Forest degradation early warning platform fusing vegetation spectral features and reinforcement learning

Abstract

The invention relates to the technical field of intelligent prediction of forest degradation, and discloses a forest degradation early warning platform fusing vegetation spectral features and reinforcement learning, and the platform comprises the steps: obtaining a satellite / unmanned plane image through a multi-source remote sensing data collection module; the spectral feature extraction module calculates a vegetation index, a red edge slope and a water stress index; the multi-feature fusion module integrates spectrum, texture and phenological features and outputs a degradation probability; the reinforcement learning prediction module constructs a state transition model and predicts a degradation gradient; the early warning decision module generates a dynamic early warning level based on the gradient value; the feedback optimization module updates network parameters through a loss function; therefore, the problems of difficult early recognition and early warning lag are solved, the recognition rate is improved by 40%, and the early warning response is shortened to 15 days.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology for forest degradation, and in particular to a forest degradation early warning platform that integrates vegetation spectral characteristics and reinforcement learning. Background Technology

[0002] Currently, forest degradation monitoring mainly relies on two types of technologies: First, manual patrols: forest rangers regularly record vegetation status, with a coverage rate of less than 30% (especially in remote areas), relying on subjective experience, and the data update cycle is long, generally exceeding 3 months; Second, single remote sensing technologies: using multispectral satellites (such as Landsat) to provide NDVI indices, but with low band resolution (30m), they cannot capture early physiological stresses (such as water deficit). Hyperspectral UAVs can identify features such as red edge slope, but the data is isolated and not associated with phenological and textural features, resulting in a high false alarm rate; for example, in Amazon rainforest monitoring, a single NDVI model only issued an early warning 6 months after an insect outbreak (a lag rate of 78%).

[0003] The aforementioned monitoring technologies are insufficient to meet increasingly demanding requirements and have significant shortcomings in forest degradation detection, as detailed below:

[0004] First, the aforementioned monitoring technologies have blind spots in early identification. Manual inspections cannot quantify canopy photosynthetic efficiency (PRI index) and water stress (requires 531 / 570nm band). Single NDVI has low sensitivity to early degradation (such as slight decrease in chlorophyll) (cannot be detected when the change is <0.05).

[0005] Secondly, the aforementioned monitoring technologies suffer from the problem of delayed early warning. Existing technologies generally output static risk levels without predicting trends, resulting in a significant delay in response and delayed early warning.

[0006] Finally, the aforementioned monitoring technologies have poor dynamic adaptability, and traditional machine learning models (such as SVM) require regular manual data annotation and cannot autonomously optimize parameters through environmental feedback.

[0007] In summary, those skilled in the art urgently need a forest degradation early warning platform that integrates multispectral / hyperspectral data and can predict degradation trajectories and gradient risks through reinforcement learning. Summary of the Invention

[0008] The purpose of this invention is to solve the above problems by designing a forest degradation early warning platform that integrates vegetation spectral features and reinforcement learning.

[0009] The technical solution of the present invention to achieve the above objectives is a forest degradation early warning platform that integrates vegetation spectral features and reinforcement learning, the platform comprising the following parts:

[0010] The multi-source remote sensing data acquisition module is used to acquire multispectral impact data, hyperspectral image data and environmental data collected by satellites or UAVs in real time, and output the above data as raw data.

[0011] The spectral feature extraction module calculates the vegetation index based on the original data using a vegetation index fusion model, and outputs the vegetation index and spectral decomposition features as spectral features.

[0012] The multi-feature fusion module integrates spectral features, texture features, and phenological features through a decision-level fusion function to calculate the current degradation probability.

[0013] The reinforcement learning prediction module constructs a degradation trend prediction model, predicts the state vector of the forest at the next time step based on the current degradation probability, and calculates the degradation probability gradient over a period of time in the future.

[0014] The early warning decision module generates dynamic early warning levels based on the degradation probability gradient and early warning triggering logic, and sends the dynamic early warning levels to the user terminal.

[0015] The feedback optimization module updates the network parameters in the degradation trend prediction model based on the loss function, thereby achieving iterative optimization of the degradation trend prediction model.

[0016] The mathematical expression for the vegetation index fusion model is:

[0017] VI fused =α·NDVI+β·EVI+γ·PRI

[0018] In the formula, NDVI is the normalized vegetation index, EVI is the enhanced vegetation index, PRI is the photochemical reflectance index, and α, β and γ are the weighting coefficients, α = 0.4, β = 0.3 and γ = 0.3.

