Forest degeneration process identification and degeneration degree division method based on time sequence

Abstract

The invention provides a forest degeneration process identification and degeneration degree division method based on a time sequence, and relates to the field of forest resource degeneration restoration. The method comprises the following steps: acquiring surface reflectance data, and calculating an NBR index after preprocessing; fitting an NBR time sequence through a LandTrendr algorithm, and extracting sample place time sequence data based on random sampling and visual interpretation; generating an adversarial network and expanding sample point fitting data into a simulation data set in combination with a self-supervised learning technology; training CNN, random forest and BOSSVS time sequence classification models, and constructing a hybrid classification model through an integrated voting strategy; and identifying a forest degeneration process by using a hybrid classification model, and generating a degeneration degree diagram through reclassification. The method can achieve the efficient recognition of the complex degradation process of the forest region, has good time sequence adaptability and regional applicability, and provides technical support for the monitoring and management of ecological restoration of regional forests.

Description

Technical Field

[0001] This invention relates to the field of forest ecological restoration technology, specifically to a method for identifying forest degradation processes and classifying degradation levels based on time series data. Background Technology

[0002] Forests, as the largest carbon sink on land, play a crucial role in maintaining the function of global ecosystems. In recent years, with the increasing frequency of extreme weather events, and the intensification of agricultural expansion, urbanization, and human activities such as logging, forest cover and quality have been severely affected, and disturbance events have become increasingly frequent. The resulting forest degradation problem is becoming increasingly widespread, not only disrupting the carbon cycle balance and climate regulation mechanisms, but also leading to a significant decline in forest productivity, biodiversity, and carbon sink functions. Unlike typical disturbance events, forest degradation often manifests as gradual changes or subtle abrupt changes, making its identification and monitoring more challenging. Therefore, there is an urgent need to develop an efficient and precise technological system to accurately identify forest degradation processes and obtain information on the degree of forest degradation.

[0003] Domestic and international scholars have conducted extensive research in the field of forest degradation, with the main focus being on the quantitative assessment of the degree of forest degradation, its causes, processes, mechanisms, and ecological impacts. Traditional research methods largely rely on field surveys and fixed-point monitoring. While these methods offer high data accuracy and can objectively reflect the state of forest degradation, they suffer from limitations such as large workload, high labor intensity, high cost, long cycle time, low efficiency, and poor timeliness. In contrast, the Landsat satellite dataset provides the longest-running land surface observation record to date. Its inclusion of spectral bands highly sensitive to changes in forest structure significantly expands the forest observation database, providing a solid data foundation for forest ecosystem assessment. Currently used methods for forest degradation detection include: spectral mixture analysis (SMA), high-resolution image visual interpretation, LiDAR, radar and UAV monitoring, and degradation measurement based on landscape fragmentation. The SMA method utilizes the spectral reflectance characteristics of different land cover types to decompose pixels into proportions of different cover types and identifies degraded areas by calculating the Normalized Degradation Index (NDFI). Visual interpretation identifies degradation distribution on a large scale by recognizing features such as canopy gaps, forest roads, burned areas, and urban land use. LiDAR can accurately measure tree height and estimate biomass, assessing the degree of degradation from the perspective of forest ecological function loss; landscape fragmentation-based assessment methods quantify degradation levels based on landscape features such as reduced patch size, increased patch number, and forest edge expansion. However, most of these methods focus on binary state mapping (degraded / non-degraded), lacking in-depth characterization of the dynamic characteristics of the degradation process and struggling to distinguish degradation intensity.

[0004] Therefore, there is an urgent need for a method that can accurately identify vegetation degradation processes based on pixel-level time series information, and then classify the degree of degradation. Summary of the Invention

[0005] The purpose of this invention is to provide a time-series-based method for identifying forest degradation processes and classifying degradation levels, which solves the problem that existing methods cannot capture forest degradation processes and classify their levels, and provides information on forest degradation processes and levels for areas with frequent disturbances to forest ecosystems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying forest degradation processes and classifying degradation levels based on time series data includes the following steps:

[0008] S1: Obtain Landsat surface reflectance data from the GEE platform, and calculate the NBR index after preprocessing;

[0009] S2: Fit the NBR time series using the LandTrendr algorithm, and extract the fitted NBR time series data from the sampling locations based on random sampling and visual interpretation;

[0010] S3: Generative adversarial networks combined with self-supervised learning techniques are used to expand the NBR time series data fitted at sample sites into a simulated dataset containing 7 types of degradation recovery.

[0011] S4: Train CNN, Random Forest and BOSSVS time series classification models, and build a hybrid classification model through an ensemble voting strategy;

[0012] S5: Apply a hybrid classification model to identify forest degradation processes and generate degradation degree maps through reclassification.

[0013] Furthermore, the preprocessing described in S1 includes:

[0014] The CFmask algorithm was used to mask the selected images for clouds, shadows, and snow.

[0015] The ee.mosaic and ee.clip functions in the GEE platform are used to precisely crop and seamlessly stitch the processed images according to the boundaries of the study area, forming a complete image dataset of the study area.

[0016] Furthermore, the parameters of the LandTrendr algorithm described in S2 are set as follows: maximum number of segments (maxSegments) = 6, peak threshold (spikeThreshold) = 0.9, vertexCountOvershoot = 3, recovery threshold (recoveryThreshold) = 0.25, minimum number of observations (minObservationsNeeded) = 6, preventOneYearRecovery = true, p-value threshold (pvalThreshold) = 0.05, and bestModelProportion = 0.75. Furthermore, the degradation recovery types described in S3 include: degradation-complete recovery, degradation-partial recovery, degradation-no recovery, post-recovery degradation, bare land at the start of the study period becoming woodland, stable vegetation, and bare land.

[0017] Furthermore, the CNN described in S4 has the following training parameters for this time series classification model: 20,000 training epochs, 16 batch size, Adam optimizer, 0.01 learning rate, [6, 12] convolutional layer filters, auto padding, and cross-entropy loss function. Furthermore, the random forest described in S4 has the following training parameters for this time series classification model: 100 decision trees (n_estimators), information gain (entropy) as the node splitting criterion, unlimited maximum tree depth (max_depth), 2 minimum samples per node split (min_samples_split), 1 minimum sample size per leaf node (min_samples_leaf), sqrt square root feature selection strategy (max_features), True bootstrap, 42 random seed (random_state), and -1 parallel processing (n_jobs).

