A Remote Sensing Classification Method for Forest Types Based on Historical Sample Migration

Abstract

The present invention provides a method for remote sensing classification of forest types based on historical sample migration, belonging to the field of forest monitoring. The method first determines the monitoring area, historical years and target year, obtains the corresponding sample points, then obtains the long-term multi-source remote sensing data of the historical sample points and characterizes the forest growth and succession characteristics, then determines whether the forest type has changed, and migrates the sample data that has not changed to the target year; extracts classification features related to remote sensing classification using the multi-source remote sensing data of the target year; at the same time, divides the monitoring area into multiple floristic regions, and selects the optimal classification features for each floristic region from the relevant classification features; finally, classifies the forest types of each floristic region based on the optimal classification features and the migrated historical forestry field survey sample point data that have not changed. The present invention increases the number of sample points in the target year and improves the accuracy of remote sensing classification of mountain forest types under the condition of small samples.

Description

Technical Field

[0001] The present invention belongs to the field of forest monitoring, and particularly relates to a remote sensing classification method for forest types based on historical sample migration. Background Art

[0002] Mountain forest types are rich in variety, being the main body of the ecological security barrier, an important repository of natural resources, and a treasure house of biodiversity. Accurate classification and mapping of mountain forest types have important scientific significance for sustainable forest resource management, biodiversity conservation, and ecosystem service evaluation, etc. Traditional mountain forest type classification relies on field survey data; with the development of satellite remote sensing technology, due to its advantages such as wide coverage, short revisit period, and low data acquisition cost, it has become the main means for mountain forest type mapping.

[0003] In recent years, when classifying mountain forests through remote sensing technology, traditional machine learning and deep learning methods are generally adopted. However, the classification accuracy of such methods highly depends on the quality and quantity of real samples. The mountain terrain conditions are complex, environmental factors are variable, and traffic accessibility is low, making it often difficult to obtain a large number of forestry field survey samples, resulting in problems such as low classification accuracy and large spatial differences in credibility commonly existing in the remote sensing classification of mountain forest types by the above methods.

[0004] In the prior art, generally, the historical forestry field survey samples are migrated to the target year to expand the sample quantity, thereby improving the accuracy of remote sensing classification. The national forestry field survey data over the years contains a large number of historical known samples. If these sample information can be migrated to the target year to expand the sample quantity, the small sample problem faced in the remote sensing classification of mountain forest types will be fundamentally solved. However, the existing sample migration methods are not yet mature, with low migration efficiency and poor quality, and still cannot effectively improve the accuracy and precision of forest type remote sensing classification. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, the present invention aims to provide a remote sensing classification method for forest types based on historical sample migration, which uses the growth and succession characteristics during the period from the historical year to the target year to migrate the historical forestry field survey sample point data where the forest type has not changed to the target year, so as to expand the sample quantity of the target year, and combines the overall samples after sample expansion with the classification requirements of the flora to conduct remote sensing classification of forest types, improving the classification accuracy and precision of forest types under small sample conditions.

[0006] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0007] A remote sensing classification method for forest types based on historical sample migration, the method comprising the following steps:

[0008] Step S1, determine the monitoring area, historical years, and target year, and obtain the historical sample points of the forestry field survey in the historical years and the target sample points of the forestry field survey in the target year within the monitoring area;

[0009] Step S2, obtain the long-term multi-source remote sensing data of the historical sample points from the historical year to the target year;

[0010] Step S3, use the long-term multi-source remote sensing data to characterize the forest growth and succession characteristics of the historical sample points;

[0011] Step S4, based on the forest growth and succession characteristics, determine whether the forest type of the historical sample points has changed from the historical year to the target year; if the forest type label in the target year is the same as the forest type label in the historical year, it is determined that there is no change; if the forest type label in the target year is different from the forest historical label in the historical year, it is determined that there is a change;

[0012] Step S5, according to whether the forest type has changed, migrate the forestry field survey sample data of the historical sample points that have not changed to the target year, and delete the historical sample points that have changed;

[0013] Step S6, use the multi-source remote sensing data in the target year to extract classification features related to the remote sensing classification of the regional forest type;

[0014] Step S7, according to the spatial distribution characteristics of the flora in the monitoring area, divide the monitoring area into multiple floras, and use the associated hierarchical clustering method to separately select the optimal classification features for the forest type classification of each flora from the relevant classification features;

[0015] Step S8, based on the optimal classification features and the forestry field survey sample data of the historical sample points that have not changed and have been migrated to the target year, use the random forest algorithm to classify the forest type of each flora;

[0016] Step S9, fuse the classification results of the forest type classification of all floras to obtain the spatial distribution map of the forest type in the entire monitoring area.

[0017] As a preferred embodiment of the present invention, for the step of obtaining the forestry field survey sample point data in the historical years and the target year in step S1, first obtain the forestry field survey data of the monitoring area in the historical years and the target year, and then use the method of converting area to points to generate the historical sample point data and the target sample point data map respectively.

[0018] As a preferred embodiment of the present invention, the long-time-series multi-source remote sensing data in step S2 includes all available Landsat Collection2 Tier1 surface reflectance data images, Sentinel-1 / 2 data, SRTM-DEM data, climate data, and soil moisture data corresponding to the period from historical years to the target year on the Google Earth Engine (GEE) platform.