[0019] The mathematical expression for NDVI is:

[0020]

[0021] The mathematical expression for EVI is:

[0022]

[0023] The mathematical expression for PRI is:

[0024]

[0025] In the formula, R NIR R represents the near-infrared reflectance. Red R represents the reflectivity in the red light band. Blue R represents the reflectivity in the blue light band. 531For the 531nm band reflectivity, R 570 The reflectivity is 570nm.

[0026] The mathematical expression for the decision-level fusion function is:

[0027] F(x)=σ(w s ·S(x)+w t ·T(x)+w p ·P(x))

[0028] In the formula, F(x) is the current degradation probability, S(x) is the spectral feature vector, T(x) is the texture feature vector, P(x) is the phenological feature vector, σ is the activation function, and w s w t and w p S(x), T(x), and P(x) are the feature weight coefficients, respectively, and F(x) is the forest degradation probability, where F(x) ∈ [0,1].

[0029] The S(x) = [VI] fused [Red edge slope, water stress index] T The VI fused It is a subset of S(x).

[0030] The mathematical expression for the degradation trend prediction model is:

[0031] s t+1 =f θ (s t ,a t )+∈ t

[0032] In the formula, s t+1 Let s be the state vector at time t+1. t Let F(x) be the state vector at time t, containing the degradation probability F(x) at time t. t The red edge slope at time t t Moisture harmonic index at time t t , that is s t =[F(x)] t Red edge slope t Water stress index t ] T ;a t The action space includes (enhanced monitoring, maintenance, and reduced monitoring); f θ Let θ be the state transition function, and θ be the network parameters of the state transition function; ∈ t Gaussian noise characterizes environmental uncertainty.

[0033] The mathematical expression for the degradation probability gradient is:

[0034]

[0035] In the formula, Let F(x) be the degradation probability gradient during the time interval t→k. t+k Let F(x) be the degradation probability at time t+k. t Let t be the degradation probability at time t.

[0036] The spectral decomposition features include, but are not limited to: red edge slope and moisture stress index.

[0037] The mathematical expression for the loss function is:

[0038]

[0039]

[0040] In the formula, Let θ be the network parameters of the state transition function and φ be the model parameters of the policy network. E represents the loss function, which can simultaneously optimize the network parameters of the state transition function and the model parameters of the policy network. (st,at,st+1) This represents the expectation of the transferred samples obtained from the experience replay pool D, where D is the experience replay pool used to store historical observation data, r(s) t ,a t ) is the immediate reward function, γ is the discount factor used to calculate the cumulative reward, and V ψ Let ψ be the state-value function, ψ be the parameter of the state-value function, KL be the KL divergence, used to measure the difference between the new policy and the old policy, and π be the state-value function. φ For policy networks, π old For the policy network before the update, α * is the regularization coefficient, which controls the weight of the KL divergence term.

[0041] The warning triggering logic is as follows:

[0042]

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This invention combines multispectral satellite data with ground vegetation index sensors, and optimizes the degradation model through reinforcement learning to achieve automatic identification of forest cover changes and prediction of degradation trends, thereby improving accuracy and enabling early warning.

[0045] 2. This invention combines a spectral feature extraction module with a multi-feature fusion module to effectively improve the early recognition rate, which can increase the early recognition rate by 40%.

[0046] 3. This invention utilizes a reinforcement learning prediction module to predict the state transition of the forest and uses an early warning decision module to calculate the degradation probability gradient, effectively solving the problem of early warning lag and controlling the early warning time within 15 days.

[0047] 4. This invention iteratively optimizes the network parameters of the state transition function and the model parameters of the policy network through a feedback optimization module and an experience replay pool, thereby achieving automated updates of the platform's monitoring function and improving the prediction accuracy of forest degradation status. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the principle of a forest degradation early warning platform that integrates vegetation spectral features and reinforcement learning as described in this invention.

