A time-series forest age mapping method that incorporates forest disturbance and recovery events

Abstract

The application discloses a time series forest age mapping method combining forest disturbance and recovery events, comprising the following steps: (1) obtaining a remote sensing satellite image data source; (2) constructing a forest age estimation model; (3) estimating the spatial distribution of forest age in a benchmark year by a random forest algorithm; (4) improving a LandTrendr algorithm; (5) detecting forest disturbance and recovery information by a Landsat disturbance and trend detection (LandTrendr) algorithm; (6) integrating different change year results; and (7) calculating a time series forest age. The application can detect forest changes by the LandTrendr algorithm, can recognize wide disturbance intensity, can complete forest age mapping of natural forests and artificial forests, and can improve the mapping precision by combining the detected forest and recovery events into forest age time series estimation.

Description

Technical Field

[0001] This invention relates to the field of forest age mapping, and more particularly to a time-series forest age mapping method that combines forest disturbance and restoration events. Background Technology

[0002] Forest age is a key parameter determining the carbon sequestration status and potential of forest ecosystems. The lack of a time-series spatial pattern of forest age makes it impossible to capture forest disturbance and restoration history within sufficient spatial and temporal domains, thus increasing the uncertainty in estimating past and future forest carbon sinks. Therefore, considering forest disturbance and restoration history and exploring the dynamic changes in forest age is of significant scientific importance for accurately estimating forest ecosystem carbon sinks.

[0003] Combining remote sensing with ground-based observations provides an effective and feasible method for obtaining the spatiotemporal distribution of forest age. Currently, remote sensing-based forest age estimation methods tend to establish regression relationships between forest age and vegetation features such as spectral and textural characteristics to predict the spatial distribution of forest age. The basic physical mechanism of this method lies in the fact that the different optical properties of tree species, branches, leaves, bark, trunks, and canopies within a forest stand lead to variations in the light reflectance and transmittance of the canopy structure. Furthermore, as forest age increases, the image texture of the forest canopy changes, shifting from a uniform style to a clump-like pattern, which also allows for the establishment of a relationship between canopy texture features and forest age. Based on these reasons, researchers both domestically and internationally have utilized regression methods such as multiple linear regression, artificial neural networks, and support vector machines, combined with spectral and textural features, to invert the spatial distribution of forest age.

[0004] However, the above methods can only obtain the spatial distribution of forest age for a specific year. The key to extrapolating forest age over time series lies in combining the forest age of a specific year with disturbance and restoration history. Therefore, by using a specific year as a base year and combining disturbance and restoration events to extrapolate forest age backward, it is possible to map the spatial distribution of forest age over time series.

[0005] Currently, the only similar prior art invention is "A New Method for Spatial Mapping of Plantation Forest Age" (Application No.: CN201710592204.2, Authorization Announcement No.: CN107247809B). It first uses a random forest algorithm to simulate the forest age distribution for a given year, then uses a vegetation change tracking (VCT) algorithm to detect forest disturbance events, and finally calculates the spatial distribution of the time-series plantation age affected by the disturbance based on the first two methods.

[0006] Compared to the VCT algorithm, the Landsat Disturbance and Trend Detection (LandTrendr) algorithm is widely used globally as an effective method for detecting forest disturbance and recovery events. The advantages of the LandTrendr algorithm lie in its ability to detect both abrupt disturbance events (clear-cutting) and long-term, slow events (forest growth, post-disturbance recovery) through line segmentation. Furthermore, the LandTrendr algorithm can be configured with different spectral indices to obtain varying detection results, unlike the VCT algorithm which is limited to the IFZ index. Therefore, this invention attempts to utilize the Google Earth Engine (GEE) cloud platform to more accurately determine the years of forest disturbance and recovery based on the LandTrendr algorithm with different spectral indices. This allows for coupling with a random forest model to estimate the age of time-series forests, and the entire process is fully programmed to automate the mapping of time-series forest ages at different scales and time ranges.

[0007] The existing Vegetation Change Tracking (VCT) algorithm used to detect forest changes can only identify limited disturbance intensities and can only complete the mapping of the age of plantation forests. The forest restoration events detected by the VCT algorithm are not incorporated into the forest age estimation. Or, it can only run forest disturbance and restoration detection algorithms based on a single spectral index to obtain the corresponding forest change information. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing a time-series age mapping method that combines forest disturbance and restoration events. This method solves the problem that previous methods could only map the age of small- to medium-scale plantations, enabling large-scale age mapping of both plantations and natural forests. It obtains more accurate information on forest disturbance and restoration, improves the accuracy of time-series age spatial mapping, and allows users to freely set various spectral indices, thereby combining the results of each index to further enhance the accuracy of age spatial mapping.

