Long-time-sequence forest coverage mapping method based on multi-source remote sensing image

Abstract

The invention discloses a long-time-sequence forest coverage mapping method based on a multi-source remote sensing image, and belongs to the technical field of long-time-sequence forest coverage mapping. The method comprises the following steps: S1, acquiring and processing remote sensing data; s2, constructing a high-quality dense time sequence time sequence cube; s3, phenological feature extraction is carried out based on the constructed time sequence cube; s4, generating a land coverage map by using a random forest classification method based on the extracted phenological features in combination with topographic features; and S5, combining the land coverage maps to obtain a forest coverage map. According to the method, the advantages of multi-source and multi-temporal data are exerted through the constructed high-quality dense time sequence data set, interference caused by factors such as weather, cloud layers and noise can be effectively reduced, and the forest coverage monitoring precision is further improved.

Description

Technical Field

[0001] The invention belongs to the technical field of long-term forest cover mapping, and in particular relates to a long-term forest cover mapping method based on multi-source remote sensing images. Background Art

[0002] As an important part of terrestrial ecosystems, forests not only play a key role in carbon storage, climate regulation, and biodiversity conservation, but also directly affect the operation of the global climate change mechanism. High-precision forest cover data can accurately and timely monitor forest changes, especially long-term forest changes, which is of great significance for assessing forest health, predicting ecosystem service capabilities, and responding to global climate change challenges. However, traditional forest survey methods mainly rely on field sampling and field surveys, which not only consume a lot of manpower and time, but also make it difficult to obtain high-precision data over a large scale and over a long period of time. To address these limitations, remote sensing technology, with its wide coverage and high timeliness, has become a more efficient solution.

[0003] With the development of remote sensing technology, using satellite images to map forest cover to accurately monitor forest changes has gradually become an effective means. Remote sensing technology can not only conduct large-scale repeated observations, with the advantages of strong timeliness and rich historical data, but also provide monitoring capabilities from regional to global scales through multi-sensor, multi-temporal and multi-spectral characteristics. Among them, the US Landsat satellite provides up to 50 years of image data. Its 30-meter spatial resolution and 16-day revisit cycle have laid a solid foundation for large-scale and long-term forest change research. At the same time, the Sentinel-2 images with 10-meter / 20-meter spatial resolution and 5-day revisit cycle have increased the amount of observation data, further improving the timeliness and accuracy of forest change monitoring. The C-band polarization data of Sentinel-1 has enhanced the coverage and reliability of forest change research through all-day and all-weather monitoring. Based on the use of Landsat and Sentinel data, a variety of methods have been developed to extract long-term forest cover change information.

[0004] The core of long-term forest change monitoring lies in data acquisition and processing methods. Existing technologies mainly use three methods to generate long-term forest change datasets: (1) based on automated or semi-automated remote sensing time series algorithms; (2) based on existing land cover datasets; and (3) multi-source remote sensing data integration algorithms. Among them, the first two use Landsat images as the main data source, and use long-term images and single-year images to analyze forest changes respectively; while the multi-source remote sensing data integration algorithm achieves higher-precision forest change monitoring by fusing data from different sensors and combining multiple features such as spectrum, texture, and polarization. The use of multi-source and multi-temporal remote sensing data can provide effective and sufficient data for long-term forest change monitoring on the one hand, and can well reflect the seasonal changes of vegetation, thereby increasing the reliability of forest change monitoring on the other hand. However, multi-source data integration also faces certain challenges, such as image quality being limited by clouds, shadows, and sensor problems, as well as errors caused by differences in spectral reflectance between different sensors. In order to further improve the accuracy and reliability of long-term forest change monitoring, scholars have begun to explore land cover change monitoring methods based on phenological characteristics.

[0005] The multi-source remote sensing data integration algorithm is based on the extraction of pixel spectral, texture, and polarization characteristics, combined with knowledge and decision criteria, to construct a hierarchical classification method or use the substitution method and combination method to realize the identification of forest distribution. Compared with single-phase data, this method makes up for the limitations of single sensor and discontinuous phase monitoring by integrating multi-source and multi-phase features. However, problems such as cloud layer, cloud shadow, and sensor differences limit the number of available images and cannot meet the demand for high-quality data for long-term forest change monitoring. In addition, the differences in wavelength and spectral reflectance between different sensors may lead to errors in monitoring results. Therefore, when conducting long-term forest change monitoring, the advantages of multi-source and multi-phase data should be fully utilized. At the same time, it is necessary to comprehensively consider the time differences of images and the differences of sensors, combine the phenological characteristics of forests and their seasonal changes, reduce the uncertainty caused by limited observation data, and ensure the accuracy of monitoring results.

