Vegetation coverage evaluation method and device based on cooperative attention state space network

Abstract

The present application provides a kind of vegetation coverage evaluation method and device based on collaborative attention state space network, solve the technical problem of low identification ability of existing vegetation monitoring evaluation model to vegetation growth disturbance factor.Method includes: the following optimization of state space network model: increase the attention weight of vegetation global feature by transfer equation;Dynamic aggregation of model parameters is formed according to observation data;According to the observation equation, form the discrete channel to identify disturbance events and the continuous channel of gradual change process;In the process of multi-source fusion of observation data, frequency domain feature enhancement is carried out;Vegetation index optimization and correction are formed in the process of observation data processing by state space network model.Break the sequence constraint of traditional model, realize global vegetation state perception.Construct discrete and continuous double channel to distinguish drought stress and pest and disease and other ecological semantics.Using parameter dynamic aggregation improves the season adaptability to observation data.Improve the measurement dimension and accuracy of observation data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vegetation coverage evaluation, and in particular to a vegetation coverage evaluation method and device based on a collaborative attention state space network. BACKGROUND

[0002] In the prior art, ecological quality evaluation usually adopts a state space network model to process and classify vegetation monitoring data, and evaluates the ground vegetation coverage degree according to the characteristic data (such as infrared, near-infrared, etc.) of the spatial form of vegetation growth. For example, the normalized vegetation index (NDVI) formed by processing remote sensing telemetry data through a vegetation state space network model is taken as an important index representing the growth condition of ground vegetation. However, vegetation growth is jointly affected by natural conditions (such as rainfall, temperature, soil moisture, etc.) and human activities (such as returning farmland to forest, urbanization, etc.), and extreme climate events such as drought or flood can cause abnormal vegetation growth, resulting in a decrease in the vegetation coverage index value. This decrease reflects the influence of climate disasters, rather than the influence of human activities on the ecological environment, so that the ecological quality evaluation result for human activities is distorted, and the effectiveness of ecological protection cannot be truly reflected. The technical reason for the above evaluation defects is that:

[0003] Vegetation monitoring faces special challenges such as multi-scale feature coupling, nonlinear response, and confusion of interference sources, and needs to process remote sensing data of different scales at the same time. However, the state space network model of the prior art is limited by sequence modeling, cannot realize global interaction of vegetation features, and is difficult to balance local details and long-range dependencies. At the same time, the static parameters of the model are difficult to adapt to the seasonal changes of vegetation growth, and there is a lagging effect in response to climate change, and it is also impossible to separate natural disturbances such as drought, flood and pest from the influence of human activities. SUMMARY

[0004] In view of the above problems, the embodiments of the present application provide a vegetation coverage evaluation method and device based on a collaborative attention state space network, which solves the technical problem of low identification ability of existing vegetation monitoring evaluation models for vegetation growth interference factors.

[0005] The vegetation coverage evaluation method based on the collaborative attention state space network of the embodiments of the present application comprises:

[0006] The following optimizations of the state space network model are performed:

[0007] The attention weight of the global feature of the vegetation is increased through the transfer equation;

[0008] The model parameter dynamic aggregation is formed according to the observation data;

[0009] According to the observation equation, a discrete channel for identifying interference events and a continuous channel for gradual processes are formed;

[0010] Frequency domain feature enhancement is performed in the process of multi-source fusion of observation data;

[0011] Vegetation index optimization and correction are formed in the process of state space network model processing observation data.

[0012] In an embodiment of the present application, the optimization includes:

[0013] The attention position offset weight is embedded in the state transition equation of the state space network, and the contribution degree of different global features in the historical state to the state update is dynamically adjusted through the attention position offset weight;

[0014] A parameter dynamic aggregation parallel framework is formed, and a dynamic convolution kernel is used to extract observation data features;

[0015] Discrete channels and continuous channels of state space network decoding output are formed, interference factors are classified through the discrete channels, and the gradual change degree of vegetation is quantified through the continuous channels.

[0016] In an embodiment of the present application, the frequency domain feature enhancement in the process of multi-source fusion of observation data includes:

[0017] The multi-source remote sensing data is unified into observation data through bilinear interpolation;

[0018] The frequency domain feature enhancement is formed by frequency domain conversion-noise suppression-multiscale feature fusion on the observation data.

[0019] In an embodiment of the present application, the vegetation index optimization and correction formed in the process of state space network model processing observation data includes:

[0020] The state space network model is trained and deployed through historical observation data;

[0021] The vegetation index formed by the state space network model is optimized;

[0022] The vegetation index is corrected by interference factors.

