Natural surface feature extraction method for remote sensing image of natural reserve

Abstract

The invention relates to the technical field of environment image processing, in particular to a natural surface feature extraction method for a remote sensing image of a natural reserve. On the basis of multi-source remote sensing data, namely a digital elevation model, a multi-temporal remote sensing image and real-time meteorological data, high-precision semantic segmentation is directly completed on an original low-quality image through spatial-temporal feature decoupling, degradation perception joint training, a gating fusion mechanism and an environmental safety hidden variable modeling technology; and unified monitoring and early warning of various types of environmental safety problems are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental image processing, and in particular to a method for extracting natural objects from remote sensing images of nature reserves. Background Art

[0002] With the rapid development of remote sensing technology, semantic segmentation models have been widely used in natural feature classification and ecological monitoring. However, in long-term dynamic monitoring scenarios, existing technologies still have the following significant shortcomings:

[0003] Insufficient adaptability to seasonal and environmental changes. The texture and spectral characteristics of natural features such as wetlands and grasslands change dynamically with seasons and climatic conditions. Traditional models rely on fixed training data and struggle to adapt to different seasonal scenarios, leading to fluctuations in segmentation accuracy. For example, the spectral differences between dense vegetation in the rainy season and bare soil in the dry season can easily lead to model misjudgments, requiring frequent data re-labeling and model fine-tuning, significantly increasing operational costs.

[0004] Low robustness to data degradation. Remote sensing images are often affected by degradation factors such as cloud cover, sensor noise, and low resolution. Existing methods typically use separate preprocessing steps such as cloud removal and denoising before segmentation. However, preprocessing can result in loss of ground object information or the introduction of artifacts, and the process is lengthy.

[0005] Lack of multi-source data fusion and real-time adjustment capabilities. Ecological monitoring requires the integration of remote sensing imagery, terrain data, and real-time meteorological information. However, traditional models rely solely on image data and lack effective mechanisms for integrating heterogeneous multi-source data. Furthermore, existing technologies are unable to dynamically adjust model parameters based on real-time environmental changes, resulting in degraded segmentation performance under extreme weather conditions or sudden human activity.

[0006] Inadequate environmental safety assurance capabilities. Current technologies primarily focus on natural feature classification and ecological monitoring, lacking the ability to uniformly model and conduct real-time monitoring of diverse environmental safety issues (such as droughts, floods, and fires). Especially in the context of intensified climate change and the frequent occurrence of natural disasters, there is an urgent need for a technical solution that can integrate multi-source data, rapidly identify anomalies, and provide risk warnings to enhance the overall level of ecological protection and environmental safety management. Summary of the Invention

[0007] The present invention discloses a method for extracting natural objects from remote sensing images of nature reserves. The specific method is as follows:

[0008] Obtain digital elevation models, remote sensing image data, and meteorological time series data of the nature reserve to be processed;

[0009] Extract static features from the digital elevation model to obtain static feature information;

[0010] Perform dynamic feature extraction and fusion based on environmental degradation perception on remote sensing image data and meteorological time series data to obtain dynamic feature information;

[0011] Fuse static feature information and dynamic feature information based on gated fusion mechanism;

[0012] The fused features are input into the feature extraction model to obtain a feature extraction image;

[0013] Based on the feature extraction image, determine whether the nature reserve area to be processed is abnormal.

[0014] Furthermore, static features of the digital elevation model are extracted. The specific method is as follows:

[0015] The specific information of the digital elevation model D(x,y) is obtained as follows:

[0016] Slope information

[0017] Aspect information

[0018] Surface curvature information

[0019] Terrain relief information

[0020] Splice slope information, gradient information, surface curvature information, and terrain relief information:

[0021]

[0022] ConvNeXt and SwinTransformer hybrid encoder extracts static features. The formula for extracting local features in the ConvNeXt stage is as follows:

[0023] F local =ConvNeXt-Stage1-2(X static )

[0024] Where, X static For static terrain data, ConvNeXt-Stage1-2 represents the first two stages of the ConvNeXt network;

[0025] SwinTransformer extracts global features as follows:

[0026] F static =Swin-Stage3-4(F local )

[0027] Where Swin-Stage3-4 represents the last two stages of the SwinTransformer network.

