A remote sensing forest land change detection method based on timing consistency judgment

Abstract

The present application relates to a kind of remote sensing forest change detection method based on timing consistency judgment, it is related to high-resolution remote sensing image processing technology and the remote sensing image data mining technology field based on deep learning, the method includes the following steps: making forest change detection data set based on three time series remote sensing image;Three time phase remote sensing forest change detection model based on timing attention fusion is built;Loss function and optimizer are set;The model is trained using the constructed data set;Based on test time enhancement is predicted, and the result quality is improved using post-processing.This application enhances the inhibition ability to seasonal interference by inputting three time series remote sensing images: this application uses the previous, current and same period three high-resolution remote sensing images as input, comprehensively considers the historical state and the change trend of before and after time node in the same area, effectively suppresses the seasonal false change caused by the difference of phenology cycle.

Description

Technical Field

[0001] This invention relates to high-resolution remote sensing image processing technology and deep learning-based remote sensing image data mining technology, and in particular to a remote sensing forest change detection method based on temporal consistency judgment. Background Technology

[0002] Forest resources, as an important component of terrestrial ecosystems, possess multiple ecological functions, including water conservation, windbreak and sand fixation, climate regulation, and biodiversity maintenance. Their dynamic changes directly impact regional ecological security and the achievement of national ecological civilization goals. Therefore, conducting dynamic monitoring of forest resources, especially the timely and accurate identification of areas of change, is of great significance for natural resource management, ecological protection, and national spatial planning.

[0003] With the rapid development of remote sensing technology, forest land change detection based on satellite remote sensing data has become one of the mainstream methods. Remote sensing technology has advantages such as wide coverage, high acquisition frequency, and rich information dimensions, and is especially suitable for monitoring forest land resource changes over large areas and at multiple time scales.

[0004] Currently, forest land change detection methods are mainly divided into two categories: the first category is time-series modeling methods based on medium- and low-resolution remote sensing imagery. These methods generally use medium- and low-resolution data such as MODIS and Landsat, and identify changes by constructing long-term vegetation index (NDVI / EVI) trajectories. They show significant advantages in monitoring large-scale, gradual changes, for example, they can effectively capture annual-scale forest degradation processes. However, due to the low resolution of the images used, they are unable to meet the requirements for accurate identification of small-scale forest land changes and extraction of patch-level changes, thus limiting their application value in refined ecological monitoring and land use regeneration.

[0005] The second category is change detection methods based on two periods of high-resolution remote sensing imagery. These methods typically use satellite imagery data with a resolution better than 1 meter, combined with deep learning models (such as Siamese neural networks and Transformer structures) to extract differences and perform semantic recognition between the two periods of imagery. They offer advantages such as clear delineation of change boundaries and high accuracy in patch extraction, and have gradually become the main technical method for refined forest land change detection. However, as a natural land cover, forest land's image characteristics are easily affected by phenological changes. When the two periods of imagery are collected during the dormant and growing seasons of trees, significant differences in indicators such as canopy status and leaf area index can lead to numerous seasonal pseudo-changes in the model, severely affecting detection accuracy. This problem is particularly prominent in mid-to-high latitude regions (such as northern my country), severely limiting the effectiveness of this method.

[0006] Currently, there is a lack of forest land change detection methods that can integrate the fine representation capabilities of high-resolution imagery with the robust discrimination capabilities of temporal information. How to construct a deep learning framework that can fully utilize the spatial details of high-resolution remote sensing imagery while incorporating temporal information to suppress seasonal interference has become a pressing technical problem in the field of forest land change detection. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing forest land change detection algorithms by providing a remote sensing forest land change detection method based on temporal consistency judgment.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] A remote sensing method for detecting forest land change based on temporal consistency judgment includes the following steps:

[0010] Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images;

[0011] Step 2: Construct a three-temporal remote sensing forest change detection model based on temporal attention fusion;

[0012] Step 3: Set the loss function and optimizer;

[0013] Step 4: Train the model using the constructed dataset;

[0014] Step 5: Make predictions based on test-time enhancements and use post-processing to improve the quality of results.