[0023] The vegetation coverage evaluation device based on the collaborative attention state space network in the embodiment of the present application includes:

[0024] The model optimization module is used for the following optimization of the state space network model:

[0025] The attention weight of the global feature of the vegetation is increased through the transition equation;

[0026] The model parameter dynamic aggregation is formed according to the observation data;

[0027] According to the observation equation, a discrete channel for identifying interference events and a continuous channel for gradual processes are formed;

[0028] The data enhancement module is used for frequency domain feature enhancement in the process of multi-source fusion of observation data;

[0029] The optimization correction module is used for vegetation index optimization and correction in the process of state space network model processing observation data.

[0030] In an embodiment of the present application, the model optimization module comprises:

[0031] The global attention construction unit is used for embedding attention position offset weight in the state transition equation of the state space network, and dynamically adjusting the contribution degree of different global features in the historical state to the state update through the attention position offset weight;

[0032] The parameter aggregation construction unit is used for forming a parameter dynamic aggregation parallel framework, and extracting observation data features by using a dynamic convolution kernel;

[0033] The dual-type channel construction unit is used for forming a discrete channel and a continuous channel of the state space network decoding output, classifying interference factors through the discrete channel, and quantifying the vegetation gradual change degree through the continuous channel.

[0034] In an embodiment of the present application, the data enhancement module comprises:

[0035] The observation data formation unit is used for unifying multi-source remote sensing data into observation data through bilinear interpolation;

[0036] The frequency domain feature enhancement unit is used for frequency domain conversion-noise suppression-multiscale feature fusion to form frequency domain feature enhancement on the observation data.

[0037] In an embodiment of the present application, the optimization correction module comprises:

[0038] The model training unit is used for training and deploying the state space network model through historical observation data

[0039] The index optimization unit is used for optimizing the vegetation index formed by the state space network model

[0040] The interference correction unit is used for interference factor correction on the vegetation index.

[0041] The electronic device of the embodiment of the present application comprises:

[0042] The processor, the memory, and the interface in communication with the gateway;

[0043] The memory is used to store programs and data, and the processor calls the programs stored in the memory to execute the above method.

[0044] The computer readable storage medium of the embodiment of the present application comprises a program, which is used for the above method when executed by a processor.

[0045] The vegetation coverage evaluation method based on the collaborative attention state space network of the embodiment of the present application breaks through the sequence modeling constraint by forming a non-causal attention mechanism through targeted optimization of the existing evaluation model and observation data, breaks the sequence constraint of the traditional state space model, and realizes global vegetation state perception. The discrete and continuous dual-channel hidden variable architecture is constructed to realize the decoupling of the interference type and intensity, and to realize the differentiation of ecological semantics such as drought stress and pest and disease. The parameter dynamic aggregation is used to improve the seasonal adaptability of the observation data to the response lag effect of climate change. At the same time, a multi-frequency domain feature extraction process of vegetation response characteristics is formed, which improves the measurement dimension and accuracy of the observation data. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The flowchart of the vegetation coverage evaluation method based on the collaborative attention state space network of the embodiment of the present application is shown.

[0047] Figure 2 The application process schematic diagram of the vegetation coverage evaluation method based on the collaborative attention state space network of the embodiment of the present application is shown.

[0048] Figure 3 The architecture schematic diagram of the vegetation coverage evaluation device based on the collaborative attention state space network of the embodiment of the present application is shown.

[0049] Figure 4 The architecture schematic diagram of the electronic device of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The vegetation coverage evaluation method based on the collaborative attention state space network of the embodiment of the present application is shown as Figure 1 In Figure 1 , the embodiment includes:

[0052] Step 100: the following optimization of the state space network model is performed:

[0053] The attention weight of the global feature of the vegetation is increased through the transition equation;

[0054] The model parameter dynamics are aggregated according to the observation data;

[0055] The discrete channel for identifying interference events and the continuous channel for gradual processes are formed according to the observation equation.

[0056] As can be understood by those skilled in the art, the state space network comprises a state transition equation and an observation equation. The state transition equation is used to update the current state according to the state of the last moment and the current input, and the transition between states is realized through a nonlinear function, which can be realized by a neural network. The observation equation maps the state vector to the output at the corresponding moment. Since the dependence of state update on the historical state and the current input is fixed (according to predefinition or static learning), the existing state space network cannot realize global perception of vegetation features.

[0057] The attention weight of the global feature of the vegetation is increased through the transition equation, a collaborative attention mechanism is established, so that the state update dynamically adjusts the contribution of each part of the data according to the importance of the features at different positions (or time points) in the global data, which can adapt to the vegetation change pattern under different space-time conditions.

[0058] The convolution kernel of the state space network model is self-adapted to the input data through parameter dynamic aggregation, so that the convolution kernel focuses on key information and improves the flexibility of feature extraction, and more accurately adapts to different climates and seasonal scenes.