[0018] Furthermore, the training parameters of the BOSSVS time series classification model described in S4 are as follows: alphabet size n_bins is 4, sliding step window_step is 1, window size window_size is 10, feature normalization flags norm_mean and norm_std are False, and IDF weighting use_idf is True.

[0019] Furthermore, the integrated voting strategy described in S4 employs a hard voting mechanism based on the mode to integrate the prediction results of the three models: CNN, Random Forest, and BOSSVS.

[0020] Furthermore, the reclassification described in S5 has the following rules:

[0021] The terms "degradation-no-recovery" and "recovery-post-degradation" are merged into a single severe degradation type.

[0022] "Degradation-partial recovery" is defined as a mild degradation type;

[0023] The categories “bare land at the start of the study period that was later converted to forest” and “degradation-complete restoration” are combined into a single restoration type.

[0024] In summary, the present invention has at least one of the following beneficial technical effects:

[0025] Compared with traditional forest degradation monitoring algorithms, this method can acquire degradation process information with high accuracy in identifying degradation processes. By expanding the simulated degradation process dataset through generative adversarial networks combined with self-supervised learning and training a hybrid model integrating CNN time series classification, random forest time series classification, and BOSSVS time series classification, it can effectively capture complex degradation-restoration time series features, achieving refined identification of forest degradation processes. By fitting the original NBR time series with the LandTrendr algorithm, high-frequency noise is effectively smoothed while retaining key change points, improving the robustness of degradation process identification. By reclassifying the degradation process results, the degradation intensity of degraded areas is extracted, providing a scientific basis for regional-scale forest health assessment and ecological restoration. This method relies on Landsat historical image sequences and can conduct degradation monitoring on a large scale and over long time series, making it particularly suitable for natural forest areas with frequent disturbances, complex degradation processes, and a lack of sample plot data, thus demonstrating high applicability. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention.

[0027] Figure 2 This is a schematic diagram illustrating the principle of LandTrendr fitting time series in this invention.

[0028] Figure 3 This is a two-dimensional visualization comparison chart of real and synthetic time series data in this invention;

[0029] Figure 4 This is a schematic diagram of the dataset simulating the degradation process of the present invention;

[0030] Figure 5 This is an accuracy diagram of the degradation process classification model of the present invention;

[0031] Figure 6 This is a classification diagram of the degradation process of the present invention;

[0032] Figure 7 This is a classification diagram of the degree of degradation of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] This embodiment provides a method for identifying forest degradation processes and classifying degradation levels based on time series data. Figure 1 As shown, it includes the following steps:

[0035] S1: Obtain Landsat surface reflectance data from the GEE platform, and calculate the NBR index after preprocessing;

[0036] S2: Fit the NBR time series using the LandTrendr algorithm, and extract the fitted NBR time series data from the sampling locations based on random sampling and visual interpretation;

[0037] S3: Generative adversarial networks combined with self-supervised learning techniques are used to expand the NBR time series data fitted at sample sites into a simulated dataset containing 7 types of degradation recovery.

[0038] S4: Train CNN, Random Forest and BOSSVS time series classification models, and build a hybrid classification model through an ensemble voting strategy;

[0039] S5: Apply a hybrid classification model to identify forest degradation processes and generate degradation degree maps through reclassification.

[0040] The following is a detailed description with reference to specific embodiments.

[0041] S1: Obtain Landsat surface reflectance data from the GEE platform, and calculate the NBR index after preprocessing.

[0042] 1. Obtain Landsat surface reflectance data.

[0043] First, all available LandsatCollection 2Tier1 surface reflectance imagery data from the Google Earth Engine (GEE) platform are retrieved using image asset IDs. This dataset contains imagery collected by three main sensors: Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI).

[0044] The `ee.Filter` function in GEE was used to filter out all available imagery covering the study area between 1995 and 2022, ensuring the continuity and completeness of the time series. While the remote sensing data used in the study area covers the entire Greater Khingan Mountains region, detailed analysis and validation were primarily concentrated within the Huzhong Forestry Bureau and Huzhong Nature Reserve. This design ensured both the representativeness of the study and the depth and availability of the data.

[0045] 2. Preprocessing of surface reflectance data

[0046] The CFmask algorithm was used to mask clouds, shadows, and snow in the selected images, effectively eliminating atmospheric interference. Then, the ee.mosaic and ee.clip functions from the GEE platform were used to precisely crop and seamlessly stitch the processed images according to the study area boundaries, forming a complete image dataset for the study area.

[0047] 3. NBR index calculation.

[0048] Compared with other vegetation indices (such as NDVI), NBR has higher sensitivity to chlorophyll content in vegetation, and has significant advantages such as low noise interference, high detection sensitivity and low error rate, making it particularly suitable for monitoring and identifying forest disturbance and degradation processes.

[0049] Normalized Burn Ratio (NBR) was chosen as the core indicator for forest disturbance detection. The NBR index is calculated using the reflectance in the near-infrared (NIR) and short-wave infrared (SWIR) bands, and its formula is as follows:

[0050]

[0051] NIR stands for near-infrared band, which is Band 4 in Landsat TM / ETM+ imagery and Band 5 in OLI; SWIR stands for short-wave infrared band, which is Band 7 in Landsat imagery.

[0052] This data processing method provides a high-quality data foundation for subsequent time series analysis and identification of forest degradation processes.

[0053] S2: Fit the NBR time series using the LandTrendr algorithm, and extract the fitted NBR time series data from the sampling locations based on random sampling and visual interpretation.

[0054] 1. Fit the NBR time series using the LandTrendr algorithm, such as... Figure 2 As shown.

[0055] (1) Optimize the parameters of the LandTrendr algorithm.

[0056] LandTrendr is an algorithm specifically designed for analyzing Landsat time-series data, effectively detecting disturbances and recovery trends in land cover change processes. This algorithm uses piecewise linear regression to fit remote sensing indices over long time series, thereby identifying key time points and magnitudes of land cover change.