[0015] In the above technical solution, the steps for creating the forest land change detection dataset in step 1 are as follows:

[0016] (1) Collect three time-series remote sensing images of the same area and perform preprocessing, including radiometric correction, geometric correction, orthorectification, image fusion, etc., to ensure that the resolution of the three images is consistent and the geographical location is aligned; the three time-series remote sensing images include: images from the same period last year, previous images and current images;

[0017] (2) Mark the actual forest land change patches through manual visual interpretation. Only areas where changes have occurred between the current image and previous images, and where the current image is inconsistent with the image of the same period last year, should be marked. Changes where the current image is consistent with the image of the same period last year, or where the changes have occurred between the previous image and the image of the same period last year, should not be marked.

[0018] (3) The three-phase time-series remote sensing images and change labels are constructed into sample pairs. The images are cropped using a fixed-size sliding window, and the cropping overlap rate is set. Only sample pairs with a change pixel ratio greater than the threshold in the label image are retained to alleviate the class imbalance problem.

[0019] (4) Divide the dataset according to the ratio of training set: validation set: test set = 4:1:1.

[0020] In the above technical solution, the main structure of the three-temporal remote sensing forest land change detection model in step 2 is as follows:

[0021] (1) Input mechanism: The three-phase time-series remote sensing images are used as input, corresponding to the reference phase, the previous phase and the current phase respectively; the three images are stitched together according to the channel dimension to form a three-branch structure input tensor with a size of B×9×H×W; the three-phase time-series remote sensing images include: the same period last year image, the previous image and the current image;

[0022] (2) Backbone network: a three-branch weight-sharing structure is used for feature extraction; each branch uses a pre-trained Vision Transformer model with shared parameters;

[0023] (3) Neck network: Since the output of ViT is a non-pyramid feature structure, in order to better adapt to the multi-scale decoding requirements, a multi-scale feature transformation module is introduced to adjust and fuse the features of ViT output at different levels and construct a feature representation with scale hierarchy.

[0024] (4) Temporal attention fusion module: For features at each scale, a cross-temporal feature fusion strategy based on attention mechanism is introduced; the features of the current time phase are used as the query, the features of the previous time phase are used as the key, and the features of the reference time phase are used as the value to calculate the cross-temporal attention matrix and perform feature weighted fusion.

[0025] (5) Decoding head: Based on the fused multi-scale temporal features, the UPerHead decoder is used to decode features and integrate contextual information, and finally outputs a binary classification change map to achieve accurate identification and extraction of forest change areas.

[0026] In the above technical solution, step 3 involves designing the loss function and optimizer, as detailed below:

[0027] (1) Loss function: The cross-entropy loss function is used to supervise the learning of the change category; this loss function is used to effectively measure the difference between the model's predicted map and the true label in the binary classification task, and promote the model's accurate discrimination of the change region;

[0028] (2) Optimizer: AdamW optimizer is used to update model parameters; the initial learning rate is set to a small value; a linear warm-up strategy is introduced in the early stage of training; a cosine annealing learning rate decay strategy is adopted to enable the model to converge quickly in the early stage and be gradually refined in the later stage.

[0029] In the above technical solution, step 4 specifically includes:

[0030] The proposed model was trained end-to-end using a constructed three-phase time-series remote sensing image change detection dataset.

[0031] Data augmentation strategies introduced during training include: random rotation, random cropping, random horizontal and vertical flipping, and photometric transformations with brightness, contrast, saturation, and hue perturbations.

[0032] After each training round, the model performance is evaluated on the validation set. The evaluation metric is the average intersection-union ratio, and the weights of the best-performing model on the validation set are saved.

[0033] In the above technical solution, in step 4: the training batch size is set to 4, and the total number of training rounds is set to 100 epochs.

[0034] In the above technical solution, step 5 specifically includes:

[0035] (1) Data augmentation during testing: During the inference stage, the input image is augmented in multiple scales and directions to generate multiple augmented samples; the model predicts each augmented sample separately and uses a fusion strategy to integrate the multiple prediction results to obtain a more robust and stable final change detection output.

[0036] (2) Post-processing of results: Morphological closing operation is used to fill the hole areas in the prediction and smooth the boundary contours; small spot filtering strategy is combined to remove scattered noise and suppress pseudo changes; the change detection results are converted from raster format to vector spot form to realize object-level expression of the changed area.

