: one is a long time series dataset of single wetland type, which has long-term annual observation capability but cannot comprehensively represent multi-type ecosystems; the other is a multi-type wetland classification dataset, which provides more comprehensive category information, but due to its insufficient time resolution, it cannot reveal the year-by-year dynamic changes, which seriously restricts its in-depth application in carbon cycle simulation and ecological protection management.

[0011] In summary, the existing tidal wetland monitoring technology and data products are obviously unbalanced between "time frequency (year-by-year dynamic)" and "type richness (multi-type identification)", making it difficult to meet the needs of both aspects. This technical bottleneck highlights the urgency and necessity of developing a new method that can balance multi-type accurate identification and high time consistency year-by-year monitoring. SUMMARY

[0012] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a Landsat-based tidal wetland spatio-temporal evolution monitoring method, which can classify multiple types of tidal wetlands such as tidal flats, mangroves and salt marshes, effectively improve the temporal consistency and stability of the entire time series classification results, and generate a set of high-quality data sets that can accurately reflect the dynamic evolution of tidal wetlands year by year.

[0013] The application adopts the following technical solutions:

[0014] The Landsat-based tidal wetland spatio-temporal evolution monitoring method comprises the following steps:

[0015] S1, obtaining multi-temporal Landsat remote sensing image data of a target area, and performing consistency preprocessing;

[0016] S2, based on the preprocessed multi-temporal Landsat remote sensing image data, constructing a spectral index time series including NDWI, NDVI, PSRI and NIRv;

[0017] S3, based on the spectral index time series, generating key period images;

[0018] S4, based on the key period images, performing initial classification of tidal flats and wetland vegetation to obtain initial classification maps for each year;

[0019] S5, based on the spectral index time series, using LandTrendr algorithm for segmentation, and performing temporal consistency optimization on the initial classification maps according to the obtained segmentation results to obtain optimized classification results;

[0020] S6, post-processing the optimized classification results to generate a tidal wetland dynamic classification data set.

[0021] Further, the consistency preprocessing in S1 comprises atmospheric correction, terrain correction and cirrus correction of Landsat remote sensing images, and cloud and cloud shadow mask processing based on optical satellite quality evaluation band and Fmask algorithm, and removing images with cloud cover greater than 70%.

[0022] Further, the spectral index time series is composed of 6 spectral bands and 4 spectral indexes of Landsat images, totaling 10 features.

[0023] Further, the key period images in S3 include low-tide flat images, high-tide flat images, green leaf period images and vegetation senescence period images.

[0024] Further, the low-tide flat image is based on the NDVI value sorting of the tidal flat sample, and the top-ranked images are selected for maximum NDVI synthesis.

[0025] And / or, the high tide beach images are sorted based on the NDWI values ​​of the beach samples, and the images with the highest ranking are selected for maximum NDWI synthesis;

[0026] And / or, the vegetation green leaf period images are sorted based on the NIRv values ​​of swamp samples, and the top-ranked images are selected for median composite;

[0027] And / or, the vegetation aging images are sorted based on the PSRI values ​​of the swamp samples, and the top-ranked images are selected for median composite analysis.

[0028] Furthermore, the initial classification in S4 uses a machine learning classification model to extract tidal flats and wetland vegetation respectively; wherein, tidal flat extraction includes the classification of tidal flats and non-tidal flats, and wetland vegetation extraction includes the classification of mangroves, salt marshes, and non-tidal wetland vegetation.

[0029] Furthermore, the classification features used in the tidal flat extraction include: six spectral band values ​​of the low tide image and four spectral indices: PSRI, NDWI, NDVI, and NIRv; the difference between the low tide image and the high tide image in the corresponding spectral bands; and the difference between the low tide image and the high tide image in the corresponding spectral indices, totaling 20 features.

[0030] Furthermore, the classification features used in the wetland vegetation extraction include: 40 spectral bands of images during low tide, high tide, green leaf stage, and senescence stage, as well as NIRv, PSRI, and their differences between images during the green leaf stage and senescence stage, totaling 42 features.

[0031] Furthermore, the timing consistency optimization in S5 specifically includes:

[0032] Low tide NDWI and low tide NDVI sequences were constructed using the 95th percentile of NDWI and the 5th percentile of NDVI for each year, respectively.

[0033] The sequence is segmented using the LandTrendr algorithm to obtain multiple segments;

[0034] Based on the mode principle, the classification results within each segment are adjusted for consistency, using the following formula:

[0035]

[0036] In the formula Indicates from arrive The classification results by year segment, Represents a set of categories. Represents a subset of the category set. Representing the the classification result of the year 2018.

