A method, system and computer program product for detecting cross-seasonal changes in forest areas

Abstract

The present invention discloses a forest area inter-seasonal change detection method, system and computer program product, the method comprising: obtaining forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data; constructing a forest area phenology and environmental disturbance detection model based on a convolutional neural network and a temporal attention mechanism; constructing a forest area vegetation type change detection model based on a convolutional neural network and a multi-task learning method; obtaining forest area phenology and environmental disturbance detection results, and fine-tuning and retraining the forest area vegetation type change detection model based on the results to obtain a forest area inter-seasonal change detection model; and obtaining forest area inter-seasonal change detection results. The technical solution provided by the present invention is based on forest area vegetation type change detection, integrates forest area phenology and environmental disturbance change information, extracts detailed change information of vegetation coverage at a fine scale, and realizes accurate extraction of fine change information of vegetation coverage in multiple periods within a year; and comprehensively evaluates and fine-tunes the network model based on multiple indicators.

Description

Technical Field

[0001] The present invention relates to the field of forest area dynamic change detection, and in particular to a forest area cross-season change detection method. Background Art

[0002] Dynamic change detection in forest areas can timely understand the quantity and quality of forest resources, grasp the laws and trends of the growth and decline of forest resources, analyze the natural, economic, and social objective conditions that affect and restrict forest growth, and establish or update forest resource archives. It is of great significance to revealing ecosystem environmental changes, vegetation restoration, and reconstruction layout.

[0003] Remote sensing technology provides a new means for detecting dynamic changes in forest areas. Compared with traditional forest resource inventories, remote sensing technology can obtain a large amount of forest image data with different spatial resolutions and multiple temporal phases, providing a rich data source for forest change detection.

[0004] Through medium-resolution imaging spectral equipment, low spatial resolution forest remote sensing time series data can be obtained. Forest remote sensing time series data has the characteristics of high temporal resolution and can be used in large-scale forest change monitoring to obtain surface coverage information at more time points. However, these data have low spatial resolution and can only monitor changes in land use types over large areas, making it difficult to monitor areas with relatively small changes.

[0005] High-resolution imaging spectroscopy equipment can be used to obtain multi-temporal high-spatial resolution remote sensing image data of forest areas. These data have higher spatial resolution and can detect more subtle changes in forest areas, but cannot detect inter-seasonal changes in forest areas well. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the current technology, the present invention provides a method for detecting cross-seasonal changes in forest areas. Based on the detection of changes in vegetation types in forest areas, it integrates the information on changes in forest phenology and environmental disturbances, extracts detailed change information on fine-scale vegetation coverage, and realizes accurate extraction of fine-scale change information on vegetation coverage in multiple periods within a year.

[0007] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0008] A method for detecting cross-seasonal changes in forest areas comprises the following steps:

[0009] Obtain forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data;

[0010] A sample data set was constructed based on remote sensing time series data. A time series full convolutional neural network model was constructed and trained based on convolutional neural network and temporal attention mechanism to obtain a forest phenology and environmental disturbance detection model.

[0011] A sample data set was constructed based on multi-temporal and high-spatial-resolution remote sensing image data. A multi-task full convolutional neural network model was constructed and trained based on the convolutional neural network model and multi-task learning to obtain a forest vegetation type change detection model.

[0012] Input remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain forest phenology and environmental disturbance detection results;

[0013] According to the results of forest phenology and environmental disturbance detection, the forest vegetation type change detection model is fine-tuned and retrained to obtain the forest cross-season change detection model;

[0014] The multi-temporal and high-spatial-resolution remote sensing image data are input into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results.

[0015] According to one aspect of the present invention, the method for detecting cross-seasonal changes in forest areas further includes: preprocessing forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data.

[0016] According to one aspect of the present invention, the preprocessing of forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data includes: radiation calibration, atmospheric correction, cloud and shadow removal, geometric correction, and image registration.

[0017] According to one aspect of the present invention, when constructing a sample data set, data enhancement is performed on image data, and the sample data set is constructed based on a time window sliding technology.

[0018] According to one aspect of the present invention, the method of constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training the model comprises:

[0019] Using the layered deep convolutional network architecture as the spatial encoder, by setting the convolution layer, ReLU activation function and batch normalization layer step by step, combined with the stride convolution module, the feature extraction is deepened layer by layer, the spatial features of the image are extracted, and the feature map stacked along the time axis and processed by spatial encoding is obtained;

[0020] Integrate the temporal attention encoder and apply the temporal attention mechanism to assign different weights to the feature maps at each time point in the sequence, achieve efficient encoding of the dataset in the temporal dimension, and identify the temporal dynamic characteristics of the forest area;

[0021] The model is built based on U-Net, and the spatial resolution of the feature map is gradually improved by using the layered strided transposed convolution technology. The skip connection technology and attention mechanism are introduced to fuse the feature map with the temporal dynamic features to reconstruct the spatial details of the original remote sensing time series data.

[0022] According to one aspect of the present invention, the method of constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training the model further includes:

[0023] The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small objects and balance the problem of uneven categories;

[0024] Use Adam optimizer to improve training efficiency and model convergence speed;

[0025] Implement a learning rate decay strategy to achieve fast learning in the early stages of training and maintain stable convergence in the later stages;

[0026] An early stopping mechanism is used to prevent overfitting.

[0027] According to one aspect of the present invention, the method of constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training the model further includes:

[0028] Comprehensively evaluate the model based on multiple indicators on the validation set, including at least precision, recall, and F1 score;

[0029] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0030] According to one aspect of the present invention, the multi-task full convolutional neural network model is constructed based on the convolutional neural network model and multi-task learning and trained, including:

[0031] A hierarchical feature extraction network is constructed as an encoder to deeply analyze the complex spatial information contained in high spatial resolution remote sensing image data, integrate the spatiotemporal attention mechanism, integrate the change information of the time dimension, and extract comprehensive multi-scale spatiotemporal features from multi-phase high spatial resolution remote sensing image data;

[0032] A converter is constructed based on the Transformer architecture, a spatiotemporal feature integration module is designed, a self-attention mechanism is used, and a multi-task learning strategy for change detection and semantic segmentation is adopted to refine and strengthen the multi-scale spatial features output by the encoder to identify an aspect of vegetation cover change over time. The multi-task full convolutional neural network model is constructed based on the convolutional neural network model and multi-task learning and trained, and further includes:

[0033] A fusion loss function is used to integrate the cross entropy loss of semantic segmentation and the binary cross entropy loss of change detection. An adjustable weight coefficient is introduced to dynamically adjust the weight according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance.

[0034] Use the Adam optimizer with adaptive learning rate to improve training efficiency and model convergence speed;

[0035] Integrate dynamic loss weight adjustment and learning rate scheduling strategies to help the model maintain a balance in task learning between training stages and optimize the overall performance of the model;

[0036] The early stopping mechanism and model regularization measures are adopted to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process.

[0037] According to one aspect of the present invention, the constructing a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning and training it further includes:

[0038] Use multiple indicators, including at least precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance;

[0039] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0040] According to one aspect of the present invention, the method for detecting cross-seasonal changes in forest areas further comprises the following steps:

[0041] Based on the detection results of inter-seasonal changes in forest areas, a spatiotemporal distribution map of inter-seasonal changes in forest areas is constructed.

[0042] A forest area inter-seasonal change detection system, based on the forest area inter-seasonal change detection method as described above, comprises:

[0043] Data acquisition module, used to obtain forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data;

[0044] The phenological disturbance model construction module is used to construct a sample data set based on remote sensing time series data, build a time series full convolutional neural network model based on convolutional neural network and temporal attention mechanism, and train it to obtain a forest phenology and environmental disturbance detection model;

[0045] The vegetation change model building module is used to build a sample data set based on multi-temporal high-spatial resolution remote sensing image data, build a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning, and train it to obtain a forest vegetation type change detection model;

[0046] The phenological disturbance acquisition module is used to input remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain the forest phenology and environmental disturbance detection results;

[0047] The model adjustment module is used to fine-tune and retrain the forest vegetation type change detection model based on the forest phenology and environmental disturbance detection results to obtain the forest cross-season change detection model;

[0048] The detection module is used to input multi-temporal and high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results.