[0057] The LandTrendr algorithm application programming interface (API) built into the Google Earth Engine (GEE) cloud platform was used to call its runLT function to perform segmented fitting processing on the original NBR time series images of the study area. This processing mainly includes the following key steps:

[0058] First, the algorithm performs piecewise linear fitting on the NBR time series of each pixel, identifying potential surface change events by automatically detecting abrupt change points (vertices) in the sequence. This piecewise regression method can effectively capture potential change processes in the time series while eliminating random noise caused by factors such as clouds and atmospheric conditions.

[0059] To ensure the accuracy and stability of the fitting results, the parameters of the LandTrendr algorithm were optimized:

[0060] maxSegments (maximum number of segments): 6, which limits the time series to a maximum of 6 segments to avoid overfitting;

[0061] spikeThreshold: 0.9, used to identify and remove outliers;

[0062] vertexCountOvershoot: 3, controls the sensitivity of vertex detection;

[0063] recoveryThreshold: 0.25, defines the minimum amplitude of the recovery process;

[0064] minObservationsNeeded (minimum number of observations): 6, ensuring that each line segment has enough observation data to support it;

[0065] preventOneYearRecovery: true, avoids misinterpreting short-term changes as part of the recovery process;

[0066] pvalThreshold (p-value threshold): 0.05, the criterion for judging statistical significance;

[0067] bestModelProportion: 0.75, the criterion for selecting the best-fitting model;

[0068] By carefully tuning these parameters, we can ensure the reliability and stability of the fitting results while maintaining the sensitivity of the algorithm.

[0069] (2) Extract and process time series data.

[0070] The band selection function ee.Image.select() is used to extract the fitted-to-vertex image data from the output. This step obtains the NBR fitting trajectory over continuous time periods within the study area, clearly showing the variation pattern of each pixel in the time dimension.

[0071] The extracted fitted time series data are saved in TIFF format, which offers good compatibility and data integrity, facilitating subsequent analysis and processing. The fitted trajectory of each pixel contains rich information, reflecting not only the spatiotemporal variation characteristics of vegetation but also recording the occurrence time of disturbance events and the dynamic changes in the recovery process.

[0072] 2. Establish and implement a random sampling strategy.

[0073] Based on the ArcGIS software platform, a systematic random sampling was conducted on the study area. The sampling process specifically included the following steps:

[0074] Two hundred sampling points were generated within the study area using the "Create Random Points" tool. To ensure the spatial uniformity and representativeness of the sampling points, a minimum spacing of 500 meters was set between them. This spacing was determined based on the following considerations: firstly, to ensure that the sampling points cover different ecological types and topographical conditions within the study area; and secondly, to avoid the impact of spatial autocorrelation between sampling points on the analysis results.

[0075] The spatial distribution strategy for sampling points employs a systematic random sampling method, which combines the comprehensiveness of regular grids with the unbiasedness of random sampling. First, the study area is divided into several 500×500 meter grids, and then a sampling point is randomly generated within each grid. This method ensures both the spatial representativeness of the sampling points and satisfies the independence requirement of statistical analysis.

[0076] 3. Perform visual interpretation and type determination.

[0077] Based on the acquired sampling points, Landsat true-color composite surface reflectance imagery and Sentinel hyperspectral imagery data were further integrated to perform detailed visual interpretation and analysis of each sampling point. This process employed multi-temporal and multi-source remote sensing data comparative analysis to ensure the accuracy of type determination.

[0078] During visual interpretation, the focus is on the spectral characteristic changes of each sampling point over time, combined with spatial context information for comprehensive judgment. Interpreters undergo professional training and are familiar with the spectral characteristics and spatiotemporal manifestations of various vegetation degradation and recovery patterns, enabling them to accurately identify different types of vegetation change processes.

[0079] The system of degradation and restoration types is established based on long-term observation and theoretical foundation. A comprehensive classification system of degradation and restoration types has been established, specifically including the following seven types:

[0080] Degradation-Complete Recovery Type: This refers to vegetation that, after being degraded by natural or human disturbances, undergoes natural or artificial restoration measures and ultimately fully recovers to its pre-disturbance state and ecological functions. This type is characterized by a NBR value that first decreases, then increases, and finally returns to its original level.

[0081] Degraded-partially-recovered type: This refers to vegetation that shows a certain recovery trend after degradation, but fails to fully recover to its pre-disturbance state due to various limiting factors. This type is characterized in NBR time series as an initial decrease followed by an increase, but the final value remains below the original level.

[0082] Degraded-Non-Recovery Type: This refers to a situation where vegetation has not been effectively restored after degradation, remaining in a degraded state or continuing to worsen. This type is characterized by a decline in the NBR value without a significant upward trend, remaining at a low level.

[0083] Post-recovery degradation type: This specifically refers to special patterns where the land was bare before the study period, initially showed a trend of vegetation recovery during the study period, but subsequently suffered disturbance and degraded again. This type reflects the fragility of the ecosystem and its tendency to be repeatedly disturbed;

[0084] The study begins with bare land and then transforms into forest: This describes a situation where the land was initially bare or had very low vegetation cover, and subsequently gradually developed into stable forest through natural succession or afforestation. This type reflects the positive effects of vegetation succession and ecological restoration.

[0085] Stable vegetation and bare land types: These include vegetation and bare land areas that remained relatively stable throughout the study period. The NBR values ​​of these areas oscillated slightly around the vegetation threshold without showing a clear trend, reflecting the stability of the ecosystem.

[0086] 4. Extracting fitted NBR time series data from sampling sites

[0087] After classification, LandTrendr-fitted NBR time series data from the sample sites were extracted to establish a sample database. Each sample contains the following information: spatial coordinates, its corresponding degradation recovery type, complete NBR time series data, and visual interpretation confidence ratings to aid in verification (such as field verification photos). The sample database is organized in a structured manner to facilitate subsequent model training and validation. Simultaneously, sample metadata records were established, detailing the interpretation process, quality control information, and relevant notes for each sample to ensure data traceability and transparency.