[0028] Furthermore, dynamic feature extraction and fusion based on environmental degradation perception are performed on remote sensing image data and meteorological time series data. The specific method is as follows:

[0029] Perform degradation area detection and generate degradation mask;

[0030] For degraded regions, uncertainty-guided label smoothing is performed;

[0031] Combined with the degradation mask, the weights of remote sensing image data and meteorological time series data are adjusted;

[0032] The adjusted remote sensing image data and meteorological time series data are fused as dynamic feature information.

[0033] Furthermore, degradation area detection is performed to generate a degradation mask. The specific method is as follows:

[0034] The degradation detection network D is used to mark the degradation area of the input remote sensing image and generate the degradation mask M degrade , the specific formula is as follows:

[0035] M degrade =D(I)∈[0,1] H×W

[0036] in, is the input remote sensing image, M degrade The value range is [0,1]. A value close to 1 indicates that the location is a degraded area, and a value close to 0 indicates that the location is a normal area.

[0037] Furthermore, uncertainty-guided label smoothing is performed. The specific method is as follows:

[0038] The original label formula is as follows:

[0039]

[0040] Where H is the height, W is the width, C classes Represents the number of categories in the semantic segmentation task. The label matrix is a three-dimensional tensor with a size of H×W×C classes ; Y(i,j,k)∈{0,1} indicates whether the position (i,j) belongs to the kth category;

[0041] The smoothed label formula is as follows:

[0042]

[0043] For degenerate regions, labels are smoothed to a uniform distribution For normal regions, the labels remain unchanged.

[0044] Furthermore, the degradation mask is combined to adjust the weight of the remote sensing image data, and the adjusted spatial features are:

[0045]

[0046] In the formula, MaskedPatchify is a block operation. When generating a sequence of small blocks, a lower weight is given to the small blocks in the degraded area, that is:

[0047] P t (i,j)=MaskedPatchify(I t ,M degrade )=(1-M degrade (i,j))·I t (i,j)

[0048] Among them, I t (i, j) represents the pixel value of the remote sensing image at position (i, j) at the tth time step;

[0049] Use LSTM for time series fusion to fuse the spatial features of multiple temporal phases into a global time feature to capture changes in the time dimension. The formula is as follows:

[0050]

[0051] In the formula, the input is the spatial feature of each time step The output is the hidden state h t , the hidden state h of the next time step T is used as the global temporal feature F time .

[0052] Furthermore, the degradation mask is combined to adjust the weight of the meteorological time series data. The specific formula is as follows:

[0053]

[0054] Where ⊙ is element-by-element multiplication, 1-M degrade The weight of degraded areas in meteorological features is reduced.

[0055] Furthermore, the adjusted remote sensing image data and meteorological time series data are fused, and the specific formula is as follows:

[0056]

[0057] Furthermore, static feature information and dynamic feature information are fused based on the gated fusion mechanism. The specific formula is as follows:

[0058] F fused =α·F static +(1-α)·F dynamic

[0059] Where, is the static feature information, is the dynamic feature information, α∈[0,1] is the dynamic fusion weight generated by real-time meteorological time series data;

[0060] The dynamic fusion weight α is generated as follows:

[0061] Encode meteorological data. The specific formula is as follows:

[0062]

[0063] Among them, W current is the meteorological data of the current time step. MLP performs nonlinear mapping on the meteorological data to generate meteorological features h weather , whose dimension is 64;

[0064] Generate terrain category weights based on static features. The specific formula is as follows:

[0065]

[0066] Concat concatenates the five static features and then maps the concatenated static features to C through convolution operation. types For each terrain category, Softmax normalizes the convolution result and generates a probability distribution TypeMap;

[0067] Dynamically generate fusion weights to control the contribution ratio of static features and dynamic features at each position. The specific formula is as follows:

[0068] α=σ(W2·ReLU(W1·Concat(h weather ,TypeMap)+b1)+b2)

[0069] Where W1 and W2 are learnable weight matrices, ReLU is the activation function, σ is the Sigmoid function, and the output α∈[0,1] H×W is the pixel-by-pixel fusion weight.