[0019] As a preferred embodiment of the present invention, the forest growth succession characteristics in step S4 include the spectral similarity of historical sample points between the historical year and the target year, the temporal trajectory of historical sample points from the historical year to the target year, and the environmental similarity between historical sample points and target sample points of the same type.

[0020] As a preferred embodiment of the present invention, to determine whether the forest type of historical sample points has changed from the historical year to the target year in step S4, the threshold segmentation method is adopted.

[0021] As a preferred embodiment of the present invention, the threshold segmentation formula is as follows:

[0022]

[0023] And, Environmental similarity = max(S i1 , S i2 , …, S ij )

[0024] In formula (8), Foresttype is the forest type label of historical sample points; SAD and ED are spectral similarity indexes quantifying the similarity between historical sample points in the historical year and the target year, σ1 and σ2 are the segmentation thresholds of SAD and ED respectively; disturbance year is the forest disturbance detection result of historical sample points during the period from the historical year to the target year, Historical year is the historical year; Environnmental similarity is the environmental similarity between historical sample points and target sample points of the same type, σ3 is the segmentation threshold of environmental similarity, and S i,j is the environmental similarity between all sample points of the same forest type between historical sample point i and target sample point j.

[0025] As a preferred embodiment of the present invention, σ1 takes a value of 0.95; σ2 takes a value of 0.1.

[0026] As a preferred embodiment of the present invention, the migration model of the forestry field survey sample data of historical sample points in step S5 is:

[0027]

[0028] In formula (9), the forest type 历史样本点 is the forest type of the historical sample points in each historical forestry field survey sample during the historical year to the target year. If there is no forest type replacement for the historical sample points, it means that the forest type label of this sample point has not changed, and it is migrated to the target year for forest type remote sensing classification; otherwise, if the type label information of this sample point has changed, then this sample point is deleted.

[0029] As a preferred embodiment of the present invention, the relevant classification features extracted in step S6 include: spectrum, polarization, texture, phenology, and environment.

[0030] As a preferred embodiment of the present invention, the method further includes:

[0031] Step S10, using the forestry field survey sample data of the target year to evaluate the mapping accuracy.

[0032] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0033] A forest type remote sensing classification method based on historical sample migration provided by the present invention, based on the growth succession information of historical samples during the historical year to the target year, migrates the historical samples without forest type replacement to the target year to achieve forest type remote sensing classification of the target year. As is well known, even under the condition of no natural or human interference, forest communities will continuously update, grow, and succeed over time; this growth succession process is usually accompanied by changes in characteristics such as tree species composition, stand structure, canopy coverage, and vegetation functional traits, and these characteristic changes often cause changes in the spectral characteristics of the forest canopy and the local microclimate. For example, in the pioneer stage (such as after a forest fire or after a glacier retreat), the stand patches are mainly dominated by fast-growing tree species (such as poplar and birch); while in the secondary forest and mature forest stages, they are gradually replaced by taller shade-tolerant tree species (such as oak and maple), gradually forming a stable stand structure and forest ecosystem; the over-mature forest stage may contain endangered or special environment-adapted tree species, and the stand structure is more complex and diverse. The present invention overcomes the influence of the growth succession of forests on the quality of sample migration ignored by the existing methods, can effectively improve the quality of migration of historical forestry field survey samples, and improve the classification accuracy of forest types in the target year.

[0034] Of course, it is not necessary for any product or method implementing the present invention to simultaneously achieve all the above-mentioned advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a flowchart of the remote sensing classification method of forest types based on historical sample migration provided by the embodiments of the present invention;

[0037] Figure 2 It is a spatial distribution pattern diagram of the forest types in the monitoring area in terms of altitude in 2023 inferred from a specific application example of the embodiments of the present invention;

[0038] Figure 3 It is a comparison diagram of PA indicators for remote sensing classification of forest types between a specific application example of the embodiments of the present invention and other existing methods;

[0039] Figure 4 It is a comparison diagram of UA indicators for remote sensing classification of forest types between a specific application example of the embodiments of the present invention and other existing methods;

[0040] Figure 5 It is a comparison diagram of F1 scores for remote sensing classification of forest types between a specific application example of the embodiments of the present invention and other existing methods;

[0041] Figure 6 It is a comparison diagram of the number of migrated samples of different forest types for historical sample migration between a specific application example of the embodiments of the present invention and other existing methods. Detailed implementation manners

[0042] After the inventors of the present application discovered the above problems, they conducted a detailed study on the existing remote sensing classification methods of forest types, especially the methods using historical migration samples for classification. The study found that there are currently several methods for classifying the surface cover in the target year by migrating historical field survey samples for remote sensing classification of forest types, which mainly include historical sample reuse methods, historical sample weighting methods, and dual-temporal image change detection methods, etc., but all have certain deficiencies.

[0043] For example, the historical sample weighting method dynamically adjusts the weights of samples based on the contribution of historical samples to the target model and uses the weighted historical samples for land cover classification in the target year. For example, Zhou et al. (2016) introduced a transfer term into the target classification model and iteratively adjusted the sample weights based on the classification accuracy of historical and target samples in the model (if a historical sample was misclassified, it indicated that the spatial distribution of the sample was inconsistent with the remote sensing data in the target year, and its weight should be reduced; if a target sample was misclassified, it meant that the sample had not been fully learned by the model, and its weight should be increased) until all target samples were correctly classified, and then used the weighted samples for land cover classification. Compared with the historical sample reuse method, although it reduces the impact of historical samples with inconsistent spatial distributions with the remote sensing data in the target year on the classification accuracy, it does not consider the potential impact of the replacement of historical sample label information caused by land cover changes on the land cover classification accuracy in the target year.