[0006] At present, there are two main methods for obtaining phenological parameters to identify land cover changes based on long-term image features, such as time-series spectral indices: one is seasonal image synthesis; the other is percentile image synthesis based on indicators. The former relies on the phenological calendar, selects images with matching time series, and captures seasonal characteristics through synthesis rules for land cover identification; the latter determines percentiles through the statistical distribution of characteristic variables to generate percentile images reflecting the phenological characteristics of land cover. Existing technologies have shown that these two methods have similar overall accuracy in multi-type land cover classification mapping. However, the seasonal synthesis-based method relies on prior knowledge of the phenological calendar, while the percentile synthesis method does not require any prior information or explicit assumptions about seasonal time, so it has a wider range of application and stronger applicability. At present, the extraction of these phenological characteristics mainly relies on time-series spectral indices generated by single sensors and low-resolution images, such as MODIS, SPOT VEGETATION, and GIMMS AVHRR. In recent years, phenological feature extraction based on medium and high resolution images (such as Landsat and Sentinel) has been successfully applied to multi-type land cover classification, as well as mapping of single vegetation types, wetlands, water bodies and crops. However, affected by clouds and shadows, these images will produce extreme values ​​after mask processing. When using the original percentile synthesis method to extract phenology, especially the maximum and minimum percentile indicators, these erroneous extreme values ​​will be retained, resulting in deviations in the extraction of phenological features. Existing technologies usually choose to discard these extreme values ​​or replace them with adjacent percentile indicators, but this may ignore the real extreme phenomena in the image. In order to solve these problems, it is particularly important to reduce polluted images and increase effective observation data, so as to consolidate the data foundation of the percentile synthesis method, and the importance of the joint use of multi-platform data (such as Landsat and Sentinel-1 / 2) is even more prominent. On the basis of solving the time differences of images and sensor differences, integrating multi-sensor data to extract phenological and land cover characteristics can maximize the advantages of long-term, multi-source, multi-phase data and algorithm synthesis, thereby obtaining more reliable forest cover data to more effectively monitor temporal forest changes.

[0007] The existing forest cover extraction methods have the following main problems: (1) Relying on a single remote sensing data source or a single temporal data, the forest cover monitoring results are easily affected by external interference such as weather and clouds. Optical remote sensing data are easily affected by weather conditions, while SAR data often interferes with accuracy due to large noise. In addition, the existing methods often ignore the temporal characteristics of the objects and fail to make full use of the temporal information contained in multi-temporal and multi-source data, which in turn affects the final monitoring accuracy; (2) Although multi-temporal data increases the observation frequency, the challenges brought by image quality (such as cloud cover, shadows, etc.) and different sensor characteristics (such as spectral reflectance differences) make the number of available images insufficient, resulting in increased uncertainty in forest cover extraction; (3) When performing extreme value percentile synthesis, the extraction of phenological features is often biased due to improper handling of extreme value errors. In order to solve these problems, the temporal and spectral characteristics in multi-temporal and multi-source data should be further explored to improve data utilization and enhance the discriminant characteristics of objects, so as to improve the accuracy and reliability of forest cover extraction. Summary of the invention

[0008] In response to the problems mentioned in the background technology, the present invention proposes a long-term forest cover mapping method based on multi-source remote sensing images. It constructs a high-quality dense time series image stack based on multi-source remote sensing data, analyzes the time series change characteristics of different landforms, and characterizes the phenological characteristics. At the same time, combined with terrain characteristics, a forest cover map is drawn based on the random forest method to describe the spatial pattern of forest distribution, thereby improving the accuracy and reliability of forest cover extraction.

[0009] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0010] A long-term forest cover mapping method based on multi-source remote sensing images includes the following steps:

[0011] S1: Remote sensing data acquisition and processing;

[0012] S2: High-quality dense time series time cube construction;

[0013] S3: Extract phenological features based on the constructed time series cube;

[0014] S4: Generate land cover map using random forest classification method based on extracted phenological features combined with terrain features;

[0015] S5: Merge the land cover map to obtain the forest cover map.

[0016] Preferably, in S1, the specific contents of remote sensing data acquisition and processing are:

[0017] Filter the Landsat-8-OLI SR, Sentinel-2-MSI SR satellite surface reflectivity optical remote sensing data and Sentinel-1-SAR GRD synthetic aperture radar remote sensing data from past years.

[0018] As a preference, the collected Landsat-8-OLI SR and Sentinel-2-MSI SR satellite surface reflectance optical remote sensing data are masked with CFmask, and then all image data are cropped and uniformly resampled to a resolution of 30 meters.