[0070] Furthermore, the degradation detection network D, static feature extraction, and dynamic feature extraction are jointly trained, and the joint loss function is as follows:

[0071] L total =L seg +α·L degrade +β·L uncertainty

[0072]

[0073] L degrade =BCE(M degrade ,G)

[0074]

[0075] Among them, α and β are hyperparameters; L seg Segmentation loss is the segmentation result predicted by the model, is the smoothed label; L degrade is the degradation detection loss, G∈[0,1] H×W is the true annotation of the degraded area, BCE is the binary cross entropy loss; L uncertainty is uncertainty regularization, and λ is the regularization coefficient.

[0076] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0077] By combining digital elevation models (DEMs), multi-temporal remote sensing imagery, and real-time meteorological data, this method can effectively address the complex terrain and dynamic environmental changes in nature reserves. For example, seasonal changes in vegetation cover and fluctuations in land features under extreme weather conditions can be adaptively processed through multi-source data fusion, significantly improving the model's robustness and segmentation accuracy in complex scenarios.

[0078] 2. Traditional methods typically require complex preprocessing of remote sensing images, which not only increases the computational burden but can also lead to information loss and artifacts. This new approach utilizes a degradation-aware mechanism to perform semantic segmentation directly on raw, low-quality images, reducing reliance on independent preprocessing steps and streamlining the overall workflow. This design avoids the error accumulation associated with traditional preprocessing and segmentation, improving the model's overall performance and generalization capabilities.

[0079] 3. Through the uncertainty-guided label smoothing mechanism and joint training strategy, the present invention can effectively utilize imprecisely labeled data and reduce the dependence on large-scale high-quality labeled data, thereby reducing the cost of manual labeling and shortening the model development cycle.

[0080] 4. This invention combines real-time meteorological data with time-series feature extraction technology to rapidly respond to environmental changes and is suitable for long-term dynamic monitoring scenarios. This real-time feature is crucial for rapid early warning of sudden natural disasters.

[0081] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings of the present invention are described below.

[0083] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0084] The present invention will be further described below with reference to the accompanying drawings and examples.

[0085] A natural feature extraction method for remote sensing images of nature reserves, such as Figure 1 The specific steps are as follows:

[0086] S1. Obtain the digital elevation model, remote sensing image data and meteorological time series data of the nature reserve area to be processed.

[0087] S2. Extract static features from the digital elevation model to obtain static feature information.

[0088] In step S2, the input data is the digital elevation model (DEM) of the nature reserve under study. Then the input data is preprocessed.

[0089] Five topographic features were selected to cover the ecological significance of multiple terrains in the nature reserve, including slope, aspect, surface curvature, and terrain relief, as follows:

[0090] Digital Elevation Model (DEM): D(x,y)

[0091] slope:

[0092] Slope direction:

[0093] Surface curvature:

[0094] Terrain relief:

[0095] Normalize and concatenate the above static features:

[0096]

[0097] This paper uses an improved SwinTransformer+ConvNeXt hybrid architecture for feature extraction, balancing global attention and local feature extraction. SwinTransformer captures global context through windowed multi-head self-attention (Window-MSA), making it suitable for modeling long-range dependencies in terrain features, such as the continuous orientation of mountains and the overall shape of watersheds. ConvNeXt, a modern CNN architecture, utilizes large-kernel deep convolution to extract detailed local features, such as micro-topography curvature and slope abrupt changes. It is computationally efficient and adaptable to high-resolution remote sensing data.