[0037] Further, the post-processing in S6 includes classification result merging, spatial filtering and spatial smoothing;

[0038] The classification result merging is that when the same pixel is classified as both beach and wetland vegetation, it is marked as beach first;

[0039] And / or, the spatial filtering is to remove isolated patches that are not intersected with the 500-meter buffer of the maximum seawater range;

[0040] And / or, the spatial smoothing is to smooth the classification result by using a 3x3 window majority filter.

[0041] Compared with the prior art, the beneficial effects of the present application are that: the Landsat-based tidal wetland spatio-temporal evolution monitoring method of the present application effectively improves the time consistency of the dynamic monitoring result of the tidal wetland, the method can identify and correct the abnormal jump of the classification result caused by poor early image quality, low observation frequency or uneven cloud coverage by introducing the core step of "time series analysis-based time series consistency optimization", which solves the common problem of the existing method that the classification results of different years are inconsistent as pointed out in the background technology, thereby generating a more coherent and stable year-by-year dynamic data set in the time dimension, which more truly reflects the evolution process of the tidal wetland; the Landsat-based tidal wetland spatio-temporal evolution monitoring method of the present application realizes the year-by-year dynamic monitoring capability based on multi-type identification, the method constructs a complete process from feature construction, initial classification to time series optimization, so that it can distinguish between beach, mangrove, salt marsh and other multi-type tidal wetlands while providing year-by-year dynamic monitoring results, which directly addresses the imbalance between "type richness" and "time frequency" of the existing data products as pointed out in the background technology, and makes up for the lack of comprehensive representation of single-type data set and the defect that multi-type snapshot cannot reflect the year-by-year change; the Landsat-based tidal wetland spatio-temporal evolution monitoring method of the present application provides an objective and reproducible standardized solution, the method establishes a structured technical process, reduces the dependence on manual selection of "best image" or artificial setting of consistency test threshold, and reduces the uncertainty and regional limitations introduced by subjective factors, which makes the scheme have stronger universality and operability, and provides a stable and reliable technical path for large-scale, long-term tidal wetland mapping tasks. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the Landsat-based tidal wetland spatio-temporal evolution monitoring method in the embodiments of the present application;

[0043] Figure 2A comparison diagram of beach classification results of a typical area in the low tide wave band (b) and the low tide-difference (c) combined wave band in the embodiment of the present application;

[0044] Figure 3 A comparison diagram of beach classification results of a typical area in the low tide wave band (b) and the low tide-difference (c) combined wave band in the embodiment of the present application;

[0045] Figure 4 A comparison diagram of beach classification results of a typical area in the low tide wave band (b) and the low tide-difference (c) combined wave band in the embodiment of the present application; DETAILED DESCRIPTION

[0046] The present application will be further described below in conjunction with the drawings and specific embodiments, and it should be noted that the following described embodiments or technical features can be combined in any manner to form new embodiments without conflict.

[0047] As shown in Figure 1 The embodiment of the present application provides a Landsat-based tidal wetland spatiotemporal evolution monitoring method, which comprises the following steps:

[0048] S1, obtaining multi-temporal Landsat remote sensing image data of a target area, and performing consistency preprocessing.

[0049] The remote sensing image used in the embodiment of the present application is derived from three series of Landsat 8 (OLI), Landsat 7 (ETM+) and Landsat 5 (TM). First, the above-mentioned image is subjected to preprocessing operations such as atmospheric correction, terrain correction and cirrus correction, so as to ensure the consistency of pixel positions among different series of images. Secondly, in view of the cloud interference problem, based on the mask variable of the optical satellite quality evaluation band, the image with a cloud cover greater than 70% is removed. On this basis, since the mask result may still have omissions or errors, the embodiment of the present application further adopts the Fmask algorithm to mask process the residual cloud and cloud shadow covered area, so as to obtain the Landsat image data meeting the classification requirements. The finally obtained image band information is shown in the following table:

[0050] Landsat band information table:

[0051]

[0052] To construct a classification model with high precision, the embodiment of the present application adopts multi-source data to randomly collect and systemically label the tidal wetland samples, and ensures that there are a certain number of samples in each year, thereby ensuring the continuity and consistency of the sample data in the time dimension. In the high-resolution GoogleEarth image, the embodiment of the present application preferentially selects the tidal wetland samples with obvious interpretation signs, and such samples have clear ground feature boundaries and typical interpretation characteristics, which can effectively support the accurate determination of the wetland type. Meanwhile, the embodiment of the present application also collects a large number of tidal wetland samples with accurate time information and geographic coordinate information by using the sample data disclosed by the Earth Observation and Modeling Center, and defines the wetland type of the samples in combination with the photographic image.