[0044] Another example is the dual-temporal image change detection method, which assumes that changes in land cover types will cause changes in land surface radiation characteristics. First, calculate the spectral feature similarity between historical and target-year remote sensing data, and then determine whether the label information of historical samples has changed by setting a similarity threshold, and transfer the historical samples without label information change to the target year. Finally, based on the transferred samples and the remote sensing data of the target year, use machine learning algorithms to achieve land cover classification in the target year. For example, Lin et al. (2019) transferred historical samples without label information change to the target year by comparing the spectral feature differences (Change Vector Analysis, CVA) of historical samples (in 2010) between the historical year and the target year, and used the transferred samples and Landsat data to achieve automatic classification of land cover types in rapidly urbanized areas. Another example is that Huang et al. (2020) based on the global land cover type sample dataset (First all-season sample set) collected by the FROM-GLC project in 2015, transferred historical samples without label information change to the target year by calculating the spectral similarity and spectral distance between the target year and Landsat data in 2015, and used these transferred samples and Landsat data to conduct global land cover classification in 1990, 1995, 2000, 2005, and 2010 respectively. Although this type of method reduces the impact of the change of historical sample label information on the quality of transferred samples to a certain extent, there are still the following two limitations in the process of transferring historical forestry field survey samples: 1) The spectral feature differences of different forest types in multi-spectral remote sensing images are small, and only using dual-temporal remote sensing images usually can only identify the replacement of forest-non-forest, and it is difficult to effectively identify the replacement information of forest types at the stand scale, such as historical samples that have changed from coniferous / broad-leaved forests to coniferous and broad-leaved mixed forests; 2) This method does not fully consider the potential impact of the natural growth succession of forests from the historical year to the target year on the quality of sample transfer.

[0045] The above analysis shows that the historical sample weighting method is rather cumbersome, which requires a cumbersome weight dynamic adjustment process to generate samples suitable for land cover classification in the target year. Although the double-temporal remote sensing image change detection method has been improved on the basis of reusing historical samples, these methods do not consider the impact of the natural growth and succession of forests during the period from the historical year to the target year on the quality of sample migration, resulting in low quality of the migrated forestry field survey samples. As is well known, even under the condition of no natural or human interference, forest communities will continuously update, grow and succeed over time; this growth and succession process is usually accompanied by changes in characteristics such as tree species composition, stand structure, canopy cover and vegetation functional traits, and these characteristic changes often cause changes in the spectral characteristics of the forest canopy and the local microclimate. For example, in the pioneer stage (such as after a forest fire or after a glacier retreat), the stand patches are mainly dominated by fast-growing tree species (such as poplar and birch); while in the secondary forest and mature forest stages, they are gradually replaced by taller shade-tolerant tree species (such as oak and maple), gradually forming a stable stand structure and forest ecosystem; the over-mature forest stage may contain endangered or specially environment-adapted tree species, and the stand structure is more complex and diverse. Therefore, in the process of migrating historical forestry field survey samples, it is necessary to comprehensively consider the impact of forest growth and succession on the quality of sample migration, so as to improve the quality of the migrated samples, increase the number of effective samples in the target year, and improve the accuracy of remote sensing classification of mountain forest types under the condition of small samples. However, the monitoring methods of the existing technologies do not consider the above problems.

[0046] It should be noted that the defects existing in the above solutions of the prior art are all the results obtained by the inventors through practice and careful research. Therefore, the process of discovering the above problems and the solutions proposed by the embodiments of the present invention below for the above problems should both be the contributions made by the inventors to the present invention in the process of the present invention.

[0047] Next, the technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can also be combined with each other.

[0048] It should be noted that: similar reference numerals and letters indicate similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0049] After the above in-depth analysis, the embodiment of the present invention provides a remote sensing classification method for forest types based on historical sample migration. By using the growth succession information of historical samples from historical years to the target year, the historical samples in the same monitoring area where the forest type has not changed are migrated to the target year to achieve remote sensing classification of forest types in the target year. First, multi-source remote sensing data from historical years to the target year are used to quantitatively characterize the growth succession characteristics of forests; then, the growth succession characteristics are used to determine whether the forest type labels of historical samples have changed, and the historical samples that have not changed are migrated to the target year; finally, based on multi-source remote sensing data and migrated samples, a machine learning algorithm is used for remote sensing classification of forest types in the target year, and the field forestry survey data in the target year is used to quantitatively evaluate the effectiveness of the automatic migration method of the historical forestry field survey samples for remote sensing classification of mountain forest types.

[0050] As Figure 1 shown, the remote sensing classification method for forest types based on historical sample migration includes the following steps:

[0051] Step S1, determine the monitoring area, historical year and target year, and obtain historical sample points of forestry field surveys in the historical year and target sample points of forestry field surveys in the target year within the monitoring area.