[0088] This systematic sampling and analysis process can yield high-quality sample datasets. This method not only ensures the scientific validity and representativeness of the samples, but also provides a technical route and methodological system that can be referenced for similar studies.

[0089] S3: Generative adversarial networks (GANs) combined with self-supervised learning techniques are used to expand the NBR time-series data fitted to sample sites into a simulated dataset containing seven types of degradation recovery, such as... Figure 4 As shown.

[0090] By employing Generative Adversarial Networks (GANs) combined with self-supervised learning techniques, LandTrendr-fitted NBR time series data from sample sites are efficiently augmented to construct a simulated dataset containing seven types of degradation processes, which is then divided into training and validation sets.

[0091] 1. Data Preprocessing and Preparation

[0092] First, read the fitted NBR time series data of the sampling points and convert it into a three-dimensional tensor format suitable for deep learning model processing, with the shape (N, seq_len, feature_dim), where N represents the number of samples, seq_len represents the time series length, and feature_dim is the feature dimension.

[0093] To ensure the stability and convergence efficiency of model training, the input data is standardized preprocessed to normalize the NBR values ​​to the [0,1] interval. The normalization formula is as follows:

[0094]

[0095] Where X min and Xmax These represent the minimum and maximum NBR values ​​in the training set, respectively. The processed data is converted to PyTorch tensor format to fully utilize the parallel computing power of the GPU to accelerate the training process.

[0096] 2. Construct the encoder structure

[0097] An autoencoder structure based on a recurrent neural network consists of two parts: an encoder (Embedder) and a decoder (Recovery). The encoder is constructed using gated recurrent units (GRUs) and is responsible for mapping the input time series to a low-dimensional latent space. Its mathematical representation is as follows:

[0098] H = GRU(X) norm )

[0099] Where H represents the latent space.

[0100] This process can effectively capture long-term dependencies and dynamic change characteristics in NBR time series.

[0101] The decoder is responsible for reconstructing the latent space representation back into the original sequence space. Through a series of upsampling or inverse transformation operations, it restores the shape and key information of the input sequence as accurately as possible. The decoder receives the latent representation H and attempts to reconstruct the original input sequence:

[0102]

[0103] By minimizing the reconstruction error, i.e., the mean square error The autoencoder learns key feature representations of the input data. This design enables the model to capture long-term dependencies and dynamic change patterns in NBR time series.

[0104] The training objective of an autoencoder is to minimize the reconstruction error, making the difference between the decoder output sequence and the original input sequence as small as possible. This optimization process forces the autoencoder to learn the most essential feature representations of the NBR time series in the latent space, while filtering out noise and redundant information. For time series data, the autoencoder structure of a recurrent neural network is particularly adept at capturing its temporal dependencies, and can effectively model the dynamic process of vegetation indices evolving over time.

[0105] Through this unsupervised learning approach, autoencoders can extract representative features from limited NBR time-series data, providing high-quality input features for subsequent forest degradation process identification and classification tasks. Latent space representation not only achieves data dimensionality reduction but also enhances the model's robustness to noise and outliers, improving the overall performance of the forest degradation monitoring system.

[0106] 3. Establish a self-supervised learning module.

[0107] A self-supervised module is introduced into the latent space to learn the dynamic dependencies of the time series by predicting the changes of the latent vectors in adjacent time steps. The training objective of this module is to make the predicted sequence temporally consistent with the true latent sequence. Specifically, for the latent representation h... t The self-supervised module predicts the representation h of the next time step. t+1 The objective function is:

[0108]

[0109] This self-supervised task does not rely on manual annotation, but uses the inherent temporal structure of the data as a supervision signal, enabling the model to learn the continuous evolution of forest degradation processes.

[0110] 4. Generator design and discriminator optimization.

[0111] The core architecture of Generative Adversarial Networks (GANs) consists of two competing neural networks: a generator and a discriminator. The generator learns the distribution of real data to generate indistinguishable data; the discriminator distinguishes whether the input data is real or fabricated by the generator.

[0112] (1) Generator Design

[0113] The generator receives a random noise vector z and a class condition c (degradation recovery type), and generates a simulated temporal representation in the latent space. The generator employs a multilayer perceptron structure, first generating an initial sequence, and then refining the temporal dynamics through a recurrent neural network. The generation process can be represented as:

[0114] H gen =G(z,c)

[0115] The category condition c is injected into the generation process in the form of an embedding vector to ensure that the generated sequence has a clear category attribute.

[0116] (2) Discriminator optimization

[0117] The discriminator is responsible for distinguishing between genuine latent representations and spurious representations generated by the generator. This discriminator combines a convolutional neural network and an attention mechanism to effectively capture both local and global features in the time series. The output of the discriminator is a probability estimate of the sequence's authenticity.

[0118] D(H) = P(H is the true sequence)

[0119] During training, the generator attempts to create more realistic data to deceive the discriminator, while the discriminator continuously improves its discriminative ability, forming a dynamic game equilibrium. The generator and discriminator are continuously optimized in adversarial training, thereby achieving the generation of high-fidelity time-series data with categorical attributes.

[0120] 5. Establish a multi-objective loss function.

[0121] Three loss functions are used to jointly train the encoding, decoding, generation, and discrimination processes. The model training employs a multi-objective loss function to balance the optimization objectives of different tasks. The overall loss function is expressed as:

[0122] L=λ E L E +λ S L S +λ U L U

[0123] Among them, L E For reconstruction loss, L S To monitor the loss, L U To combat losses, λ E , λ S and λ U For weight hyperparameters.

[0124] Combat loss L U It consists of both generator loss and discriminator loss:

[0125]

[0126] In the formula: D(x) represents the discriminator network, D(G(z)) is the generator network, and p data(x) p represents the true data distribution. z (z) is the prior distribution of noise.

[0127] 6. Simulated time series data generation and verification.

[0128] After model training is complete, the system invokes a dedicated data generation module to generate a specified number of simulated time-series samples for each of the seven degradation process labels. In this embodiment, 500 time-series samples are generated for each degradation process, totaling 3500 samples, ensuring a balanced sample size across all categories. Results show that the generated synthetic data maintains the statistical characteristics of the original data while possessing sufficient diversity, such as... Figure 3 As shown.

