Forest disturbance prediction method and system based on optics and SAR dense time sequence Informer model

Abstract

The invention discloses a forest disturbance prediction method and system based on an optics and SAR dense time sequence Informer model, and the method comprises the following steps: S1, obtaining a remote sensing data set of which the time span is greater than ten years in an experimental area, and the remote sensing data set comprises optical satellite data and SAR satellite data; s2, collecting meteorological data, related to forest disturbance, of the experimental area; s3, using an MOGP model to fill the data value missing at the tj moment in the optical time sequence and the data value missing at the ti moment in the SAR time sequence; s4, obtaining a data organization form table of pixels in the Informer model; S5, constructing a forest disturbance prediction model based on an Informer network, designing an Informer network framework, and completing a forest disturbance prediction model construction module; and S6, carrying out forest disturbance prediction based on an Informer network. According to the method, on one hand, the problem of unreliable change detection caused by sparse single optical or single SAR remote sensing data is solved, and on the other hand, the development of a forest disturbance monitoring technology is promoted, and the timeliness of forest disturbance prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image time series change detection, and in particular relates to a forest disturbance prediction method and system based on an optical and SAR dense time series Informer model. Background Art

[0002] Forests are the largest carbon reservoirs in terrestrial ecosystems. Forest disturbances are changes in forests caused by natural and human factors. They refer to events that lead to a decrease or increase in forest coverage, including forest disturbances caused by human logging, fires, pests and diseases, meteorological disasters, etc., as well as forest recovery caused by artificial afforestation and natural regeneration. Since forest disturbances caused by forest fires are sudden disturbances, and forest disturbances caused by natural regeneration are gradual disturbances, it can be seen that forest disturbances caused by different factors are complex and diverse, which poses a challenge to the accurate detection of information such as the time of occurrence, intensity, and duration of forest disturbances.

[0003] Remote sensing technology has the characteristics of wide coverage, repeated observations, and many satellites in orbit. It is an efficient technical means to detect forest disturbance. Forest disturbance detection is interfered by complex factors such as climate fluctuations, vegetation phenology changes, and differences in atmospheric conditions, making it difficult to effectively extract forest disturbance information using dual-phase or multi-phase remote sensing images. Accordingly, the use of long-term remote sensing images can take into account the temporal information of remote sensing images, which helps to distinguish interference factors such as climate fluctuations and vegetation phenology changes from real forest disturbances.

[0004] In 2020, the Informer network was proposed, which improved the Transformer network algorithm in terms of long time series input, long time series output, and computational complexity, and has the ability to output prediction results of continuous time points at one time. In the past three years, it has been applied in wind speed forecasting, power system load forecasting, stock forecasting, etc. The data used in these application fields are single-point two-dimensional data, while remote sensing image data are spatiotemporal three-dimensional data. The resulting Informer model data organization problem is the first technical problem faced. In addition, under the Informer model data time equidistant requirement, single optical or SAR remote sensing image data is relatively sparse, and it is difficult to obtain equally spaced remote sensing data through a single optical image or a single SAR image. Therefore, the sparsity problem of single remote sensing image data is another technical problem faced by time series data analysis. Summary of the invention

[0005] In response to the above technical problems, the present invention proposes a forest disturbance prediction method and system based on optical and SAR dense time series Informer model. The method proposes a missing value filling method for optical and SAR image data based on the MOGP model to fill in sparse remote sensing time series data sets, optimizes the Informer model for spatiotemporal three-dimensional data, and takes into account the inducing factors of forest disturbance. It can provide sufficient data volume for subsequent forest disturbance prediction and improve the timeliness of forest disturbance prediction.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A forest disturbance prediction method based on optical and SAR dense time series informer model includes the following steps:

[0008] S1, obtain remote sensing data sets with a time span of more than ten years in the experimental area, including optical satellite data and SAR satellite data. The collected optical time series is expressed as The SAR timing is expressed as Among them, t i ,t j is the i and j moments;

[0009] S2, collect meteorological data related to forest disturbance in the experimental area;

[0010] S3, using MOGP model for optical timing In t j Missing data values ​​at the moment, SAR time series In t i If both optical and SAR have data at the same time, that is, t i =t j There are no missing values ​​for the temporal and SAR data. The specific steps are as follows:

[0011] S31, analyze and obtain optical timing and SAR timing Moderately strong correlation feature pairs are used as input data for the MOGP model;

[0012] S32, secondly, taking the strongly correlated feature pairs as input, the MOGP model based on the composite covariance function is used for training to learn the correlation between the optical time series and the SAR time series, and to fill in the optical feature values ​​and SAR feature values ​​of the missing time nodes;

[0013] In MOGP modeling, the covariance function is:

[0014] k MOGP (x, x′, l, l′) = k c (l, l′)×kt (x, x′),

[0015] Where, l, l' represents the index of SAR time series and optical time series, k c and k t Model the correlation of multiple outputs and the covariance in a single sequence respectively, where x,x' is the time index of output l,l';

[0016] S33, finally, we get the padded time series data sets TS{t 1 ,t 2 ,…,t j ,…t new-1};

[0017] S4, using the time series analysis method, obtain the historical disturbance type information of the forest in the experimental area. Next, using the pixel as the analysis unit, organize the filled time series image set, meteorological data and the historical disturbance type information of the forest in the experimental area to obtain the data organization form table of the pixel in the Informer model:

[0018]

[0019]

[0020] According to the data organization form in the table, all pixels in the experimental area are processed row by row and column by column to complete the data input task of the Informer model;

[0021] S5, build a forest disturbance prediction model based on the Informer network, design the Informer network framework, which consists of an encoder and a decoder. The encoder adopts a multi-head probabilistic sparse self-attention mechanism, and the input of the decoder is the time to be predicted t new and its previous short time series TS{t j ,…t new-1}, complete the forest disturbance prediction model construction module;

[0022] S6, Forest disturbance prediction based on Informer network.

[0023] In step S2, the meteorological data related to forest disturbance include: monthly maximum sustained wind speed data, monthly average wind speed data, monthly minimum temperature data, monthly maximum temperature data, monthly average temperature data and monthly precipitation data of the experimental area.

[0024] In step S4, the time series analysis method adopts the CCDC algorithm.

[0025] In step S4, the historical disturbance types of the forest in the experimental area include no disturbance, forest converted to non-forest, and non-forest converted to forest.

[0026] In step S5, the specific method of the multi-head probabilistic sparse self-attention mechanism in the encoder of the Informer network is to measure the distance between the attention probability and the unified probability according to the KL divergence, so that the key matrix only pays attention to u dominant Query matrices, where u can be set, and obtains u Query matrices under the sparsity measurement, thereby reducing the complexity of the calculation of each layer of the attention mechanism.

[0027] In step S5, the decoder of the Informer network is a generative reasoning decoder, which decodes the input data using a multi-head probabilistic sparse self-attention mechanism, and then combines the Informer model encoder and decoder through multi-head attention to obtain the prediction result of the time to be predicted at one time, thereby achieving the purpose of fast prediction.

[0028] In step S5, the input of the Informer network is the padded time series dataset TS{t 1 ,t 2 ,…,t j ,…t new-1}, we also need to fill in the time series dataset TS{t 1 ,t 2 ,…,t j ,…t new-1} performs position encoding, and position encoding is implemented using a sinusoidal function.

[0029] The specific operations of step S6 are:

[0030] Based on Python language, the Informer network model input module is input into the forest disturbance prediction model construction module of the Informer network, and the forest disturbance prediction model dimension, the number of heads of the multi-head attention mechanism, the dimension of the fully connected layer, the probability sparse attention mechanism factor, the encoding and decoding layers are set to obtain the time t to be predicted. new Forest disturbance prediction results.

[0031] The present invention further discloses a forest disturbance prediction system based on an optical and SAR dense time series informer model, comprising:

[0032] Data processing module: collect existing optical and SAR satellite image data, extract optical image features and SAR image features, use the MOGP model to fill the missing values ​​of optical and SAR sparse time series, and obtain the filled time series data set TS{t 1 ,t 2 ,…,tj ,…t new-1};

[0033] Model input module: CCDC algorithm is used to fill in the time series data set TS{t 1 ,t 2 ,…,t j ,…t new-1} to obtain the forest disturbance type of historical data; the filled time series dataset TS{t 1 ,t 2 ,…,t j ,…t new-1} image features, meteorological data, and forest disturbance type data are organized to obtain the model input data of the pixels, and all pixels in the experimental area are processed to complete the data input task of the Informer model;

[0034] Model building module: build the Informer network and design the encoder and decoder of the Informer network;

[0035] Model application module: Set the Informer model running parameters to obtain the time t to be predicted new Forest disturbance prediction results.