[0059] The discrete channel output of the probability distribution of the corresponding main interference factors and the continuous channel output of the quantitative vegetation sustained change are formed through the decoding process of the state vector formed by the observation equation. The natural interference and human activity influence are effectively separated. The distortion of the evaluation result caused by the existing compound interference is avoided.

[0060] Step 200: Frequency domain feature enhancement in the process of multi-source fusion of observation data.

[0061] Different data sources of observation data have significant time sequence characteristics and contain time domain features. Through frequency domain feature enhancement of multi-source fused observation data, additional multi-scale vegetation response features of observation data can be obtained. Different frequencies correspond to different vegetation dynamic laws. Through frequency domain transformation and processing, target features are highlighted and interference is suppressed.

[0062] Step 300: Vegetation index optimization and correction formed in the process of state space network model processing observation data.

[0063] The processing of the observation data by the collaborative attention-based state space network formed after training needs to optimize the vegetation index output by the observation equation according to actual conditions, improve performance, and adapt to specific needs. Interference is eliminated by correction to restore the true ground signal. The optimization methods include non-causal attention calculation, parallel processing of discrete and continuous channels, dynamic benchmark construction, etc. The correction methods include interference quantification and index correction.

[0064] The vegetation coverage evaluation method based on the collaborative attention state space network of the embodiment of the application breaks through the sequence modeling constraint and breaks the sequence constraint of the traditional state space model by forming a non-causal attention mechanism through targeted optimization of the existing evaluation model and observation data, realizing global vegetation state perception. The discrete and continuous dual-channel hidden variable architecture is constructed to realize the decoupling of interference types and intensity, and to realize the differentiation of ecological semantics such as drought stress and pest and disease. The parameter dynamic aggregation is used to improve the seasonal adaptability of the observation data to the response lag effect of climate change. At the same time, a multi-frequency domain feature extraction process of vegetation response features is formed, which improves the measurement dimension and accuracy of the observation data.

[0065] As shown in Figure 1 , in an embodiment of the application, step 100 includes:

[0066] Step 110: embedding attention position offset weights in the state transition equation of the state space network, and dynamically adjusting the contribution degree of different global features in the historical state to the state update through the attention position offset weights.

[0067] The attention position offset weights dynamically calculate the influence of various global features in the historical state on the local features of the changing observation data. The more critical historical experience / information is selected from the global features, and the judgment of the present state is optimized comprehensively. The attention mechanism formed by the attention position offset weights enables the model to better capture long-distance, non-local useful associations when processing long sequences (such as multi-period remote sensing data of vegetation and time series monitoring signals).

[0068] In an embodiment of the application,

[0069]

[0070] wherein s t is the updated state, s t-1 is the historical state, s j is a set of related states, Ω is the global pixel set, and α tj is the attention position offset weight, which is dynamically generated by the feature similarity.

[0071] Step 120: forming a parameter dynamic aggregation parallel framework, and using a dynamic convolution kernel to extract observation data features.

[0072] The dynamic aggregation of multiple convolution kernels combines different feature extraction methods in a nonlinear manner, thereby having stronger representation ability and being able to better capture complex features in input data.

[0073] In an embodiment of the present application, the dynamic weight of the convolution kernel is

[0074] W dy =Sigmoid(g(x t ))⊙W base

[0075] wherein W dy is the dynamic weight of the convolution kernel adapted to the current input, W base is a predefined basic weight, x t is the feature of the current input observation data (processed by function g), and the Sigmoid function is used to generate a coefficient between 0 and 1.

[0076] The framework includes multiple parallel convolution kernels with different weights and characteristics. The weight of each convolution kernel is calculated through an attention mechanism. The attention mechanism generates an attention weight for each convolution kernel according to the features of the input data, representing the importance of the convolution kernel for the current input. Then, the multiple convolution kernels are dynamically aggregated according to the attention weights to obtain the final weight used for convolution operation. Since the attention weight is dynamically calculated according to the observation data, the convolution kernels participating in aggregation and their weights will change for different inputs, thereby realizing the dynamic adjustment of the convolution kernel. This mechanism enables the model to automatically select the most suitable combination of convolution kernels according to the characteristics of the input data, enhancing the model's ability to capture complex data patterns.

[0077] Step 130: Forming discrete channels and continuous channels of state space network decoding output, classifying interference factors through discrete channels, and quantifying vegetation gradual change degree through continuous channels.

[0078] Through the cooperative work of discrete channels and continuous channels, fine modeling of feature data output by the state space network model observation equation is realized.