[0129] After generating the original sequences using the `generate_synthetic` function, the `squeeze` function is used to adjust the data shape, ensuring that each time series maintains a consistent one-dimensional vector format. Then, the data from each category are concatenated sequentially to construct a complete simulated sample set.

[0130] A category label field is appended to the end of the sample set to form structured labeled tabular data. Finally, the `pandas.to_csv` function is called to output the data as a CSV file, where the first 20 columns are time series feature data and the last column is the category label for the degradation process. This standardized output format ensures that the generated dataset can be directly used as training and validation data for downstream classification models, realizing a complete closed-loop process from time series modeling to data augmentation. The simulated time series is divided into an 80% training set and a 20% validation set. The training set is used to train a hybrid model, and the validation set is used to validate the classification model's accuracy.

[0131] S4: Train CNN, Random Forest and BOSSVS time series classification models, and build a hybrid classification model through an ensemble voting strategy.

[0132] 1. Build and train a CNN time series classification model.

[0133] A CNN time series classification model is constructed using the CNNClassifier module from the sktime library. This classifier consists of multiple functional modules: the input layer dimension is determined based on the time series features; the feature extraction part includes two one-dimensional convolutional layers and two average pooling layers; and the output layer is a fully connected layer that uses the softmax activation function to output the class probability. This end-to-end deep learning architecture can directly learn feature representations from raw time series data and complete the classification task.

[0134] (1) Input layer design

[0135] The input layer fully considers the characteristics of time series data, employing N×k neurons to form the network entry point. Here, k represents the number of variables in the input time series, specifically the NBR index in this study; N represents the time step length corresponding to each variable, i.e., the time span within the study period, corresponding to the time series length between 1995 and 2022. This design ensures that the network can process sequence data with different time spans while maintaining the integrity of the temporal information.

[0136] (2) Feature extraction mechanism

[0137] The feature extraction section employs two one-dimensional convolutional layers, a design particularly well-suited for processing time-series data. One-dimensional convolutional layers effectively capture local patterns and temporal dependencies within the time series. The kernel size is set to 7, a carefully tuned choice that captures sufficiently long temporal patterns without introducing excessive parameter overhead.

[0138] The first convolutional layer uses 6 filters, and the second layer uses 12 filters. This progressively increasing number of filters aligns with the hierarchical principle of feature extraction in deep learning. Shallow convolutions capture basic features, while deeper convolutions combine these features to form higher-level abstract representations.

[0139] The pooling layer configuration consists of an average pooling layer following each convolutional layer, with a pooling window size of 3. Average pooling reduces the dimensionality of the feature maps while preserving the most important feature information. This design effectively controls model complexity, prevents overfitting, and improves the model's translation invariance.

[0140] (3) Feature fusion strategy

[0141] After multiple convolution and pooling operations, the original time series is transformed into a series of feature maps with rich semantic information. Subsequently, the network intelligently concatenates these feature maps along the channel dimension to generate a new comprehensive feature vector sequence. This feature vector sequence serves as the final abstract representation of the original input, preserving the key discriminative information of the time series while eliminating redundant noise.

[0142] (4) Output layer and classification decision

[0143] The output layer adopts a fully connected layer structure, and the number of its neurons is strictly equal to the set number of time series classification categories, n. In this study, n=7, corresponding to 7 types of degradation processes. This layer receives the high-level feature representation output by the feature extraction layer and maps it to the category space.

[0144] The output is probabilistically converted using the softmax activation function, transforming the neurons' raw output values ​​into a class probability distribution. The softmax function ensures that the sum of all output probabilities is 1, with the probability value for each class ranging from 0 to 1, providing an intuitive measure of classification confidence.

[0145] The classification decision mechanism ultimately selects the class with the highest probability value from the output neurons. This maximum probability selection strategy is simple and effective, and the predicted probability distribution of each sample also provides important information for subsequent uncertainty analysis.

[0146] (5) Network training strategy

[0147] The training parameters are configured using a carefully optimized training strategy:

[0148] The training epochs are set to 20,000 to ensure that the model converges fully.

[0149] With a batch size of 16, a good balance is achieved between computational efficiency and gradient stability;

[0150] The Adam optimizer was chosen, which combines the advantages of momentum and RMSProp and can automatically adjust the learning rate.

[0151] The learning rate was set to 0.01. This value has been experimentally verified to ensure convergence speed and model stability.

[0152] The number of filters in the convolutional layer is [6, 12];

[0153] The padding strategy, with the padding mode set to "auto," enables intelligent padding selection: when the input sequence length is less than 60, "same" padding is used to maintain the feature map length; otherwise, "valid" padding is used to avoid noise interference from edge effects. This adaptive padding strategy ensures effective processing of time series of different lengths.

[0154] For loss function selection, the cross-entropy function is chosen, as it is particularly suitable for multi-class classification problems. Cross-entropy loss effectively measures the difference between the predicted probability distribution and the true distribution, exhibits robustness to class imbalance problems, and fully meets the requirements of multi-class output in the degradation process.

[0155] Input data requirements: The classification input consists of multi-dimensional, equal-length NBR time series. All input sequences must be normalized and maintain the same time length. The dimensions of the input data are [batch_size, sequence_length, num_features], where num_features is 1 (NBR exponent) in this study.

[0156] The output format is a degradation process type label, represented by seven categories using one-hot encoding. The network output is the probability of each sample belonging to each category, and the final classification result is used to obtain the specific category label through an argmax operation.

[0157] 2. Construct and train a random forest time series classification model.

[0158] (1) Construction of random forest time series classification model.

[0159] TimeSeriesForestClassifier is a classifier specifically designed for time series data based on the random forest algorithm. Its core idea combines global feature extraction of time series with ensemble learning. This classifier constructs multiple decision trees, each focusing on different random sub-intervals of the time series, thereby comprehensively capturing both local and global feature patterns of the time series.

[0160] The basic architecture of this algorithm follows the ensemble learning framework of random forests, but it has been specifically optimized for the characteristics of time series data. Unlike traditional random forests that process independent and identically distributed tabular data, TimeSeriesForestClassifier fully considers the temporal dependencies and pattern continuity of time series data, making it perform excellently in time series classification tasks such as forest degradation process identification.