[0049] A computer program product comprises a computer program, wherein when the computer program is executed, the steps of the above-mentioned method for detecting cross-seasonal changes in forest areas are implemented.

[0050] Advantages of the present invention:

[0051] The present invention provides a method for detecting cross-seasonal changes in forest areas. The method utilizes change detection based on forest vegetation types, integrates forest phenology and environmental disturbance change information, extracts detailed change information on fine-scale vegetation coverage, solves the problem of difficulty in distinguishing high spatial resolution remote sensing phenology and disturbance, and makes up for the problem of poor detection effect of fine-scale changes in remote sensing with medium spatial resolution, thereby realizing accurate extraction of fine vegetation coverage change information for multiple periods within a year.

[0052] This method can also conduct a comprehensive evaluation of the network model based on multiple indicators, and fine-tune and improve the network model according to the evaluation results; based on the detection results of inter-seasonal changes in forest areas, the data is integrated to construct a spatiotemporal distribution map of inter-seasonal changes in vegetation types in forest areas, providing important information for environmental monitoring and decision-making analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 This is a flow chart of a method for detecting cross-seasonal changes in forest areas according to Embodiment 1 of the present invention;

[0055] Figure 2 This is a flow chart of a method for detecting cross-seasonal changes in forest areas described in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Embodiment 1

[0058] like Figure 1As shown, a method for detecting cross-seasonal changes in forest areas comprises the following steps:

[0059] S1: Acquire forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data.

[0060] (1) Remote sensing time series data of forest areas

[0061] The forest area remote sensing time series data mainly uses Gaofen (GF) high time series data for research, and multispectral remote sensing data such as Sentinel-2 (S2) and Landsat satellites as auxiliary data sources for research. Among them, Gaofen series satellites (such as GF-1 / 6) provide high temporal and spatial resolution remote sensing images, which are helpful for accurately monitoring surface changes; S2 provides 10 to 20 meters of high-resolution multispectral images, which are suitable for capturing vegetation cover changes; Landsat series provides long-term continuous observation records, which are suitable for analyzing long-term trends.

[0062] Specifically, multispectral remote sensing time series data must meet the following requirements: spatial resolution is better than 15 meters, and better than 10 meters in key monitoring areas; to ensure that multispectral remote sensing data fully reflect the changes in each stage within the key growing season and capture key phenological change points, based on data availability, data are acquired every two weeks to one month over a one-year span to form a continuous time series with consistent radiation and spatial alignment; in the selection of multispectral sensors, priority is given to using data acquired by the same sensor at different time points. If conditions limit the use of different multispectral sensors, sensors with similar band characteristics should be selected as much as possible to ensure the comparability of the obtained data; cloud cover area is less than 5%.

[0063] Different geographical features, cloud cover and surface characteristics directly affect the selection of data sources. Therefore, we will carefully select appropriate remote sensing data according to the specific conditions of the test area to ensure that the collected data can accurately reflect the actual conditions of each place and effectively respond to the challenges brought by climate factors, ensuring the accuracy of research data and the practicality of results.

[0064] (2) Multi-temporal and high spatial resolution remote sensing image data of forest areas

[0065] Acquire multi-temporal and high-spatial-resolution remote sensing image data of forest areas to ensure that the data used can effectively reflect changes in vegetation coverage and provide a basis for model construction and analysis.

[0066] The GF-2 and GF-1 PMS images are selected as the main data sources, which can provide high spatial resolution surface feature information and high accessibility. Among them, GF-2 images have high spatial resolution and can provide detailed vegetation coverage features; while GF-1 PMS images effectively supplement the problem of insufficient coverage of GF-2 in specific areas and enhance the data foundation.

[0067] The forest area cross-seasonal change detection method also includes: preprocessing the forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data.

[0068] The preprocessing of forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data includes: radiation calibration, atmospheric correction, cloud and shadow removal, geometric correction, image registration, etc.

[0069] Preprocessing aims to improve data quality and provide a solid foundation for subsequent analysis.

[0070] S2: Construct a sample data set based on remote sensing time series data, build and train a time series full convolutional neural network model based on convolutional neural network and temporal attention mechanism, and obtain a forest phenology and environmental disturbance detection model.

[0071] When constructing the sample data set, the image data is enhanced and the sample data set is constructed based on the time window sliding technology.

[0072] The sample data construction first generates a sample set reflecting the change spots based on the existing remote sensing images and spot data. The calibrated samples include the vegetation cover type before and after and the change type samples.

[0073] Using two phases of remote sensing images from the same sensor in adjacent quarters, vegetation cover remote sensing image change calibration is performed by province or region based on remote sensing image features and vegetation cover changes. Based on the vegetation cover type reflected by the attributes of fixed sample plots or forest map patches, the remote sensing image features of the location are used for comparative analysis, and representative plots with relatively consistent remote sensing image features are selected. The vegetation cover type before and after, remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to a 512m×512m square vector frame.

[0074] Based on the established remote sensing images and data labels of each vegetation cover type, combined with the remote sensing images and forest and grass resource data of the previous and next two periods, and referring to the attribute information of forest and grass resource spots, the characteristic change areas of the remote sensing images of the previous and next two periods are compared and determined, and the change type and the corresponding remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to the 512m×512m square vector frame. The specific change types include at least construction project use, logging, reclamation, disaster events, and ecological protection and restoration measures.

[0075] This step aims to establish a large-scale, diverse and spatially distributed sample database to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. Models that do not meet this accuracy standard will be inspected in detail, supplemented with labels for interpretation signs and change types, and training parameters will be adjusted and iterative training will be repeated to achieve higher accuracy.

[0076] By carefully dividing the pre-processed time series remote sensing data, we create training and validation sets that reflect the diversity of vegetation coverage and environmental changes. At the same time, we use data enhancement methods such as random rotation, flipping, and cropping to enhance the model's generalization ability for diverse surface features. The training samples constructed based on the time window sliding technology not only capture the vegetation coverage information at a single moment, but also integrate the change trend on the time axis, providing the model with dynamic change information on vegetation coverage.

[0077] The time series fully convolutional neural network model is dedicated to improving the recognition accuracy of vegetation cover changes and environmental disturbance events, and emphasizes the ability to locate disturbances in space and time. The model consists of three main parts: first, the spatial encoding module uses deep convolutional layers to extract the spatial features of the image; second, the temporal encoding module introduces an attention mechanism to identify key change moments in the time series and enhance sensitivity to temporal changes; finally, the spatial decoding module is responsible for accurately reconstructing the spatial details of the semantic segmentation results while maintaining high spatial resolution.

[0078] The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a time attention mechanism and training the model includes:

[0079] S21: Use the layered deep convolutional network architecture as the spatial encoder. By setting the convolution layer, ReLU activation function and batch normalization layer step by step, combined with the strided convolution module, the feature extraction is deepened layer by layer to extract the spatial features of the image and obtain the feature map stacked along the time axis and processed by spatial encoding.

[0080] A layered deep convolutional network architecture is used as a spatial encoder to extract spatial features of remote sensing images. By setting convolutional layers, ReLU activation functions, and batch normalization layers step by step, combined with the step convolution module, feature extraction is deepened layer by layer, effectively reducing feature dimensions and compressing remote sensing data information, providing effective features for the temporal encoding and spatial decoding stages.

[0081] In view of the inherent complexity of time series data, a strategy of stacking feature maps processed by spatial encoding along the time axis is adopted, so that the model can comprehensively utilize the information of the entire time series in subsequent analysis, effectively identify and capture various vegetation cover change patterns from seasonal changes to sudden environmental disturbances, enhance the model's ability to analyze remote sensing image time series data, and provide reliable data support for accurately identifying and locating vegetation changes and environmental disturbances.

[0082] S22: Integrate the temporal attention encoder and apply the temporal attention mechanism to assign different weights to the feature maps at each time point in the sequence, achieve efficient encoding of the dataset in the temporal dimension, and identify the temporal dynamic characteristics of the forest area.

[0083] Temporal encoding is a key step in step S4, which aims to accurately capture and analyze the dynamic changes of remote sensing time series data. By integrating the temporal attention encoder and applying the temporal attention mechanism, the analysis of feature sequences is optimized. This mechanism enhances the model's ability to identify key time points by assigning different weights to the feature graphs of each time point in the sequence, while achieving effective compression of redundant time information and improving processing efficiency.