[0079] In an embodiment of the present application, the discrete channel is used for classification and identification of disaster states, and the construction process includes:

[0080] According to the observation equation output, extract key indicators such as spectral index (e.g. NDVI), precipitation outliers, and temperature fluctuations;

[0081] Use clustering methods (such as K-means or fuzzy C-means clustering (FCM)) to pre-classify historical disaster data and generate initial state labels such as drought, flood, and pest;

[0082] Introduce attention mechanism with clustering labels as supervision signal, dynamically focus on features strongly related to disaster types (e.g. NDVI drop for drought, precipitation anomaly for flood) through e.g. convolutional neural network;

[0083] Output discrete state probability distribution (e.g. the probability of a region being in drought state is 85%)

[0084] In an embodiment of the invention, continuous channels are used for the quantitative identification of vegetation gradual change processes, and the construction process includes:

[0085] Form a variational autoencoder (VAE) architecture:

[0086] Encoder: map time series data (e.g. NDVI time series) to latent variable space z, learn low-dimensional representation of continuous features such as growth trend, recovery rate, etc.

[0087] Form latent variable constraint: force latent variable to follow standard normal distribution through KL divergence, ensure continuity and interpretability;

[0088] Decoder: reconstruct the original data from the latent variable, minimize reconstruction loss (e.g. mean square error), while preserving gradual change information (e.g. changes in vegetation recovery rate);

[0089] Feature enhancement: combine wavelet transform to decompose time series data into different frequency components, highlight short-term fluctuations (e.g. drought stress) and long-term trends (e.g. ecological recovery);

[0090] Gated recurrent unit (GRU): capture dynamic dependencies in time series, improve modeling accuracy of gradual change processes;

[0091] Output continuous latent variable z (e.g. quantitative value of vegetation recovery rate).

[0092] As shown in Figure 1 In an embodiment of the invention, step 200 includes:

[0093] Step 210: unify multi-source remote sensing data into observation data through bilinear interpolation.

[0094] In two-dimensional space, bilinear interpolation is used to calculate the interpolation of any position (non-grid point) by using the weighted average of the surrounding four known points, and to smoothly unify multi-source remote sensing data (such as satellite images of different resolutions) into a fixed scale spatial grid to form spatio-temporal alignment, and to form a data set with fixed spatial grid (such as ten meters).

[0095] In an embodiment of the invention, the source data includes:

[0096] MODIS optical remote sensing data, spatial resolution 500 meters, temporal resolution daily;

[0097] Landsat optical remote sensing data, spatial resolution 30 meters, time resolution 16 days;

[0098] ERA5 meteorological data, spatial resolution 10 kilometers, time resolution every hour;

[0099] CMADS meteorological data, spatial resolution 1 kilometer, time resolution every hour;

[0100] Ground ecological monitoring station data, point scale measurement, daily collection;

[0101] SRTM topographic data, spatial resolution 30 meters, static data.

[0102] In an embodiment of the present application, the pre-processing of the observation data includes:

[0103] Cloud mask processing to remove the interference of clouds on optical data;

[0104] Atmospheric correction to eliminate the effects of atmospheric scattering and absorption;

[0105] Feature extraction to calculate twelve vegetation indices such as NDVI, EVI and LSWI.

[0106] Step 220: Frequency domain conversion-noise suppression-multiscale feature fusion to form frequency domain feature enhancement.

[0107] By combining frequency analysis from time domain to frequency domain with dynamic purification of spatial features and multiscale fusion, and by adjusting dynamic parameters to adapt to different data scenarios (such as vegetation growing season / dormant season, normal / disaster state), long-term trends, short-term dynamics and spatial heterogeneity of data can be accurately captured. Dynamic adaptive data features are especially suitable for vegetation growth monitoring, disaster detection and other scenarios. Finally, high-value features with optimized noise and dimension and clear physical meaning are output.

[0108] In an embodiment of the present application, the frequency domain feature enhancement process includes:

[0109] a. Using frequency domain conversion to decompose continuous / discrete signals in time domain into different frequency components. Specifically, it includes:

[0110] Input time domain sequence features (such as vegetation NDVI time series data: time step T x feature dimension D, D can include NDVI, EVI, precipitation, etc.);

[0111] Divide long time series data (such as 10-year NDVI) according to "sliding window" (such as window length 12 months, corresponding to the annual growth cycle of vegetation);

[0112] - Perform FFT on each window, convert time-domain signal to "amplitude spectrum (frequency intensity) + phase spectrum (timing phase)", decompose three types of frequency components:

[0113] Low-frequency components (such as interannual scale vegetation degradation trend);

[0114] Medium-frequency components (such as seasonal scale "greening-peak-yellowing" cycle);

[0115] High-frequency components (such as short-term drought, disease and pest-induced NDVI drop, or sensor noise);

[0116] In an embodiment of the present application, the time-domain features are converted to the frequency domain using the following mathematical transformation:

[0117]

[0118] Extract 0-3Hz low-frequency signals to represent seasonal changes in vegetation; extract 3-10Hz medium-frequency signals to represent inter-monthly changes in vegetation; extract 10Hz or higher high-frequency signals to represent sudden interference events;

[0119] - Output multi-window frequency domain feature map (window number x frequency dimension x feature dimension).