[0037] The present invention has the following beneficial effects:

[0038] This invention presents a remote sensing forest land change detection method based on temporal consistency judgment. By inputting three-phase temporal remote sensing images, it enhances the ability to suppress seasonal interference. The invention uses high-resolution remote sensing images from previous, current, and concurrent periods as input, comprehensively considering the historical state of the same area and the changing trends at different time points, effectively suppressing seasonal spurious changes caused by differences in phenological cycles. By introducing concurrent images as an auxiliary reference, areas of non-realistic change caused by differences between dormant and growing seasons can be identified, thereby improving the model's sensitivity and robustness to actual forest land changes, and enhancing the reliability and temporal stability of the detection.

[0039] This invention presents a remote sensing forest change detection method based on temporal consistency judgment. By constructing a deep learning model based on temporal consistency judgment, it achieves accurate extraction of real changes. The invention proposes a forest change detection model based on a temporal consistency strategy, innovatively introducing a "previous-current-same" three-period discrimination mechanism. Only when a change simultaneously meets the conditions of "difference between the previous period and the current period" and "difference between the current period and the same period" is it determined to be a real forest change. This significantly reduces false detections and missed detections caused by a single or lack of temporal reference. This model not only possesses high-precision perception capabilities for forest boundary changes but also supports accurate extraction of details at the patch level, making it particularly suitable for the refined needs of forest dynamic monitoring and ecological management in mid-to-high latitude regions. Attached Figure Description

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0041] Figure 1 This is an overall flowchart of the remote sensing forest land change detection method based on temporal consistency judgment of the present invention;

[0042] Figure 2 This is a flowchart of the calculation of the temporal attention fusion module in the remote sensing forest change detection method based on temporal consistency judgment of the present invention;

[0043] Figure 3 This is a comparative experimental result of forest land change detection using the remote sensing forest land change detection method based on temporal consistency judgment of the present invention. Detailed Implementation

[0044] The inventive concept of this invention is as follows:

[0045] This invention presents a remote sensing forest change detection method based on temporal consistency judgment. It innovatively incorporates three periods of high-resolution remote sensing imagery: images from the same period last year, previous periods, and the current period. Temporal comparative analysis is performed during the change identification process. Only when a change occurs between the current period's image and previous periods' images, and this change is inconsistent with the image from the same period last year, is it considered a genuine change. This effectively eliminates false change interference caused by seasonal differences, cyclical vegetation changes, and other factors.

[0046] Furthermore, this invention constructs a deep learning model architecture that integrates the Vision Transformer (ViT) backbone network structure with the Temporal Attention Fusion module. By introducing a temporal attention mechanism into the multi-scale feature layer, it achieves in-depth mining and dynamic alignment of three phases of image information, thereby enhancing the model's ability to perceive real forest land change areas.

[0047] The method of this invention can achieve high-precision, patch-level extraction of forest land change areas, and has good generalization performance and engineering application value.

[0048] The remote sensing forest change detection method based on temporal consistency judgment of the present invention includes the following steps:

[0049] Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images;

[0050] Step 2: Construct a three-temporal remote sensing forest change detection model based on temporal attention fusion;

[0051] Step 3: Set the loss function and optimizer;

[0052] Step 4: Train the three-temporal remote sensing forest change detection model using the constructed dataset;

[0053] Step 5: Make predictions based on test-time enhancements and use post-processing to improve the quality of results.

[0054] Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images; wherein, the steps for creating the forest land change detection dataset are as follows:

[0055] (1) Collect three high-resolution remote sensing images of the same area (last year's image, previous image, and current image). The image resolution should be better than 1 meter. Preprocessing should be performed, including radiometric correction, geometric correction, orthorectification, and image fusion, to ensure that the resolution of the three images is consistent and the geographical location is aligned.

[0056] (2) Mark the actual forest land change patches through manual visual interpretation. Only areas where changes have occurred between the current image and previous images, and where the current image is inconsistent with the image of the same period last year, should be marked. Changes where the current image is consistent with the image of the same period last year, or where the changes have occurred between the previous image and the image of the same period last year, should not be marked.

[0057] (3) The three-phase time-series remote sensing images and change labels are constructed into sample pairs. The images are cropped using a fixed-size (512×512 pixels) sliding window, and the cropping overlap rate (10%) is set. Only sample pairs with a change pixel ratio greater than the threshold (0.1%) in the label image are retained to alleviate the class imbalance problem.

[0058] (4) Divide the dataset according to the ratio of training set: validation set: test set = 4:1:1.