[0049] Figure 2 This is a flowchart of the prediction process for the state vector and degradation probability gradient described in this invention;

[0050] Figure 3 This is a flowchart of the iterative optimization of the degradation trend prediction model described in this invention;

[0051] Figure 4 This is a data comparison table of Embodiment 1 of the present invention;

[0052] Figure 5 This is a data comparison table of Embodiment 2 of the present invention. Detailed Implementation

[0053] The present invention will now be described in detail with reference to the accompanying drawings;

[0054] A forest degradation early warning platform that integrates vegetation spectral features with reinforcement learning, such as Figure 1-3 As shown, the platform includes the following components:

[0055] The multi-source remote sensing data acquisition module is used to acquire multispectral impact data, hyperspectral image data and environmental data collected by satellites or UAVs in real time, and output the above data as raw data.

[0056] The spectral feature extraction module can receive raw data and use it as input to calculate vegetation indices (NDVI - Normalized Difference Vegetation Index, EVI - Enhanced Vegetation Index, and PRI - Photochemical Reflectance Index, a spectral index used to detect vegetation photosynthetic efficiency and stress status) and spectral decomposition features (such as red edge slope and water stress index) through a vegetation index fusion model, and output the vegetation indices and spectral decomposition features as spectral features.

[0057] The multi-feature fusion module can receive spectral features, integrate spectral features, texture features and phenological features through a decision-level fusion function, calculate the current degradation probability and output it;

[0058] The reinforcement learning prediction module can receive the current degradation probability, build a degradation trend prediction model, predict the state vector of the forest at the next time step (i.e. the forest's transition pattern) based on the current degradation probability, and calculate the degradation probability gradient over a future period of time (i.e., the time period between the current time and the next time step).

[0059] The early warning decision module generates dynamic early warning levels (red / orange / yellow) based on the degradation probability gradient and early warning triggering logic, and sends the dynamic early warning levels to the user terminal.

[0060] The feedback optimization module updates the network parameters in the degradation trend prediction model based on the loss function, thereby achieving iterative optimization of the degradation trend prediction model.

[0061] The mathematical expression for the vegetation index fusion model is:

[0062] VI fused =α·NDVI+β·EVI+γ·PRI

[0063] In the formula, NDVI is the normalized vegetation index, EVI is the enhanced vegetation index, PRI is the photochemical reflectance index, and α, β and γ are the weighting coefficients, α = 0.4, β = 0.3 and γ = 0.3.

[0064] The mathematical expression for NDVI is:

[0065]

[0066] The mathematical expression for EVI is:

[0067]

[0068] The mathematical expression for PRI is:

[0069]

[0070] In the formula, R NIR R represents the near-infrared reflectance. Red R represents the reflectivity in the red light band. Blue R represents the reflectivity in the blue light band. 531 For the 531nm band reflectivity, R 570 The reflectivity is 570nm.

[0071] The mathematical expression for the decision-level fusion function is:

[0072] F(x)=σ(w s·S(x)+w t ·T(x)+w p ·P(x))

[0073] In the formula, F(x) is the current degradation probability, S(x) is the spectral feature vector, T(x) is the texture feature vector, P(x) is the phenological feature vector, σ is the activation function (i.e., the Sigmoid function), and w s w t and w p Let w be the feature weight coefficients of S(x), T(x), and P(x), respectively. s w t and w p After optimization with CNN, we can obtain w. s =0.6, w t =0.3, w p =0.1, F(x) is the probability of forest degradation, F(x)∈[0,1].

[0074] S(x) = [VI fused [Red edge slope, water stress index] T VI fused It is a subset of S(x).

[0075] The mathematical expression for the degradation trend prediction model is:

[0076] s t+1 =f θ (s t ,a t )+∈ t

[0077] In the formula, s t+1 Let s be the state vector at time t+1 (i.e., the next time step). t Let F(x) be the state vector at time t, containing the degradation probability F(x) at time t. t (From the decision-level fusion function), the red edge slope at time t t Moisture harmonic index at time t t Isospectral decomposition characteristics, i.e., s t =[F(x)] t Red edge slope t Water stress index t ] T ;a t The action space is primarily used to represent management strategies, including (increased monitoring, maintenance, reduced monitoring); f θ Let θ be the state transition function (implemented by a neural network), and θ be the network parameters of the state transition function; ∈ t Gaussian noise characterizes environmental uncertainty.

[0078] The mathematical expression for the degradation probability gradient is:

[0079]

[0080] In the formula, Let F(x) be the degradation probability gradient during the time interval t→k. t+k Let F(x) be the degradation probability at time t+k. t Let t be the degradation probability at time t.