[0098] ConvNeXt and Swin Transformer hybrid encoder extracts static features. The formula for extracting local features in the ConvNeXt stage is as follows:

[0099] F local =ConvNeXt-Stage1-2(X static )

[0100] X static For static terrain data, ConvNeXt-Stage1-2 represents the first two stages of the ConvNeXt network.

[0101] Swin Transformer extracts global features as follows:

[0102] F static =Swin-Stage3-4(F local )

[0103] Swin-Stage3-4 represents the last two stages of the SwinTransformer network.

[0104] S3. Perform dynamic feature extraction and fusion based on environmental degradation perception on remote sensing image data and meteorological time series data to obtain dynamic feature information.

[0105] In step S3, dynamic feature extraction and fusion based on environmental degradation perception are performed on the remote sensing image data and the meteorological time series data. The specific method is as follows:

[0106] S31. Perform degradation area detection and generate a degradation mask.

[0107] In step S31, the degraded areas in the remote sensing image will significantly affect the feature extraction and segmentation accuracy. Through any lightweight degradation detection network D, the input image is annotated with degraded areas to generate a degradation mask M. degrade This mask will be used to adjust the weights of spatiotemporal feature extraction and meteorological condition annotation, that is, to guide the subsequent spatiotemporal feature extraction and gated fusion mechanism to avoid the interference of degraded areas on global feature modeling. The formula is as follows:

[0108] M degrade =D(I)∈[0,1] H×W

[0109] in, is the input remote sensing image, M degrade The value range is [0,1]. A value close to 1 indicates that the location is a degraded area, and a value close to 0 indicates that the location is a normal area.

[0110] S32. Perform uncertainty-guided label smoothing on the degraded region.

[0111] In step S32, degraded regions are often associated with high uncertainty. Directly using the labels of these regions may lead to model overfitting or performance degradation. This section introduces an uncertainty-guided label smoothing mechanism to smooth the labels of degraded regions and reduce their contribution to the loss function.

[0112] The original label formula is as follows:

[0113]

[0114] Indicates the spatial resolution of the input image, where H is the height, W is the width, and C classes Represents the number of categories in the semantic segmentation task. The label matrix is a three-dimensional tensor with a size of H×W×C classes Y(i,j,k)∈{0,1} indicates whether the position (i,j) belongs to the kth category.

[0115] The smoothed label formula is as follows:

[0116]

[0117] For degenerate regions, labels are smoothed to a uniform distribution For normal regions, the labels remain unchanged. Using smoothed labels in the loss function can reduce the negative impact of degraded regions on model training.

[0118] S33. Combine the degradation mask to adjust the weights of the remote sensing image data and the meteorological time series data.

[0119] S331. Adjust the weight of remote sensing image data:

[0120] In the spatiotemporal feature extraction stage, this part combines the degradation mask M degrade The spatial features of multi-temporal images are weighted and adjusted. The adjusted spatial features are:

[0121]

[0122] MaskedPatchify is an improved version of the block operation. When generating a sequence of small blocks, it gives lower weights to small blocks in the degraded area, that is:

[0123] P t (i,j)=MaskedPatchify(I t ,M degrade )=(1-M degrade (i,j))·I t (i,j)

[0124] Among them, I t (i, j) represents the pixel value of the remote sensing image at position (i, j) at the tth time step.

[0125] Then, LSTM is used for time series fusion to fuse the spatial features of multiple time phases into a global time feature to capture the changes in the time dimension. The formula is as follows:

[0126]

[0127] The input is the spatial features at each time step The output is the hidden state h t The hidden state h at the last time step T is used as the global temporal feature F time .

[0128] S332. Adjust the weight of meteorological time series data:

[0129] During the meteorological condition injection phase, this part uses the degradation mask to adjust the weight of meteorological features to reduce the impact of the degradation area on the dynamic features. The adjusted meteorological features are:

[0130]

[0131] ⊙ is element-by-element multiplication, 1-M degrade The weight of degraded areas in meteorological features is reduced.