[0053] The embodiment of the present application divides the tidal wetland samples into four categories of beach, mangrove, salt marsh and other, and ensures that there are sufficient samples in each year to meet the requirements of the sample quantity and category balance for the classifier training.

[0054] S2, based on the pre-processed multi-temporal Landsat remote sensing image data, a time sequence of spectral indices including NDWI, NDVI, PSRI and NIRv is constructed.

[0055] Based on the spectral band values listed in the Landsat band information table, the normalized vegetation index (NDVI), the normalized water body index (NDWI), the near-infrared vegetation reflectance index (NIRv) and the vegetation attenuation index (PSRI) of each Landsat remote sensing image are calculated. Further, based on the six spectral bands and the above four spectral indices, ten time sequence characteristics are constructed for subsequent image synthesis, supervised classification and time sequence consistency optimization. The specific spectral index acquisition method is as follows:

[0056] 1. Normalized Difference Vegetation Index (NDVI)

[0057] It can be used to describe the change of biomass index and can separate vegetation from water and soil. It is very effective for vegetation extraction and is calculated by the difference and ratio of near-infrared band and red band. The calculation formula is as follows:

[0058]

[0059] 2. Normalized Difference Water Index (NDWI)

[0060] In the normalized difference processing, the specific wave band of the remote sensing image is applied, so that the water body information display effect in the remote sensing image is enhanced. The NDWI is the normalized ratio index of the middle infrared wave band and the near infrared wave band. Compared with the normalized vegetation index, it has obvious extraction effect on the water content of the vegetation canopy; when the vegetation canopy is unable to be identified due to water stress, the NDWI index has significant effect on water extraction, and has great importance in wetland monitoring. The calculation formula is as follows:

[0061]

[0062] 3. Plant Senescence Reflectance Index (PSRI)

[0063] It can predict the canopy stress, and the increase of PSRI indicates the beginning of the senescence of vegetation and the maturity of plant fruits. The calculation formula is as follows:

[0064]

[0065] 4. Near-Infrared Reflectance of Vegetation (NIRv)

[0066] It is highly related to the photosynthesis of vegetation and has superior performance in simulating vegetation phenology, and the calculation formula is as follows:

[0067]

[0068] S3, based on the time sequence of the spectral index, a key period image is generated.

[0069] The embodiment of the application provides a tidal wetland image synthesis method based on tide and phenology characteristics, which aims to generate characteristic images of four key periods, i.e., low-tide beach, high-tide beach, vegetation green leaf period and vegetation senescence period, to support subsequent wetland classification and dynamic monitoring. The method filters and synthesizes representative images by analyzing the spectral differences in different periods, so as to ensure that the results can accurately reflect the ground feature and ecological environment change of the tidal wetland in the key stage. Specifically, the method comprises the following steps:

[0070] 1. Low-tide beach image synthesis: the negative correlation between the tidal beach normalized vegetation index (NDVI) and the tide height is used to filter the low-tide period image. First, the spatial average NDVI value of the sample point is calculated to mark the wetland area in the remote sensing image; then, the images are sorted in descending order of NDVI value, and the top-ranked images are selected as the low-tide observation images; finally, the maximum NDVI synthesis method is used to generate the low-tide beach composite image (Median Low-Tide Composite, MLTC), which effectively represents the ground features in the low-tide period.

[0071] 2. High-tide composite: The normalized difference water index (NDWI) of tidal flat is positively correlated with the tidal height. The high-tide period image is selected by using this correlation. Specifically, the spatially averaged NDWI value of the sample point is calculated to mark the wet area. The images are sorted in descending order of NDWI value, and the top-ranked images are selected as the high-tide observation images. Then, the maximum NDWI composite method is used to generate the high-tide composite image (Median High-Tide Composite, MHTC), thereby accurately capturing the characteristics of the high-tide period.

[0072] 3. Green-leaf period composite: For vegetation with significant seasonal changes, the green-leaf period image is selected by using the high sensitivity of the red edge band to plant reflectivity changes. The NIRv value of the sample point in the marsh area is calculated to mark the wet area. The images are sorted in descending order of NIRv value, and the top-ranked images are selected as the green-leaf period images. The median spectral index composite method is used to generate the green-leaf period composite image (Cloud-free Green Composite, CFGC). This method is not sensitive to outliers and can effectively reduce noise to ensure accurate expression of green-leaf period characteristics.

[0073] 4. Senescence period composite: The senescence period image is selected by using the plant senescence index (PSRI). Specifically, the average PSRI value of the sample point in the marsh area is calculated to mark the wet area. The images are sorted in descending order of PSRI value, and the top-ranked images are selected as the senescence period images. The median spectral index composite method is used to generate the senescence period composite image (Cloud-free Senescence Composite, CFSC). This method can effectively depict the characteristics of the vegetation senescence stage and improve the reliability of life cycle monitoring.