[0019] As a preference, in S2, the specific contents of constructing a high-quality dense time series cube are as follows:

[0020] S21: spectral reflectance consistency matching;

[0021] The linear regression equation between the bands of Landsat-8-OLI SR and Sentinel-2-MSI SR images was constructed. The spectral bands of Sentinel-2-MSI SR images were linearly transformed using Landsat-8-OLI SR data as a reference. Then, the images in the image collection were median synthesized.

[0022] S22: Time series linear interpolation;

[0023] The missing data are filled by combining the adjacent high-quality observations in the time series data through the time series linear interpolation method, and finally a high-quality dense time series data set is obtained.

[0024] Preferably, in S3, the specific content of extracting phenological features based on the constructed time series cube is:

[0025] S31: Extraction and analysis of time series variation curves of different features of different landforms;

[0026] S32: Phenological feature extraction.

[0027] Preferably, in S31, the specific contents of extracting and analyzing the time series variation curves of different features of different ground objects are:

[0028] Based on the optical remote sensing data and SAR remote sensing data time series cube, the spectral index, NDVI texture characteristics and polarization characteristics of each image in the time series cube are calculated;

[0029] The calculation process of the characteristic index of the time series cube of optical remote sensing data is as follows:

[0030]

[0031] in, , , , , , They represent the blue, green, red, near infrared, shortwave infrared 1, and shortwave infrared 2 bands of the images in the optical remote sensing data time series cube respectively; represents the normalized difference vegetation index, represents the normalized building index, represents the improved normalized difference water index, represents the surface water index; Indicates contrast, represents the variance, Indicates relevance, represents entropy, represents the average value; L / S2 represents the optical remote sensing data time series cube;

[0032] The calculation process of polarization characteristic data is:

[0033]

[0034] Wherein, VV represents the vertical polarization mode of Sentinel1; VH represents the vertical horizontal polarization mode of Sentinel1; VV / VH represents the ratio of VV to VH polarization modes; VH / VV represents the ratio of VH to VV polarization modes; Represents the SAR remote sensing data time series cube built based on Sentinel1;

[0035] By using the constructed high-quality dense time series dataset, we extracted the curves of different land objects changing over time, deeply analyzed the differences in time series of different land objects, and analyzed the curves of spectral bands, spectral indices, polarization characteristics, and NDVI texture characteristics changing over time.

[0036] Preferably, in S32, the specific content of phenological feature extraction is:

[0037] The spectral bands, spectral indices, NDVI texture features in the optical remote sensing data time series cube constructed by Landsat-8-OLI SR and Sentinel-2-MSI SR and the polarization feature data in the Sentinel-1-SAR GRD synthetic aperture radar remote sensing data time series cube are realized 10 th , 25 th , 50 th , 75 th , 90 th Percentile statistics of multi-temporal indicators are used to characterize the phenological characteristics of landforms with long time series and medium to high resolution;

[0038] The calculation process of phenological characteristics is:

[0039]

[0040] Among them, M S1_S2_T Represents the characteristic indicators in the optical remote sensing data time series cube, S P_R Represents polarization feature data in the time series cube based on SAR remote sensing data; It means that the spectral bands, spectral indices, texture features, and polarization features in the time series cube of optical remote sensing data and SAR remote sensing data are further analyzed by 10 th , 25 th , 50 th , 75 th , 90 th A total of five multi-temporal indices were calculated to characterize the phenological characteristics of different land features.

[0041] Preferably, in S4, the specific content of generating a land cover map using a random forest classification method based on the extracted phenological features combined with terrain features is:

[0042] Based on the GEE cloud platform and Google Earth high-resolution remote sensing images, different ground feature sample points were selected and the samples were divided into 7:3. The percentile statistical indicators extracted from S3 were combined with terrain features for random forest classification.

[0043] Construct a confusion matrix and calculate the overall accuracy, Kappa coefficient, producer accuracy, and user accuracy indicators to evaluate the classification accuracy.

[0044] As a preference, the overall accuracy P o The calculation process is:

[0045]

[0046] Among them, N ii Indicates the number of samples correctly classified into the true category; N represents the total number of samples; k represents the number of categories;

[0047] The calculation process of Kappa coefficient is:

[0048]

[0049]

[0050] Among them, P e It represents the sum of the product of the actual and predicted number of samples for all categories divided by the square of the total number of samples; N represents the total number of samples; the number of real samples of each category is a 1 , a 2 , ..., ak , and the predicted number of samples of each category is b 1 , b 2 , ..., b k ; k represents the number of categories;

[0051] The calculation process of producer precision is:

[0052]

[0053] Among them, a represents the number of correctly classified samples in the ground object category; b represents the number of samples that are actually in this category but are misclassified as a certain category;

[0054] The calculation process of user precision is:

[0055]

[0056] Among them, a represents the number of samples that are correctly classified; c represents the number of samples that are actually of a certain category but are misclassified as that category.