[0132] S34. Fuse the adjusted remote sensing image data and meteorological time series data as dynamic feature information to generate the final feature representation of the dynamic branch:

[0133]

[0134] S4. Fuse static feature information and dynamic feature information based on the gated fusion mechanism.

[0135] The spatiotemporal feature decoupling module consists of a static feature branch and a dynamic feature branch. Feature fusion is achieved through a gated fusion mechanism. The final feature representation is mathematically expressed as:

[0136] F fused =α·F static +(1-α)·F dynamic

[0137] in, For static terrain features, is the dynamic environment feature, and α∈[0,1] is the fusion weight generated by real-time meteorological data.

[0138] There is a dynamic fusion weight α in the feature fusion expression. The following are the steps to generate the dynamic weight:

[0139] First, meteorological data encoding is performed to provide meteorological context information for dynamic weight generation. The formula is:

[0140]

[0141] Among them, W current is the meteorological data of the current time step. MLP performs nonlinear mapping on the meteorological data to generate meteorological features h weather , whose dimension is 64.

[0142] Then, terrain category weights are generated based on static features to guide the generation of dynamic weights. The formula is as follows:

[0143] Concat concatenates the five static features and then maps the concatenated static features to C through convolution operation. types For each terrain category, Softmax normalizes the convolution result and generates a probability distribution TypeMap.

[0144] Finally, the fusion weight is dynamically generated to control the contribution ratio of static features and dynamic features at each position. The formula is as follows:

[0145] α=σ(W2·ReLU(W1·Concat(h weather ,TypeMap)+b1)+b2)

[0146] W1 and W2 are learnable weight matrices, ReLU is the activation function, σ is the Sigmoid function, and the output α∈[0,1] H×W is the pixel-by-pixel fusion weight.

[0147] S5. Input the fused features into a feature extraction model to obtain a feature extraction image.

[0148] Traditional methods use a two-stage process of preprocessing and segmentation, which cannot achieve end-to-end optimization. This part will jointly train the degradation detection network D, the dynamic feature branch, and the static feature branch to perform semantic segmentation directly on the original low-quality image.

[0149] The joint loss function consists of the following three parts:

[0150] Segmentation loss: is the segmentation result predicted by the model, is the smoothed label.

[0151] Degradation detection loss: L degrade =BCE(M degrade ,G),G∈[0,1] H×W is the true annotation of the degraded region, and BCE is the binary cross entropy loss.

[0152] Uncertainty Regularization: The regularization term constrains the dynamic feature gradient of the degraded region to avoid the model's over-reliance on it. λ is the regularization coefficient.

[0153] The total loss is: L total =L seg +α·L degrade +β·L uncertainty , where α and β are hyperparameters. All modules pass L total Implement joint optimization to ensure that the model can complete semantic segmentation on original low-quality images.

[0154] S6. Determine whether the nature reserve area to be processed is abnormal based on the feature extraction image.

[0155] In step S6, by introducing environmental safety latent variables, multiple types of environmental safety issues, such as drought, flood, fire, etc., are modeled as anomaly detection tasks in the feature space.

[0156] The latent variables are generated by jointly encoding static features (topographic information) and dynamic features (meteorological and remote sensing time series data). The core idea is to extract high-level features that reflect the regional environmental state by fusing information from multiple heterogeneous data sources. These features include not only long-term, static information such as topography and vegetation, but also dynamic information that changes over time and meteorological conditions.

[0157] In this way, latent variables can capture the common patterns and distinctive characteristics of different environmental safety issues, such as droughts, floods, and fires, within a unified feature space. For example, droughts are often manifested as decreased vegetation cover and soil moisture; floods may be associated with abnormally increased precipitation and an expansion of surface water bodies. These phenomena will manifest themselves in distinct abnormal patterns within the feature space, and latent variables are designed to capture these patterns.