[0084] In the implementation, the method of combining the temporal attention encoder with the multi-scale analysis strategy is adopted to achieve efficient encoding of the time dimension on remote sensing datasets with uniform spatial resolution characteristics. This strategy aims to significantly improve the model's ability to identify and analyze the intrinsic changes in time series data by comprehensively utilizing various time scales, comprehensively capture periodic phenological changes and short-term environmental disturbances in time series data, and optimize the model's effectiveness in accurately identifying and locating vegetation cover changes and environmental disturbance events.

[0085] S23: Build a model based on U-Net, use layered strided transposed convolution technology to gradually improve the spatial resolution of feature maps, introduce skip connection technology and attention mechanism, fuse feature maps with temporal dynamic features, and reconstruct the spatial details of the original remote sensing time series data.

[0086] The spatial decoding stage adopts a design that is symmetrical to the spatial encoding stage. By using the layered strided transposed convolution technology to gradually improve the spatial resolution of the feature map, the spatial details of the original remote sensing time series data can be accurately reconstructed. At the same time, by introducing the jump connection technology, the deep features extracted in the spatial encoding stage and the temporal dynamic features identified in the temporal encoding stage are effectively integrated, the model's ability to capture spatial details and the comprehensive expression of temporal information is optimized, and the recognition accuracy of vegetation cover changes and environmental disturbances in spatiotemporal distribution information is enhanced.

[0087] This step builds a model based on U-Net. U-Net is chosen because its symmetrical design and feature fusion capabilities have been proven in medical image processing and remote sensing image analysis. This step customizes the U-Net architecture, including optimizing the transposed convolution layer to better restore high-resolution spatial information, and adjusting the jump connection strategy to promote the fusion of spatial and temporal dimension features. In addition, considering the characteristics of remote sensing time series data, the attention mechanism is introduced to enhance the model's ability to recognize key spatiotemporal change features, and improve the model's ability to identify and accurately locate vegetation cover changes and environmental disturbance events.

[0088] The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a time attention mechanism and training the model further includes:

[0089] The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small objects and balance the problem of uneven categories;

[0090] Use Adam optimizer to improve training efficiency and model convergence speed;

[0091] Implement a learning rate decay strategy to achieve fast learning in the early stages of training and maintain stable convergence in the later stages;

[0092] An early stopping mechanism is used to prevent overfitting.

[0093] Model training is the key to achieving accurate semantic segmentation based on time series remote sensing data, which includes multiple links such as loss function design, optimizer selection and learning rate management.

[0094] The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small targets and balance the problem of uneven categories. The Adam optimizer is selected to improve training efficiency and model convergence speed. By implementing learning rate decay, the model can achieve rapid learning in the early stage of training and maintain stable convergence in the later stage. The early stopping mechanism can prevent overfitting of the model.

[0095] S3: Construct a sample data set based on multi-temporal and high-spatial-resolution remote sensing image data. Based on the convolutional neural network model and multi-task learning, build and train a multi-task full convolutional neural network model to obtain a forest vegetation type change detection model.

[0096] When constructing the sample data set, the image data is enhanced and the sample data set is constructed based on the time window sliding technology.

[0097] The sample data construction first generates a sample set reflecting the change spots based on the existing remote sensing images and spot data. The calibrated samples include the vegetation cover type before and after and the change type samples.

[0098] Using two periods of high spatial resolution remote sensing images from the same sensor in adjacent quarters, vegetation cover remote sensing image change calibration is carried out by province or region based on remote sensing image features and vegetation cover changes. According to the vegetation cover type reflected by the attributes of fixed sample plots or forest map patches, the remote sensing image features of the location are used for comparative analysis, and representative plots with relatively consistent remote sensing image features are selected. The vegetation cover type before and after, remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to a 512m×512m square vector frame.

[0099] Based on the established remote sensing images and data labels of each vegetation cover type, combined with the remote sensing images and forest and grass resource data of the previous and next two periods, and referring to the attribute information of forest and grass resource spots, the characteristic change areas of the remote sensing images of the previous and next two periods are compared and determined, and the change type and the corresponding remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to the 512m×512m square vector frame. The specific change types include at least construction project use, logging, reclamation, disaster events, and ecological protection and restoration measures.

[0100] This step aims to establish a large-scale, diverse and spatially distributed sample database to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. Models that do not meet this accuracy standard will be inspected in detail, supplemented with labels for interpretation signs and change types, and training parameters will be adjusted and iterative training will be repeated to achieve higher accuracy.

[0101] The data annotation strategy focuses on the identification of changes in remote sensing images at different time points for each pixel and the accurate classification of vegetation types. By comparing image data at two or more time points, we can determine the change status of a specific pixel and the vegetation cover type before and after the change. This method aims to simplify the data annotation process while ensuring that the model can accurately learn to identify vegetation change patterns in multi-temporal data, providing effective data for subsequent accurate change detection and semantic segmentation tasks.

[0102] The multi-task fully convolutional neural network model mainly consists of three parts: encoder, converter and decoder.

[0103] The method of constructing a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning and training the model includes:

[0104] S31: Construct a hierarchical feature extraction network as an encoder to perform in-depth analysis on the complex spatial information contained in high spatial resolution remote sensing image data, integrate the spatiotemporal attention mechanism, integrate the change information of the time dimension, and extract comprehensive multi-scale spatiotemporal features from multi-phase high spatial resolution remote sensing image data.

[0105] The encoder uses a deep learning framework, such as U-Net, a model designed specifically for semantic segmentation tasks. By applying hierarchical feature extraction technology, the complex spatial information contained in high-resolution remote sensing data is deeply analyzed, thereby achieving overall recognition of vegetation coverage areas from local to global. In order to adapt to the variability of remote sensing images in spatial scale, the encoder extracts features at multiple spatial scales to ensure accurate recognition of vegetation coverage features under different spatial resolution conditions. Considering the multi-temporal properties of the input data, the encoder further designs a fusion spatiotemporal attention mechanism to effectively integrate the change information in the time dimension in order to accurately reveal the dynamic process of vegetation changes over time.

[0106] During the implementation process, the preprocessed multi-temporal high-spatial resolution remote sensing image data is first subdivided into small blocks, and a convolutional neural network architecture such as EfficientNet or U-Net is used as the basic framework to build a hierarchical feature extraction network. This network design can extract multi-scale features at different levels, and then build a hierarchical feature system. The integration of features at different levels is achieved through the feature pyramid network technology. At the same time, by integrating the spatiotemporal attention mechanism, such as Transformer or a specially designed spatiotemporal fusion module, the significant changes in multi-temporal remote sensing data are focused on to more accurately identify vegetation features that change over time. With the help of this deep learning architecture, the encoder can effectively extract comprehensive multi-scale and spatiotemporal features from multi-temporal high-spatial resolution remote sensing image data, providing a solid feature foundation for vegetation cover change detection and classification tasks.

[0107] S32: Build a converter based on the Transformer architecture, design a spatiotemporal feature integration module, use the self-attention mechanism, and adopt a multi-task learning strategy for change detection and semantic segmentation to refine and strengthen the multi-scale spatial features of the encoder output to identify changes in vegetation cover over time.

[0108] The key to the design of the converter is to refine and strengthen the multi-scale spatial features obtained from the encoder based on the Transformer architecture, with the goal of accurately identifying changes in vegetation cover over time. This architecture uses the Transformer mechanism to achieve efficient integration of features at different spatial scales, ensuring that key vegetation cover change information can be identified. The built-in self-attention mechanism enhances the model's ability to interpret complex relationships between features, effectively reveals the spatial change pattern of vegetation cover, and is an indispensable component in the task of vegetation cover change detection.

[0109] In the implementation process, the converter accurately processes the multi-scale spatial feature maps output by the encoder by designing a spatiotemporal feature integration module to identify the temporal changes in vegetation cover. The spatiotemporal feature integration module uses the self-attention mechanism to fuse and analyze the features of the same geographical location at different time points, enhancing the perception of temporal changes, especially in identifying significant change areas, while retaining key spatial details and extracting key spatiotemporal change features. By implementing a multi-task learning strategy for change detection and semantic segmentation, the converter not only independently processes each detection target, but also optimizes the detection performance through the information exchange mechanism between tasks.

