A method for constructing a large spatio-temporal spectral model of forestry resources for the forestry vertical domain

Abstract

The present invention discloses a method for constructing a large spatiotemporal spectrum model of forestry resources for the forestry vertical field, belonging to the field of image data processing technology. The method comprises constructing a spatiotemporal spectrum dataset of forestry resources, which includes spectral graphs, geographic information, climate data, and real fire probability maps at different time steps within the monitoring area; constructing an ST-SegNet network, including a SegNet encoder, a SegNet decoder, a ConvLSTM model, a feature fusion layer, and a merging layer; constructing the total loss of the ST-SegNet network, and training the ST-SegNet network to obtain a large spatiotemporal spectrum model of forestry resources. The present invention improves the spatiotemporal consistency and classification accuracy of remote sensing image data, ensuring higher-quality data processing; through spatiotemporal spectrum data fusion, it enhances adaptability to complex scenarios and improves detail processing capabilities; provides a new technical path for environmental monitoring and disaster response, and significantly improves the efficiency of emergency warning systems.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image technology, and in particular to a method for constructing a large spatiotemporal spectrum model of forestry resources oriented to the forestry vertical field. Background Art

[0002] With the rapid development of large models, spatiotemporal feature extraction methods that include temporal information have shown great potential in remote sensing, geographic information systems, environmental monitoring and other fields. As a deep learning model, the spatiotemporal spectral large model not only significantly improves the effect of target recognition and disaster prediction by introducing the fusion of temporal, spatial and spectral information, but also ensures the accuracy and consistency of data processing, thereby promoting the application and development of remote sensing technology. In remote sensing monitoring, ensuring the spatiotemporal consistency and accuracy of image data is crucial for performing tasks such as target detection and resource monitoring. Specific application requirements often require efficient data analysis, real-time disaster warning systems, etc. to meet the needs of environmental protection and disaster response. However, the limitations of existing technologies make disaster prediction models face great challenges in processing high-resolution spatiotemporal data.

[0003] The SegNet network is a deep convolutional encoding-decoding architecture for image semantic segmentation, in which the encoder network is responsible for extracting image features and generating feature maps, while the decoder is mainly responsible for restoring the feature maps to the same spatial resolution as the original input image and generating segmentation results.

[0004] The ConvLSTM (Convolutional Long Short-Term Memory) network is a structure that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). It is primarily used to process spatiotemporal data and is particularly suitable for tasks that require considering spatial features and temporal dependencies, such as weather forecasting, video analysis, and traffic flow prediction. Unlike traditional LSTMs, ConvLSTMs apply a convolution operation at each time step, helping to capture spatial information in time series data. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for constructing a large spatiotemporal spectrum model of forestry resources in the vertical field of forestry, which can solve the above problems, improve the spatiotemporal consistency and classification accuracy of remote sensing image data, enhance the adaptability to complex scenes, and improve the warning efficiency.

[0006] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows: a method for constructing a large spatiotemporal spectrum model of forestry resources in the forestry vertical field, comprising the following steps;

[0007] S1, constructing a spatiotemporal spectrum dataset of forestry resources, including steps S11 to S14;

[0008] S11, obtain remote sensing images of the monitoring area at different time steps and different spectral bands. The remote sensing image obtained at time step t and spectral band λ is I(t,x,y,λ), t∈[1,T], T is the total number of time steps, λ∈[1,B], B is the total number of spectral bands, x and y are the width and height of a pixel in the remote sensing image, respectively, x∈[1,r x ], y∈[1,r y ],r x 、r y are the width and height of the remote sensing image respectively;

[0009] S12, merge the channels of B remote sensing images at time step t into λ all , get the spectrum X(t,x,y,λ at time step t all );

[0010] S13, obtain X(t,x,y,λ all ) auxiliary data, including geographic information G(x,y) and climate data C(t,x,y);

[0011] S14, for X(t,x,y,λ all ) each pixel is labeled with the true category and the true fire probability, the category includes fire, and the true fire probability of all pixels constitutes the true fire probability map, where X(t,x,y,λ all ) is the true fire probability map P true (t,x,y), X(t,x,y,λ all ) and its corresponding auxiliary data constitute samples, and all samples constitute a spatiotemporal spectrum dataset;

[0012] S2, construct the ST-SegNet network, including SegNet encoder, SegNet decoder, ConvLSTM model, feature fusion layer and merging layer;