[0057] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0058] (1) The present invention collects high-quality dense time series data sets constructed by Landsat-8-OLI SR (L), Sentinel-1-SAR GRD (S1) and Sentinel-2-MSI SR (S2) remote sensing images, giving full play to the advantages of multi-source and multi-temporal data. Different from using a single remote sensing data source or a single temporal data, the method of the present invention can effectively reduce the interference caused by factors such as weather, clouds and noise, and further improve the accuracy of forest cover monitoring.

[0059] (2) The method of the present invention takes into account the temporal differences of images, sensor differences, forest phenological characteristics and seasonal change characteristics, and maximizes the benefits of the synthesis of long-term multi-source and multi-phase data and algorithms by extracting "pseudo-change" features, thereby ensuring the accuracy of forest cover information characterization and solving the problem in the prior art that the time distribution of effective observation values ​​at different pixel points is extremely uneven due to insufficient effective data. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a technical flow chart of the present invention;

[0061] Figure 2 It is a schematic diagram of the synthesis of 8-day step length data of the present invention;

[0062] Figure 3It is the 8-day time-series interval change curve of different ground objects in different spectral bands of the present invention (where af represents the time-series change curve of different ground objects in blue, green, red, near infrared, short-wave infrared 1, and short-wave infrared 2 bands respectively);

[0063] Figure 4 is the 8-day time series interval variation curve of different ground objects in different spectral indexes of the present invention (where ad represents the time series variation curve of different ground objects in NDVI, NDBI, MNDWI, and LSWI respectively);

[0064] Figure 5 is the 8-day time series variation curve of different ground objects in different polarization characteristics of the present invention (where ad represents the time series variation curve of different ground objects in VV, VH, VV / VH ratio, and VH / VV ratio, respectively);

[0065] Figure 6 is the 8-day time series interval change curve of different ground objects in different texture features of the present invention (where ae represents the time series change curve of different ground objects in contrast, variance, mean, variance, entropy, and correlation, respectively);

[0066] Figure 7 This is the land cover map (a) and forest cover map (b) of Nanjing in 2023. DETAILED DESCRIPTION

[0067] The present invention is further illustrated below in conjunction with specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0068] The long-term forest cover mapping method based on multi-source remote sensing images provided in this embodiment includes collecting and processing Landsat-8-OLI SR (L), Sentinel-1-SAR GRD (S1) and Sentinel-2-MSI SR (S2) multi-source remote sensing data in the study area, firstly median synthesis of images in the image set at an interval of 8 days, and then linearly interpolating the synthesized images to fill the pixel vacancies caused by cloud mask processing in the synthesized images, and finally obtaining a high-quality dense time series data set at an interval of 8 days. Secondly, using the constructed high-quality dense time series data set, extract the time-varying curves of different ground objects, deeply analyze the differences of different ground objects in the time series, and analyze the time-varying curves of spectral bands, spectral indices, polarization characteristics, and texture characteristics. Secondly, based on the constructed dense time series data, the percentage statistics method of a single image is further improved to improve the accurate description of forest cover, that is, the six spectral bands, four spectral indices, five texture features in L / S2 and four polarization feature data in S1 are respectively realized 10 th , 25 th , 50 th , 75 th , 90 th (5P) Percentile statistics of five multi-temporal indicators are used to characterize the phenological characteristics of forests and other landforms with long time series and medium and high resolution. Based on the extracted phenological characteristics and terrain characteristics, a random forest classification method is used to generate a land cover map containing 6 categories, and other categories except forests are further merged to obtain a forest cover map. The specific implementation steps are as follows:

[0069] S1: Remote sensing data acquisition and processing;

[0070] Taking Nanjing as an example, all 2023 Landsat-8-OLI SR (L), Sentinel-2-MSI SR (S2) satellite surface reflectivity (SR) optical remote sensing data and Sentinel-1-SAR GRD (S1) synthetic aperture radar (SAR) remote sensing data covering the study area were screened based on the Google Earth Engine (GEE) cloud platform.

[0071] All optical remote sensing data were masked with Cfmask to eliminate interference caused by clouds, shadows, etc., to ensure the quality of available data. All images were cropped to the size of the study area and uniformly resampled to a resolution of 30 meters.