[0110] S33: Construct a decoder based on multi-task processing to perform change detection and semantic segmentation on vegetation cover, embed a self-attention mechanism, conduct in-depth analysis of spatiotemporal features, optimize the feature fusion process, accurately analyze feature maps based on rich details, fuse the multi-scale spatiotemporal features output by the encoder and the composite features of vegetation cover changes over time output by the converter, and accurately detect vegetation cover changes.

[0111] The decoder design is based on the principle of multi-tasking, aiming to accurately perform vegetation cover change detection and semantic segmentation. The design comprehensively applies the multi-layer spatial features and temporal features generated by the encoder and converter to enhance the accurate detection of vegetation cover changes. Through the embedded self-attention mechanism, the analysis of spatiotemporal features is deepened, and the feature fusion process is optimized to ensure accurate analysis based on detailed feature maps. At the same time, the decoder adopts a specific multi-task learning strategy to independently process detection and classification tasks by customizing the task-related output layers, and establishes an information exchange mechanism to promote mutual learning between tasks.

[0112] In implementation, the decoder will fuse the composite features from the encoder and converter, and accurately analyze through the self-attention mechanism to highlight key changes and details. Applying the multi-task learning framework, the decoder is equipped with specific output layers to independently optimize each task for vegetation change detection and semantic segmentation. An information exchange mechanism is set up to promote mutual learning between tasks and improve the overall performance of the model. The performance of the decoder is continuously monitored and evaluated. Based on performance feedback, the parameters of the self-attention mechanism, the weight distribution of the multi-task learning strategy, and the details of feature fusion and information exchange are adjusted to maximize the efficiency and accuracy of the decoder.

[0113] The method of constructing a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning and training the model further includes:

[0114] A fusion loss function is used to integrate the cross entropy loss of semantic segmentation and the binary cross entropy loss of change detection. An adjustable weight coefficient is introduced to dynamically adjust the weight according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance.

[0115] Use the Adam optimizer with adaptive learning rate to improve training efficiency and model convergence speed;

[0116] Integrate dynamic loss weight adjustment and learning rate scheduling strategies to help the model maintain a balance in task learning between training stages and optimize the overall performance of the model;

[0117] The early stopping mechanism and model regularization measures are adopted to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process.

[0118] Model training includes multiple steps such as loss function selection, optimizer selection and learning rate management.

[0119] In order to effectively handle the dual tasks of semantic segmentation and change detection, this method adopts a fusion loss function. This function combines the cross entropy loss of semantic segmentation with the binary cross entropy loss of change detection, and introduces adjustable weight coefficients. These coefficients allow the model to dynamically adjust the weights according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance.

[0120] When training the model, the Adam optimizer with adaptive learning rate is selected to accelerate model convergence and improve efficiency. By integrating dynamic loss weight adjustment and learning rate scheduling strategy, the model is assisted to maintain a balance of task learning between each training stage and optimize the overall model performance. To prevent overfitting, early stopping mechanism and model regularization measures such as Dropout and batch normalization are adopted to enhance the generalization ability of the model and ensure the stability of the training process.

[0121] S4: Input the remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain the forest phenology and environmental disturbance detection results.

[0122] This method focuses on two types of change detection in forest areas. First, according to the vegetation coverage type classification system of the "National Technical Regulations for Comprehensive Monitoring of Forest and Grass Ecology", combined with the specific vegetation coverage types of the project technology integration demonstration base, the change patches within the vegetation coverage are extracted; second, according to the remote sensing interpretation and verification change types in the "National Technical Regulations for Comprehensive Monitoring of Forest and Grass Ecology", the change types of the change patches obtained in the previous step are extracted, including construction project use, logging, reclamation, disasters, and ecological protection and restoration measures.

[0123] The output of the forest phenology and environmental disturbance detection model aims to generate a feature map with the same resolution as the input image, in which each pixel is carefully classified according to its disturbance state. This classification operation is implemented through the spatial decoder inside the model, which maps high-dimensional features to a predefined category space through a 1x1 convolutional layer to achieve accurate recognition of the surface state.

[0124] With the help of the model's accurate classification of each pixel, the forest phenology and environmental disturbance detection model focuses on determining the specific time and location of disturbance events.

[0125] S5: According to the results of forest phenology and environmental disturbance detection, the forest vegetation type change detection model is fine-tuned and retrained to obtain the forest cross-season change detection model.

[0126] The output of the forest vegetation type change detection model consists of two parts: one is the semantic segmentation result for each pixel, that is, identifying which vegetation type each pixel belongs to; the other is the change detection result, which identifies whether each pixel has changed between the time points under investigation and the nature of the change. These two parts of the output together constitute the model's comprehensive analysis of multi-temporal remote sensing data, which can provide rich information for subsequent application scenarios.

[0127] This method is based on the forest phenology and environmental disturbance detection method and the forest vegetation type change detection method. It compares the fusion strategy of feature level and decision level and the model fine-tuning and retraining strategy to analyze the effect of cross-season vegetation cover change detection.

[0128] Among them, the fusion strategy of feature level and decision level is to combine the phenological disturbance characteristics from time series remote sensing and the vegetation change characteristics from multi-temporal remote sensing data to perform feature-level fusion in the feature extraction stage; the decision-level fusion strategy is to comprehensively analyze the output results of the two independent models to achieve a more accurate classification of vegetation cover changes.

[0129] The detection model fine-tuning and retraining strategy utilizes the phenological disturbance identification results obtained from the forest phenology and environmental disturbance detection method to fine-tune and retrain the forest vegetation type change detection model in the forest vegetation type change detection method, so that the model pays more attention to those areas with significant phenological disturbances.

[0130] Based on the forest vegetation type change detection model, phenological disturbance information is integrated to detect cross-seasonal changes in forest areas and further improve the accuracy of vegetation cover change detection.

[0131] S6: Input the multi-temporal and high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results.

[0132] Ultimately, the output of the forest area inter-seasonal change detection model is: one part is the semantic segmentation result for each pixel, that is, identifying which vegetation type each pixel belongs to; the other part is the inter-seasonal change detection result, identifying the inter-seasonal changes of each pixel and the nature of the changes.

[0133] The beneficial effects of this embodiment are as follows: this method is based on the detection of changes in forest vegetation types, and integrates the spatiotemporal distribution information of forest phenology and environmental disturbance changes, so that the model can better understand the background information of vegetation cover changes, and enhance the detail recognition ability and accuracy of vegetation cover change detection; extract detailed change information on fine scale of vegetation cover, solve the problem of difficulty in distinguishing high spatial resolution remote sensing phenology and disturbance, make up for the problem of poor effect of medium spatial resolution remote sensing fine scale change detection, and realize accurate extraction of fine vegetation cover change information for multiple periods within a year.

[0134] Embodiment 2

[0135] like Figure 2 As shown, a method for detecting cross-seasonal changes in forest areas comprises the following steps:

[0136] S1: Acquire forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data.

[0137] (1) Remote sensing time series data of forest areas

[0138] The forest area remote sensing time series data mainly uses Gaofen (GF) high time series data for research, and multispectral remote sensing data such as Sentinel-2 (S2) and Landsat satellites as auxiliary data sources for research. Among them, Gaofen series satellites (such as GF-1 / 6) provide high temporal and spatial resolution remote sensing images, which are helpful for accurately monitoring surface changes; S2 provides 10 to 20 meters of high-resolution multispectral images, which are suitable for capturing vegetation cover changes; Landsat series provides long-term continuous observation records, which are suitable for analyzing long-term trends.

[0139] Specifically, multispectral remote sensing time series data must meet the following requirements: spatial resolution is better than 15 meters, and better than 10 meters in key monitoring areas; to ensure that multispectral remote sensing data fully reflect the changes in each stage within the key growing season and capture key phenological change points, based on data availability, data are acquired every two weeks to one month over a one-year span to form a continuous time series with consistent radiation and spatial alignment; in the selection of multispectral sensors, priority is given to using data acquired by the same sensor at different time points. If conditions limit the use of different multispectral sensors, sensors with similar band characteristics should be selected as much as possible to ensure the comparability of the obtained data; cloud cover area is less than 5%.