[0074] Through the above methods, the characteristic images of the four key periods can be obtained, which not only considers the spatio-temporal characteristics of tidal wetlands, but also reflects the phenological changes of vegetation, providing high-quality support data for wetland classification and dynamic monitoring, and significantly improving the monitoring accuracy and efficiency.

[0075] S4, based on the key period images, an initial classification of the tidal flat and wetland vegetation is performed to obtain an initial classification map for each year.

[0076] This step uses the spectral information of the four key period images as classification features and uses the tidal wetland sample set as training data to train a machine learning classification model, thereby extracting the initial classification map of the tidal flat and wetland vegetation for each year. Specifically, it includes:

[0077] 1. Tidal flat extraction:

[0078] Determination of the maximum sea water extent: select the high tide image, and use the threshold segmentation formula to generate the maximum sea water extent (MSE, MaxSeaExtent). This method identifies potential water bodies through the OSTU algorithm, and combines the comparison of NDVI and NDWI to reduce the possibility of salt marshes being misjudged as water bodies. The threshold segmentation formula is as follows:

[0079]

[0080] wherein, is the threshold value obtained based on the OSTU algorithm using the NDWI distribution.

[0081] Sample redistribution: the "permanent water", "mangrove", "salt marsh" and "other" training samples are uniformly divided into the "other" category, and are input into the classifier together with the "tidal flat" samples.

[0082] Feature construction: the classification features include 6 spectral values of the low tide image, and PSRI, NDWI, NDVI, NIRv four spectral indices, and the difference values of the corresponding spectral values and indices of the low tide and high tide images, totaling 20 features.

[0083] Classification and synthesis: 10 independent machine learning classifiers are used for classification, and the majority synthesis method (taking the most frequent category of pixels) is used to obtain the classification map, and the tidal flat result within the maximum sea water extent is taken as the final tidal flat distribution map.

[0084] 2. Wetland vegetation extraction:

[0085] Feature construction: the median low tide composite image (MLTC), the median high tide composite image (MHTC), the green leaf period composite image (CFGC) and the senescence period composite image (CFSC) are spliced to form a composite image containing 40 spectral bands, and each pixel contains the spectral response information of low tide, high tide, growing season and dormant season. In addition to the spectral bands, the NIRv, PSRI and their difference values of the CFGC and CFSC images are added as features to reflect the seasonal differences of the vegetation

[0086] Sample classification: the training samples are divided into "mangrove", "salt marsh" and "other" three categories, and are input into the classification respectively.

[0087] Classification and synthesis: 10 independent machine learning classifiers are used to map the mangrove and salt marsh, and the majority synthesis method is used to obtain the final wetland vegetation classification result.

[0088] S5, based on the spectral index time series, the LandTrendr algorithm is used for segmentation, and the obtained segmentation result is used to optimize the temporal consistency of the initial classification map, and the optimized classification result is obtained.

[0089] To improve the temporal consistency of the initial classification results, this invention introduces an optimization module based on temporal segmentation into the remote sensing index time series. The specific process is as follows:

[0090] 1. Time series segmentation extraction:

[0091] The evolution of tidal wetlands is mainly manifested in the transformation between water bodies and mudflats, as well as the transformation between mudflats and vegetation. To capture these dynamic changes, this invention constructs two key sequences: the low tide NDWI sequence, constructed using the 95th percentile of the annual NDWI, to characterize water body dynamics; and the low tide NDVI sequence, constructed using the 5th percentile of the annual NDVI, to characterize vegetation dynamics.

[0092] Based on these two time series curves, this invention introduces the LandTrendr algorithm for segmentation. This algorithm extracts the spectral variation trajectory of the Landsat time series, simplifying complex variation features into several connected straight line segments, thereby capturing relatively stable, undisturbed time segments and obtaining their start and end time information. Since this method only focuses on the trend component of the time series data, it can capture short-term changes while smoothing long-term trends and effectively eliminating noise without losing key details. This invention applies it to the monitoring of tidal wetland disturbance processes, automatically identifying disturbance points in the NDWI and NDVI sequences—that is, the moments most likely to involve type transitions—and generating multiple time series segments for each pixel accordingly. Within each time segment, since no disturbance occurs, the mudflat classification results of the NDWI segment should remain consistent, and the tidal wetland vegetation classification results of the NDVI segment should also remain consistent.