[0009] The objective of this invention is achieved through the following technical solution: a time-series forest age mapping method combining forest disturbance and recovery events, wherein the technical solution for time-series forest age spatial distribution mapping includes the following steps:

[0010] (1) Acquisition of remote sensing satellite image data sources:

[0011] Based on the user's required mapping area, year, and resolution, obtain remote sensing satellite data of surface reflectance imagery for the corresponding location, year, and resolution; and obtain corresponding ground survey data on forest age.

[0012] (2) Constructing a forest age estimation model:

[0013] Based on the original spectral bands of surface reflectance data, principal component analysis variables, spectral indices, tassel transformation, and texture feature variables with gray-level co-occurrence matrix are calculated. Combined with forest age ground survey data, a training set is constructed, and the training set is used to train a random forest regression model.

[0014] (3) Random forest algorithm for estimating the spatial distribution of forest age in the baseline year:

[0015] After the random forest regression model is trained, the non-forest areas within the mapping range are masked out, and the forest age of the baseline year within the mapping range is estimated based on the random forest regression model.

[0016] (4) Improved LandTrendr algorithm:

[0017] The input data of the LandTrenr algorithm is expanded to enable it to retrieve surface reflectance image data from Sentinel and Moderate Resolution Imaging Spectroradiometer (MODIS), construct corresponding time series stack data, namely the ImageCollection object in GEE, and then calculate and segment the spectral trajectory through this object. Based on the trajectory segmentation algorithm, the year of change is obtained.

[0018] (5) Landsat Disturbance and Trend Detection (LandTrendr) algorithm detects forest disturbance and restoration information:

[0019] By setting different spectral indices, and based on the improved LandTrendr algorithm, we can obtain different years of forest disturbance and restoration.

[0020] (6) Integrating results from different years of change:

[0021] (6.1) The results of the changes in different spectral indices over the years were combined into a single image using a majority voting method;

[0022] (6.2) The last five band variables in the results of different spectral indices are used as independent variables of the random forest classification model, and the interference and recovery information judged by the user's visual interpretation are used as labels to create a forest interference and recovery training set. The training set is used to train the random forest classification model to simulate the boundary region of forest interference and recovery. The model output results are two binary raster images.

[0023] (6.3) Finally, the results of the year-by-year changes are masked using binary raster images to obtain the spatial distribution maps of the years in which forest disturbance and restoration occurred.

[0024] (7) Estimating the age of forests over time series:

[0025] First, the forest age for the following year is calculated using the forest age of the baseline year. Corresponding pixels are retrieved from the LandTrendr detection results of years of forest disturbance and recovery. Pixels affected by forest disturbance in the following year are assigned a value of 0, pixels recovering from forest damage are assigned a value of 1, and pixels unchanged in the following year are assigned the age value of the previous year by adding 1. This process creates the spatial distribution of forest age for the following year. Then, the forest age for the following year is iteratively obtained using the forest age of the previous year. Finally, the time-series forest age mapping required by the user is completed.

[0026] Furthermore, the detection results of the spectral index include six bands: year of change, magnitude of change, duration of change, spectral value before change, rate of spectral change, and signal-to-noise ratio.

[0027] Furthermore, the majority voting method specifically works as follows: when the number of majority elements is greater than n / 2, the value of the majority element is assigned to the pixel. Conversely, when there is no majority element, or the number of majority elements is less than n / 2, the value with the highest detection accuracy in the time series is assigned to the pixel.

[0028] Furthermore, in step (6.2), the labels are: "Forest Loss / Forest Unchanged" and "Forest Recovery / Forest Unchanged".

[0029] The beneficial effects of this invention are:

[0030] (1) The LandTrendr algorithm is used to detect forest changes, which can identify a wide range of disturbance intensities and complete the mapping of the forest age of natural forests and plantations.

[0031] (2) Improve the accuracy of mapping by combining the forest and restoration events detected by the LandTrendr algorithm to the time series estimation of forest age.