[0036] Compared with the prior art, the present invention has the advantages of:

[0037] (1) A method for filling missing values ​​in optical and SAR image data based on the MOGP model was proposed to obtain a dense long-term series image set, which solved the problem of sparse single remote sensing data and provided sufficient data volume for subsequent forest disturbance prediction.

[0038] (2) Under the Informer network, taking into account the meteorological factors related to forest disturbance, the model input data is scientifically organized to construct the Informer model, and a forest disturbance prediction method based on the Informer model is proposed. It not only tracks the historical laws of forest disturbance, but also improves the timeliness of forest disturbance monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 :Flowchart of the forest disturbance prediction method based on optical and SAR dense time series Informer model of the present invention;

[0040] Figure 2 : The Informer model network framework of the present invention;

[0041] Figure 3 : Experimental area. DETAILED DESCRIPTION

[0042] The specific implementation process of the present invention is described in detail below in conjunction with the accompanying drawings and examples. It should be noted that the specific embodiments described herein are only used to explain the present invention, and various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0043] Embodiment 1:

[0044] This embodiment provides a forest disturbance prediction method based on optical and SAR dense time series informer model, see Figures 1 to 3 . The following steps are included:

[0045] S1, obtain long-term remote sensing data sets, including optical satellite data and SAR satellite data. Optical satellites use the Landsat series, including Landsat 4 / 5 / 7 / 8 / 9 satellites, which have accumulated a large amount of historical archived data since July 1982, providing a rich data volume for obtaining historical data sets for forest disturbance monitoring. Therefore, the blue band, green band, red band, near-infrared band, and short-wave infrared band of Landsat4 / 5 / 7 / 8 / 9 were collected, including the NDVI (Normalized Difference Vegetation Index), NBR (Normalized Burning Ratio), EVI (Enhanced Vegetation Index), EVI2 (Enhanced Vegetation Index 2), and the brightness, greenness, and humidity of the tasseled cap transformation obtained by mathematical calculation of the above spectral bands. The SAR satellites used are European Space Agency SAR satellites, including ERS-1, ERS-2, ENVISAT, and Sentinel satellites, all of which are C-band SAR data. SAR satellite image data have been collected since 1991. After pre-processing such as filtering, SAR amplitude images are extracted.

[0046] S2, collects meteorological data related to forest disturbance.

[0047] The monthly maximum sustained wind speed data, monthly average wind speed data, monthly minimum temperature data, monthly maximum temperature data, monthly average temperature data and monthly precipitation data of the experimental area from 1982 to the present were collected.

[0048] S3, the MOGP model is used to fill the missing values ​​of optical and SAR sparse time series data.

[0049] Landsat satellite timing Indicates that the ESA SAR satellite timing is Using the MOGP model and The specific steps are as follows:

[0050] S31, first, analyze and obtain and Moderately strong correlation feature pairs are used as input data for the MOGP model.

[0051] S32, secondly, take the strongly correlated feature pairs as input, and use the MOGP model based on the composite covariance function for training, learn the correlation between the optical time series and the SAR time series, and fill in the optical and SAR feature values ​​of the missing time nodes. The composite covariance function not only considers the correlation between the same output corresponding to different input image features, but also considers the correlation between different outputs.

[0052] S33, finally, we obtain the time series datasets TS{t 1 ,t 2 ,…,t j ,…t new-1}.

[0053] S4, the CCDC algorithm is used to obtain the historical disturbance dataset of the forest in the experimental area.

[0054] The CCDC algorithm was used to obtain the historical forest disturbance monitoring information in the experimental area and obtain the disturbance types, including no disturbance, forest to non-forest, and non-forest to forest.

[0055] S5, construct a forest disturbance prediction model based on Informer network.