[0093] 2. Segmentation type determination:

[0094] Each time series segment should not contain significant disturbance points, therefore its tidal wetland type should remain stable. This invention uses the mode principle to determine the segment type, that is, selecting the type with the highest frequency within the segment as the representative category, as shown in the following formula:

[0095]

[0096] In the formula Indicates from arrive The classification results by year segment, Represents the set of all possible categories. Represents a subset of the category set. Representing the The classification result of the year, when multiple categories have the same frequency, is determined according to the priority of the classification accuracy, and the priority order is in turn mudflat, mangrove, salt marsh and other. Among them, the NDWI sequence in the low tide period is mainly used to optimize the mudflat classification result, and the NDVI sequence in the low tide period is used to optimize the tidal wetland vegetation classification result. Through the method, the present application effectively reduces the classification jump phenomenon caused by early low-quality images, significantly improves the stability and consistency of the annual sequence, and thus obtains more accurate annual dynamic mapping results of the tidal wetland.

[0097] S6, post-processing the optimized classification result to generate a tidal wetland annual dynamic classification data set.

[0098] Since the vegetation-free tidal flat and the vegetated tidal wetland are drawn respectively, integration is needed when the results are superimposed. For the possible classification conflicts (i.e. the same pixel is labeled as mudflat and wetland vegetation in a certain year), according to the rule that the mudflat recognition accuracy is higher, it is uniformly corrected to the mudflat type. To further improve the data consistency, two types of post-processing are implemented: one is spatial filtering, which removes isolated patches that do not intersect with the maximum sea water range (MSE) 500m buffer zone, to reduce the misclassification of inland areas; the other is spatial smoothing, which uses a 3*3 window majority filter to correct isolated pixels that are not surrounded by the same type of pixels to the surrounding dominant type, thereby improving the spatial continuity and overall stability of the classification result.

[0099] Further, the consistency preprocessing in S1 includes: atmospheric correction, terrain correction and cirrus correction of the Landsat remote sensing image, and cloud and cloud shadow mask processing based on optical satellite quality evaluation band and Fmask algorithm, to remove images with cloud cover greater than 70%.

[0100] Further, the spectral index time sequence is composed of 6 spectral bands and 4 spectral indexes of Landsat images, totaling 10 features.

[0101] Further, the key period image in S3 includes low tide beach image, high tide beach image, vegetation green leaf period image and vegetation senescence period image.

[0102] Further, the low tide beach image is sorted based on the NDVI value of the mudflat sample, and the top-ranked image is selected for maximum NDVI synthesis; and / or, the high tide beach image is sorted based on the NDWI value of the mudflat sample, and the top-ranked image is selected for maximum NDWI synthesis; and / or, the vegetation green leaf period image is sorted based on the NIRv value of the marsh sample, and the top-ranked image is selected for median synthesis; and / or, the vegetation senescence period image is sorted based on the PSRI value of the marsh sample, and the top-ranked image is selected for median synthesis.

[0103] Further, the initial classification in S4 adopts a machine learning classification model to respectively perform mudflat extraction and wetland vegetation extraction; wherein the mudflat extraction includes classification of mudflat and non-mudflat, and the wetland vegetation extraction includes classification of mangrove, salt marsh, and non-tidal wetland vegetation.

[0104] Further, the classification features used in the mudflat extraction include 6 spectral band values and 4 spectral indices (PSRI, NDWI, NDVI, and NIRv) of the low-tide image, and the difference between the low-tide image and the high-tide image in the corresponding spectral band and spectral index, totaling 20 features.

[0105] Further, the classification features used in the wetland vegetation extraction include 40 spectral bands of the low-tide, high-tide, green-leaf, and senescence images, and the difference between the green-leaf image and the senescence image in NIRv and PSRI, totaling 42 features.

[0106] Further, the time sequence consistency optimization in S5 specifically includes:

[0107] Constructing a low-tide period NDWI sequence and a low-tide period NDVI sequence, and respectively adopting the 95th percentile value of NDWI and the 5th percentile value of NDVI each year;

[0108] Segmenting the sequences using the LandTrendr algorithm to obtain multiple segments;

[0109] Adjusting the classification results within each segment based on the principle of mode, and the formula is:

[0110]

[0111] wherein represents the classification results of the segment from to year, represents a category set, represents a subset in the category set, represents the classification results of the year.

[0112] Further, the post-processing in S6 includes classification result merging, spatial filtering, and spatial smoothing;

[0113] The classification result merging is to preferentially mark as mudflat when the same pixel is simultaneously classified as mudflat and wetland vegetation; and / or the spatial filtering is to remove isolated patches that do not intersect with the 500-meter buffer of the maximum seawater range; and / or the spatial smoothing is to perform smoothing processing on the classification results using a 3x3 window mode filter.