[0120] b. Suppress high-dimensional feature (containing noise, redundant information) fluctuations, focus on effective signals, dynamically allocate weights by modeling the spatial / frequency dependence between features, suppress fluctuating noise, and strengthen effective features. Specifically, it includes:

[0121] - Input high-dimensional frequency domain features (containing high-frequency noise, such as abnormal high-frequency signals caused by cloud cover in vegetation data);

[0122] - Dependence matrix calculation: calculate the "spatial-frequency similarity" between features (such as the similarity of the medium-frequency seasonal features of a certain pixel to adjacent pixels), generate a dependence matrix, and depict "which features are related and useful";

[0123] - Attention weight allocation: based on the dependence matrix, use Softmax to generate attention weights, assign high weights to "features strongly related to the target task" (such as high-frequency abnormal features of vegetation diseases and pests), and assign low weights to "fluctuating noise" (such as random sensor interference);

[0124] - Feature purification output: multiply the frequency domain features and attention weights element by element, suppress high-dimensional fluctuations (noise weight tends to 0), and retain effective features (such as medium-frequency features of seasonal cycles and high-frequency abnormal features of disasters).

[0125] c. Weighted fusion of multi-dimensional features enhances task adaptability. Multi-source features at both the "frequency scale" (low, medium, and high frequency) and "spatial scale" (e.g., remote sensing features at different resolutions) are adaptively weighted and fused to highlight key scale information relevant to the current task, avoiding the limitations of single-scale features. Specifically, this includes:

[0126] - Input the purified multi-frequency scale features (low-frequency trend, mid-frequency seasonality, high-frequency anomaly), combined with the original data's "multi-spatial scale features" (such as Sentinel-2's 10m resolution and Landsat's 30m resolution features).

[0127] - Perform scale feature mapping: unify features of different scales to the same dimension (e.g., adjust the number of channels through 1×1 convolution) to ensure fusion;

[0128] - Perform adaptive weight calculation: dynamically generate weights based on "feature contribution"; for example, give high weight to "mid-frequency seasonal scale features" in phenological extraction tasks (highlighting the greening and withering cycles); give high weight to "high-frequency anomaly scale features" in drought detection tasks (highlighting the short-term dynamics of NDVI drop); give high weight to "high spatial resolution features" in areas with strong spatial heterogeneity (such as distinguishing the boundaries between farmland and forest land).

[0129] - Weighted fusion and output: Multi-scale features are summed according to weights to output the final encoded features with reduced dimensions and focused information (such as the comprehensive feature vector of vegetation growth status).

[0130] like Figure 1 As shown, in one embodiment of the present invention, step 300 includes:

[0131] Step 310: Train and deploy the state-space network model using historical observation data.

[0132] In one embodiment of the present invention, observational data formed from historical remote sensing and telemetry data is first used for pre-training. Specifically, ten years of historical data are used, spanning from 2013 to 2024, with a total data volume of 120TB; the batch size is set to 128 spatial units, each unit covering an area of ​​10×10 square kilometers; the AdamW optimizer is used, with a learning rate set to 0.001 and a weight decay coefficient of 0.01; training is performed for 500 epochs, and an early stopping mechanism is enabled, terminating training when the verification loss does not decrease for 20 consecutive epochs.

[0133] Secondly, a stable model is deployed on edge nodes. Specifically, model optimization is performed for edge devices, employing FP8 quantization technology to compress the model size to 38% of the original version; convolutional layers and activation functions are fused to reduce input / output operations; and 16-way concurrent processing is supported to improve system throughput. In one embodiment of this invention, the edge node is located in the monitoring area, and its hardware configuration includes an NVIDIA Jetson AGX Orin processor, equipped with 32GB of memory; 1TB NVMe solid-state drive and 4TB mechanical hard drive storage; dual redundant network connections of 5G and Gigabit Ethernet; total power consumption of less than 60 watts; and support for solar power. An Ubuntu 22.04LTS operating system is installed; the TensorRT 8.6 inference engine is deployed; gRPC data interface service is configured to support MODIS, Landsat, and Sentinel data stream access; and a Python 3.10 and PyTorch 2.1 model runtime environment is established.

[0134] Secondly, online incremental learning is conducted. Specifically, the model is continuously optimized after deployment. The optimization process includes, but is not limited to, collecting new data, detecting concept drift, updating model parameters, and validating model performance. The sliding window is set to 90 days of time-series data; incremental updates are performed weekly; and an elastic memory mechanism is adopted, with a regularization coefficient β set to 0.01 to prevent catastrophic forgetting.