[0059] Step 2: Construct a three-temporal remote sensing forest land change detection model based on temporal attention fusion; the main structure of the three-temporal remote sensing forest land change detection model is as follows:

[0060] (1) Input mechanism: Three time-series remote sensing images (last year's image, previous image, and current image) are used as input, corresponding to the reference time phase, previous time phase, and current time phase, respectively. The three images are stitched together according to the channel dimension to form a three-branch structure input tensor with a size of B×9×H×W, thereby fully introducing cross-temporal contextual information and improving the model's ability to perceive temporal changes.

[0061] (2) Backbone network: A three-branch weight-sharing structure is used for feature extraction. Each branch uses a pre-trained Vision Transformer (ViT-Base, Patch-14) model with shared parameters to fully leverage its advantages in long-distance dependency modeling and high semantic expression, and to obtain multi-scale features with rich contextual information.

[0062] (3) Neck Network: Since the output of ViT is a non-pyramid feature structure, in order to better adapt to the multi-scale decoding requirements, a multi-scale feature transformation module (MultiLevelNeck) is introduced to adjust and fuse the features of ViT outputs at different levels and construct a feature representation with scale hierarchy.

[0063] (4) Temporal Attention Fusion Module: For features at each scale, a cross-temporal feature fusion strategy based on an attention mechanism is introduced. Specifically, features from the current time phase are used as the query, features from the previous time phase are used as the key, and features from the reference time phase (last year's same period) are used as the value. A cross-temporal attention matrix is ​​calculated, and feature weighting fusion is performed. This mechanism enhances the representation of features in truly changed areas while effectively suppressing spurious changes caused by non-real changes such as seasonal variations, thus improving the accuracy and reliability of change detection.

[0064] (5) Decoding head: Based on the fused multi-scale temporal features, the UPerHead decoder is used to decode features and integrate contextual information, and finally outputs a binary classification change map to achieve accurate identification and extraction of forest change areas.

[0065] Step 3: Setting the loss function and optimizer; specifically, the design of the loss function and optimizer is as follows:

[0066] (1) Loss function: The cross-entropy loss function is used for supervised learning of the change category. This loss function can effectively measure the difference between the model's predicted map and the true label in the binary classification task, and promote the model's accurate discrimination of the change region.

[0067] (2) Optimizer: The AdamW optimizer was selected to update the model parameters. To ensure training stability and accuracy, the initial learning rate was set to a small value (1e-5). In the early stages of training, a linear warm-up strategy was introduced, such as gradually increasing the learning rate over the first 10 training epochs to prevent gradient oscillations. Subsequently, a cosine annealing learning rate decay strategy was adopted to enable the model to converge rapidly in the early stages and be gradually refined in the later stages, thereby improving the final performance.

[0068] Step 4: Train the three-temporal remote sensing forest land change detection model using the constructed dataset; specifically:

[0069] The proposed three-temporal remote sensing forest change detection model was trained end-to-end using a constructed three-phase time-series remote sensing image change detection dataset. During training, various data augmentation strategies were introduced to enhance the model's generalization ability, including random rotation, random cropping, random horizontal and vertical flipping, and photometric transformations such as brightness, contrast, saturation, and hue perturbations. The training batch size was set to 4, and the total number of training epochs was set to 100. After each training epoch, the model performance was evaluated on the validation set using the mean intersection-over-union (mIoU) metric, and the weights of the best-performing model on the validation set were saved.

[0070] Step 5: Predict based on test-time augmentation and improve result quality using post-processing; specifically:

[0071] (1) Data augmentation during testing: During the inference phase, the input image is augmented at multiple scales and in multiple directions to generate multiple augmented samples. The model predicts each augmented sample separately and uses a fusion strategy (simple averaging) to integrate the multiple prediction results to obtain a more robust and stable final change detection output.

[0072] (2) Post-processing of results: To further improve the spatial coherence and interpretability of the prediction results, morphological closing operations are used to fill the hole areas in the prediction and smooth the boundary contours. At the same time, a small spot filtering strategy is combined to remove scattered noise and suppress spurious changes. Finally, the change detection results are converted from raster format to vector spot format to achieve object-level representation of the changed areas.

[0073] The present invention will now be described in detail with reference to the accompanying drawings.