[0140] Different geographical features, cloud cover and surface characteristics directly affect the selection of data sources. Therefore, we will carefully select appropriate remote sensing data according to the specific conditions of the test area to ensure that the collected data can accurately reflect the actual conditions of each place and effectively respond to the challenges brought by climate factors, ensuring the accuracy of research data and the practicality of results.

[0141] (2) Multi-temporal and high spatial resolution remote sensing image data of forest areas

[0142] Acquire multi-temporal and high-spatial-resolution remote sensing image data of forest areas to ensure that the data used can effectively reflect changes in vegetation coverage and provide a basis for model construction and analysis.

[0143] The GF-2 and GF-1 PMS images are selected as the main data sources, which can provide high spatial resolution surface feature information and high accessibility. Among them, GF-2 images have high spatial resolution and can provide detailed vegetation coverage features; while GF-1 PMS images effectively supplement the problem of insufficient coverage of GF-2 in specific areas and enhance the data foundation.

[0144] The forest area cross-seasonal change detection method also includes: preprocessing the forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data.

[0145] The preprocessing of forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data includes: radiation calibration, atmospheric correction, cloud and shadow removal, geometric correction, image registration, etc.

[0146] Preprocessing aims to improve data quality and provide a solid foundation for subsequent analysis.

[0147] S2: Construct a sample data set based on remote sensing time series data, build and train a time series full convolutional neural network model based on convolutional neural network and temporal attention mechanism, and obtain a forest phenology and environmental disturbance detection model.

[0148] When constructing the sample data set, the image data is enhanced and the sample data set is constructed based on the time window sliding technology.

[0149] The sample data construction first generates a sample set reflecting the change spots based on the existing remote sensing images and spot data. The calibrated samples include the vegetation cover type before and after and the change type samples.

[0150] Using two phases of remote sensing images from the same sensor in adjacent quarters, vegetation cover remote sensing image change calibration is performed by province or region based on remote sensing image features and vegetation cover changes. Based on the vegetation cover type reflected by the attributes of fixed sample plots or forest map patches, the remote sensing image features of the location are used for comparative analysis, and representative plots with relatively consistent remote sensing image features are selected. The vegetation cover type before and after, remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to a 512m×512m square vector frame.

[0151] Based on the established remote sensing images and data labels of each vegetation cover type, combined with the remote sensing images and forest and grass resource data of the previous and next two periods, and referring to the attribute information of forest and grass resource spots, the characteristic change areas of the remote sensing images of the previous and next two periods are compared and determined, and the change type and the corresponding remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to the 512m×512m square vector frame. The specific change types include at least construction project use, logging, reclamation, disaster events, and ecological protection and restoration measures.

[0152] This step aims to establish a large-scale, diverse and spatially distributed sample database to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. Models that do not meet this accuracy standard will be inspected in detail, supplemented with labels for interpretation signs and change types, and training parameters will be adjusted and iterative training will be repeated to achieve higher accuracy.

[0153] By carefully dividing the pre-processed time series remote sensing data, we create training and validation sets that reflect the diversity of vegetation coverage and environmental changes. At the same time, we use data enhancement methods such as random rotation, flipping, and cropping to enhance the model's generalization ability for diverse surface features. The training samples constructed based on the time window sliding technology not only capture the vegetation coverage information at a single moment, but also integrate the change trend on the time axis, providing the model with dynamic change information on vegetation coverage.

[0154] The time series fully convolutional neural network model is dedicated to improving the recognition accuracy of vegetation cover changes and environmental disturbance events, and emphasizes the ability to locate disturbances in space and time. The model consists of three main parts: first, the spatial encoding module uses deep convolutional layers to extract the spatial features of the image; second, the temporal encoding module introduces an attention mechanism to identify key change moments in the time series and enhance sensitivity to temporal changes; finally, the spatial decoding module is responsible for accurately reconstructing the spatial details of the semantic segmentation results while maintaining high spatial resolution.

[0155] The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a time attention mechanism and training the model includes:

[0156] S21: Use the layered deep convolutional network architecture as the spatial encoder. By setting the convolution layer, ReLU activation function and batch normalization layer step by step, combined with the strided convolution module, the feature extraction is deepened layer by layer to extract the spatial features of the image and obtain the feature map stacked along the time axis and processed by spatial encoding.

[0157] A layered deep convolutional network architecture is used as a spatial encoder to extract spatial features of remote sensing images. By setting convolutional layers, ReLU activation functions, and batch normalization layers step by step, combined with the step convolution module, feature extraction is deepened layer by layer, effectively reducing feature dimensions and compressing remote sensing data information, providing effective features for the temporal encoding and spatial decoding stages.

[0158] In view of the inherent complexity of time series data, a strategy of stacking feature maps processed by spatial encoding along the time axis is adopted, so that the model can comprehensively utilize the information of the entire time series in subsequent analysis, effectively identify and capture various vegetation cover change patterns from seasonal changes to sudden environmental disturbances, enhance the model's ability to analyze remote sensing image time series data, and provide reliable data support for accurately identifying and locating vegetation changes and environmental disturbances.

[0159] S22: Integrate the temporal attention encoder and apply the temporal attention mechanism to assign different weights to the feature maps at each time point in the sequence, achieve efficient encoding of the dataset in the temporal dimension, and identify the temporal dynamic characteristics of the forest area.

[0160] Temporal encoding is a key step in step S4, which aims to accurately capture and analyze the dynamic changes of remote sensing time series data. By integrating the temporal attention encoder and applying the temporal attention mechanism, the analysis of feature sequences is optimized. This mechanism enhances the model's ability to identify key time points by assigning different weights to the feature graphs of each time point in the sequence, while achieving effective compression of redundant time information and improving processing efficiency.

[0161] In the implementation, the method of combining the temporal attention encoder with the multi-scale analysis strategy is adopted to achieve efficient encoding of the time dimension on remote sensing datasets with uniform spatial resolution characteristics. This strategy aims to significantly improve the model's ability to identify and analyze the intrinsic changes in time series data by comprehensively utilizing various time scales, comprehensively capture periodic phenological changes and short-term environmental disturbances in time series data, and optimize the model's effectiveness in accurately identifying and locating vegetation cover changes and environmental disturbance events.

[0162] S23: Build a model based on U-Net, use layered strided transposed convolution technology to gradually improve the spatial resolution of feature maps, introduce skip connection technology and attention mechanism, fuse feature maps with temporal dynamic features, and reconstruct the spatial details of the original remote sensing time series data.

[0163] The spatial decoding stage adopts a design that is symmetrical to the spatial encoding stage. By using the layered strided transposed convolution technology to gradually improve the spatial resolution of the feature map, the spatial details of the original remote sensing time series data can be accurately reconstructed. At the same time, by introducing the jump connection technology, the deep features extracted in the spatial encoding stage and the temporal dynamic features identified in the temporal encoding stage are effectively integrated, the model's ability to capture spatial details and the comprehensive expression of temporal information is optimized, and the recognition accuracy of vegetation cover changes and environmental disturbances in spatiotemporal distribution information is enhanced.

[0164] This step builds a model based on U-Net. U-Net is chosen because its symmetrical design and feature fusion capabilities have been proven in medical image processing and remote sensing image analysis. This step customizes the U-Net architecture, including optimizing the transposed convolution layer to better restore high-resolution spatial information, and adjusting the jump connection strategy to promote the fusion of spatial and temporal dimension features. In addition, considering the characteristics of remote sensing time series data, the attention mechanism is introduced to enhance the model's ability to recognize key spatiotemporal change features, and improve the model's ability to identify and accurately locate vegetation cover changes and environmental disturbance events.

[0165] The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a time attention mechanism and training the model further includes:

[0166] The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small objects and balance the problem of uneven categories;

[0167] Use Adam optimizer to improve training efficiency and model convergence speed;

[0168] Implement a learning rate decay strategy to achieve fast learning in the early stages of training and maintain stable convergence in the later stages;

[0169] Use early stopping mechanism to prevent overfitting;

[0170] Comprehensively evaluate the model based on multiple indicators on the validation set, including at least precision, recall, and F1 score;

[0171] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0172] Model training is the key to achieving accurate semantic segmentation based on time series remote sensing data, which includes multiple links such as loss function design, optimizer selection and learning rate management.