[0135] Step 320: Optimize the vegetation index generated by the state-space network model.

[0136] In one embodiment of the present invention, non-causal attention calculation is performed; specifically, global cell association is calculated using the following attention weight formula:

[0137]

[0138] Wherein, molecule: exp(φ(s) i )Tψ(s j The correlation between global pixels si and sj is amplified by taking their inner product after feature mapping (φ and ψ are feature transformation functions) and then applying an exponential function (exp). This can be simply understood as global pixel s... i and global pixels s j The matching score indicates a higher relevance. Denominator: ∑k ∈Ω exp(φ(s i )Tψ(s k )) is to convert the global pixel s i and all candidate elements s in set Ω k Calculate all the matching scores and then sum them up.

[0139] α ij The final global pixel si The overall attention weight of the global pixel sj is a coefficient between 0 and 1, and the greater the value, the more the global pixel sj is concerned about the global pixel sj, and the subsequent processing will allocate more importance to the global pixel sj. In an embodiment of the present application, the spatial range is set to a twenty-kilometer radius neighborhood; the feature dimension is one hundred and twenty-eight-dimensional latent variable; and a sparse matrix is used to accelerate the calculation process. i

[0140] In an embodiment of the present application, dual-channel state decoding is performed; specifically, the parallel processing flow of discrete and continuous channels includes input feature processing, discrete channel classification, continuous channel regression, and result fusion; the discrete channel outputs several categories of ecological disturbance probability distribution; the continuous channel outputs continuous values such as recovery rate and stress intensity; and the fusion weight is automatically adjusted through a dynamic learning layer.

[0141] In an embodiment of the present application, a dynamic benchmark is constructed; specifically, a vegetation growth benchmark is constructed using the following formula:

[0142] Benchmark t =α·LT t +(1-α)·ST t

[0143] Wherein, Benchmark t : "Benchmark value" at time t (such as a reference standard for a certain indicator, a comprehensive evaluation value).

[0144] LT t : "Long-term term" at time t (which can be understood as a long-term trend, a long-term performance value, such as a long-term average, a long-term regularity).

[0145] ST t : "Short-term term" at time t (which can be understood as a short-term fluctuation, an immediate performance value, such as the latest value at the moment, a short-term change).

[0146] α: "Weight coefficient" (range between 0 and 1), used to adjust the influence degree of long-term and short-term factors on the benchmark value. In an embodiment of the present application, the long-term benchmark is calculated based on twenty-year climate average; the short-term benchmark uses an adaptive time window, ranging from thirty days to three hundred and sixty-five days; and the fusion coefficient α is dynamically determined according to the vegetation type, with a value range of zero point three to zero point seven.

[0147] Step 330: Perform interference factor correction on the vegetation index.

[0148] In an embodiment of the present application, interference quantization is performed. The natural interference intensity is calculated using the following weighted sum formula:

[0149] Δ t =λ​d ·I d +λ c ·I c

[0150] Wherein Id represents discrete interference intensity, the value range is 0 to 1;Ic represents continuous interference quantity, after standardization processing;λ coefficient is the regional adaptive weight.

[0151] In an embodiment of the present application, exponential correction is carried out. Perform vegetation coverage index correction, use the following formula:

[0152] VCI corr =VCI obs -Δ t

[0153] Specifically, the corrected vegetation condition index (VCI corr ) is equal to the observed vegetation condition index (VCI obs ) minus a time-dependent correction (Δt). That is, from the actual observed vegetation condition (such as the health degree, coverage degree index of vegetation), deduct the interference or error caused by time factor, so as to obtain more accurate corrected vegetation condition index, and evaluate the vegetation state more truly. In an embodiment of the present application, the spatial scale is ten-meter grid unit;The time scale is daily update;The uncertainty of propagation is simulated by Monte Carlo simulation one thousand times.

[0154] In an embodiment of the present application, the result is verified. Adopt multi-source verification method, including ground station verification, using fifty ecological monitoring stations synchronous data;Unmanned aerial vehicle verification, monthly high-resolution aerial photography is carried out;Temporal and spatial consistency test, analyze the difference between adjacent pixels;Expert evaluation, blind evaluation verification is carried out by ecologists.

[0155] The vegetation coverage evaluation method based on collaborative attention state space network of the embodiment of the present application solves the problem that the existing model cannot effectively separate natural interference and human activity influence by using the collaborative attention state space network model formed after optimization. Through the double-channel hidden variable architecture, discrete interference events and continuous gradual processes are modeled respectively, and accurate separation is realized. The problem of insufficient dynamic adaptability of the model is solved. In view of the defects that the existing technology adopts static parameters and is difficult to adapt to climate change and vegetation growth season conversion, a parameter dynamic aggregation framework is designed, the model parameters are adjusted in real time according to the input features, and the system robustness is improved. The problem of poor model interpretability is solved. Through discrete state classification and continuous gradual variable, the interpretable ecological state is output. At the same time, based on the collaborative attention mechanism, the sequence constraint of the traditional state space model is broken, the global vegetation state perception is realized;Through frequency domain feature enhancement, multi-scale vegetation response features are extracted;Use double-channel architecture to process sudden interference and gradual process respectively;Finally, the vegetation coverage index is optimized through dynamic correction algorithm.