[0013] The SegNet encoder is used to all ) to extract features and obtain the encoding feature F spatial (t,x,y,λ all );

[0014] The encoding features are constructed into a time series, and the ConvLSTM model is used to capture the time series features, where the encoding feature F spatial (t,x,y,λ all ) The corresponding time series feature is F temporal (t,x,y);

[0015] The feature fusion layer is used to extract features from geographic information and climate data and then fuse them to obtain auxiliary data features, where X(t, x, y, λ all )’s auxiliary data feature is F aux (t,x,y);

[0016] The merging layer is used to transform F spatial (t,x,y,λ all ), F temporal (t,x,y),F aux (t,x,y) are merged to obtain X(t,x,y,λ all )'s comprehensive spatiotemporal characteristics F spatio-temporal (t,x,y);

[0017] The SegNet decoder is used to spatio-temporal (t,x,y) decoding, outputting the predicted category of each pixel at time step t;

[0018] The disaster predictor is used to spatio-temporal (t,x,y) is predicted, and the fire probability of each pixel at time step t is output to form the predicted fire probability map P pred (t,x,y);

[0019] S3, construct the total loss L of the ST-SegNet network total , L total =λ1L seg +λ2L disaster ;

[0020] Where, L seg is the category loss, L disaster is the probability loss of fire, λ1 and λ2 are L seg 、L disaster The weighting coefficient of

[0021] S4, using the spatiotemporal dataset to minimize L total The ST-SegNet network was trained until convergence to obtain a large spatiotemporal spectrum model of forestry resources.

[0022] Preferably, the categories include vegetation, trees, wildlife, water bodies, roads, buildings, farmland, bushes and grasslands;

[0023] The geographic information G(x, y) is obtained by labeling the land cover types of the remote sensing images of the monitoring area, and the land cover types include forests, farmlands, water bodies and urban buildings;

[0024] The climate data C(t, x, y) is obtained by labeling the climate characteristics of the remote sensing image of the monitoring area at time step t, and the climate characteristics include temperature, humidity, precipitation, wind direction and air quality;

[0025] The geographic information and climate data are all marked at the pixel level.

[0026] As a preference: the SegNet encoder is based on the ViT network to X(t,x,y,λ all ) to perform feature extraction;

[0027] The feature fusion layer includes a feature extractor and a fusion layer. The feature extractor is used to extract features of geographic information G(x, y) and climate data C(t, x, y) respectively. The fusion layer is used to fuse the extracted features by weighting or merging.

[0028] The disaster predictor consists of a convolutional layer, a nonlinear activation function layer and a Sigmoid regression head.

[0029] As a preference: the dimension of the remote sensing image I(t,x,y,λ) is (T,r x ,r y ,B), spectrum graph X(t,x,y,λ all ) has the dimension (T,r x ,r y ,B), encoding feature F spatial (t,x,y,λ all ) has a dimension of (r x ,r y ,f dim ), time series feature F temporal The dimension of (t,x,y) is (T,r x ,r y ,f dim ), auxiliary data features F aux The dimension of (t,x,y) is (r x ,r y ,f dim ), comprehensive spatiotemporal features F spatio-temporal The dimension (T,r) of (t,x,y) x ,r y ,f dim ), f dim is the number of output channels of the SegNet encoder.

[0030] As a preference: L seg 、L disaster Calculate according to the following formula;

[0031] ,

[0032] In the formula, for the pixel point with coordinates (x, y), y true (x,y),y pred (x, y) are the true category and predicted category of the pixel respectively;

[0033] ,

[0034] Where N is the number of samples in a batch, n is the nth sample in a batch, 、 are the true fire probability map and predicted fire probability map of the nth sample respectively.

[0035] Compared with the prior art, the advantages of the present invention are:

[0036] (1) The present invention has good spatiotemporal consistency: By introducing the technology of spatiotemporal spectral data fusion, the present invention not only processes the image at each moment, but also effectively captures the relationship between spatial and temporal dimensions through the model. It also has good spatiotemporal consistency for the changes in continuous time series images. This method avoids the temporal disorder and spatial inconsistency problems that traditional methods easily produce in dynamic scenes, significantly improves the spatiotemporal consistency and classification accuracy of remote sensing image data, and improves the image processing quality.