[0072] S2: High-quality dense time series time cube construction;

[0073] S21: spectral reflectance consistency matching;

[0074] The linear regression method was selected to construct the linear regression equations between the bands of the L and S2 images. Taking the L data as a reference, the blue, green, red, near-infrared, short-wave infrared 1, and short-wave infrared 2 of the S2 image were linearly converted to ensure the consistency of the spectral reflectance of different sensor data (Landsat-8-OLI SR and Sentinel-2-MSI SR).

[0075] Based on the GEE cloud platform, the above preprocessed optical remote sensing images (L and S2) and SAR remote sensing images (S1) were synthesized by median with a time step of 8 days to obtain comparable dense time series data cubes. Figure 2 as shown).

[0076] S22: Time series linear interpolation;

[0077] When there is no high-quality observation data in the L / S2 time series cube within 8 days or the image is missing due to cloud mask processing, the missing data is filled by combining the adjacent high-quality observations in the time series data through the time series linear interpolation method. Finally, a high-quality dense time series dataset (8 days) is obtained.

[0078] S3: Extract phenological features based on the constructed time series cube;

[0079] S31: Extraction and analysis of time series variation curves of different features of different landforms;

[0080] The spectral index (NDVI, NDBI, MNDWI, LSWI), NDVI texture features (variance, contrast, correlation, entropy, mean) and polarization features (VV / VH, VH / VV) of each image in the time series cube are calculated based on the optical remote sensing data (L and S2) and SAR remote sensing data (S1) time series cubes for the subsequent extraction of different time series features.

[0081] The calculation process of the characteristic index of the time series cube of optical remote sensing data is as follows:

[0082]

[0083] in, , , , , , They represent the blue, green, red, near infrared, shortwave infrared 1, and shortwave infrared 2 bands of the images in the optical remote sensing data time series cube respectively; , , , They represent the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Built-up Index (NDBI), the Modified Normalized Difference Water Index (MNDWI), and the Land Surface Water Index (LSWI) among the spectral indices calculated in the time series cube of optical remote sensing data. , , , , They represent the contrast (cont), variance (var), correlation (corr), entropy (ent), and mean of the texture feature indicators calculated based on the spectral index NDVI obtained in the optical remote sensing data time series cube. L / S2 represents the optical remote sensing data time series cube.

[0084] The calculation process of polarization characteristic data is:

[0085]

[0086] Wherein, VV represents the vertical polarization mode of Sentinel 1. VH represents the vertical horizontal polarization mode of Sentinel 1. VV / VH represents the ratio of VV to VH polarization modes. VH / VV represents the ratio of VH to VV polarization modes. Represents a SAR remote sensing data time series cube built based on Sentinel1.

[0087] Different features of different landforms will vary at different times. Based on the GEE platform, we use the constructed high-quality dense time series data set to extract the temporal change curves of different landforms (forests, water bodies, building land, grasslands, cultivated land, bare land), and deeply analyze the characteristic differences of different landforms in the time series of previous years (8-day step length). We extract and analyze the spectral bands (blue, green, red, near-infrared, short-wave infrared 1, short-wave infrared 2), spectral indices, polarization features, and texture features.

[0088] Spectral indices include Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), Modified Normalized Difference Water Index (MNDWI), Land Surface Water Index (LSWI), polarization features (VV, VH, VV / VH, VH / VV), and texture features (variance, contrast, correlation, entropy, and mean).

[0089] S32: phenological feature extraction;

[0090] Based on the high-quality dense time series data cube constructed above, the shortcomings of the percentage statistics method of a single remote sensing data source are further improved to improve the accurate description of forest cover. The six spectral bands, four spectral indices, and five texture features in the 8-day step optical remote sensing data time series cube constructed by L / S2 (Landsat-8-OLI SR and Sentinel-2-MSI SR) and the four polarization feature data (S P_R ) respectively achieve 10 th , 25 th , 50 th , 75 th , 90 th (5P) Percentile statistics of five multi-temporal indices are used to characterize the phenological characteristics of forests and other landforms with long time series and medium to high resolution (P).

[0091] The calculation process of phenological characteristics is:

[0092]

[0093] Among them, M S1_S2_T Represents the characteristic indicators contained in the optical remote sensing data time series cube, such as spectral bands and calculated spectral indexes and texture features, S P_R Represents the polarization feature data in the SAR remote sensing data time series cube; It means that the spectral band, spectral index, texture feature and polarization feature in the optical remote sensing data time series cube and the SAR remote sensing data time series cube are calculated 10 th , 25 th , 50 th , 75 th , 90 thThere are five multi-temporal indicators in total.