[0156] The application process of the vegetation coverage evaluation method based on the collaborative attention state space network embodiment of the present application is as shown in Figure 2 In Figure 2 , the application process includes:

[0157] - Perform multi-source data collection, data sources include MODIS data (500m / day); Landsat data (30m / 16 days); ERA5 meteorological data; ground monitoring station data; topographic data (SRTM).

[0158] - Perform data preprocessing, the processing process includes cloud mask processing; atmospheric correction; spatio-temporal alignment (10m grid); feature extraction (NDVI / EVI, etc.).

[0159] - Perform frequency domain feature enhancement, the processing process includes forming low frequency signals (seasonal changes) by FFT time-frequency conversion; medium frequency signals (inter-monthly changes); high frequency signals (sudden disturbances).

[0160] - Model training is performed on the formed collaborative attention state space network, the training process includes 120TB of historical observation data; using AdamW optimizer; completing 500 training cycles; and performing online incremental learning.

[0161] - Perform non-causal attention calculation to achieve global pixel correlation;

[0162] - Perform dual-channel state decoding to form dynamic weight fusion and cross-entropy / KL constraints through discrete channels (ecological disturbances), continuous channels (gradual processes)

[0163] - Perform dynamic benchmarking, the construction parameters include long-term benchmarking (20 years); short-term benchmarking (30-365 days); vegetation type adaptation; dynamic adjustment of fusion coefficients.

[0164] - Perform interference correction output, the correction process includes interference intensity quantification; vegetation index correction; uncertainty propagation; multi-source verification.

[0165] The vegetation coverage evaluation device based on the collaborative attention state space network embodiment of the present application is as shown in Figure 3 In Figure 3 , the present embodiment includes:

[0166] The model optimization module 10 is used to perform the following optimization of the state space network model:

[0167] Increase the attention weight of the global feature of the vegetation through the transfer equation;

[0168] Form a dynamic aggregation of model parameters according to observation data;

[0169] According to the observation equation, a discrete channel for identifying interference events and a continuous channel for gradual processes are formed;

[0170] The data enhancement module 20 is configured to perform frequency domain feature enhancement in the process of multi-source fusion of observation data.

[0171] The optimization correction module 30 is configured to perform vegetation index optimization and correction in the process of processing observation data by the state space network model.

[0172] As shown in Figure 3 In an embodiment of the present application, the model optimization module 10 includes:

[0173] The global attention construction unit 11 is configured to embed attention position offset weights in the state transition equation of the state space network, and dynamically adjust the contribution degree of different global features in the historical state to the state update through the attention position offset weights.

[0174] The parameter aggregation construction unit 12 is configured to form a parameter dynamic aggregation parallel framework, and extract observation data features by using a dynamic convolution kernel.

[0175] The dual-type channel construction unit 13 is configured to form discrete channels and continuous channels of the state space network decoding output, classify interference factors through the discrete channels, and quantify the vegetation gradual change degree through the continuous channels.

[0176] As shown in Figure 3 In an embodiment of the present application, the data enhancement module 20 includes:

[0177] The observation data formation unit 21 is configured to unify multi-source remote sensing data into observation data through bilinear interpolation.

[0178] The frequency domain feature enhancement unit 22 is configured to perform frequency domain conversion-noise suppression-multi-scale feature fusion to form frequency domain feature enhancement on the observation data.

[0179] As shown in Figure 3 In an embodiment of the present application, the optimization correction module 30 includes:

[0180] The model training unit 31 is configured to train and deploy the state space network model by using historical observation data.

[0181] The index optimization unit 32 is configured to optimize the vegetation index formed by the state space network model.

[0182] The interference correction unit 33 is configured to correct the vegetation index by interference factors.

[0183] The embodiment of the present application also provides a specific implementation of an electronic device capable of implementing all steps in the method in the above-mentioned embodiments, which is described in detail as follows.Figure 4 The electronic device 600 specifically includes the following contents:

[0184] The processor 610, the memory 620, the communication unit 630 and the bus 640;

[0185] The processor 610, the memory 620 and the communication unit 630 can communicate with each other through the bus 640, and the communication unit 630 can be used to realize information transmission between the server-side device and the terminal device and other related devices.