[0074] like Figure 1 As shown, the remote sensing forest land change detection method based on temporal consistency judgment of the present invention includes the following steps:

[0075] Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images;

[0076] Step 2: Construct a three-temporal remote sensing forest change detection model based on temporal attention fusion;

[0077] Step 3: Set the loss function and optimizer;

[0078] Step 4: Train the three-temporal remote sensing forest change detection model using the constructed dataset;

[0079] Step 5: Make predictions based on test-time enhancements and use post-processing to improve the quality of results.

[0080] The specific steps of the remote sensing forest land change detection method based on temporal consistency judgment of the present invention are described below:

[0081] Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images;

[0082] Specifically, the following steps are included:

[0083] (1) Three high-resolution remote sensing images of the same area were acquired and preprocessed. The three images are from the same period last year, the previous period, and the current period. All three were acquired from high-resolution remote sensing image data of Jilin-1 satellite with a resolution better than 1 meter. To ensure the spatial consistency and comparability of the three images, radiometric correction, geometric correction, orthorectification, and image fusion were performed to ensure that the images are consistent in radiometric characteristics, spatial resolution, and geographical location.

[0084] (2) Labeling of actual forest land change patches. Based on the preprocessed three-phase high-resolution remote sensing images, change areas were identified and labeled through manual visual interpretation. To improve the quality of the labeled data, the following strategy was followed during the labeling process: only areas that changed between previous and current images, and whose current image is inconsistent with the image from the same period last year, were labeled. Specifically, areas whose current image is consistent with the image from the same period last year (e.g., seasonal vegetation changes), and areas whose changes occurred between the image from the same period last year and previous images, were not labeled. This strategy can effectively avoid false differences caused by non-structural changes such as seasonal variations and periodic agricultural activities, and improve the temporal consistency and authenticity of change labels.

[0085] (3) Constructing slice samples. The three-phase time-series remote sensing images and their corresponding change labels were organized into a unified training sample structure, and batch cropping was performed using a sliding window method. The window size was fixed at 512×512 pixels, and the overlap rate between windows was set to 10% during the cropping process to enhance the coverage and diversity of the samples. To further alleviate the problem of scarce change category samples in the data, only sample windows with a change pixel ratio exceeding 0.1% in the label map were retained. Each valid sample pair contains slices of the three-phase image and the corresponding change label map.

[0086] (4) Divide the dataset into training, validation, and test sets. After the dataset is constructed, it is divided into training, validation, and test sets in a 4:1:1 ratio. During the division process, it is ensured that samples of various types are evenly distributed in different subsets, and that samples from the same region do not appear in multiple subsets to avoid data leakage. The training set is used for learning model parameters, the validation set is used for parameter tuning and early stopping judgment, and the test set is used for the final performance evaluation of the model.

[0087] Step 2: Construct a three-temporal remote sensing forest change detection model based on temporal attention fusion;

[0088] This invention proposes a forest land change detection model based on three-phase time-series remote sensing imagery. The model's core design principle is time-series consistency discrimination. It utilizes the differences between the current and previous time phases to determine changes and incorporates imagery from the same period of the previous year as a time-series reference to eliminate spurious differences caused by seasonal fluctuations. The overall model structure includes the following steps:

[0089] (1) Input mechanism: The input to the model is three time-series remote sensing images of the same region, namely, the same period last year's image I. s Preliminary images I p And current image I c All three images are RGB three-channel images. To improve the model's ability to perceive temporal change patterns, the three images are stitched together along the channel dimension to form a unified three-branch input tensor, denoted as X = Concat(I s ,I p ,I c )∈R B×9×H×W Where B represents the batch size, and H and W are the image space dimensions.

[0090] (2) Backbone Network: To extract deep features with rich contextual semantics, a three-branch weight-sharing Vision Transformer is used as the feature extraction backbone network. Specifically, the three branches process three phases of images respectively and share model parameters. Each branch uses a ViT-Base pre-trained model, containing 12 Transformer encoder layers, each with a channel dimension of 768. The input image is divided into 14×14 pixel patches, and a sequence input is constructed through patch embedding. Multi-scale intermediate features are extracted from the 3rd, 6th, 9th, and 12th layers of the Transformer to construct a pyramid feature output. To improve training stability and efficiency, all Transformer layer weights are frozen during the training phase, and only the downstream modules are trained to enhance stability. The network is initialized with weights pre-trained using the externally loaded Jilin-1 large model. Through this shared backbone network structure, unified semantic encoding of forest change areas in the input images can be achieved, providing a consistent and highly expressive basic representation for subsequent feature fusion.