[0158] Environmental safety issues are essentially a phenomenon that deviates from the normal state. For example, under normal circumstances, indicators such as vegetation coverage, soil moisture, and temperature in a certain area will fluctuate within a certain range; when extreme events occur, these indicators may deviate significantly from the normal range. Hidden variables construct a high-dimensional feature space by fusing static and dynamic features. In this space, the state of each area can be represented by a hidden variable vector z. If the hidden variable z of a certain area new Compared with its historical baseline distribution {z1,z2...,z N If the distance between the two is large, it means that the current state of the area may be abnormal. The specific solution is as follows:

[0159] S61. Generation of environmental safety hidden variables

[0160] First input the static feature branch output F static , dynamic feature branch intermediate feature F time 、Meteorological characteristics embed , and then concatenate these features through a lightweight autoencoder to generate the latent variable z. The formula is as follows:

[0161] z=f enc (Concat(F static ,F time ,W embed ))

[0162] Among them, the encoder f enc It is a three-layer fully connected network.

[0163] S62. Historical baseline construction

[0164] Collect historical data of the study area, N time steps, and generate a set of latent variables {z1,z2...,z N}, and calculate the mean μ and covariance matrix ∑:

[0165]

[0166] For the current latent variable z new , calculate the Mahalanobis distance as the anomaly score:

[0167]

[0168] Taking into account the impact of seasonal changes on anomaly scores, a seasonal adjustment coefficient is introduced:

[0169] Threshold=α·s

[0170] For example, it can be set to 1.2 during the rainy season and 0.8 during the dry season.

[0171] The adjusted anomaly score Threshold is normalized and superimposed on the semantic segmentation result to generate a risk heat map:

[0172] R = Sigmoid(Threshold)·M seg

[0173] Among them, M seg is the semantic segmentation mask, R∈[0,1] H×W Represents a risk heat map, ranging from 0 to 1. The risk heat map can visually display the security risk level of different locations in the area.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for extracting natural features from remote sensing images of nature reserves, characterized in that: The specific method is as follows: Obtain digital elevation models, remote sensing image data, and meteorological time series data of the nature reserve to be processed; Extract static features from the digital elevation model to obtain static feature information; Perform dynamic feature extraction and fusion based on environmental degradation perception on remote sensing image data and meteorological time series data to obtain dynamic feature information; Fuse static feature information and dynamic feature information based on gated fusion mechanism; The fused features are input into the feature extraction model to obtain a feature extraction image; Based on the feature extraction image, determine whether the nature reserve area to be processed is abnormal.

2. The method for extracting natural features from remote sensing images of nature reserves according to claim 1, wherein: The static feature extraction of the digital elevation model is performed as follows: The specific information of the digital elevation model D(x,y) is obtained as follows: Slope information Aspect information Surface curvature information Terrain relief information Splice slope information, gradient information, surface curvature information, and terrain relief information: ConvNeXt and Swin Transformer hybrid encoder extracts static features. The formula for extracting local features in the ConvNeXt stage is as follows: F local =ConvNeXt-Stage1-2(X static ) Where, X static For static terrain data, ConvNeXt-Stage1-2 represents the first two stages of the ConvNeXt network; Swin Transformer extracts global features as follows: FAVORITE static =Swin-Stage3-4(F local ) Where Swin-Stage3-4 represents the last two stages of the Swin Transformer network.

3. The method for extracting natural features from remote sensing images of nature reserves according to claim 1, wherein: Dynamic feature extraction and fusion based on environmental degradation perception are performed on remote sensing image data and meteorological time series data. The specific method is as follows: Perform degradation area detection and generate degradation mask; For degraded regions, uncertainty-guided label smoothing is performed; Combined with the degradation mask, the weights of remote sensing image data and meteorological time series data are adjusted; The adjusted remote sensing image data and meteorological time series data are fused as dynamic feature information.

4. The method for extracting natural features from remote sensing images of nature reserves according to claim 3, wherein: Perform degradation area detection to generate degradation mask. The specific method is as follows: The degradation detection network D is used to mark the degradation area of the input remote sensing image and generate the degradation mask M degrade , the specific formula is as follows: M degrade =D(I)∈[0,1] H×W in, is the input remote sensing image, M degrade The value range is [0,1]. A value close to 1 indicates that the location is a degraded area, and a value close to 0 indicates that the location is a normal area.