[0186] The processor 610 is configured to invoke the computer program in the memory 620, and the processor executes the computer program to realize all steps in the vegetation coverage evaluation method based on the collaborative attention state space network in the above embodiments.

[0187] Those skilled in the art should understand that the memory can be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM) and the like. The memory is used to store programs, and the processor executes the programs after receiving the execution instructions. Further, the software programs and modules in the memory can also include an operating system, which can include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.), and can communicate with various hardware or software components to provide an operating environment for other software components.

[0188] The processor can be an integrated circuit chip with processing capability. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP) and the like. The general-purpose processor can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0189] The application further provides a computer readable storage medium, which comprises a program used for executing the vegetation coverage evaluation method based on the collaborative attention state space network according to any one of the preceding method embodiments when executed by a processor.

[0190] Those skilled in the art should understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the foregoing storage medium includes various storage media capable of storing program codes, such as ROM, RAM, magnetic disk or optical disk, and the specific type of the medium is not limited in the application.

[0191] The above description is only the preferred embodiment of the application, but the protection scope of the application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A vegetation cover assessment method based on a cooperative attention state space network, characterized in that, The method comprises the following steps: The following optimizations of the state space network model are performed: The attention weight of the global feature of the vegetation is increased through the transition equation; The model parameter dynamic aggregation is formed according to the observation data; The discrete channel for identifying interference events and the continuous channel for gradual change processes are formed according to the observation equation; Frequency domain feature enhancement is performed in the process of multi-source fusion of the observation data; Vegetation index optimization and correction are formed in the process of observation data processing by the state space network model. The optimization comprises: The attention position offset weight is embedded in the state transition equation of the state space network, and the attention position offset weight is used to dynamically adjust the contribution degree of different global features in the historical state to the state update; The parameter dynamic aggregation parallel framework is formed, and the dynamic convolution kernel is used to extract the observation data features; The discrete channel and the continuous channel of the state space network decoding output are formed, the interference factor classification is performed through the discrete channel, and the vegetation gradual change degree quantification is performed through the continuous channel.

2. The vegetation cover assessment method of claim 1, wherein, The frequency domain feature enhancement in the process of multi-source fusion of the observation data comprises: The multi-source remote sensing data is unified into the observation data through bilinear interpolation; The frequency domain feature enhancement is formed through frequency domain conversion-noise suppression-multiscale feature fusion of the observation data.

3. The method of vegetation cover assessment according to claim 1, characterized in that, The vegetation index optimization and correction in the process of observation data processing by the state space network model comprises: The state space network model is trained and deployed through historical observation data; The vegetation index formed by the state space network model is optimized; The vegetation index is corrected according to the interference factor.

4. A vegetation cover assessment device based on a co-attention state space network, characterized by, The method comprises the following steps: A model optimization module is configured to perform the following optimizations of the state space network model: The attention weight of the global feature of the vegetation is increased through the transition equation; The model parameter dynamic aggregation is formed according to the observation data; The discrete channel for identifying interference events and the continuous channel for gradual change processes are formed according to the observation equation; A data enhancement module is configured to perform frequency domain feature enhancement in the process of multi-source fusion of the observation data; An optimization correction module is configured to form vegetation index optimization and correction in the process of observation data processing by the state space network model. The model optimization module comprises: A global attention construction unit is configured to embed the attention position offset weight in the state transition equation of the state space network, and the attention position offset weight is used to dynamically adjust the contribution degree of different global features in the historical state to the state update; A parameter aggregation construction unit is configured to form the parameter dynamic aggregation parallel framework, and the dynamic convolution kernel is used to extract the observation data features; A double-type channel construction unit is configured to form the discrete channel and the continuous channel of the state space network decoding output, the interference factor classification is performed through the discrete channel, and the vegetation gradual change degree quantification is performed through the continuous channel.

5. The vegetation cover assessment apparatus of claim 4, wherein, The data enhancement module comprises: An observation data formation unit is configured to unify the multi-source remote sensing data into the observation data through bilinear interpolation; A frequency domain feature enhancement unit is configured to perform frequency domain feature enhancement through frequency domain conversion-noise suppression-multiscale feature fusion of the observation data.

6. The vegetation cover assessment device of claim 4, wherein, The optimization correction module comprises: A model training unit is configured to train and deploy the state space network model through historical observation data An index optimization unit is configured to optimize the vegetation index formed by the state space network model An index optimization unit is configured to optimize the vegetation index formed by the state space network model An interference correction unit is configured to correct the vegetation index by interference factors.

7. An electronic device, comprising: The method comprises the steps of: A processor, a memory, and an interface for communicating with a gateway; The memory is configured to store programs and data, and the processor is configured to invoke the programs stored in the memory to execute the method according to any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a program, which when executed by a processor, is configured to execute the method according to any one of claims 1 to 3.

Images