[0091] Each branch output contains feature representations at multiple scale levels:

[0092] F s =ViT(I s ), F p =ViT(I p ), F c =ViT(I c )

[0093] (3) Neck Network: Since the features output by ViT typically do not have a pyramidal hierarchical structure, they are difficult to use directly for multi-scale decoding. Therefore, this invention introduces a multi-scale feature transformation module (MultiLevelNeck). This module adjusts the outputs of different layers in the Transformer to a uniform number of channels and constructs a multi-level feature pyramid through convolution and upsampling operations: Each feature map, *∈{s,p,c}, contains the outputs corresponding to the three phase images and is fed into the temporal fusion module of the next stage. This structure enhances the model's robustness to inconsistent target scales (such as different sizes of forest patches) and improves its ability to identify changes in edge regions.

[0094] (4) Temporal Attention Fusion Module: To fuse features from three phases along the temporal dimension, extract true change patterns, and suppress spurious changes, this invention designs a cross-temporal feature fusion module based on an attention mechanism. At each scale, the current temporal feature is used as the query (Q), the previous temporal feature as the key (K), and the reference temporal feature (from the same period last year) as the value (V), to construct attention weights and perform weighted fusion.

[0095] The specific process is as follows: First, the three input features are mapped to query, key, and value representations, respectively. Taking a feature at a certain scale as an example, Q = W q *F c K = W k *F p v = W v *F s Among them, W q W k W v The convolution weights are then flattened into a two-dimensional matrix form: Q′∈R N×C , K′∈R C×N , V′∈R N×C Where N = H × W, and the scaled dot product attention is calculated:

[0096]

[0097] Where C represents the feature channel dimension. The fused result is then obtained:

[0098] O′=A·V′∈R N×C O = reshape(O′) ∈ R C×H×W

[0099] Finally, through the output convolution W out The fused temporal features are obtained as follows:

[0100] F fused =W out *O

[0101] This module, while maintaining semantic enhancement of real-world change regions, can suppress false detections caused by seasonal spectral fluctuations, thereby improving the overall accuracy and reliability of the model's change detection.

[0102] (5) Decoding Head: The fused multi-scale temporal features are fed into the decoding head module for semantic decoding and context integration. This invention uses the UPerHead decoder, which combines the advantages of Feature Pyramid Network (FPN) and Pyramid Pooling Module (PPM), and can fully aggregate context information at different scales. The output is a single-channel binary classification image P. out Each pixel location represents the category label for either forest change (1) or no change (0):

[0103]

[0104] Step 3: Set the loss function and optimizer;

[0105] To achieve effective training of the forest change detection model, this step includes two aspects: loss function design and optimizer configuration, the details of which are as follows:

[0106] (1) Loss Function: This invention employs a binary cross-entropy loss function to supervise learning the difference between the model output and the true change label. This loss function effectively guides the model to learn the discrimination boundary between "changed" and "unchanged" pixels. Specifically, let H and W represent the height and width of the image, respectively, and Y... ij This represents the true label value at the (i, j)th pixel position (1 if it changes, 0 if it doesn't). This represents the predicted probability value output by the model at this pixel location. The loss function is defined as follows:

[0107]

[0108] The design of this loss function can enhance the model's sensitivity to small changes in pixels, thereby improving the overall detection accuracy.

[0109] (2) Optimizer: In the model optimization stage, this invention uses the AdamW optimizer to iteratively update the network parameters. AdamW is an improved algorithm developed based on the standard Adam optimizer. Its main feature is the decoupling of the L2 regularization term (weight decay), making it more suitable for training deep network models containing Transformer modules. To ensure the stability and accuracy of training, the initial learning rate is set to a small value (1e-5). In the early stage of training, a linear warm-up strategy is introduced, such as gradually increasing the learning rate in the first 10 training epochs to prevent gradient oscillations. Subsequently, a cosine annealing learning rate decay strategy is adopted to enable the model to converge quickly in the early stage and be gradually refined in the later stage, thereby improving the final performance.