[0173] The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small targets and balance the problem of uneven categories. The Adam optimizer is selected to improve training efficiency and model convergence speed. By implementing learning rate decay, the model can achieve rapid learning in the early stage of training and maintain stable convergence in the later stage. The early stopping mechanism can prevent overfitting of the model.

[0174] Through comprehensive evaluation based on key indicators such as precision, recall rate and F1 score on the validation set, the efficiency and accuracy of the model in high-precision semantic segmentation tasks are ensured.

[0175] During the performance evaluation, the segmentation prediction results of the model were compared with the field verification data to further explore the performance of the model in capturing seasonal changes and identifying disturbance events. Based on the evaluation results, this study fine-tuned the network architecture, refined the temporal attention mechanism, and adjusted the training strategy to enhance the model's ability to process remote sensing time series data and improve the accuracy of semantic segmentation, thereby enhancing the model's ability to reveal the spatial and temporal distribution characteristics of disturbances.

[0176] S3: Construct a sample data set based on multi-temporal and high-spatial-resolution remote sensing image data. Based on the convolutional neural network model and multi-task learning, build and train a multi-task full convolutional neural network model to obtain a forest vegetation type change detection model.

[0177] When constructing the sample data set, the image data is enhanced and the sample data set is constructed based on the time window sliding technology.

[0178] The sample data construction first generates a sample set reflecting the change spots based on the existing remote sensing images and spot data. The calibrated samples include the vegetation cover type before and after and the change type samples.

[0179] Using two periods of high spatial resolution remote sensing images from the same sensor in adjacent quarters, vegetation cover remote sensing image change calibration is carried out by province or region based on remote sensing image features and vegetation cover changes. According to the vegetation cover type reflected by the attributes of fixed sample plots or forest map patches, the remote sensing image features of the location are used for comparative analysis, and representative plots with relatively consistent remote sensing image features are selected. The vegetation cover type before and after, remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to a 512m×512m square vector frame.

[0180] Based on the established remote sensing images and data labels of each vegetation cover type, combined with the remote sensing images and forest and grass resource data of the previous and next two periods, and referring to the attribute information of forest and grass resource spots, the characteristic change areas of the remote sensing images of the previous and next two periods are compared and determined, and the change type and the corresponding remote sensing data source, remote sensing data acquisition time, spatial resolution and other attribute information are calibrated according to the 512m×512m square vector frame. The specific change types include at least construction project use, logging, reclamation, disaster events, and ecological protection and restoration measures.

[0181] This step aims to establish a large-scale, diverse and spatially distributed sample database to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. Models that do not meet this accuracy standard will be inspected in detail, supplemented with labels for interpretation signs and change types, and training parameters will be adjusted and iterative training will be repeated to achieve higher accuracy.

[0182] The data annotation strategy focuses on the identification of changes in remote sensing images at different time points for each pixel and the accurate classification of vegetation types. By comparing image data at two or more time points, we can determine the change status of a specific pixel and the vegetation cover type before and after the change. This method aims to simplify the data annotation process while ensuring that the model can accurately learn to identify vegetation change patterns in multi-temporal data, providing effective data for subsequent accurate change detection and semantic segmentation tasks.

[0183] The multi-task fully convolutional neural network model mainly consists of three parts: encoder, converter and decoder.

[0184] The method of constructing a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning and training the model includes:

[0185] S31: Construct a hierarchical feature extraction network as an encoder to perform in-depth analysis on the complex spatial information contained in high spatial resolution remote sensing image data, integrate the spatiotemporal attention mechanism, integrate the change information of the time dimension, and extract comprehensive multi-scale spatiotemporal features from multi-phase high spatial resolution remote sensing image data.

[0186] The encoder uses a deep learning framework, such as U-Net, a model designed specifically for semantic segmentation tasks. By applying hierarchical feature extraction technology, the complex spatial information contained in high-resolution remote sensing data is deeply analyzed, thereby achieving overall recognition of vegetation coverage areas from local to global. In order to adapt to the variability of remote sensing images in spatial scale, the encoder extracts features at multiple spatial scales to ensure accurate recognition of vegetation coverage features under different spatial resolution conditions. Considering the multi-temporal properties of the input data, the encoder further designs a fusion spatiotemporal attention mechanism to effectively integrate the change information in the time dimension in order to accurately reveal the dynamic process of vegetation changes over time.

[0187] During the implementation process, the preprocessed multi-temporal high-spatial resolution remote sensing image data is first subdivided into small blocks, and a convolutional neural network architecture such as EfficientNet or U-Net is used as the basic framework to build a hierarchical feature extraction network. This network design can extract multi-scale features at different levels, and then build a hierarchical feature system. The integration of features at different levels is achieved through the feature pyramid network technology. At the same time, by integrating the spatiotemporal attention mechanism, such as Transformer or a specially designed spatiotemporal fusion module, the significant changes in multi-temporal remote sensing data are focused on to more accurately identify vegetation features that change over time. With the help of this deep learning architecture, the encoder can effectively extract comprehensive multi-scale and spatiotemporal features from multi-temporal high-spatial resolution remote sensing image data, providing a solid feature foundation for vegetation cover change detection and classification tasks.

[0188] S32: Build a converter based on the Transformer architecture, design a spatiotemporal feature integration module, use the self-attention mechanism, and adopt a multi-task learning strategy for change detection and semantic segmentation to refine and strengthen the multi-scale spatial features of the encoder output to identify changes in vegetation cover over time.

[0189] The key to the design of the converter is to refine and strengthen the multi-scale spatial features obtained from the encoder based on the Transformer architecture, with the goal of accurately identifying changes in vegetation cover over time. This architecture uses the Transformer mechanism to achieve efficient integration of features at different spatial scales, ensuring that key vegetation cover change information can be identified. The built-in self-attention mechanism enhances the model's ability to interpret complex relationships between features, effectively reveals the spatial change pattern of vegetation cover, and is an indispensable component in the task of vegetation cover change detection.

[0190] In the implementation process, the converter accurately processes the multi-scale spatial feature maps output by the encoder by designing a spatiotemporal feature integration module to identify the temporal changes in vegetation cover. The spatiotemporal feature integration module uses the self-attention mechanism to fuse and analyze the features of the same geographical location at different time points, enhancing the perception of temporal changes, especially in identifying significant change areas, while retaining key spatial details and extracting key spatiotemporal change features. By implementing a multi-task learning strategy for change detection and semantic segmentation, the converter not only independently processes each detection target, but also optimizes the detection performance through the information exchange mechanism between tasks.

[0191] S33: Construct a decoder based on multi-task processing to detect vegetation cover changes and semantic segmentation. The decoder embeds a self-attention mechanism to deepen the analysis of spatiotemporal features, optimize the feature fusion process, accurately analyze feature maps based on rich details, and fuse the multi-scale spatiotemporal features output by the encoder and the composite features of vegetation cover changes over time output by the converter to accurately detect vegetation cover changes. This design combines the multi-layer spatial features and spatiotemporal features generated by the encoder and converter to enhance the accurate detection of vegetation cover changes. By embedding a self-attention mechanism, the analysis of spatiotemporal features is deepened, and the feature fusion process is optimized to ensure accurate analysis based on feature maps with rich details. At the same time, the decoder adopts a specific multi-task learning strategy to independently handle detection and classification tasks by customizing the task-related output layers, and establishes an information exchange mechanism to promote mutual learning between tasks.

[0192] In implementation, the decoder will fuse the composite features from the encoder and converter, and accurately analyze through the self-attention mechanism to highlight key changes and details. Applying the multi-task learning framework, the decoder is equipped with specific output layers to independently optimize each task for vegetation change detection and semantic segmentation. An information exchange mechanism is set up to promote mutual learning between tasks and improve the overall performance of the model. The performance of the decoder is continuously monitored and evaluated. Based on performance feedback, the parameters of the self-attention mechanism, the weight distribution of the multi-task learning strategy, and the details of feature fusion and information exchange are adjusted to maximize the efficiency and accuracy of the decoder.

[0193] The method of constructing a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning and training the model further includes:

[0194] A fusion loss function is used to integrate the cross entropy loss of semantic segmentation and the binary cross entropy loss of change detection. An adjustable weight coefficient is introduced to dynamically adjust the weight according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance.