[0052] In this step, to obtain the data of forestry field survey sample points in the historical year and target year in the monitoring area, first obtain the forestry field survey data of the monitoring area in the historical year and target year, and then use the method of converting surface to points to generate historical sample point data and target sample point data respectively.

[0053] Step S2, obtain long-term multi-source remote sensing data of the historical sample points from the historical year to the target year.

[0054] In this step, the long-term multi-source remote sensing data includes all available Landsat Collection 2 Tier1 surface reflectance data images, Sentinel-1 / 2 data, SRTM-DEM data, climate data, and soil moisture data corresponding to the period from the historical year to the target year on the Google Earth Engine (GEE) platform.

[0055] Step S3, use the long-term multi-source remote sensing data to characterize the forest growth succession characteristics of the historical sample points; the forest growth succession characteristics include the spectral similarity of the historical sample points between the historical year and the target year, the temporal trajectory of the historical sample points from the historical year to the target year, and the environmental similarity between the historical sample points and the target sample points of the same type.

[0056] In this step, methods such as spectral angle distance (SAD) and Euclidean distance (ED) can be used to calculate the spectral similarity of historical sample points between historical years and the target year; the improved LandTrendr algorithm (iLandTrendr) can be used to calculate the temporal trajectory of historical sample points from historical years to the target year; methods such as the random forest regression algorithm and using the relationship between the Gaussian similarity function and various geographical environment covariates can be used to calculate the environmental similarity between historical sample points and target sample points of the same type.

[0057] In a specific embodiment, to calculate the spectral similarity of historical sample points between historical years and the target year, Landsat data of the monitoring area in historical years and the target year are collected, and two metrics, spectral angle distance (SAD) and Euclidean distance (ED), are used to quantify the spectral similarity between historical samples in optical remote sensing images of historical years and the target year. The specific steps are as follows:

[0058] Step S311: Collect Landsat data corresponding to historical sample points in historical years and the target year;

[0059] Step S312: Use SAD and ED to calculate the spectral similarity between historical sample points in Landsat images of historical years and the target year. The calculation formula is as follows:

[0060]

[0061]

[0062] In the formula, θ is the spectral angle, is the reference spectral feature corresponding to the sample point in historical year t1, is the spectral feature corresponding to the historical sample in target year t2; i is the corresponding spectral band, which represents bands B1 - B5, and B7 of Landsat5 TM data or bands B2 - B7 of Landsat8 OLI; if there is no change in the spectral features between historical and target years, then SAD is equal to 1 and ED is equal to 0. Preferably, the segmentation thresholds of SAD and ED are set to 0.95 and 0.1 respectively.

[0063] In yet another specific embodiment, the temporal trajectory of the historical sample points from the historical year to the target year is calculated. The Landsat data for each vegetation growing season from the historical year to the target year is collected, and then the areas covered by clouds and cloud shadows are masked. Then, the masked images are synthesized to generate an image that can well capture the Landsat observations. Finally, based on the long-term Landsat data, the iLandTrendr (Improved Landsat-based detection of trends in disturbance and recovery) algorithm is used to detect whether forest replacement has occurred for the historical samples during this period, that is, to detect whether forest disturbance events such as forest fires, deforestation, and pests and diseases have occurred in the historical sample data. Specifically, it includes the following steps:

[0064] Step S321: Collect the Landsat data for each vegetation growing season within the range of long-term years, mask the areas covered by clouds and cloud shadows, and then synthesize the masked images to generate an image that can well capture the Landsat observations;

[0065] Step S322: Calculate the Normalized Burn Ratio (NBR) for each pixel, then use the linear interpolation method to fill in the corresponding NBR values for the years covered by clouds and cloud shadows pixel by pixel, and finally use the SG filtering algorithm to filter out the data noise to obtain the NBR time series trajectory of each pixel changing over time;

[0066] Step S323: Based on the NBR time series trajectory within the range of long-term years, use the iLandTrendr algorithm to detect the years when forest disturbance occurs for each pixel; that is, assume that when forest disturbance occurs, the NBR time series trajectory will show a decrease in value, while when the forest recovers, the NBR index will show a gradually increasing trend. Based on the iLandTrendr algorithm, the year corresponding to the node where the "decrease - increase" change in the NBR value occurs is considered the year when forest disturbance occurs, and the corresponding pixel is used as the pixel where forest disturbance occurs within the range of long-term years.

[0067] In another specific embodiment, for calculating the environmental similarity between the historical sample points and the target sample points of the same type, first, multi-source remote sensing data of the target year is used to characterize the environmental conditions related to the spatial distribution of forest types, such as forest canopy characteristics (spectrum, polarization, phenology, spatial texture, etc.), geographical environment (precipitation, temperature, and soil, etc.); second, the Gaussian similarity function is used to compare the similarity of environmental conditions between the historical sample points and the target sample points. Specifically, it includes the following steps:

[0068] Step S331: Use the multi-source remote sensing data obtained in the target year to comprehensively represent the environmental conditions of each geographical location;

[0069] Step S332, calculate the environmental similarity between the historical sample points and the target sample points, and its calculation formula is as follows:

[0070]