[0114] The Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application solves the problems of single type of tidal wetland, limited spatiotemporal resolution and inconsistent classification results in time series in the existing tidal wetland spatiotemporal evolution monitoring. The Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application constructs the time series features of remote sensing indexes to capture the differences reflected by different tidal heights and vegetation phenology changes, and then generates the initial classification results of tidal wetland year by year based on the machine learning classification model. On this basis, the time series segmentation optimization module is designed and introduced to adjust and correct the initial classification results, so as to reduce the missing and misclassification caused by insufficient or insufficient observation quantity. Through the method, the dynamic monitoring of three types of tidal wetland, namely, beach, mangrove and salt marsh, year by year can be realized.

[0115] The Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application is based on the existing research status. There is still a lack of long-time series data set with year-by-year time resolution and multi-type classification capability in the field of tidal wetland dynamic monitoring. The Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application is a method for monitoring the spatiotemporal evolution of tidal wetland year by year and multi-type based on public open Landsat time series data. On the one hand, the annual tidal characteristics and vegetation phenology characteristics are fused, the spectral features in different tidal and phenology periods are extracted, and the multi-type classification results of tidal wetland year by year are generated. On the other hand, a completely objective time series segmentation optimization method is designed to segment the time series of remote sensing indexes, and the segmentation information is used to adjust the consistency of the initial classification year by year, so as to improve the consistency and classification accuracy in the time dimension. Through the above method, the dynamic monitoring of three types of tidal wetland, namely, beach, mangrove and salt marsh, year by year (the time resolution is 1 year) can be realized, and the deficiencies of the existing methods in type richness and time continuity are effectively made up.

[0116] Specifically, the Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application adds the spectral feature difference value of the high tide period and the low tide period when selecting features, so that the beach classification accuracy of the classifier is higher. As shown in Figure 2 , adding the band difference feature can reduce the missing (yellow box) and misclassification (purple box) areas of the beach.

[0117] Based on the time series segmentation optimization method, the Landsat-based tidal wetland spatiotemporal evolution monitoring method of the embodiment of the present application can reduce the interference of noise areas in image synthesis on the classification results, obtain tidal wetland classification results more in line with the actual situation, and improve the spatial classification accuracy and time accuracy. As shown in Figure 3 , the optimized results can avoid large-area misclassification (red box) of the beach and inland misclassification (yellow box) of the tidal wetland vegetation.

[0118] Compared with the prior art, the Landsat-based tidal wetland spatio-temporal evolution monitoring method has the following outstanding advantages: the Landsat-based tidal wetland spatio-temporal evolution monitoring method adds the difference between the high-tide period and the low-tide wetland spectral feature index in feature selection when classifying the tidal flat, which greatly improves the extraction accuracy of the tidal flat. Existing research mainly uses the original spectrum as the classification feature, or constructs a new remote sensing index based on the spectral curve feature to prompt the discrimination of the classifier, without considering the time difference information of the spectral difference between high tide and low tide; the Landsat-based tidal wetland spatio-temporal evolution monitoring method proposes a time series segmentation optimization method without subjective threshold, which optimizes the extraction results of the tidal flat and the wetland vegetation every year by using the correlation of the time series, and improves the time accuracy of the tidal wetland dynamics. After obtaining the annual classification dynamics, existing research often sets a series of rules for time consistency test, but the rules have regional limitations. The present application starts from the time series segmentation result, and the test process is more objective and universal; the Landsat-based tidal wetland spatio-temporal evolution monitoring method integrates the tidal level information, the vegetation phenology feature and the time series information to construct a multi-year tidal wetland classification method, which can draw the long-time series of the tidal flat, the mangrove forest and the marsh distribution dynamics every year. Compared with previous research, the time interval frequency and the richness of the classification type are significantly improved.

[0119] Experimental example

[0120] Step one: remote sensing image preprocessing

[0121] Landsat images with less than 70% cloud cover in the study area and study period are obtained on the GEE platform. The obtained images are masked by using the Fmask layer to remove clouds, snow and cloud shadows, so as to ensure the effectiveness and consistency of the subsequent analysis images.

[0122] Step two: sample collection

[0123] Google Maps reference area, CEOM database and Landsat images are used as ground truth data sources. In the study period, a certain number of tidal flat samples, mangrove samples, salt marsh samples and other types of samples are collected for each year to ensure that the samples are representative and comprehensive.

[0124] Step three: constructing Landsat index time series dataset

[0125] Based on each Landsat image, four types of remote sensing indices are calculated: NDVI, NDWI, PSRI, and NIRv. The four indices are combined with the six original bands of Landsat to form a time series dataset containing 10 features, which are used for subsequent modeling and analysis.