[0094] S4: Generate land cover map using random forest classification method based on extracted phenological features combined with terrain features;

[0095] Based on the GEE cloud platform and Google Earth high-resolution remote sensing images, different ground feature sample points (including 530 forests, 468 built-up land, 136 water bodies, 466 cultivated land, 310 grasslands, and 220 bare land, a total of 2130) were selected. The samples were divided into 7:3 for model training and accuracy evaluation. The percentage statistical indicators extracted above, namely, 10 of the 6 spectral bands in L / S2, 4 spectral indices, 5 texture features, and 4 polarization features in S1, were used to train the model and evaluate the accuracy. th , 25 th , 50 th , 75 th , 90 th (5P) Percentile statistics of five multitemporal indices, a total of 95 multitemporal features combined with three topographic features (slope, aspect, elevation), a total of 98 features were used for classification by the random forest classification method.

[0096] Construct a confusion matrix to calculate the overall accuracy (OA), Kappa coefficient (Kappa), producer accuracy (PA), and user accuracy (UA) to evaluate the classification accuracy.

[0097] Overall accuracy (P o ) is calculated as:

[0098]

[0099] Among them, N ii represents the number of samples correctly classified into their true categories; N represents the total number of samples; k represents the number of categories.

[0100] The calculation process of Kappa coefficient is:

[0101]

[0102]

[0103] Among them, P e It means the sum of the product of the actual and predicted number of samples for all categories divided by the square of the total number of samples; the number of real samples of each category is a 1 , a 2 , ..., a k , and the predicted number of samples of each category is b 1 , b 2 , ..., b k; N represents the total number of samples; k represents the number of categories.

[0104] The calculation process of producer accuracy (PA) is:

[0105]

[0106] Among them, a represents the number of samples that are actually a certain category (such as forest) and are correctly classified as forest; b represents the number of samples that are actually that category but are misclassified as other categories.

[0107] The calculation process of user accuracy (UA) is:

[0108]

[0109] Among them, a represents the number of samples that are actually a certain category (such as forest) and are correctly classified as forest; c represents the number of samples that are actually other categories but are misclassified as this category.

[0110] Table 1 Confusion matrix of land cover classification results

[0111]

[0112] Results and analysis in this embodiment:

[0113] 1. Analysis of time series change curves of different landforms;

[0114] like Figure 3 As shown ( Figure 3 a- Figure 3 f in the figure represents blue, green, red, near infrared, shortwave infrared 1, and shortwave infrared 2, respectively), showing the time series variation curves of 6 different land objects (forest, water body, construction land, cultivated land, grassland, and bare land) in different spectral bands. The reflectance of forest and grassland is low in the visible light (blue, green, and red light) bands, but high in the near infrared band, and has obvious seasonal changes. The reflectance of water bodies is low in all bands, especially in the near infrared and shortwave infrared bands. The reflectance of construction land is relatively stable in all bands, and is relatively high in the visible light and shortwave infrared bands. The reflectance of cultivated land changes significantly, especially in different growth stages, and the response to the red light and near infrared bands is particularly obvious. The reflectance of bare land is high in the visible light and shortwave infrared bands, but relatively low in the near infrared band.

[0115] like Figure 4 As shown ( Figure 4 a- Figure 4The d in the figure represents NDVI, NDBI, MNDWI, and LSWI, respectively, showing the time series change curves of different ground objects in different spectral indices. The NDVI and LSWI values ​​of forests and grasslands are relatively high and show obvious seasonal changes, especially in the growing season, when both NDVI and LSWI values ​​increase, which are suitable for identifying and monitoring seasonal vegetation. The NDVI value of water bodies is always low, while the MNDWI value is high and stable, which is very suitable for identifying water bodies. The NDBI value of construction land is high and stable, which is suitable for distinguishing artificial surfaces. The NDVI, NDBI, and LSWI values ​​of cultivated land change significantly at different growth stages, which helps to identify the planting cycles of different crops. The NDVI value of bare land is low and the NDBI value is high, which is reflected in drier months, and the LSWI value fluctuates with moisture changes, reflecting the sensitivity of bare soil to shortwave infrared and moisture.

[0116] like Figure 5 As shown ( Figure 5 a- Figure 5 d in represents VV, VH, VV / VH, and VH / VV, respectively), showing the time series variation curves of different ground objects in different polarization characteristics. The VV and VH polarization reflections of forests and grasslands are higher during the growing season, while the reflections in autumn and winter are relatively low. The VV / VH and VH / VV ratios reflect the seasonal changes of forests, grasslands, and cultivated land. The VV and VH polarization scattering of water bodies are both low, and the VV / VH ratio is high, while the VH / VV is low, which can effectively distinguish water bodies. The VV and VH scattering of construction land is high, and the VV / VH ratio is high and relatively stable. The polarization reflection of cultivated land changes significantly under different growth cycles, especially in the VV / VH and VH / VV ratios, which show obvious seasonal differences. The bare land has low reflection under VV and VH polarizations, and the VV / VH ratio is high, while the VH / VV is low, reflecting changes in soil moisture and soil structure.