5. The method for extracting natural features from remote sensing images of nature reserves according to claim 4, wherein: Perform uncertainty-guided label smoothing. The specific method is as follows: The original label formula is as follows: Where H is the height, W is the width, C classes Represents the number of categories in the semantic segmentation task. The label matrix is a three-dimensional tensor with a size of H×W×C classes ; Y(i,j,k)∈{0,1} indicates whether the position (i,j) belongs to the kth category; The smoothed label formula is as follows: For degenerate regions, labels are smoothed to a uniform distribution For normal regions, the labels remain unchanged.

6. The method for extracting natural features from remote sensing images of nature reserves according to claim 5, wherein: The weight of remote sensing image data is adjusted in combination with the degradation mask, and the adjusted spatial features are: In the formula, MaskedPatchify is a block operation. When generating a sequence of small blocks, a lower weight is given to the small blocks in the degraded area, that is: P t (i,j)=MaskedPatchify(I t ,M degrade )=(1-M degrade (i,j))·I t (i,j) Among them, I t (i, j) represents the pixel value of the remote sensing image at position (i, j) at the tth time step; Use LSTM for time series fusion to fuse the spatial features of multiple temporal phases into a global time feature to capture changes in the time dimension. The formula is as follows: In the formula, the input is the spatial feature of each time step The output is the hidden state h t , the hidden state h of the next time step T is used as the global temporal feature F time .

7. The method for extracting natural features from remote sensing images of nature reserves according to claim 6, wherein: Combined with the degradation mask, the weight of meteorological time series data is adjusted. The specific formula is as follows: Where ⊙ is element-by-element multiplication, 1-M degrade The weight of degraded areas in meteorological features is reduced.

8. The method for extracting natural features from remote sensing images of nature reserves according to claim 7, wherein: The specific formula for integrating the adjusted remote sensing image data and meteorological time series data is as follows: 。 9. The method for extracting natural features from remote sensing images of nature reserves according to claim 8, wherein: The static feature information and dynamic feature information are fused based on the gated fusion mechanism. The specific formula is as follows: F fused =α·F static +(1-α)·F dynamic Where, is the static feature information, is the dynamic feature information, α∈[0,1] is the dynamic fusion weight generated by real-time meteorological time series data; The dynamic fusion weight α is generated as follows: Encode meteorological data. The specific formula is as follows: Among them, W current is the meteorological data of the current time step. MLP performs nonlinear mapping on the meteorological data to generate meteorological features h weather , whose dimension is 64; Generate terrain category weights based on static features. The specific formula is as follows: Concat concatenates the five static features and then maps the concatenated static features to C through convolution operation. types terrain categories, Softmax normalizes the convolution results and generates a probability distribution TypeMap; Dynamically generate fusion weights to control the contribution ratio of static features and dynamic features at each position. The specific formula is as follows: α=σ(W2·ReLU(W1·Concat(h weather ,TypeMap)+b1)+b2) Where W1 and W2 are learnable weight matrices, ReLU is the activation function, σ is the Sigmoid function, and the output α∈[0,1] H×W is the pixel-by-pixel fusion weight.

10. The method for extracting natural features from remote sensing images of nature reserves according to claim 1, wherein: The degradation detection network D, static feature extraction and dynamic feature extraction are jointly trained, and the joint loss function is as follows: L total =L seg +α·L degrade +β·L uncertainty L degrade =BCE(M degrade ,G) Among them, α and β are hyperparameters; L seg Segmentation loss is the segmentation result predicted by the model, is the smoothed label; L degrade is the degradation detection loss, G∈[0,1] H×W is the true annotation of the degraded area, BCE is the binary cross entropy loss; L uncertainty is uncertainty regularization, and λ is the regularization coefficient.

Images