[0126] Step Four: Generate Low Tide Period Image

[0127] For all images within a year, the NDVI mean value of the beach sample on the image is calculated. All images are sorted by the NDVI mean value, and the top 20% of the sorted results are selected for median synthesis, thereby generating the low tide period image for that year.

[0128] Step Five: Generate High Tide Period Image

[0129] For all images within a year, the NDWI mean value of the beach sample on the image is calculated. All images are sorted by the NDWI mean value, and the top 20% of the sorted results are selected for median synthesis, thereby generating the high tide period image for that year.

[0130] Step Six: Generate Decay Period Image

[0131] For all images within a year, the PSRI mean value of the salt marsh sample on the image is calculated. All images are sorted by the PSRI mean value, and the top 20% of the sorted results are selected for median synthesis, thereby generating the decay period image for that year.

[0132] Step Seven: Generate Green Leaf Period Image

[0133] For all images within a year, the NIRv mean value of the beach sample on the image is calculated. All images are sorted by the NIRv mean value, and the top 20% of the sorted results are selected for median synthesis, thereby generating the green leaf period image for that year.

[0134] Step Eight: Obtain MSE

[0135] Based on the high tide period image generated in Step Five, the threshold segmentation method is used to extract water body information, obtaining the maximum sea extent (MSE) for each year.

[0136] Step Nine: Construct Beach Extraction Dataset

[0137] The samples collected in Step Two are divided into beach and non-beach types. The 10 spectral feature values of each sample in the low tide period image, as well as the 10 spectral feature values of the high-low tide difference, are extracted, thereby constructing the beach extraction dataset.

[0138] Step Ten: Train Beach Classifier

[0139] The data set constructed in step nine is divided into a training set and a test set in a ratio of 7:3, a random forest classifier is trained using the training set, and accuracy verification is performed based on the test set.

[0140] Step eleven: generate initial tidal flat classification results each year

[0141] The tidal flat classifier trained in step ten is used, the low-tide spectral feature values and the high-low tide difference feature values of each pixel in the study area are taken as input, and the corresponding classification results are output. Combined with the MSE range obtained in step eight, mask cropping is performed to obtain the initial tidal flat classification results each year.

[0142] Step twelve: build wetland vegetation classification data set

[0143] The samples collected in step two are divided into three categories: mangrove, salt marsh, and non-tidal wetland vegetation. 40 spectral feature values of the samples on the low-tide, high-tide, green leaf, and decay period images are extracted, and the NIRv and PSRI difference values of the green leaf and decay period are calculated, a total of 42 features are obtained, and a wetland vegetation classification data set is constructed.

[0144] Step thirteen: train tidal wetland vegetation classifier

[0145] The data set constructed in step twelve is divided into a training set and a test set in a ratio of 7:3, a random forest classifier is trained using the training set, and accuracy verification is performed based on the test set.

[0146] Step fourteen: generate initial tidal wetland vegetation classification results each year

[0147] The tidal wetland vegetation classifier trained in step thirteen is used, the 42 spectral features of each pixel in the study area are taken as input, and the corresponding classification results of each pixel are obtained, thereby generating the initial wetland vegetation classification each year.

[0148] Step fifteen: build annual NDVI and NDWI time series

[0149] Based on the Landsat time series data in step three, the NDWI and NDVI index sequences of each year are calculated. The 5% percentile synthesis method is used to generate the NDWI time series, and the 95% percentile synthesis method is used to generate the NDVI time series, which are used to represent the distribution of water bodies and vegetation in the low-tide period each year.

[0150] Step sixteen: Landtrendr segmentation of annual time series

[0151] The NDWI time series and NDVI time series obtained in step fifteen are input into the Landtrendr algorithm, and the segmentation results of NDWI and NDVI are output.

[0152] Step 17: Optimize the initial tidal flat classification results

[0153] Based on the NDWI segmentation results obtained in step sixteen, and combined with the temporal segmentation optimization formula:

[0154]

[0155] The initial tidal flat classification results obtained in step eleven are optimized to improve the consistency of the tidal flats over time.

[0156] Step 18: Optimize the initial wetland vegetation classification results

[0157] Based on the NDVI segmentation results obtained in step sixteen, and combined with the temporal segmentation optimization formula:

[0158]

[0159] The initial wetland vegetation classification results obtained in step fourteen are optimized to improve the consistency of wetland vegetation over time.

[0160] Step 19: Merging the classification results of mudflat and wetland vegetation

[0161] The tidal flat sequences obtained in steps seventeen and eighteen are merged with the wetland vegetation sequences. If a pixel is classified as both tidal flat and vegetation in the same year, it is preferentially labeled as tidal flat to ensure the accuracy of the results.