[0110] Step 4: Train the three-temporal remote sensing forest change detection model using the constructed dataset;

[0111] This invention uses a constructed three-phase time-series remote sensing image change detection dataset to perform end-to-end training on the proposed three-phase remote sensing forest land change detection model. The training method and configuration are as follows:

[0112] (1) Training method: Supervised learning is adopted. The input is three-phase time-series remote sensing images, and the label is a change mask. The parameters are optimized by minimizing the difference between the prediction results and the true labels to achieve accurate identification of forest change areas.

[0113] (2) Data Augmentation: To enhance the model's generalization ability, various data augmentation strategies are introduced during the training phase, including random rotation, random cropping, horizontal flipping, vertical flipping, and photometric transformations such as brightness, contrast, saturation, and hue perturbation. Augmentation operations are applied randomly in each training round to improve the model's adaptability to different imaging conditions and spatial variations.

[0114] (3) Training parameters: The training adopts a mini-batch update strategy with a batch size of 4 and a total of 100 rounds. After each round of training, the model performance is evaluated on the validation set, and the input image size is uniformly 512×512 pixels.

[0115] (4) Performance evaluation: The average intersection-union ratio (mIoU) is used as the main evaluation index. The mIoU is calculated as follows:

[0116]

[0117] Among them, TP c FP c 、FN c These represent the number of true positive, false positive, and false negative pixels for category c, respectively.

[0118] (5) Model saving: During training, the mIoU value on the validation set is recorded in each round, and the parameters of the best performing model are saved for testing and deployment.

[0119] Step 5: Make predictions based on test-time augmentations and use post-processing to improve the quality of results;

[0120] After the model training is completed, this invention employs enhanced inference and post-processing methods to improve the stability and spatial coherence of change detection results, specifically including:

[0121] (1) Data augmentation during testing: During the inference phase, the input image is augmented at multiple scales (1.0×, 0.75×, 1.25×) and in multiple directions (horizontal and vertical flipping) to generate multiple augmented samples. The model predicts each augmented sample separately and uses a fusion strategy (simple averaging) to integrate the multiple prediction results to obtain a more robust and stable final change detection output.

[0122] (2) Post-processing of results: To further improve the spatial coherence and interpretability of the prediction results, morphological closing operations are used to fill the hole areas in the prediction and smooth the boundary contours. At the same time, a small spot filtering strategy is combined to remove scattered noise and suppress spurious changes. Finally, the change detection results are converted from raster format to vector spot format to achieve object-level representation of the changed areas.

[0123] The remote sensing forest land change detection method of this invention, based on temporal consistency judgment, enhances the ability to suppress seasonal interference by inputting three-phase time-series remote sensing images: This invention uses high-resolution remote sensing images from the previous period, the current period, and the same period as input, comprehensively considering the historical state of the same area and the changing trends of previous and subsequent time points, effectively suppressing seasonal pseudo-changes caused by differences in phenological cycles. By introducing contemporaneous images as auxiliary references, areas of non-realistic change caused by differences between dormancy and growth periods can be identified, thereby improving the model's sensitivity and robustness to actual forest land changes, and enhancing the reliability and temporal stability of the detection.

[0124] This invention presents a remote sensing forest change detection method based on temporal consistency judgment. By constructing a deep learning model based on temporal consistency judgment, it achieves accurate extraction of real changes. The invention proposes a forest change detection model based on a temporal consistency strategy, innovatively introducing a "previous-current-same" three-period discrimination mechanism. Only when a change simultaneously meets the conditions of "difference between the previous period and the current period" and "difference between the current period and the same period" is it determined to be a real forest change. This significantly reduces false detections and missed detections caused by a single or lack of temporal reference. This model not only possesses high-precision perception capabilities for forest boundary changes but also supports accurate extraction of details at the patch level, making it particularly suitable for the refined needs of forest dynamic monitoring and ecological management in mid-to-high latitude regions.