[0195] Use the Adam optimizer with adaptive learning rate to improve training efficiency and model convergence speed;

[0196] Integrate dynamic loss weight adjustment and learning rate scheduling strategies to help the model maintain a balance in task learning between training stages and optimize the overall performance of the model;

[0197] Use early stopping mechanism and model regularization measures to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process;

[0198] Use multiple indicators, including at least precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance;

[0199] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0200] Model training includes multiple steps such as loss function selection, optimizer selection and learning rate management.

[0201] In order to effectively handle the dual tasks of semantic segmentation and change detection, this method adopts a fusion loss function. This function combines the cross entropy loss of semantic segmentation with the binary cross entropy loss of change detection, and introduces adjustable weight coefficients. These coefficients allow the model to adjust the weights according to the task.

[0202] The decoder design is based on the principle of multi-tasking, aiming to accurately perform vegetation cover change detection characteristics and dynamically adjust weights according to complexity, so as to achieve a balance between tasks and optimize model performance.

[0203] When training the model, the Adam optimizer with adaptive learning rate is selected to accelerate model convergence and improve efficiency. By integrating dynamic loss weight adjustment and learning rate scheduling strategy, the model is assisted to maintain a balance of task learning between each training stage and optimize the overall model performance. To prevent overfitting, early stopping mechanism and model regularization measures such as Dropout and batch normalization are adopted to enhance the generalization ability of the model and ensure the stability of the training process.

[0204] In order to accurately evaluate the performance of the model, multiple indicators are used, including precision, recall rate, F1 score, and overall accuracy, etc. Based on the comprehensive evaluation results, the network model is fine-tuned and improved.

[0205] S4: Input the remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain the forest phenology and environmental disturbance detection results.

[0206] This method focuses on two types of change detection in forest areas. First, according to the vegetation coverage type classification system of the "National Technical Regulations for Comprehensive Monitoring of Forest and Grass Ecology", combined with the specific vegetation coverage types of the project technology integration demonstration base, the change patches within the vegetation coverage are extracted; second, according to the remote sensing interpretation and verification change types in the "National Technical Regulations for Comprehensive Monitoring of Forest and Grass Ecology", the change types of the change patches obtained in the previous step are extracted, including construction project use, logging, reclamation, disasters, and ecological protection and restoration measures.

[0207] The output of the forest phenology and environmental disturbance detection model aims to generate a feature map with the same resolution as the input image, in which each pixel is carefully classified according to its disturbance state. This classification operation is implemented through the spatial decoder inside the model, which maps high-dimensional features to a predefined category space through a 1x1 convolutional layer to achieve accurate recognition of the surface state.

[0208] With the help of the model's accurate classification of each pixel, the forest phenology and environmental disturbance detection model focuses on determining the specific time and location of disturbance events.

[0209] S5: According to the results of forest phenology and environmental disturbance detection, the forest vegetation type change detection model is fine-tuned and retrained to obtain the forest cross-season change detection model.

[0210] The output of the forest vegetation type change detection model consists of two parts: one is the semantic segmentation result for each pixel, that is, identifying which vegetation type each pixel belongs to; the other is the change detection result, which identifies whether each pixel has changed between the time points under investigation and the nature of the change. These two parts of the output together constitute the model's comprehensive analysis of multi-temporal remote sensing data, which can provide rich information for subsequent application scenarios.

[0211] This method is based on the forest phenology and environmental disturbance detection method and the forest vegetation type change detection method. It compares the fusion strategy of feature level and decision level and the model fine-tuning and retraining strategy to analyze the effect of cross-season vegetation cover change detection.

[0212] Among them, the fusion strategy of feature level and decision level is to combine the phenological disturbance characteristics from time series remote sensing and the vegetation change characteristics from multi-temporal remote sensing data to perform feature-level fusion in the feature extraction stage; the decision-level fusion strategy is to comprehensively analyze the output results of the two independent models to achieve a more accurate classification of vegetation cover changes.

[0213] The detection model fine-tuning and retraining strategy utilizes the phenological disturbance identification results obtained from the forest phenology and environmental disturbance detection method to fine-tune and retrain the forest vegetation type change detection model in the forest vegetation type change detection method, so that the model pays more attention to those areas with significant phenological disturbances.

[0214] Based on the forest vegetation type change detection model, phenological disturbance information is integrated to detect cross-seasonal changes in forest areas and further improve the accuracy of vegetation cover change detection.

[0215] S6: Input the multi-temporal and high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results.

[0216] Ultimately, the output of the forest area inter-seasonal change detection model is: one part is the semantic segmentation result for each pixel, that is, identifying which vegetation type each pixel belongs to; the other part is the inter-seasonal change detection result, identifying the inter-seasonal changes of each pixel and the nature of the changes.

[0217] S7: Based on the detection results of inter-seasonal changes in forest areas, construct a spatiotemporal distribution map of inter-seasonal changes in forest areas.

[0218] After obtaining the results of forest area inter-seasonal change detection using the forest area inter-seasonal change detection model, these data were integrated to construct a spatiotemporal distribution map showing the inter-seasonal change distribution of vegetation types in the forest area and its time process, providing important information for environmental monitoring and decision-making analysis.

[0219] The beneficial effects of this embodiment are: this method can also conduct a comprehensive evaluation of the network model based on multiple indicators, and fine-tune and improve the network model according to the evaluation results; according to the detection results of inter-seasonal changes in forest areas, the data is integrated to construct a spatiotemporal distribution map of inter-seasonal changes in vegetation types in forest areas, providing important information for environmental monitoring and decision-making analysis.

[0220] Embodiment 3

[0221] A forest area inter-seasonal change detection system, based on the forest area inter-seasonal change detection method as described in Embodiment 1 or 2, comprises:

[0222] Data acquisition module, used to obtain forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data;

[0223] The phenological disturbance model construction module is used to construct a sample data set based on remote sensing time series data, build a time series full convolutional neural network model based on convolutional neural network and temporal attention mechanism, and train it to obtain a forest phenology and environmental disturbance detection model;

[0224] The vegetation change model building module is used to build a sample data set based on multi-temporal high-spatial resolution remote sensing image data, build a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning, and train it to obtain a forest vegetation type change detection model;

[0225] The phenological disturbance acquisition module is used to input remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain the forest phenology and environmental disturbance detection results;

[0226] The model adjustment module is used to fine-tune and retrain the forest vegetation type change detection model based on the forest phenology and environmental disturbance detection results to obtain the forest cross-season change detection model;

[0227] The detection module is used to input multi-temporal and high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results.

[0228] Preferably, the system further comprises:

[0229] Preprocessing module, used to preprocess forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data;

[0230] The distribution map construction module is used to construct the spatiotemporal distribution map of inter-seasonal changes in forest areas based on the detection results of inter-seasonal changes in forest areas.

[0231] Embodiment 4

[0232] A computer program which, when executed, implements the steps of the method for detecting cross-seasonal changes in forest areas as described in Embodiment 1 or 2.

[0233] Embodiment 5

[0234] A readable storage medium stores a computer program as described in Example 4, and when the computer program is executed, the steps of the method for detecting cross-seasonal changes in forest areas as described in Example 1 or 2 are implemented.