[0162] Step 20: Post-processing

[0163] Spatial processing was performed on the merging results from step nineteen: First, a 500m buffer was applied to prune the data based on the MSE range; second, a 3×3 median filter was used to smooth the results, ultimately yielding a high-precision annual tidal wetland map. This mapping clearly demonstrates the annual variation characteristics of tidal wetlands. The following figure selects two typical areas, such as... Figure 4 As shown, Figure 4 (a) reveals the trend of wetland vegetation expanding year by year on the tidal flats. Figure 4 (b) reflects the process of gradual reduction of tidal flats under land reclamation intervention.

[0164] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A Landsat-based method for monitoring spatio-temporal evolution of tidal wetlands, characterized in that, The method comprises the following steps: S1, acquiring multi-temporal Landsat remote sensing image data of a target area, and performing consistency preprocessing; S2, constructing a spectral index time series including NDWI, NDVI, PSRI and NIRv based on the preprocessed multi-temporal Landsat remote sensing image data; S3, generating key period images based on the spectral index time series; S4, performing initial classification of tidal flat and wetland vegetation based on the key period images to obtain initial classification maps of each year; S5, performing segmentation based on the spectral index time series by using a LandTrendr algorithm, and performing time sequence consistency optimization on the initial classification maps according to the obtained segmentation results to obtain optimized classification results, wherein the time sequence consistency optimization specifically comprises: constructing a low tide period NDWI sequence and a low tide period NDVI sequence, and using the 95th percentile value of NDWI and the 5th percentile value of NDVI of each year respectively; segmenting the sequences by using the LandTrendr algorithm to obtain a plurality of segments; performing consistency adjustment on the classification results in each segment based on a mode principle, wherein the NDWI segment is used for initial classification of the tidal flat, and the NDVI segment is used for initial classification of the tidal wetland vegetation, and the formula is: wherein denotes the classification result segmented from i to j years, denotes a set of classes, c denotes a subset of the set of classes, denotes the classification result of the t-th year; S6, performing post-processing on the optimized classification results to generate a tidal wetland dynamic classification data set of each year.

2. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 1, characterized in that, The consistency preprocessing in S1 comprises: performing atmospheric correction, terrain correction and cirrus correction on the Landsat remote sensing image, and performing mask processing on the cloud and cloud shadow based on an optical satellite quality evaluation band and a Fmask algorithm to remove images with a cloud amount greater than 70%.

3. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 1, characterized in that, The spectral index time series is composed of 6 spectral bands and 4 spectral indexes of the Landsat image, and a total of 10 features.

4. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 1, characterized in that, The key period images in S3 include low tide flat images, high tide flat images, vegetation green leaf period images and vegetation senescence period images.

5. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 4, characterized in that, The low tide flat images are selected by ranking the NDVI values of the tidal flat samples, and the images with high rankings are selected for maximum NDVI synthesis; and / or, the high tide flat images are selected by ranking the NDWI values of the tidal flat samples, and the images with high rankings are selected for maximum NDWI synthesis; and / or, the vegetation green leaf period images are selected by ranking the NIRv values of the marsh samples, and the images with high rankings are selected for median synthesis; and / or, the vegetation senescence period images are selected by ranking the PSRI values of the marsh samples, and the images with high rankings are selected for median synthesis.

6. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 1, characterized in that, The initial classification in S4 uses a machine learning classification model to perform tidal flat extraction and wetland vegetation extraction; wherein the tidal flat extraction includes classification of tidal flat and non-tidal flat, and the wetland vegetation extraction includes classification of mangrove, salt marsh and non-tidal wetland vegetation.

7. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 6, characterized in that, The classification features used in the tidal flat extraction include 6 spectral band values, PSRI, NDWI, NDVI and NIRv of the low tide image, and the difference between the low tide image and the high tide image in the corresponding spectral bands and the difference between the low tide image and the high tide image in the corresponding spectral indexes, a total of 20 features.

8. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 6, characterized in that, The classification features used in the extraction of the wetland vegetation include 40 spectral bands of the low tide, high tide, green leaf period and senescence period images, and NIRv, PSRI and their difference values of the green leaf period and senescence period images, totally 42 features.

9. The Landsat-based spatio-temporal evolution monitoring method of tidal wetlands according to claim 1, characterized in that, The post-processing in the S6 includes classification result merging, spatial filtering and spatial smoothing; The classification result merging is that when the same pixel is classified as both the beach and the wetland vegetation, the beach is marked preferentially; And / or, the spatial filtering is to remove isolated patches which are not intersected with the 500-meter buffer of the maximum sea water range; And / or, the spatial smoothing is to smooth the classification result by using a 3*3 window majority filter.

Images