[0071] In formula (4), and are the values of the historical sample point i and the target sample point j on the v-th environmental covariate respectively; m is the number of geographical environmental covariates used to characterize the spatial distribution of forest types, which is based on the forest type samples in the target year, and uses the random forest algorithm to evaluate the importance of all classification features in step S2, and selects the features with feature importance greater than 0.00001 to represent the geographical environmental configuration corresponding to each sample point. L v (·) is a function for calculating the geographical environmental similarity between the historical sample point i and the target sample point j on a single environmental variable v. In this paper, the Gaussian similarity function is used for calculation, and its calculation formula is shown in formula 7:

[0072]

[0073] In formula (5), is the standard deviation of the v-th environmental covariate in the whole region, is the square root of the average deviation of the j-th target sample point from all historical sample points of the same type (i = 1, 2,..., k) on the v-th environmental covariate:

[0074]

[0075] is a function for synthesizing the similarities of m environmental covariates. The similarities of each environmental covariate are synthesized by using the feature importance of each environmental covariate for forest type classification, and the comprehensive environmental similarity between the historical sample point i and the target sample point j is obtained:

[0076]

[0077] In formula (7), is to evaluate the feature importance of each environmental covariate for regional forest type classification by using the Mean Decrease Accuracy (MDA) method in the random forest algorithm.

[0078] Step S4, based on the forest growth and succession characteristics, judge whether the forest type of the historical sample point has changed from the historical year to the target year; if the forest type label in the target year is the same as the forest type label in the historical year, it is judged that there is no change; if the forest type label in the target year is different from the forest historical label in the historical year, it is judged that there has been a change.

[0079] In this step, to determine whether the forest type of the historical sample points has changed from the historical year to the target year, the threshold segmentation method is adopted. Specifically, the threshold segmentation formula is as follows:

[0080]

[0081] And, Environmental similarity = max(S i1 , S i2 , …, S ij )

[0082] In Equation (8), Foresttype is the forest type label of the historical sample point; SAD and ED are spectral similarity indexes quantifying the historical sample point between the historical year and the target year, σ1 and σ2 are the segmentation thresholds of SAD and ED respectively; the smaller the value of ED or the larger the value of SAD, the more similar the spectra, and the smaller the possibility of forest type change; disturbance year is the forest disturbance detection result of the historical sample point during the period from the historical year to the target year. If the year of forest disturbance is less than or equal to the historical year (Historical year), it indicates that there is no forest disturbance at the sample location, and the possibility of the replacement of its forest type label information is small; Environmental similarity is the environmental similarity between the historical sample point and the target sample point of the same type, σ3 is the segmentation threshold of environmental similarity, and S i,j is the environmental similarity between all the sample points of the same forest type in the historical sample point i and the target sample point j. Based on the third law of geography, the present invention assumes that similar environmental configurations have similar forest types, that is, the greater the environmental similarity between the historical sample point and the target sample point, the smaller the possibility of forest type change; therefore, the present invention takes each historical sample point collected on the spot in the forestry of the target year as a case containing a specific "forest type - environmental condition" relationship, which can represent the suitable geographical environmental conditions and forest canopy characteristics of the forest type.

[0083] Step S5: According to whether the forest type has changed, migrate the forestry field survey sample data of the historical sample points that have not changed to the target year, and delete the historical sample points that have changed.

[0084] In this step, the migration model of the forestry field survey sample data of the historical sample points is:

[0085]

[0086] In Equation (9), the forest type 历史样本点For the forest types of historical sample points in each historical forestry field survey sample during the historical year to the target year, if there is no forest type replacement for the historical sample point, it means that the forest type label of this sample point has not changed and can be migrated to the target year for remote sensing classification of forest types; otherwise, if the type label information of this sample point has changed, then this sample point is deleted.

[0087] Step S6, extract classification features related to remote sensing classification of regional forest types using multi-source remote sensing data of the target year; as shown in Table 1, the extracted relevant classification features include: spectrum, polarization, texture, phenology, and environment.

[0088] Table 1 Extraction of Classification Features of Forest Types

[0089]

[0090] Among them, the spectral features include: 10 bands of Sentinel-2, NDVI, SVVI, EVI, NDRE, REPI, REI, MCARI, NDPI, MCR, IRECI, NDVIR1, TC_Greenness, etc. The polarization features include: VV, VH, ratio, DIF, AVE, NDI, RVI, mRVI, VDDPI; the texture features include: Asm, contract, dvar…(Calculation for red_edge1 using GLCM); the phenology features include: REP(std, mean, median, min, max, Q25, Q75, IQR, cv, zf), std(mRVI, VV, VH), mRVI_summerwinter, NDVI_maxsummer; the environmental features include: Elevlation, Slope, Aspect, PrecMean, PrecMs, TemMean, TemMs, SMRI.

[0091] Step S7, divide the monitoring area into multiple floras according to the spatial distribution characteristics of the flora in the monitoring area, and use the associated hierarchical clustering method to separately select the optimal classification features for forest type classification of each flora from the relevant classification features to reduce the impact of collinearity and redundancy between classification features on the classification result.

[0092] Step S8, based on the optimal classification features and the forestry field survey sample data of the historical sample points that have not been replaced and migrated to the target year, use the random forest algorithm to classify the forest types of each flora.

[0093] Step S9: Integrate the classification results of the forest type classification for all floristic regions to obtain the spatial distribution map of forest types in the entire monitoring area.

[0094] This embodiment may further include:

[0095] Step S10: Evaluate the mapping accuracy using the field forestry survey sample data of the target year.