[0125] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A remote sensing method for detecting forest land change based on temporal consistency judgment, characterized in that, Includes the following steps: Step 1: Create a forest land change detection dataset based on three-phase time-series remote sensing images; Step 2: Construct a three-temporal remote sensing forest change detection model based on temporal attention fusion; Step 3: Set the loss function and optimizer; Step 4: Train the model using the constructed dataset; Step 5: Make predictions based on test-time augmentations and use post-processing to improve the quality of results; In step 1, the steps for creating the forest land change detection dataset are as follows: (1) Collect three time-series remote sensing images of the same area and perform preprocessing, including radiometric correction, geometric correction, orthorectification, image fusion, etc., to ensure that the resolution of the three images is consistent and the geographical location is aligned. The three phases of time-series remote sensing images include: images from the same period last year, previous images, and current images; (2) Mark the actual forest land change patches through manual visual interpretation. Only areas where changes have occurred between the current image and previous images, and where the current image is inconsistent with the image of the same period last year, should be marked. No changes should be marked for areas where the current image is consistent with the image of the same period last year, or where the changes have occurred between the previous image and the image of the same period last year. (3) The three-phase time-series remote sensing images and change labels are constructed into sample pairs. The images are cropped using a fixed-size sliding window, and the cropping overlap rate is set. Only sample pairs with a change pixel ratio greater than the threshold in the label image are retained to alleviate the class imbalance problem. (4) Divide the dataset according to the ratio of training set: validation set: test set = 4:1:

1.

2. The remote sensing forest land change detection method based on temporal consistency judgment according to claim 1, characterized in that, In step 2, the main structure of the three-temporal remote sensing forest land change detection model is as follows: (1) Input mechanism: The three-phase time-series remote sensing images are used as input, corresponding to the reference phase, the previous phase and the current phase respectively; the three images are stitched together according to the channel dimension to form a three-branch structure input tensor with a size of B×9×H×W; the three-phase time-series remote sensing images include: the same period last year, the previous phase and the current phase; (2) Backbone network: a three-branch weight-sharing structure is used for feature extraction; each branch uses a pre-trained Vision Transformer model with shared parameters; (3) Neck network: Since the output of ViT is a non-pyramid feature structure, in order to better adapt to the multi-scale decoding requirements, a multi-scale feature transformation module is introduced to adjust and fuse the features of ViT output at different levels and construct a feature representation with scale hierarchy. (4) Temporal attention fusion module: For each scale of features, an attention-based cross-temporal feature fusion strategy is introduced; the current temporal features are used as queries, the previous temporal features are used as keys, and the reference temporal features are used as values ​​to calculate the cross-temporal attention matrix and perform feature weighted fusion. (5) Decoding head: Based on the fused multi-scale temporal features, the UPerHead decoder is used to decode features and integrate contextual information, and finally outputs a binary classification change map to achieve accurate identification and extraction of forest change areas.

3. The remote sensing forest land change detection method based on temporal consistency judgment according to claim 1, characterized in that, In step 3, the loss function and optimizer are designed as follows: (1) Loss function: The cross-entropy loss function is used to supervise the learning of the change category; this loss function is used to effectively measure the difference between the model's predicted map and the true label in the binary classification task, and promote the model's accurate discrimination of the change region; (2) Optimizer: AdamW optimizer is used to update model parameters; the initial learning rate is set to a small value; a linear warm-up strategy is introduced in the early stage of training; a cosine annealing learning rate decay strategy is adopted to enable the model to converge quickly in the early stage and be gradually refined in the later stage.

4. The remote sensing forest land change detection method based on temporal consistency judgment according to claim 1, characterized in that, Step 4 specifically involves: The proposed model was trained end-to-end using a constructed three-phase time-series remote sensing image change detection dataset. Data augmentation strategies introduced during training include: random rotation, random cropping, random horizontal and vertical flipping, and photometric transformations with brightness, contrast, saturation, and hue perturbations. After each training round, the model performance is evaluated on the validation set. The evaluation metric is the average intersection-union ratio, and the weights of the best-performing model on the validation set are saved.

5. The remote sensing forest land change detection method based on temporal consistency judgment according to claim 4, characterized in that, In step 4: The training batch size is set to 4, and the total number of training epochs is set to 100.

6. The remote sensing forest land change detection method based on temporal consistency judgment according to claim 1, characterized in that, Step 5 specifically includes: (1) Data augmentation during testing: During the inference stage, the input image is augmented in multiple scales and directions to generate multiple augmented samples; the model predicts each augmented sample separately and uses a fusion strategy to integrate the multiple prediction results to obtain a more robust and stable final change detection output. (2) Post-processing of results: Morphological closing operation is used to fill the hole areas in the prediction and smooth the boundary contours; small spot filtering strategy is combined to remove scattered noise and suppress pseudo changes; the change detection results are converted from raster format to vector spot form to realize object-level expression of the changed area.

Images