[0032] (3) Use a variety of remote sensing feature variables (e.g., original bands, spectral indices) based on the LandTrendr algorithm to detect forest changes, and combine them with the majority voting method to obtain more accurate information on forest disturbance and recovery. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1The present invention provides a flowchart of a time-series forest age mapping method that combines forest disturbance and recovery events. Detailed Implementation

[0035] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, the present invention provides a time-series forest age mapping method that combines forest disturbance and recovery events, the specific steps of which are as follows:

[0037] (1) Selection of remote sensing satellite image data source:

[0038] Using Google Earth Engine (GEE), surface reflectance data from Landsat, MODIS, Sentinel, and other sources are accessed (https: / / developers.google.com / earth-engine / datasets / ). Based on the user's required mapping area, year, and resolution, surface reflectance imagery data for the corresponding location, year, and resolution is selected, along with relevant ground survey data on forest age. This is to facilitate forest age mapping at large, medium, and small scales. Remote sensing satellite imagery data can also be acquired using the PIE-Engine cloud platform, which schedules the corresponding data and runs the same JavaScript scripts (including the forest age baseline year inversion and the LandTrendr algorithm mentioned below).

[0039] (2) Constructing a forest age estimation model:

[0040] Based on the raw spectral bands of surface reflectance data, principal component analysis variables, spectral indices, tasseled cap transformation, and texture features such as the gray-level co-occurrence matrix are calculated. This data is then combined with user-collected ground survey data on forest age to construct a sample set, divided into a training set (70%) and a test set (30%). The training set is used to train a random forest regression model (or multiple linear regression, artificial neural networks, support vector machines, etc.) to determine the quantitative relationship between forest age and various feature variables through modeling.

[0041] (3) Random forest algorithm for estimating the spatial distribution of forest age in the baseline year:

[0042] After optimizing and training the random forest regression model, the model's accuracy was validated using a test set, and then the model was applied to the entire mapping area. Simultaneously, the land cover dataset from GEE (https: / / developers.google.com / earth-engine / datasets / tags / landcover) was used to mask out non-forested areas within the mapping area, allowing for simulation and estimation of the forest age for the baseline year within the mapping area based on the random forest regression model.

[0043] (4) Improved LandTrendr algorithm:

[0044] The LandTrendr algorithm's input data has been expanded: The original LandTrendr algorithm first constructed a time-series stacked data based on Landsat imagery data provided by GEE, then extracted the spectral trajectories of the original bands or spectral indices, and finally detected the year of change using a trajectory segmentation algorithm. The improved LandTrendr algorithm expands its input data source by replacing the previously called Landsat imagery data with other surface reflectance imagery data (such as Sentinel and MODIS) using the ee.ImageCollection() function on GEE. This constructs corresponding time-series stacked data (ImageCollection image collection objects in GEE). This object is then used to calculate and obtain the corresponding segmented spectral trajectories using LandTrendr's spectral trajectory segmentation algorithm, thereby obtaining the corresponding year of change. This enables the detection of disturbances and restoration in forests of various scales (small, medium, and large).

[0045] (5) Landsat Disturbance and Trend Detection (LandTrendr) algorithm detects forest disturbance and restoration information:

[0046] By setting different spectral indices and using the improved LandTrendr algorithm, we can obtain different years of forest disturbance and restoration. The detection result of a certain spectral index includes 6 bands: year of change, magnitude of change, duration of change, spectral value before change, rate of spectral change, and signal-to-noise ratio.

[0047] (6) Integrate the results from different years of change. The specific steps are as follows:

[0048] (6.1) The results of the changes in different spectral indices over different years are integrated into an image using a majority voting method. Specifically, the majority voting method is as follows: when the number of majority elements is greater than n / 2, the value of the majority element is assigned to the pixel. Conversely, when there is no majority element, or the number of majority elements is less than n / 2, the value with the highest detection accuracy in the time series is assigned to the pixel.

[0049] (6.2) Using the five band variables other than "year of change" from the results of different spectral indices as independent variables of the random forest classification model, and using the interference and restoration information judged by user visual interpretation as labels, a forest interference and restoration sample set was created, with 70% used as the training set and 30% as the test set. The random forest classification model was trained using the training set, and the model was validated using the test set. After the model was established, the boundary region of forest interference and restoration was simulated, and the result was two binary raster images, labeled as "forest loss / forest unchanged" and "forest restoration / forest unchanged," respectively.

[0050] (6.3) Finally, the results of the integrated data for the different years are masked using binary raster images to obtain spatial distribution maps of the years in which forest disturbance and restoration occurred. This result can accurately show which pixels experienced forest loss or restoration and more accurately detect the years in which the changes occurred. Thus, the forest disturbance and restoration information obtained through multiple remote sensing features is more accurate and overcomes the limitation of existing technologies that can only use a single index.