[0096] Taking Yunnan Province in the southwestern part of China as the study area, the remote sensing classification method for forest types based on historical sample migration provided by the embodiment of the present invention is adopted to conduct remote sensing classification of the forest types in the selected area. The selected historical years are 2012 and 2016 respectively, and the target year is 2023.

[0097] First, migrate the historical field forestry survey samples of 2012 and 2016; then, based on the migrated samples, conduct remote sensing classification of the forest types in 2023.

[0098] The monitoring area is located in the southwestern region of China, east of the Himalayas, adjacent to Myanmar, Vietnam, and Laos. It is an important part of the Indo-Burma biodiversity hotspot. The forest coverage rate is as high as 65% and the forest stands are relatively young, with strong forest growth and large forest carbon sink potential. The altitude of the study area ranges from 76.4 to 6740 m, showing a stepped decline from north to south. The geomorphic types are complex and diverse, including plains, mountains, hills, basins, etc. The study area belongs to the subtropical plateau monsoon climate. However, due to its complex terrain conditions, the spatial heterogeneity of regional temperature and precipitation conditions is significant, and the spatial and vertical differences in local climate are large, including various ecosystems except the ocean and desert. The unique and diverse natural environmental conditions in the monitoring area provide a suitable habitat for the origin, evolution, and reproduction of various organisms, containing rich forest resource endowments, including multiple floristic regions. Among them, the eastern part is mainly the Sino-Japanese floristic region, the western part is the Sino-Himalayan floristic region, and the southern part is mainly distributed with the paleotropical floristic region. Accurate forest type classification can provide important basic data for the sustainable management of regional forest resources.

[0099] In this embodiment, the data of forest type sample points in the monitoring area are collected, including the field forestry survey data of Yunnan Province in historical years and the target year.

[0100] Among them, the historical sample data includes the forestry field survey data in 2012 and 2016 (3120), with 3699 samples in 2012 and 14065 samples in 2016; the forestry field survey samples in the target year are 3210. In this embodiment, the acquisition of forest type samples for each plot is completed in 4 steps: 1) A series of plots are arranged in Yunnan Province on the basis of comprehensively considering spatial representativeness, economy and traffic accessibility; 2) According to the preset longitude and latitude of the plot, a 30×30m forest plot is arranged using a leather rope, and the real-time kinematic GPS (RTK-GPS) is used to dynamically adjust the actual longitude and latitude information of the plot center and the four corner points to determine the plot boundary; 3) For each plot, the diameter at breast height of each tree is measured with a diameter tape, and the forest type to which each tree with a diameter at breast height ≥5cm belongs is manually identified according to the shape of the leaves, and its crown coverage area is quantitatively estimated; 4) According to the proportion of the crown coverage area of each forest type, the forest type to which the plot belongs is determined. If the proportions of coniferous forest and broad-leaved forest are both less than 65%, the plot belongs to the coniferous and broad-leaved mixed forest, otherwise it is a pure forest.

[0101] Then, long-term multi-source remote sensing data of historical sample points from historical years to the target year are obtained. The long-term multi-source remote sensing data includes all available Landsat Collection 2 Tier1 surface reflectance data images, Sentinel-1 / 2 data, SRTM-DEM data, climate data and soil moisture data corresponding to the period from the historical forest survey years (2012 / 2016) to the target year (2023) on the Google Earth Engine (GEE) platform.

[0102] After corresponding preprocessing of the above long-term multi-source remote sensing data, the forest growth and succession characteristics of historical sample points are characterized.

[0103] The characterization of the forest growth and succession characteristics in this step is mainly completed in three steps, specifically including:

[0104] The spectral angle distance (SAD) and Euclidean distance (ED) are used to quantify the spectral similarity of historical samples between historical years and the target year; the spectral similarity includes the similarity between the B1-B7 bands included in the Landsat data in the historical years (2012 / 2016) and the target year (2023);

[0105] The Landsat data of historical samples from historical years to the target year is used to quantify the temporal trajectory, and the iLandTrendr algorithm is used to detect whether forest disturbances occur during this period;

[0106] Calculate the environmental similarity between historical sample points within a region and sample points of the same type in each target year respectively using the Gaussian similarity function and the feature importance of each geographical environmental covariate.

[0107] Based on the characteristics of forest growth succession, transfer the forestry field survey sample data corresponding to the historical sample points determined not to have undergone forest type replacement to the forestry field survey sample data in 2023 to complete the migration of historical data.

[0108] Then, based on the migrated data and the multi-source remote sensing data in 2023, use machine learning algorithms to perform remote sensing classification of forest types in the target year, and execute steps S6 - S10.

[0109] According to the spatial distribution characteristics of the flora in Yunnan Province, divide Yunnan Province into multiple floras, and use the associated hierarchical clustering method to separately select the preferred features that are beneficial to the forest type classification of each plant region from the classification feature set shown in Table 1 to reduce the impact of collinearity and redundancy between classification features on the classification results. The classification features mainly include spectral, polarization, phenological, texture, and geographical environmental features; the preferred classification features of each flora are shown in Table 2:

[0110] Table 2 Preferred Classification Features of Each Flora

[0111]

[0112]

[0113] Figure 2 The spatial distribution map of forest types in the entire monitoring area is obtained by integrating the classification results of 8 floras.