[0161] Feature extraction process:

[0162] Random subsequence sampling mechanism. The algorithm first randomly samples multiple subsequences from the original time series of length m. This random sampling strategy ensures that each decision tree can focus on different local features of the time series, increasing the diversity and robustness of the model. The sampling process follows the principle of uniform distribution, ensuring that each segment of the time series has a chance to be sampled.

[0163] Statistical feature calculation. For each sampled subsequence, the algorithm extracts three core statistical features: the mean reflects the average level of the sequence, the standard deviation characterizes the degree of fluctuation of the sequence, and the slope represents the trend of the sequence. These three features together constitute a complete index system describing the local features of the time series. The feature extraction process uses a sliding window technique to ensure that pattern features at different time scales can be captured. This multi-scale feature extraction strategy enables the model to simultaneously identify short-term fluctuations and long-term trends, which is particularly important for complex processes such as forest degradation.

[0164] Feature vector construction. The three statistical features of each subsequence are concatenated sequentially to form a 3D feature vector. This process converts variable-length time series into fixed-length feature representations, laying the foundation for subsequent decision tree modeling. The fixed-dimensionality processing of the feature vectors solves the technical challenge caused by inconsistent time series lengths.

[0165] Decision tree construction and integration strategies:

[0166] Decision tree growth mechanism. Based on the feature vector dataset, the information gain criterion ("entropy") is used as the node splitting standard. Information gain effectively measures the degree to which features improve class purity, ensuring that each split maximizes classification accuracy. No depth limit is imposed during decision tree growth, allowing the model to adaptively determine the tree structure based on data complexity.

[0167] An ensemble learning framework is used to construct a time-series forest by ensembled multiple decision trees, each trained independently based on different random subsequences and feature subsets. This diversity-enhancing mechanism ensures that the ensemble model can cover various potential patterns in the time series, improving the model's generalization ability.

[0168] The final classification result employs a probability averaging strategy, combining the predictions from all decision trees to output the final category. This soft voting mechanism not only provides category labels but also classification confidence scores, offering more reference information for subsequent decisions.

[0169] (2) Model training parameter configuration and optimization.

[0170] Core parameter settings:

[0171] Number of decision trees (n_estimators = 100): Experiments have shown that 100 decision trees can achieve the best balance between model complexity and computational efficiency, ensuring that the model has sufficient expressive power while avoiding overfitting.

[0172] Node partitioning criterion ("entropy"): Information entropy is chosen instead of the Gini coefficient as the partitioning criterion because information entropy is more sensitive to probability distribution and can better handle class imbalance problems, making it suitable for tasks such as forest degradation where class distribution may be uneven.

[0173] Maximum tree depth (max_depth = None): There is no limit to the maximum tree depth, allowing the decision tree to grow freely according to the data features, fully capturing complex patterns in the data, while relying on the ensemble properties of random forests to prevent overfitting.

[0174] Regularization and stability configuration:

[0175] Sample splitting constraints: The settings of minimum number of samples for node splits (min_samples_split=2) and minimum number of samples for leaf nodes (min_samples_leaf=1) maintain the maximum flexibility of the model and are suitable for medium-sized datasets.

[0176] Feature subset selection strategy: The square root method (max_features = "sqrt") is used to determine the number of features used in each tree. This setting ensures feature diversity and avoids the curse of dimensionality, and is a classic configuration for the random forest algorithm.

[0177] Repeatability and efficiency optimization:

[0178] Bootstrap (True): Enables bootstrap sampling, where each tree uses a different training subset, increasing model diversity and providing out-of-bag (OOB) error estimation.

[0179] Random seed (random_state=42): Fixing the random seed ensures the reproducibility of the experiment and facilitates model tuning and result verification.

[0180] Parallel processing (n_jobs = -1): Utilizes multi-core processors to build decision trees in parallel, significantly improving training efficiency, and is particularly suitable for processing large-scale time series datasets.

[0181] In the task of identifying forest degradation processes, this classifier can fully utilize the temporal characteristics of NBR time series to accurately distinguish different types of degradation patterns, providing reliable technical support for forest ecosystem monitoring. A properly trained model not only achieves high classification accuracy but also possesses strong generalization ability, making it applicable to degradation monitoring tasks in different geographical regions and forest types.

[0182] 3. Construct and train the BOSSVS time series classification model

[0183] BOSSVS (Bag-of-SFA-Symbols in Vector Space) is a dictionary-based time series classification method that converts time series into symbolic representations and uses a vector space model for classification. Its core advantage lies in its ability to effectively handle local patterns in time series data while exhibiting strong robustness to noise.

[0184] (1) Method Implementation Steps

[0185] Time series segmentation and preprocessing. First, the original time series is segmented using a fixed-length sliding window. The window size (window_size) is set to 10, and the sliding step (window_step) is set to 1 to ensure that local features of the time series are captured without missing important information. Then, a first-order differencing operation is performed on each subsequence, calculating the difference between adjacent time points. This step effectively reduces the influence of trends in the series, improves the stationarity of the data, and lays the foundation for subsequent feature extraction.

[0186] Symbolic Feature Extraction. The Symbolic Fourier Approximation (SFA) method is used to transform the processed subsequences. This process first performs a Fourier transform on each subsequence, converting the time-domain signal into a frequency-domain representation, retaining the most important Fourier coefficients. Then, the continuous coefficient values ​​are discretized into a finite set of symbols. In this study, the alphabet size (n_bins) is set to 4, dividing the numerical space into four discrete intervals, each corresponding to a specific symbol. This discretization process maps the continuous value sequence to a finite symbol sequence, achieving dimensionality reduction and noise suppression.

[0187] Bag-of-Words Model Construction. The symbol sequences generated from all subsequences are combined to form a bag-of-words structure. The bag-of-words model ignores the order of symbols and focuses on counting the frequency of different symbol combinations. By counting the number of times each "word" (i.e., a specific symbol sequence) appears in the time series, a frequency histogram is constructed, transforming the original time series into a sparse vector representation. This representation captures the local pattern features of the time series while reducing data dimensionality.