[0037] (2) The present invention has good environmental adaptability: Through spatiotemporal data fusion and time series modeling technology, the present invention can capture long-term change trends and short-term emergencies. The model can not only accurately distinguish regular changes, but also quickly respond to emergencies (such as fire spread). With this advantage, the model can provide more accurate and reliable prediction results in complex scenarios such as forest fire monitoring and vegetation change identification, thereby enhancing its adaptability to complex scenarios, improving its ability to process details, and broadening its application scenarios.

[0038] (3) Emergency response capability: By introducing spatiotemporal data fusion technology into fire probability prediction, the present invention can promptly detect environmental anomalies such as forest fires or vegetation changes based on real-time remote sensing data. By analyzing time series and modeling spatial characteristics, the system can quickly identify potential threats and provide accurate fire probability and environmental change trends, thereby providing early warning in the early stages of disasters. This not only significantly improves the reaction speed of the emergency response system, but also enhances the system's prediction capabilities and decision-making support functions, which is of great significance to improving the efficiency of disaster response. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is the ST-SegNet network structure diagram of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0041] Example 1: See Figure 1 ,A method for constructing a large spatiotemporal spectrum model of forestry resources oriented to the forestry vertical field, includes the following steps;

[0042] S1, constructing a spatiotemporal spectrum dataset of forestry resources, including steps S11 to S14;

[0043] S11, obtain remote sensing images of the monitoring area at different time steps and different spectral bands. The remote sensing image obtained at time step t and spectral band λ is I(t,x,y,λ), t∈[1,T], T is the total number of time steps, λ∈[1,B], B is the total number of spectral bands, x and y are the width and height of a pixel in the remote sensing image, respectively, x∈[1,r x ], y∈[1,r y ],r x 、r y are the width and height of the remote sensing image respectively;

[0044] S12, merge the channels of B remote sensing images at time step t into λ all , get the spectrum X(t,x,y,λ at time step t all );

[0045] S13, obtain X(t,x,y,λ all ) auxiliary data, including geographic information G(x,y) and climate data C(t,x,y);

[0046] S14, for X(t,x,y,λ all ) each pixel is labeled with the true category and the true fire probability, the category includes fire, and the true fire probability of all pixels constitutes the true fire probability map, where X(t,x,y,λ all ) is the true fire probability map P true (t,x,y), X(t,x,y,λ all ) and its corresponding auxiliary data constitute samples, and all samples constitute a spatiotemporal spectrum dataset;

[0047] S2, construct the ST-SegNet network, including SegNet encoder, SegNet decoder, ConvLSTM model, feature fusion layer and merging layer;

[0048] The SegNet encoder is used to all ) to extract features and obtain the encoding feature F spatial (t,x,y,λ all );

[0049] The encoding features are constructed into a time series, and the ConvLSTM model is used to capture the time series features, where the encoding feature F spatial (t,x,y,λ all ) The corresponding time series feature is F temporal (t,x,y);

[0050] The feature fusion layer is used to extract features from geographic information and climate data and then fuse them to obtain auxiliary data features, where X(t, x, y, λ all )’s auxiliary data feature is F aux (t,x,y);

[0051] The merging layer is used to transform F spatial (t,x,y,λ all ), F temporal (t,x,y),F aux (t,x,y) are merged to obtain X(t,x,y,λ all )'s comprehensive spatiotemporal characteristics F spatio-temporal (t,x,y);

[0052] The SegNet decoder is used to spatio-temporal (t,x,y) decoding, outputting the predicted category of each pixel at time step t;

[0053] The disaster predictor is used to spatio-temporal (t,x,y) is predicted, and the fire probability of each pixel at time step t is output to form the predicted fire probability map P pred (t,x,y);

[0054] S3, construct the total loss L of the ST-SegNet network total , L total =λ1L seg +λ2L disaster ;

[0055] Where, L seg is the category loss, L disaster is the probability loss of fire, λ1 and λ2 are L seg 、L disaster The weighting coefficient of

[0056] S4, using the spatiotemporal dataset to minimize L total The ST-SegNet network was trained until convergence to obtain a large spatiotemporal spectrum model of forestry resources.