[0114] To verify the effectiveness of the method for automatically migrating historical forestry field survey samples based on remote sensing forest classification (a samples transferring method considering forest succession, STM - CFS) proposed in this paper, this embodiment uses the forestry field survey samples in the target year to quantitatively evaluate the mapping accuracy of the temporal forest canopy height distribution map inferred by this method. The quantitative evaluation indicators include Overall Accuracy (OA), Kappa coefficient, Producer’s Accuracy (PA), User’s Accuracy (UA), and F1 score (F1score), and their calculation formulas are as follows:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] Where n is the forest type, N is the total number of forest type samples in the on-site survey in 2023, and X ii represents the elements on the diagonal of the confusion matrix, that is, the total number of correctly classified pixels in each forest type, and X i+ and X +i respectively represent the column total and row total of each forest type, that is, the total number of misclassified and wrongly divided pixels.

[0121] The forest type classification map of the monitoring area in 2023 inferred by the method of this embodiment is compared with the field forestry survey sample point data in 2023 for classification accuracy, and the results are as follows:

[0122] It can be seen from Table 3 that the quality of the training samples is positively correlated with the classification accuracy. Compared with directly applying all historical samples to the forest type classification in the target year (OA: 64.94%, kappa: 0.4059), the dual-temporal image change detection method has achieved higher classification accuracy in the remote sensing classification of mountain forest types (OA: 73.27%, kappa: 0.5290). The main reason for this phenomenon may be the widespread existence of the natural growth and succession phenomenon of forests, resulting in a certain proportion of historical samples that have changed from forest to non-forest samples in the target year, that is, false positive samples; while the dual-temporal image change detection algorithm effectively filters out such samples, improving the quality of the training samples to a certain extent and increasing the classification accuracy of forest types. However, the dual-temporal image change detection algorithm can only detect samples with significant spectral feature changes, ignoring the impact of spectral feature changes brought about by the growth and replacement of forest types on the quality of sample migration, and the migrated forest type samples still contain a large number of false positive samples. Compared with only using the dual-temporal image change detection algorithm, the OA and kappa coefficients of the forest type classification results generated based on STM-CF in Yunnan Province have increased by 10.06% and 0.1855 respectively. Therefore, compared with the existing sample migration methods, the STM-CFS proposed in this paper can effectively filter out samples whose forest types have changed during the period from the historical year to the target year (such as pure forests becoming mixed forests), improve the quality of sample migration, and increase the classification accuracy of mountain forest types.

[0123] Table 3 Comparison of forest type classification accuracies of different sample migration methods

[0124]

[0125] From Figures 3 - 6 It can be seen that the sample migration method proposed in this embodiment is overall more effective than the existing sample migration methods in the remote sensing classification of mountain forest types. However, different sample migration methods have significant differences in the classification accuracy of each forest type. For example, the classification accuracy of coniferous forests and broad-leaved forests is significantly higher than that of mixed coniferous and broad-leaved forests and bamboo forests. Compared with using all historical samples, the classification accuracy of the double-temporal image change detection algorithm shows a significant upward trend for coniferous forests, broad-leaved forests, and mixed coniferous and broad-leaved forests, but the classification accuracy of bamboo forests decreases slightly. The PA (as shown in Figure 3 ), UA (as shown in Figure 4 ), and F1score (as shown in Figure 5 ) of coniferous forests increase by 8.45%, 3.39%, and 0.0644 respectively; the PA, UA, and F1score of broad-leaved forests increase by 0.95%, 15.69%, and 0.0890 respectively; the PA and F1score parameters of mixed coniferous and broad-leaved forests increase by 11.84% and 0.0018, while the UA decreases by 1.73%; the UA of bamboo forests increases by 8.11%, while the PA and F1score decrease by 24.00% and 0.0037. Compared with the double-temporal change detection method, the STM-CFS sample migration method proposed in this paper significantly improves the classification accuracy of all forest types. For example, the PA, UA, and F1score of coniferous forests increase by 12.05%, 6.29%, and 0.0963 respectively; the PA, UA, and F1score of broad-leaved forests increase by 7.87%, 5.34%, and 0.0686 respectively; the PA, UA, and F1score of mixed coniferous and broad-leaved forests increase by 29.40%, 37.37%, and 0.3810 respectively; the PA, UA, and F1score of bamboo forests increase by 12.10%, 63.19%, and 0.6478 respectively. The above research results show that compared with directly applying all historical samples to the forest type classification of the target year, using the double-temporal image change detection method to extract unchanged samples effectively improves the quality of training samples and the classification accuracy of coniferous forests, broad-leaved forests, and mixed coniferous and broad-leaved forests. However, due to the small number of bamboo forest samples (as shown in Figure 6 ), and the small spectral feature differences between bamboo forests and other forest types, the detected bamboo forest samples by this method contain a certain proportion of false positive samples in which the dominant tree species of bamboo forests have changed, resulting in a significant increase in the omission error of bamboo forests. The STM-CFS method effectively retains the samples where the forest type has not actually changed, improves the quality of training samples, and improves the classification accuracy of each forest type.