[0188] Vector Space Classification. Classification decisions are made based on the Vector Space Model. First, a TF-IDF (Term Frequency-Inverse Document Frequency) weighting mechanism is used to weight the features in the bag-of-words, increasing the weight of highly discriminative features and decreasing the weight of common features. Then, the cosine similarity between the sample to be classified and the reference templates of each category is calculated to measure their directional consistency in the high-dimensional vector space. Finally, the sample is assigned to the category with the highest similarity, completing the classification task.

[0189] (2) Parameter configuration and training

[0190] During model training, a series of optimized parameter configurations were adopted: the alphabet size was set to 4 to balance representational power and computational complexity; the sliding window parameters were kept at the default settings (window size 10, sliding step size 1) to ensure complete capture of local features; the feature standardization flags (norm_mean and norm_std) were set to False to preserve the distribution characteristics of the original data; and the IDF weighting mechanism was enabled (use_idf=True) to enhance the model's ability to identify key features.

[0191] This method provides an efficient and reliable technical means for identifying forest degradation processes, accurately capturing subtle patterns of vegetation change from long-term data, and providing strong support for ecosystem monitoring and assessment.

[0192] 4. Model Integration Prediction and Result Output

[0193] (1) Model integration prediction

[0194] In the model ensemble prediction phase, three independently trained time series classification models are first loaded in parallel: a random forest time series classifier, a CNN time series classifier, and a BOSSVS time series classifier. A parallel prediction strategy is employed, performing inference on each sample in the test set using all three models simultaneously to obtain the corresponding predicted label. This multi-model collaborative mechanism fully leverages the complementary advantages of different algorithms: random forest excels at capturing global statistical features, CNN effectively extracts local temporal patterns, and BOSSVS focuses on symbolically represented sequence features.

[0195] A mode-based hard-voting ensemble strategy is employed to synthesize the predictions of the three models. Specifically, for each sample to be classified, the category label that appears most frequently in the predictions of the three models is taken as the final classification result. In the event of a tie (i.e., the predictions of the three models are different), the prediction category with the highest confidence level is selected first. This ensemble method effectively reduces the random error and overfitting risk of a single model, significantly improving the stability and reliability of the classification results.

[0196] (2) Model accuracy verification

[0197] Validation set construction and evaluation metrics

[0198] A systematic accuracy evaluation of the ensemble model is performed using a pre-defined validation set. The validation set contains sample data with real labels to ensure the objectivity and statistical significance of the evaluation results. A multi-dimensional evaluation index system is employed, including:

[0199] Precision measures the accuracy of a model's judgment of positive samples. The formula for calculation is:

[0200]

[0201] Recall evaluates a model's ability to identify positive samples, and is calculated using the following formula:

[0202]

[0203] Overall accuracy comprehensively reflects the model's overall classification performance, and is calculated using the following formula:

[0204]

[0205] Wherein, TP (True Positive) is the number of instances correctly classified as positive by the model, FP (False Positive) is the number of instances incorrectly classified as positive by the model, TN (True Negative) is the number of instances correctly classified as negative by the model, and FN (False Negative) is the number of instances incorrectly classified as negative by the model.

[0206] (3) Spatial data preprocessing workflow

[0207] The NBR time series multi-band image, fitted by the LandTrendr algorithm, is read. Its original data dimensions are [number of bands, number of rows, number of columns]. First, the data is transposed to convert the dimensions to [number of rows, number of columns, number of bands], ensuring that each pixel corresponds to a complete time series. Then, the 3D image data is reconstructed into a 2D array with the shape [total number of pixels, time series length], preparing for subsequent batch classification processing.

[0208] The reconstructed time-series data was normalized using sklearn.preprocessing.MinMaxScaler. Min-Max scaling was applied independently to each column (i.e., the NBR values ​​of all pixels at a specific time point), uniformly transforming the numerical range to the [0,1] interval. This time-point normalization method effectively eliminated amplitude differences caused by seasonality and interannual variations, ensuring the consistency and comparability of the classifier's input data.

[0209] (4) Generation and output of classification results

[0210] Spatial classification implementation. The preprocessed LandTrendr-fitted NBR time series data is used to perform full-map classification prediction using an ensemble prediction function, such as... Figure 5 As shown, based on the time-series trajectory features of each pixel, three base classifiers are used independently to predict and generate preliminary classification results. Then, a voting mechanism is applied to integrate the predicted labels from the three models to generate a final classification result with spatial consistency.

[0211] Output and Formatting: The one-dimensional classification prediction results are reconstructed into a two-dimensional classification map according to the spatial dimensions of the original imagery, maintaining the same spatial reference system and geometric accuracy as the input data. The final output is in GeoTIFF format, which fully preserves geographic coordinate information, projection parameters, and metadata, facilitating further spatial analysis and visualization in GIS software. The output file contains complete spatial reference information, ensuring that the results can be directly applied to practical forest management and ecological assessment work.

[0212] This systematic integrated prediction and output process ensures the technical reliability and practical application value of the forest degradation process classification results, and provides a scientific basis for ecosystem monitoring.

[0213] S5: Apply a hybrid classification model to identify forest degradation processes and generate degradation degree maps through reclassification. Figure 6 and Figure 7 ).

[0214] 1. Classification of the degree of degradation.

[0215] The rasterio spatial data processing library in Python was used to read the forest degradation process classification map generated in the previous steps. This classification map contains seven detailed types of degradation and restoration processes, providing basic data for classifying the degree of degradation. The rasterio library can efficiently process geospatial raster data while preserving the original geographic coordinate system and spatial reference information.

[0216] Develop scientific reclassification rules to summarize detailed degradation process types into more practically meaningful degradation degree categories:

[0217] Severe degradation type: This integrates the "degradation-no-recovery type" and the "post-recovery degradation type". Both types are characterized by the failure of vegetation to recover effectively or the occurrence of secondary degradation after being disturbed, reflecting the continuous degradation state of the ecosystem.

[0218] Mild degradation type: "Partial recovery after degradation" is classified separately. This type is characterized by vegetation showing some recovery after being disturbed, but not fully recovering to its original state, and belongs to the moderate degree of degradation.