[0056] S51, first, after processing in step S3, it is found that the optical features NDVI and NBR are strongly correlated with the SAR amplitude features, so the three features of NDVI, NBR, and SAR amplitude are imported. Secondly, the six meteorological data obtained in step S2, the maximum sustained wind speed (MXSPD), the average wind speed (WDSP), the minimum temperature (MIN), the maximum temperature (MAX), the average temperature (AVE), and the precipitation (PRCP) are imported. Finally, the historical disturbance information (Dist) obtained in step S4 is imported to obtain the implementation case data organization form table:

[0057]

[0058] S52,Secondly, design the Informer network framework, such as Figure 2 shown.

[0059] Figure 2The encoder in adopts a multi-head probabilistic sparse self-attention mechanism, the purpose of which is to make the encoder pay attention to the important query matrix and its associated key matrix. The specific approach is to measure the distance between the attention probability and the unified probability according to the KL divergence, so that the key matrix only pays attention to u dominant query matrices, where u can be set, and obtain u query matrices under the sparsity measurement, thereby reducing the complexity of the attention mechanism calculation at each layer.

[0060] Figure 2 The input of the decoder in is the time t to be predicted new and its previous short time series TS{t j ,…t new-1 The Informer decoder is a generative reasoning decoder that decodes the input data using a multi-head probabilistic sparse self-attention mechanism. It then combines the Informer model encoder and decoder through multi-head attention to obtain t at one time. new The prediction results of the time interval are obtained to achieve the purpose of rapid prediction.

[0061] Figure 2 Positional encoding in: The input of the Informer model is in addition to the time series dataset TS{t 1 ,t 2 ,…,t j ,…t new}, we also need to embed the time series dataset TS{t 1 ,t 2 ,…,t j ,…t new} performs position encoding, and position encoding is implemented using a sinusoidal function.

[0062] S6, Forest disturbance prediction based on Informer network.

[0063] Based on Python language, set the model dimension, the number of heads of the multi-head attention mechanism, the dimension of the fully connected layer, the probabilistic sparse attention mechanism factor, the number of encoding and decoding layers and other parameters to get t new Forest disturbance prediction results at each moment.

[0064] Example 2

[0065] A forest disturbance prediction system based on optical and SAR dense time series informer model, including:

[0066] Data processing module: collect existing optical and SAR satellite image data, analyze the characteristics of optical and SAR images, and conclude that the optical image NDVI, NBR characteristics and SAR amplitude characteristics have a strong correlation. Use the MOGP model to fill the missing values ​​of the sparse time series of NDVI, NBR and SAR amplitude characteristics to obtain the filled time series data set TS{t 1 ,t 2 ,…,t j ,…t new-1}.

[0067] Model input module: CCDC algorithm is used to train the data set TS{t 1 ,t 2 ,…,t j ,…t new-1} to obtain the forest disturbance type of historical data. 1 ,t 2 ,…,t j ,…t new-1 The characteristic data of the experimental area, meteorological data (maximum sustained wind speed (MXSPD), average wind speed (WDSP), minimum temperature (MIN), maximum temperature (MAX), average temperature (AVE), precipitation (PRCP)), and forest disturbance type data (Dist) are organized to obtain the model input data of the pixel, and then processed row by row and column by column in the same way to obtain the model input data of the entire experimental area.

[0068] Model building module: build the Informer network and design the encoder and decoder of the Informer network.

[0069] Model application module: Set the Informer model operation parameters and obtain the new observation time t new Forest disturbance prediction results.

[0070] Although the preferred implementation cases of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms of specific changes without departing from the scope of protection of the invention spirit and claims, which all fall within the scope of protection of the present invention.