[0126] As can be seen from the above technical solutions, the forest type remote sensing classification method based on historical sample migration provided by the embodiments of the present invention can effectively use the forestry field investigation samples in known historical years to infer the forest type remote sensing classification map of the target year. Case studies in the monitoring area show that the method provided by the present invention can effectively improve the migration quality of historical forestry field investigation samples, increase the number of effective samples in the target year, and improve the classification accuracy of the mountain forest type under the condition of small samples by machine learning algorithms. Compared with the existing forest type remote sensing classification methods based on historical sample migration, the migration samples obtained by the forest type remote sensing classification method of the present invention have higher quality, higher consistency between the obtained samples and the forestry field investigation results in the target year, higher remote sensing classification accuracy and accuracy, and higher effectiveness for the forest type classification in a complex terrain area such as the monitoring area.

[0127] The above description is only the preferred embodiment of the present invention and the explanation of the applied technical principle, and is not intended to limit the scope of the present invention claimed, but only represents the preferred embodiment of the present invention. Those skilled in the art should understand that the scope of the invention involved in the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present invention.

Claims

1. A remote sensing classification method for forest types based on historical sample migration, characterized in that: The method comprises the following steps: Step S1, determining the monitoring area, historical years and target years, obtaining historical sample points of forestry field surveys in historical years and target sample points of forestry field surveys in the target year within the monitoring area; Step S2, obtaining long-term multi-source remote sensing data of the historical sample points from the historical year to the target year; Step S3, using long-term multi-source remote sensing data to characterize the forest growth succession characteristics of historical sample points; Step S4, based on the forest growth succession characteristics, determine whether the forest type to which the historical sample point belongs has been replaced from the historical year to the target year; if the three characteristics of the spectral similarity of the historical sample point in the historical year and the target year, the temporal trajectory similarity of the historical sample point in the historical year and the target year, and the geographical environment similarity of the historical sample point and the sample point of the same type in the target year are similar, then the forest type to which the historical sample point belongs has not been replaced; otherwise, it is determined that it has been replaced; Step S5, migrating the forestry field survey sample data of the historical sample points that have not been replaced to the target year, and deleting the historical sample points that have been replaced; Step S6, extracting classification features related to remote sensing classification of regional forest types using multi-source remote sensing data of the target year; Step S7, dividing the monitoring area into a plurality of floras according to the spatial distribution characteristics of the floras in the monitoring area, and selecting the optimal classification features for forest type classification of each flora from the relevant classification features using an association hierarchical clustering method; Step S8, based on the optimal classification features and the forestry field survey sample data of the historical sample points that have not been replaced and migrated to the target year, use the random forest algorithm to classify the forest type of each flora; Step S9, integrating the classification results of forest type classification of all floras to obtain a spatial distribution map of forest types in the entire monitoring area.

2. The forest type remote sensing classification method based on historical sample migration according to claim 1 is characterized in that: Step S1 is to obtain forestry field survey sample point data for historical years and target years in the monitoring area. First, the forestry field survey data for historical years and target years in the monitoring area are obtained, and then the historical sample point data map and the target sample point data map are generated respectively using the surface-to-point method.

3. The forest type remote sensing classification method based on historical sample migration according to claim 1 is characterized in that: The long-term multi-source remote sensing data in step S2 include all available Landsat Collection 2 Tier 1 surface reflectance data images, Sentinel-1 / 2 data, SRTM-DEM data, climate data, and soil moisture data on the Google Earth Engine GEE platform corresponding to the period from historical years to target years.

4. The forest type remote sensing classification method based on historical sample migration according to claim 1 is characterized in that: Step S4 determines whether the forest type of the historical sample point has changed from the historical year to the target year, using a threshold segmentation method.

5. The forest type remote sensing classification method based on historical sample migration according to claim 4 is characterized in that: The threshold segmentation formula is as follows: （8）， and, In formula (8), is the forest type label of the historical sample point; SAD and ED are the spectral similarity indicators that quantify the historical sample points between the historical year and the target year, and σ1 and σ2 are the segmentation thresholds of SAD and ED respectively; It is the year corresponding to the forest disturbance in the forest disturbance detection results of the historical sample points from the historical year to the target year. For historical years; is the environmental similarity between the historical sample points and the sample points of the same type in the target year, σ3 is the segmentation threshold of environmental similarity, For historical sample points Sample points of the same type as the target year The environmental similarity between all sample points of the same forest type in the .

6. The forest type remote sensing classification method based on historical sample migration according to claim 5 is characterized in that: The value of σ1 is 0.95; the value of σ2 is 0.

1.

7. The forest type remote sensing classification method based on historical sample migration according to claim 1 is characterized in that: The migration model of forestry field survey sample data of historical sample points in step S5 is: （9）， In formula (9), For each historical sample point in the historical forestry field survey sample, the forest type from the historical year to the target year is taken. If there is no forest type change at the historical sample point, it means that the forest type label of the sample point has not changed and it is migrated to the target year for forest type remote sensing classification. Otherwise, if the type label information of the sample point has changed, the sample point is deleted.

8. The forest type remote sensing classification method based on historical sample migration according to claim 1 is characterized in that: The relevant classification features extracted in step S6 include: spectrum, polarization, texture, phenology and environment.

9. The forest type remote sensing classification method based on historical sample migration according to any one of claims 1 to 8, characterized in that: The method further comprises: Step S10, evaluating mapping accuracy using forestry field survey sample data of the target year.

Images