[0235] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for detecting cross-seasonal changes in forest areas, characterized in that: The following steps are involved: Obtain forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data; A sample data set was constructed based on remote sensing time series data. A time series full convolutional neural network model was constructed and trained based on convolutional neural network and temporal attention mechanism to obtain a forest phenology and environmental disturbance detection model. A sample data set was constructed based on multi-temporal and high-spatial-resolution remote sensing image data. A multi-task full convolutional neural network model was constructed and trained based on the convolutional neural network model and multi-task learning to obtain a forest vegetation type change detection model. Input remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain forest phenology and environmental disturbance detection results; According to the results of forest phenology and environmental disturbance detection, the forest vegetation type change detection model is fine-tuned and retrained to obtain the forest cross-season change detection model; Input multi-temporal high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results; The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training the model includes: Using the layered deep convolutional network architecture as the spatial encoder, by setting the convolution layer, ReLU activation function and batch normalization layer step by step, combined with the stride convolution module, the feature extraction is deepened layer by layer, the spatial features of the image are extracted, and the feature map stacked along the time axis and processed by spatial encoding is obtained; Integrate the temporal attention encoder and apply the temporal attention mechanism to assign different weights to the feature maps at each time point in the sequence, achieve efficient encoding of the dataset in the temporal dimension, and identify the temporal dynamic characteristics of the forest area; The model is built based on U-Net, and the spatial resolution of the feature map is gradually improved by using the layered strided transposed convolution technology. The skip connection technology and attention mechanism are introduced to fuse the feature map with the time series dynamic features to reconstruct the spatial details of the original remote sensing time series data. Wherein, constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training it further includes: The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small objects and balance the problem of uneven categories; Use Adam optimizer to improve training efficiency and model convergence speed; Implement a learning rate decay strategy to achieve fast learning in the early stages of training and maintain stable convergence in the later stages; Use early stopping mechanism to prevent overfitting; Comprehensively evaluate the model based on multiple indicators on the validation set, including precision, recall, and F1 score; Fine-tune the network model based on the comprehensive evaluation results; The method of constructing a multi-task full convolutional neural network model based on a convolutional neural network model and multi-task learning and training the model includes: A hierarchical feature extraction network is constructed as an encoder to deeply analyze the complex spatial information contained in high spatial resolution remote sensing image data, integrate the spatiotemporal attention mechanism, integrate the change information of the time dimension, and extract comprehensive multi-scale spatiotemporal features from multi-phase high spatial resolution remote sensing image data; A converter is built based on the Transformer architecture, and a spatiotemporal feature integration module is designed. The self-attention mechanism is used to adopt a multi-task learning strategy for change detection and semantic segmentation to refine and strengthen the multi-scale spatial features output by the encoder to identify changes in vegetation cover over time. The decoder is constructed based on multi-task processing to detect changes in vegetation cover and perform semantic segmentation. The self-attention mechanism is embedded to conduct in-depth analysis of spatiotemporal features, optimize the feature fusion process, accurately analyze feature maps based on rich details, and fuse the multi-scale spatiotemporal features output by the encoder and the composite features of vegetation cover changes over time output by the converter to accurately detect vegetation cover changes. Wherein, the multi-task full convolutional neural network model is constructed based on the convolutional neural network model and multi-task learning and trained further includes: A fusion loss function is used to integrate the cross entropy loss of semantic segmentation and the binary cross entropy loss of change detection. An adjustable weight coefficient is introduced to dynamically adjust the weight according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance. Use the Adam optimizer with adaptive learning rate to improve training efficiency and model convergence speed; Integrate dynamic loss weight adjustment and learning rate scheduling strategies to help the model maintain a balance in task learning between training stages and optimize the overall performance of the model; Use early stopping mechanism and model regularization measures to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process; Use multiple indicators, including precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance; Based on the comprehensive evaluation results, the network model is fine-tuned.

2. The method for detecting cross-seasonal changes in forest areas according to claim 1, characterized in that: The forest area cross-seasonal change detection method also includes: preprocessing the forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data.

3. The method for detecting cross-seasonal changes in forest areas according to claim 2, characterized in that: The preprocessing of forest area remote sensing time series data and multi-phase high spatial resolution remote sensing image data includes: radiation calibration, atmospheric correction, cloud and shadow removal, geometric correction, and image registration.

4. The method for detecting cross-seasonal changes in forest areas according to claim 1, characterized in that: When constructing the sample data set, the image data is enhanced and the sample data set is constructed based on the time window sliding technology.

5. The method for detecting cross-seasonal changes in forest areas according to claim 1, characterized in that: The method for detecting cross-seasonal changes in forest areas also includes the following steps: Based on the detection results of inter-seasonal changes in forest areas, a spatiotemporal distribution map of inter-seasonal changes in forest areas is constructed.

6. A forest area seasonal change detection system, characterized in that: The method for detecting cross-seasonal changes in forest areas according to any one of claims 1 to 5 comprises: Data acquisition module, used to obtain forest area remote sensing time series data and multi-temporal high spatial resolution remote sensing image data; The phenological disturbance model construction module is used to construct a sample data set based on remote sensing time series data, build a time series full convolutional neural network model based on convolutional neural network and temporal attention mechanism, and train it to obtain a forest phenology and environmental disturbance detection model; The vegetation change model building module is used to build a sample data set based on multi-temporal high-spatial resolution remote sensing image data, build a multi-task full convolutional neural network model based on the convolutional neural network model and multi-task learning, and train it to obtain a forest vegetation type change detection model; The phenological disturbance acquisition module is used to input remote sensing time series data into the forest phenology and environmental disturbance detection model to obtain the forest phenology and environmental disturbance detection results; The model adjustment module is used to fine-tune and retrain the forest vegetation type change detection model based on the forest phenology and environmental disturbance detection results to obtain the forest cross-season change detection model; The detection module is used to input multi-temporal high-spatial-resolution remote sensing image data into the forest area inter-seasonal change detection model to obtain the forest area inter-seasonal change detection results; The method of constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training the model includes: Using the layered deep convolutional network architecture as the spatial encoder, by setting the convolution layer, ReLU activation function and batch normalization layer step by step, combined with the stride convolution module, the feature extraction is deepened layer by layer, the spatial features of the image are extracted, and the feature map stacked along the time axis and processed by spatial encoding is obtained; Integrate the temporal attention encoder and apply the temporal attention mechanism to assign different weights to the feature maps at each time point in the sequence, achieve efficient encoding of the dataset in the temporal dimension, and identify the temporal dynamic characteristics of the forest area; The model is built based on U-Net, and the spatial resolution of the feature map is gradually improved by using the layered strided transposed convolution technology. The skip connection technology and attention mechanism are introduced to fuse the feature map with the time series dynamic features to reconstruct the spatial details of the original remote sensing time series data. Wherein, constructing a time series full convolutional neural network model based on a convolutional neural network and a temporal attention mechanism and training it further includes: The cross entropy loss function is selected to handle pixel-level classification tasks, and the Dice loss is introduced to improve the recognition ability of small objects and balance the problem of uneven categories; Use Adam optimizer to improve training efficiency and model convergence speed; Implement a learning rate decay strategy to achieve fast learning in the early stages of training and maintain stable convergence in the later stages; Use early stopping mechanism to prevent overfitting; Comprehensively evaluate the model based on multiple indicators on the validation set, including precision, recall, and F1 score; Fine-tune the network model based on the comprehensive evaluation results; The method of constructing a multi-task full convolutional neural network model based on a convolutional neural network model and multi-task learning and training the model includes: A hierarchical feature extraction network is constructed as an encoder to deeply analyze the complex spatial information contained in high spatial resolution remote sensing image data, integrate the spatiotemporal attention mechanism, integrate the change information of the time dimension, and extract comprehensive multi-scale spatiotemporal features from multi-phase high spatial resolution remote sensing image data; A converter is built based on the Transformer architecture, and a spatiotemporal feature integration module is designed. The self-attention mechanism is used to adopt a multi-task learning strategy for change detection and semantic segmentation to refine and strengthen the multi-scale spatial features output by the encoder to identify changes in vegetation cover over time. The decoder is constructed based on multi-task processing to detect changes in vegetation cover and perform semantic segmentation. The self-attention mechanism is embedded to conduct in-depth analysis of spatiotemporal features, optimize the feature fusion process, accurately analyze feature maps based on rich details, and fuse the multi-scale spatiotemporal features output by the encoder and the composite features of vegetation cover changes over time output by the converter to accurately detect vegetation cover changes. Wherein, the multi-task full convolutional neural network model is constructed based on the convolutional neural network model and multi-task learning and trained further includes: A fusion loss function is used to integrate the cross entropy loss of semantic segmentation and the binary cross entropy loss of change detection. An adjustable weight coefficient is introduced to dynamically adjust the weight according to the characteristics and complexity of the task, thereby achieving a balance between tasks and optimizing model performance. Use the Adam optimizer with adaptive learning rate to improve training efficiency and model convergence speed; Integrate dynamic loss weight adjustment and learning rate scheduling strategies to help the model maintain a balance in task learning between training stages and optimize the overall performance of the model; Use early stopping mechanism and model regularization measures to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process; Use multiple indicators, including precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance; Based on the comprehensive evaluation results, the network model is fine-tuned.

7. A computer program product, characterized in that It comprises a computer program, which, when executed, implements the steps of the method for detecting cross-seasonal changes in forest areas as described in any one of claims 1 to 5.

Images