[0219] Restoration type: The study combined the "bare land at the start of the study period that turned into forest" type and the "completely restored after degradation" type. Both types showed a positive vegetation restoration trend, indicating that the ecosystem is in good condition or is improving.

[0220] 2. Generate a degradation level map through reclassification.

[0221] The reclassification function is called to batch process the original classification map, reclassifying all pixels into the three degradation level categories according to the rules described above. The final forest degradation level classification map is then generated. Figure 7 This figure shows the distribution of degradation levels at different locations within the study area in a spatially explicit manner.

[0222] Finally, the above methods were experimentally verified. The study focused on the Greater Khingan Mountains region of Northeast China, specifically the Huzhong Forestry Bureau and Huzhong Nature Reserve. This region is a typical cold-temperate coniferous forest ecosystem, sensitive to climate change and human disturbance, making it an ideal area for studying forest degradation. A trained degradation and restoration process classification model was applied to the Huzhong region, generating a classification map of forest degradation and restoration processes. Based on this map, degradation severity classification rules were further applied to generate a map of forest degradation severity in the Huzhong region. These resulting maps visually demonstrate the spatial distribution pattern and severity of forest degradation within the study area, providing a scientific basis for decision-making by local forest management departments. By identifying severely degraded areas, ecological restoration measures can be prioritized; while well-restored areas can serve as models for ecological protection. The technical process established in this study is not only applicable to the Greater Khingan Mountains region but can also be extended to the degradation monitoring and assessment of other forest ecosystems, providing effective technical support for global forest protection and ecological restoration.

[0223] In summary, this invention achieves efficient and accurate identification of large-scale forest degradation processes through a systematic methodology. First, by fitting the NBR time series generated from Landsat surface reflectance data using the LandTrendr algorithm, the continuity and stability of the time series are significantly improved, effectively reducing interference from clouds, shadows, and missing data, thus greatly enhancing the timeliness and accuracy of interference-recovery process extraction. Second, sample data is obtained through random sampling and visual interpretation, and a generative adversarial network combined with self-supervised learning technology is innovatively employed to expand the limited sample data into a simulated dataset containing seven types of degradation processes. An ensemble learning framework based on CNN, BOSSVS, and random forest time series classifiers effectively enhances the model's ability to identify diverse interference-recovery patterns, significantly improving the model's generalization performance while maintaining high classification accuracy. Finally, by scientifically reclassifying the degradation process classification results, spatial distribution maps and degradation intensity maps of the forest degradation processes in the study area are generated, providing a reliable decision-making basis for the dynamic monitoring, scientific evaluation, and precise restoration of degraded forest ecosystems. This technical solution is both innovative and practical, and has good prospects for widespread application.

[0224] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. They can also be reasonable combinations of the features described in the above embodiments. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A time series-based forest degradation process identification and degradation degree classification method, characterized in that, The method comprises the following steps: S1: obtaining Landsat surface reflectance data from the GEE platform, calculating the NBR index after preprocessing; S2: fitting the NBR time series by the LandTrendr algorithm, and extracting the fitted NBR time series data of the sample points based on random sampling and visual interpretation; S3: generating a generative adversarial network and combining a self-supervised learning technique to expand the fitted NBR time series data of the sample points into a simulated data set containing seven types of degradation and recovery; S4: training CNN, random forest and BOSSVS time series classification models, and constructing a hybrid classification model through an ensemble voting strategy; S5: applying the hybrid classification model to identify forest degradation processes, and generating a degradation degree map through reclassification.

2. The method according to claim 1, wherein, The preprocessing in S1 comprises: Using the CFmask algorithm to perform cloud, shadow and snow mask processing on the screened images; Using the ee.mosaic and ee.clip functions in the GEE platform to accurately crop and seamlessly splice the processed images according to the boundaries of the study area, forming a complete image data set of the study area.

3. The method according to claim 1, characterized in that, The LandTrendr algorithm in S2 has the following parameter settings: the maximum number of segments maxSegments is 6, the peak threshold spikeThreshold is 0.9, the vertex overshoot vertexCountOvershoot is 3, the recovery threshold recoveryThreshold is 0.25, the minimum number of observations minObservationsNeeded is 6, the preventOneYearRecovery is true, the p-value threshold pvalThreshold is 0.05, and the bestModelProportion is 0.

75.

4. The method according to claim 1, wherein, The degradation and recovery types in S3 include: degradation-completely recovery, degradation-partial recovery, degradation-no recovery, recovery degradation, bare land becoming forest at the beginning of the study period, stable vegetation and bare land.

5. The method according to claim 1, wherein, The CNN in S4 has the following training parameters for the time series classification model: the training period epoch is 20000, the batch size is 16, the optimizer is Adam, the learning rate is 0.01, the number of convolution layer filters is [6, 12], the padding method is auto, and the loss function is cross-entropy.

6. The method according to claim 1, wherein, The training parameters of the random forest in S4 are as follows: the number of decision trees n_estimators is 100, the node division criterion criterion is information gain entropy, the maximum tree depth max_depth is unlimited, the minimum number of node division samples min_samples_split is 2, the minimum number of leaf node samples min_samples_leaf is 1, the feature subset selection strategy max_features is the square root method sqrt, the bootstrap sampling bootstrap is True, the random seed random_state is 42, and the parallel processing n_jobs is-1.

7. The method according to claim 1, wherein, The training parameters of the BOSSVS in S4 are as follows: the alphabet size n_bins is 4, the sliding step window_step is 1, the window size window_size is 10, the feature normalization flags norm_mean and norm_std are False, and the IDF weighting use_idf is True. 8.The method of claim 1, wherein, The ensemble voting strategy in S4 adopts a hard voting mechanism based on the mode, and the prediction results of the three models of CNN, random forest and BOSSVS are integrated.

9. The method according to claim 1, wherein, The reclassification in S5 has the following rules: "Degradation-no recovery" and "degradation after recovery" are combined into the severe degradation type; "Degradation-partial recovery" is defined as the mild degradation type; "Converted from bare land to forest land at the beginning of the study period" and "degradation-complete recovery" are combined into the recovery type.

Images