[0057] Once the spatiotemporal spectrum model of forestry resources is obtained, it can be used for real-time monitoring of the monitoring area, and for high-precision classification and fire prediction in real time. The real-time monitoring process is as follows:

[0058] 1. Data Acquisition Base Preparation: Remote sensing images are regularly collected from platforms such as satellites, drones, or fixed sensors. These images contain both temporal dimensions (multiple time steps) and spatial dimensions (each pixel in the image). Remote sensing images at each time step contain information from multiple wavelength bands, such as red, green, blue, and near-infrared. According to step S12 of the present invention, remote sensing images at different time steps are channel-merged to generate spectrograms at different time steps. Auxiliary data is then acquired according to step S13.

[0059] 2. Feature extraction: The SegNet encoder extracts the encoding features of each spectrogram; for time series modeling of encoding features, the ConvLSTM model is used to capture the time series features, and the feature fusion layer is used to process the auxiliary data to obtain the auxiliary data features; when modeling time series, the encoding features are constructed into a time series, with the step size represented by T, and each time series contains T encoding features.

[0060] 3. Feature merging: The coding features, time series features, and auxiliary data features are merged to obtain comprehensive spatiotemporal features;

[0061] 4. Segmentation and fire prediction: Use the SegNet decoder to segment the target, generate the predicted category for each pixel in the spectral map, and use the disaster predictor to generate a predicted fire probability map;

[0062] 5. Output fire probability map: Output the probability of fire occurrence for each time step and each spatial location for fire monitoring, early warning and emergency response.

[0063] Example 2: See Figure 1 , based on Example 1, the categories include vegetation, trees, wild animals, water bodies, roads, buildings, farmland, shrubs and grasslands;

[0064] The geographic information G(x, y) is obtained by labeling the land cover types of the remote sensing images of the monitoring area, and the land cover types include forests, farmlands, water bodies and urban buildings;

[0065] The climate data C(t, x, y) is obtained by labeling the climate characteristics of the remote sensing image of the monitoring area at time step t, and the climate characteristics include temperature, humidity, precipitation, wind direction and air quality;

[0066] The geographic information and climate data are all marked at the pixel level.

[0067] The SegNet encoder is based on the ViT network to all ) to perform feature extraction;

[0068] The feature fusion layer includes a feature extractor and a fusion layer. The feature extractor is used to extract features of geographic information G(x, y) and climate data C(t, x, y) respectively. The fusion layer is used to fuse the extracted features by weighting or merging.

[0069] The disaster predictor consists of a convolutional layer, a nonlinear activation function layer and a Sigmoid regression head.

[0070] The dimension of remote sensing image I(t,x,y,λ) is (T,r x ,r y ,B), spectrum X(t,x,y,λ all ) has the dimension (T,r x ,r y ,B), encoding feature F spatial (t,x,y,λ all ) has a dimension of (r x ,r y ,f dim ), time series feature F temporal The dimension of (t,x,y) is (T,r x ,r y ,f dim ), auxiliary data features F aux The dimension of (t,x,y) is (r x ,r y ,f dim ), comprehensive spatiotemporal features F spatio-temporal The dimension (T,r) of (t,x,y) x ,r y ,f dim ), f dim is the number of output channels of the SegNet encoder.

[0071] L seg 、L disaster Calculate according to the following formula;

[0072] ,

[0073] In the formula, for the pixel point with coordinates (x, y), y true (x,y),y pred (x, y) are the true category and predicted category of the pixel respectively;