[0117] like Figure 6 As shown ( Figure 6 a- Figure 6 The e in represents contrast, variance, mean, entropy, and correlation, respectively), showing the time series change curves of different objects in different texture features. The contrast, variance, and entropy of forests and grasslands are higher in spring and summer, and the mean and correlation change with the seasons, which are suitable for vegetation monitoring. The mean of water bodies is low, and the contrast, variance, entropy, and correlation are stable, which is suitable for identifying water bodies. The contrast, mean, and variance of construction land are relatively stable, with little change throughout the year, and the entropy and correlation show obvious seasonal changes, which are suitable for distinguishing buildings. Cultivated land has higher contrast, variance, and entropy during the growing season, and the mean and correlation decrease during the fallow period. The contrast, variance, entropy, and correlation of bare land are high throughout the year, but the variance and entropy decrease under wet conditions.

[0118] 2. Analysis of forest cover extraction results;

[0119] Based on the percentile time series features constructed from Landsat-8-OLI SR (L), Sentinel-1-SAR GRD (S1) and Sentinel-2-MSI SR (S2) multi-source remote sensing image data and combined with terrain factors, the 2023 Nanjing land cover map (such as Figure 7 As shown in a in the figure), further merge the other classes except forest to obtain the forest cover map (as shown in Figure 7 Based on the confusion matrix, the overall accuracy, Kappa coefficient, producer accuracy, and user accuracy were calculated to evaluate the classification accuracy of the land cover results (Table 1).

[0120] The results show that the present invention has significant advantages in medium and high resolution forest cover mapping, with an overall accuracy of 90% and a Kappa coefficient of 88%, showing high reliability and consistency. In the identification of forest categories, both the producer accuracy and user accuracy reached 99%, achieving accurate classification of forest land object categories, indicating that the method of the present invention can accurately capture the spectral and spatial characteristics of forests, and ensure the comprehensiveness and accuracy of forest category identification even in the context of complex land objects. In addition, the classification accuracy of building land, water bodies and cultivated land is 90%, 95% and 93% respectively, which complements the high accuracy of forests and provides a solid foundation for land use classification in complex environments. Although there is some confusion in the classification of grassland and bare land, with producer accuracies of 83% and 75% respectively, the impact on the overall classification results is limited, especially there is no obvious interference with the classification accuracy of forests.

[0121] In summary, the present invention demonstrates excellent performance in forest cover mapping, can provide reliable data support for forest resource monitoring and management, and provides a powerful method for land use classification research.

[0122] The present invention obtains high-reliability forest cover data based on dense time series of multi-source remote sensing images and takes into account phenological characteristics, which can provide an accurate and reliable data source for evaluating the climate effect of forests. The high-quality dense time series data set constructed by this method can give full play to the advantages of multi-source and multi-temporal data, while taking into account the time differences and sensor differences of images, considering the phenological characteristics of forests and their seasonal change characteristics, maximizing the benefits of long-term multi-source and multi-temporal data and algorithm synthesis, thereby ensuring the accuracy of forest cover information characterization. This method can also be applied to other regions in different years to obtain high-precision forest cover data.

[0123] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A long-term forest cover mapping method based on multi-source remote sensing images, characterized by: The following steps are involved: S1: Remote sensing data acquisition and processing; S2: High-quality dense time series time cube construction; S3: Extract phenological features based on the constructed time series cube; S4: Generate land cover map using random forest classification method based on extracted phenological features combined with terrain features; S5: Merge the land cover map to obtain the forest cover map.

2. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 1 is characterized in that: In S1, the specific contents of remote sensing data acquisition and processing are: Filter the Landsat-8-OLI SR, Sentinel-2-MSI SR satellite surface reflectivity optical remote sensing data and Sentinel-1-SAR GRD synthetic aperture radar remote sensing data from past years.

3. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 2 is characterized in that: The collected Landsat-8-OLI SR and Sentinel-2-MSI SR satellite surface reflectance optical remote sensing data were masked with CFmask, and then all image data were cropped and uniformly resampled to a resolution of 30 meters.

4. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 1 is characterized in that: In S2, the specific contents of high-quality dense time series cube construction are: S21: spectral reflectance consistency matching; The linear regression equation between the bands of Landsat-8-OLI SR and Sentinel-2-MSI SR images was constructed. The spectral bands of Sentinel-2-MSI SR images were linearly transformed using Landsat-8-OLI SR data as a reference. Then, the images in the image collection were median synthesized. S22: Time series linear interpolation; The missing data are filled by combining the adjacent high-quality observations in the time series data through the time series linear interpolation method, and finally a high-quality dense time series data set is obtained.

5. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 1 is characterized in that: In S3, the specific contents of phenological feature extraction based on the constructed time series cube are as follows: S31: Extraction and analysis of time series variation curves of different features of different landforms; S32: Phenological feature extraction.

6. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 5 is characterized in that: In S31, the specific contents of the extraction and analysis of the time series variation curves of different features of different objects are as follows: Based on the optical remote sensing data and SAR remote sensing data time series cube, the spectral index, NDVI texture characteristics and polarization characteristics of each image in the time series cube are calculated; The calculation process of the characteristic index of the time series cube of optical remote sensing data is as follows: ； in, , , , , , They represent the blue, green, red, near infrared, shortwave infrared 1, and shortwave infrared 2 bands of the images in the optical remote sensing data time series cube respectively; represents the normalized difference vegetation index, represents the normalized building index, represents the improved normalized difference water index, represents the surface water index; Indicates contrast, represents the variance, Indicates relevance, represents entropy, represents the average value; L / S2 represents the optical remote sensing data time series cube; The calculation process of polarization characteristic data is: ； Wherein, VV represents the vertical polarization mode of Sentinel1; VH represents the vertical horizontal polarization mode of Sentinel1; VV / VH represents the ratio of VV to VH polarization modes; VH / VV represents the ratio of VH to VV polarization modes; Represents the SAR remote sensing data time series cube built based on Sentinel1; By using the constructed high-quality dense time series dataset, we extracted the curves of different land objects changing over time, deeply analyzed the differences in time series of different land objects, and analyzed the curves of spectral bands, spectral indices, polarization characteristics, and NDVI texture characteristics changing over time.

7. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 6 is characterized in that: In S32, the specific content of phenological feature extraction is: The spectral bands, spectral indices, NDVI texture features in the optical remote sensing data time series cube constructed by Landsat-8-OLI SR and Sentinel-2-MSI SR and the polarization feature data in the Sentinel-1-SAR GRD synthetic aperture radar remote sensing data time series cube are realized 10 th , 25 th , 50 th , 75 th , 90 th Percentile statistics of multi-temporal indicators are used to characterize the phenological characteristics of landforms with long time series and medium to high resolution; The calculation process of phenological characteristics is: ； Among them, M S1_S2_T Represents the characteristic index of optical remote sensing data time series cube, S P_R Represents polarization feature data in the time series cube based on SAR remote sensing data; It means that the spectral bands, spectral indices, texture features, and polarization features in the time series cube of optical remote sensing data and SAR remote sensing data are further analyzed by 10 th , 25 th , 50 th , 75 th , 90 th A total of five multi-temporal indices were calculated to characterize the phenological characteristics of different land features.

8. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 7 is characterized in that: In S4, the specific contents of generating land cover maps using random forest classification method based on extracted phenological features combined with terrain features are as follows: Based on the GEE cloud platform and Google Earth high-resolution remote sensing images, different ground feature sample points were selected and the samples were divided into 7:

3. The percentile statistical indicators extracted from S3 were combined with terrain features for random forest classification. Construct a confusion matrix and calculate the overall accuracy, Kappa coefficient, producer accuracy, and user accuracy indicators to evaluate the classification accuracy.

9. The long-term forest cover mapping method based on multi-source remote sensing images according to claim 8 is characterized in that: Overall accuracy P o The calculation process is: ； Among them, N ii Indicates the number of samples correctly classified into the true category; N represents the total number of samples; k represents the number of categories; The calculation process of Kappa coefficient is: ； ； Among them, P e It represents the sum of the "product of the actual and predicted number" corresponding to all categories divided by the "square of the total number of samples"; N represents the total number of samples; the number of real samples of each category is a1, a2, ..., a k , and the predicted number of samples of each category are b1, b2, ..., b k ; k represents the number of categories; The calculation process of producer precision is: ； Among them, a represents the number of correctly classified samples in the ground object category; b represents the number of samples that are actually in this category but are misclassified as a certain category; The calculation process of user precision is: ； Among them, a represents the number of samples that are correctly classified; c represents the number of samples that are actually of a certain category but are misclassified as that category.

Images