Claims

1. A forest disturbance prediction method based on optical and SAR dense time series informer model, characterized in that: The following steps are involved: S1, obtain remote sensing data sets with a time span of more than ten years in the experimental area, including optical satellite data and SAR satellite data. The collected optical time series is expressed as The SAR timing is expressed as Among them, t i ,t j is the i and j moments; S2, collect meteorological data related to forest disturbance in the experimental area; S3, using MOGP model for optical timing In t j Missing data values ​​at the moment, SAR time series In t i If both optical and SAR have data at the same time, that is, t i =t j There are no missing values ​​for the temporal and SAR data. The specific steps are as follows: S31, analyze and obtain optical timing and SAR timing Moderately strong correlation feature pairs are used as input data for the MOGP model; S32, secondly, taking the strongly correlated feature pairs as input, the MOGP model based on the composite covariance function is used for training to learn the correlation between the optical time series and the SAR time series, and to fill in the optical feature values ​​and SAR feature values ​​of the missing time nodes; In MOGP modeling, the covariance function is: k MOGP (x,x′,l,l′)=k c (l,l′)×k t (x,x′), Where, l, l' represents the index of SAR time series and optical time series, k c and k t Model the correlation of multiple outputs and the covariance in a single sequence respectively, where x,x' is the time index of output l,l'; S33, finally, we obtain the padded time series data sets TS{t1, t2, …, t j ,…t new-1 }; S4, using the time series analysis method, obtain the historical disturbance type information of the forest in the experimental area. Next, using the pixel as the analysis unit, organize the filled time series image set, meteorological data and the historical disturbance type information of the forest in the experimental area to obtain the data organization form table of the pixel in the Informer model: According to the data organization form in the table, all pixels in the experimental area are processed row by row and column by column to complete the data input task of the Informer model; S5, build a forest disturbance prediction model based on the Informer network, design the Informer network framework, which consists of an encoder and a decoder. The encoder adopts a multi-head probabilistic sparse self-attention mechanism, and the input of the decoder is the time t to be predicted. new and its previous short time series TS{t j ,…t new-1 }, complete the forest disturbance prediction model construction module; S6, Forest disturbance prediction based on Informer network.

2. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S2, the meteorological data related to forest disturbance include: monthly maximum sustained wind speed data, monthly average wind speed data, monthly minimum temperature data, monthly maximum temperature data, monthly average temperature data and monthly precipitation data of the experimental area.

3. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S4, the time series analysis method adopts the CCDC algorithm.

4. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S4, the historical disturbance types of the forest in the experimental area include no disturbance, forest converted to non-forest, and non-forest converted to forest.

5. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S5, the specific method of the multi-head probabilistic sparse self-attention mechanism in the encoder of the Informer network is to measure the distance between the attention probability and the unified probability according to the KL divergence, so that the key matrix only pays attention to u dominant Query matrices, where u can be set, and obtains u Query matrices under the sparsity measurement, thereby reducing the complexity of the calculation of each layer of the attention mechanism.

6. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S5, the decoder of the Informer network is a generative reasoning decoder, which decodes the input data using a multi-head probabilistic sparse self-attention mechanism, and then combines the Informer model encoder and decoder through multi-head attention to obtain the prediction result of the time to be predicted at one time, thereby achieving the purpose of fast prediction.

7. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: In step S5, the input of the Informer network is the padded time series dataset TS{t1,t2,…,t j ,…t new-1 }, it is also necessary to fill in the time series dataset TS{t1,t2,…,t j ,…t new-1 } performs position encoding, and position encoding is implemented using a sinusoidal function.

8. The forest disturbance prediction method based on optical and SAR dense time series informer model according to claim 1 is characterized in that: The specific operations of step S6 are: Based on Python language, the Informer network model input module is input into the forest disturbance prediction model construction module of the Informer network, and the forest disturbance prediction model dimension, the number of heads of the multi-head attention mechanism, the dimension of the fully connected layer, the probability sparse attention mechanism factor, the encoding and decoding layers are set to obtain the time t to be predicted. new Forest disturbance prediction results.

9. A forest disturbance prediction system based on optical and SAR dense time series informer model, characterized in that: include: Data processing module: collect existing optical and SAR satellite image data, extract optical image features and SAR image features, use the MOGP model to fill the missing values ​​of optical and SAR sparse time series, and obtain the filled time series data set TS{t1, t2, …, t j ,…t new-1 }; Model input module: CCDC algorithm is used to fill in the time series data set TS{t1,t2,…,t j ,…t new-1 } to obtain the forest disturbance type of historical data; the filled time series data set TS{t1,t2,…,t j ,…t new-1 } image features, meteorological data, and forest disturbance type data are organized to obtain the model input data of the pixels, and all pixels in the experimental area are processed to complete the data input task of the Informer model; Model building module: build the Informer network and design the encoder and decoder of the Informer network; Model application module: Set the Informer model running parameters to obtain the time t to be predicted new Forest disturbance prediction results.

Images