[0074] ,

[0075] Where N is the number of samples in a batch, n is the nth sample in a batch, 、 are the true fire probability map and predicted fire probability map of the nth sample respectively.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a spatiotemporal spectrum large-scale model of forestry resources in the forestry vertical field, characterized by: The following steps are included: S1, constructing a spatiotemporal spectrum dataset of forestry resources, including steps S11 to S14; S11, obtain remote sensing images of the monitoring area at different time steps and different spectral bands. The remote sensing image obtained at time step t and spectral band λ is I(t,x,y,λ), t∈[1,T], T is the total number of time steps, λ∈[1,B], B is the total number of spectral bands, x and y are the width and height of a pixel in the remote sensing image, respectively, x∈[1,r x ], y∈[1,r y ],r x 、r y are the width and height of the remote sensing image respectively; S12, merge the channels of B remote sensing images at time step t into λ all , get the spectrum X(t,x,y,λ at time step t all ); S13, obtain X(t,x,y,λ all ) auxiliary data, including geographic information G(x,y) and climate data C(t,x,y); S14, for X(t,x,y,λ all ) each pixel is labeled with the true category and the true fire probability, the category includes fire, and the true fire probability of all pixels constitutes the true fire probability map, where X(t,x,y,λ all ) is the true fire probability map P true (t,x,y), X(t,x,y,λ all ) and its corresponding auxiliary data constitute samples, and all samples constitute a spatiotemporal spectrum dataset; S2, construct the ST-SegNet network, including SegNet encoder, SegNet decoder, ConvLSTM model, feature fusion layer and merging layer; The SegNet encoder is used to all ) to extract features and obtain the encoding feature F spatial (t,x,y,λ all ); The encoding features are constructed into a time series, and the ConvLSTM model is used to capture the time series features, where the encoding feature F spatial (t,x,y,λ all ) The corresponding time series feature is F temporal (t,x,y); The feature fusion layer is used to extract features from geographic information and climate data and then fuse them to obtain auxiliary data features, where X(t, x, y, λ all )’s auxiliary data feature is F aux (t,x,y); The merging layer is used to transform F spatial (t,x,y,λ all ), F temporal (t,x,y),F aux (t,x,y) are merged to obtain X(t,x,y,λ all )'s comprehensive spatiotemporal characteristics F spatio-temporal (t,x,y); The SegNet decoder is used to spatio-temporal (t,x,y) decoding, outputting the predicted category of each pixel at time step t; The disaster predictor is used to spatio-temporal (t,x,y) is predicted, and the fire probability of each pixel at time step t is output to form the predicted fire probability map P pred (t,x,y); S3, construct the total loss L of the ST-SegNet network total , L total =λ1L seg +λ2L disaster ; Where, L seg is the category loss, L disaster is the probability loss of fire, λ1 and λ2 are L seg 、L disaster The weighting coefficient of S4, using the spatiotemporal dataset to minimize L total The ST-SegNet network was trained until convergence to obtain a large spatiotemporal spectrum model of forestry resources.

2. The method for constructing a large spatiotemporal model of forestry resources for the forestry vertical field according to claim 1 is characterized by: Said categories include vegetation, trees, wildlife, water bodies, roads, buildings, farmland, shrubs, and grasslands; The geographic information G(x, y) is obtained by labeling the land cover types of the remote sensing images of the monitoring area, and the land cover types include forests, farmlands, water bodies and urban buildings; The climate data C(t, x, y) is obtained by labeling the climate characteristics of the remote sensing image of the monitoring area at time step t, and the climate characteristics include temperature, humidity, precipitation, wind direction and air quality; The geographic information and climate data are all marked at the pixel level.

3. The method for constructing a large spatiotemporal model of forestry resources for the forestry vertical field according to claim 1 is characterized by: The SegNet encoder is based on the ViT network to all ) to perform feature extraction; The feature fusion layer includes a feature extractor and a fusion layer. The feature extractor is used to extract features of geographic information G(x, y) and climate data C(t, x, y) respectively. The fusion layer is used to fuse the extracted features by weighting or merging. The disaster predictor consists of a convolutional layer, a nonlinear activation function layer and a Sigmoid regression head.

4. The method for constructing a large-scale spatiotemporal model of forestry resources for the forestry vertical field according to claim 1 is characterized by: The dimension of remote sensing image I(t,x,y,λ) is (T,r x ,r y ,B), spectrum graph X(t,x,y,λ all ) has the dimension (T,r x ,r y ,B), encoding feature F spatial (t,x,y,λ all ) has a dimension of (r x ,r y ,f dim ), time series feature F temporal The dimension of (t,x,y) is (T,r x ,r y ,f dim ), auxiliary data features F aux The dimension of (t,x,y) is (r x ,r y ,f dim ), comprehensive spatiotemporal features F spatio-temporal The dimension (T,r) of (t,x,y) x ,r y ,f dim ), f dim is the number of output channels of the SegNet encoder.

5. The method for constructing a large spatiotemporal model of forestry resources for the forestry vertical field according to claim 1 is characterized by: L seg 、L disaster Calculate according to the following formula; ， In the formula, for the pixel point with coordinates (x, y), y true (x,y),y pred (x, y) are the true category and predicted category of the pixel respectively; ， Where N is the number of samples in a batch, n is the nth sample in a batch, 、 are the true fire probability map and predicted fire probability map of the nth sample respectively.

Images