[0051] (7) Estimating the age of forests over time series:

[0052] First, the forest age for the following year is calculated using the baseline forest age. Retrieve corresponding pixels from LandTrendr's detection results of forest disturbance and restoration years. Pixels affected by forest disturbance in the following year are assigned a value of 0, while pixels showing forest restoration are assigned a value of 1. For pixels unchanged in the following year, the age value of the previous year is incremented by 1, thus creating the spatial distribution of forest age for the following year. Then, the forest age for the following year is iteratively calculated using the forest age of the previous year. Finally, the time-series forest age mapping required by the user is completed. This approach combines both forest disturbance and restoration information, achieving a more accurate estimation of time-series forest age.

[0053] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for time-series forest age mapping combining forest disturbance and recovery events, characterized in that, The technical solution for spatial distribution mapping of forest age over time includes the following steps: (1) Acquisition of remote sensing satellite image data sources: Based on the user's required mapping area, year, and resolution, obtain remote sensing satellite data of surface reflectance imagery for the corresponding location, year, and resolution; and obtain corresponding ground survey data on forest age. (2) Constructing a forest age estimation model: Based on the original spectral bands of surface reflectance data, principal component analysis variables, multiple spectral indices, tasseled cap transformation, and texture feature variables with gray-level co-occurrence matrix are calculated. By combining ground survey data on forest age, a training set is constructed, and the training set is used to train a random forest regression model; (3) Spatial distribution of forest age in the baseline year estimated by random forest algorithm: After the random forest regression model is trained, the non-forest areas within the mapping range are masked out, and the forest age of the baseline year within the mapping range is estimated based on the random forest regression model. (4) Improved LandTrendr algorithm: The input data of the LandTrenr algorithm is expanded to allow it to retrieve surface reflectance image data from Sentinel and Moderate Resolution Imaging Spectroradiometer (MODIS), construct corresponding time-series stacked data (i.e., ImageCollection objects in GEE), calculate and segment spectral trajectories using these objects, and obtain the year of change based on the trajectory segmentation algorithm; this enables disturbance and restoration detection in forests of various scales (small, medium, and large). (5) Landsat Disturbance and Trend Detection (LandTrendr) algorithm detects forest disturbance and restoration information: By setting different spectral indices and using the improved LandTrendr algorithm, different years of forest disturbance and restoration were obtained. The detection results of the spectral indices include 6 bands: year of change, magnitude of change, duration of change, spectral value before change, rate of spectral change, and signal-to-noise ratio. (6) Integrating results from different years of change: (6.1) The results of the changes in different spectral indices over the years are integrated into an image by majority voting. Specifically, when the number of majority elements is greater than n / 2, the value of the majority element is assigned to the pixel. Conversely, when there is no majority element or the number of majority elements is less than n / 2, the value with the highest detection accuracy in the time series is assigned to the pixel. (6.2) The magnitude of change, duration of change, spectral value before change, rate of change of spectral value, and signal-to-noise ratio other than "year of change" in the results of different spectral indices are used as independent variables of the random forest classification model. Forest disturbance binary classification training set and forest recovery binary classification training set are constructed respectively. The disturbance and recovery information judged by the user's visual interpretation are used as labels. The labels other than "year of change" are as follows: the label of the forest disturbance binary classification training set is "forest loss / forest unchanged", and the label of the forest recovery binary classification training set is "forest recovery / forest unchanged". The first random forest binary classification model is trained using the forest disturbance binary classification training set to simulate the boundary area of ​​forest disturbance and output the binary raster image of forest disturbance. The second random forest binary classification model is trained using the forest recovery binary classification training set to simulate the boundary area of ​​forest recovery and output the binary raster image of forest recovery. (6.3) Finally, the integrated results of the change years obtained by the majority voting method are masked using the binary raster image of forest disturbance to obtain a spatial distribution map of the years in which forest disturbance occurred; the integrated results of the change years obtained by the majority voting method are masked using the binary raster image of forest restoration to obtain a spatial distribution map of the years in which forest restoration occurred; based on multiple remote sensing features, it can accurately represent which pixels have experienced forest loss or restoration, and more accurately detect the years in which changes occurred; (7) Estimating the age of forests over time: First, the forest age for the following year is calculated using the forest age of the baseline year. Retrieve corresponding pixels from the LandTrendr detection results of years of forest disturbance and restoration. For pixels affected by forest disturbance in the following year, assign a value of 0; for pixels restored by forest restoration, assign a value of 1; and for pixels unchanged in the following year, add 1 to the age value of the previous year. This creates the spatial distribution of the forest age for the following year. Then, iteratively use the forest age of the previous year to obtain the forest age for the following year. Finally, complete the mapping of the time-series forest age required by the user. By combining forest disturbance and restoration information, a more accurate estimation of the time-series forest age is achieved.

Images