[0081] Spectral decomposition features include, but are not limited to: red edge slope (i.e., the maximum value of the first derivative of the reflectance spectrum in the red light range (wavelength: 680nm-760nm)) and water stress index. The water stress index is estimated by measuring the canopy temperature using an infrared thermometer and combining it with the energy balance formula.

[0082] The mathematical expression for the loss function is:

[0083]

[0084] In the formula, The loss function consists of two parts, used to simultaneously optimize the network parameters of the state transition function and the model parameters of the policy network. θ represents the network parameters of the state transition function, and φ represents the model parameters of the policy network. (st,at,st+1) This represents the transition sample (current state s) obtained from the experience replay pool D. t The action taken, a t The next state s t+1 Find the expectation, where D is the experience replay pool used to store historical observation data, and r(s) t ,a t ) is the instant reward function, defined as in V represents the predicted gradient of degradation probability from time t to time t+1, where γ is the discount factor used to calculate the cumulative reward. ψ Let ψ be the state-value function, ψ be the parameter of the state-value function, KL be the KL divergence, used to measure the difference between the new policy and the old policy, and π be the state-value function. φ For policy networks, π old This refers to the policy network before the update, i.e., the parameters of the policy network before the update. Before each update, we copy the current policy network parameters to the old policy network. α * r(s) is the regularization coefficient, which controls the weight of the KL divergence term; where r(s) is the regularization coefficient. t ,a t )+γVψ(s t+1 )-Vψ(s tLet be the temporal difference error, representing the difference between the current value function estimate and the target value (the immediate reward plus the next state discount). We minimize the square of this error to update the value function (but note that in this loss function, we simultaneously optimize the policy and the state transition model, so this term also affects the parameters of the state transition model).

[0085] The logic for triggering the early warning is as follows:

[0086]

[0087] Example 1;

[0088] Application of this platform in early warning of insect pests in tropical rainforests;

[0089] Application environment: Sumatra Island, Indonesia, with an annual rainfall of >3000mm.

[0090] Implementation process:

[0091] First, data in the 531nm and 570nm bands were collected using a UAV hyperspectral analyzer.

[0092] Secondly, calculation (Exceeds threshold 0.1)

[0093] Third, by combining NDVI=0.62 and EVI=0.58, we obtain VI. fused =0.4×0.62+0.3×0.58+0.3×0.14=0.476;

[0094] Fourth, the output degradation probability F(x) = 0.67 (Sigmoid activation)

[0095] Finally, predict the gradient after 6 months.

[0096] The data comparison table for Example 1 is as follows: Figure 4 As shown.

[0097] Example 2;

[0098] Applications of this platform in monitoring freeze-thaw degradation in cold-region forests;

[0099] Application environment: Quebec, Canada, winter temperature < -30℃

[0100] Key parameters:

[0101] Red edge slope = 0.021 (frost damage caused a decrease of >50%)

[0102] Moisture stress index = 0.85 (canopy temperature anomaly),

[0103] S(x) = [0.35, 0.021, 0.85] T →F(x)=0.72

[0104] Gradient warning:

[0105] The data comparison table for Example 2 is as follows: Figure 5 As shown.

[0106] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A forest degradation early warning platform that fuses vegetation spectral features with reinforcement learning, characterized in that, The platform comprises the following parts: A multi-source remote sensing data acquisition module for real-time acquisition of raw data and output; A spectral feature extraction module for calculating vegetation index through a vegetation index fusion model based on raw data and outputting vegetation index and spectral decomposition features as spectral features; A multi-feature fusion module for integrating spectral features, texture features and phenological features through a decision-level fusion function and calculating current degradation probability; A reinforcement learning prediction module for constructing a degradation trend prediction model and predicting state vector of the forest at the next time and degradation probability gradient in the future based on current degradation probability; An early warning decision module for generating dynamic early warning level based on degradation probability gradient and early warning trigger logic and notifying users; A feedback optimization module for updating network parameters based on a loss function and realizing iterative optimization of the degradation trend prediction model.

2. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The mathematical expression of the vegetation index fusion model is: VI fused = a · NDVI + b · EVI + g · PRI In the formula, NDVI is normalized vegetation index, EVI is enhanced vegetation index, PRI is photochemical reflectance index, and a, β and γ are weight coefficients respectively.

3. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 2, characterized in that, The mathematical expression of the NDVI is: The mathematical expression of the EVI is: The mathematical expression of the PRI is: In the formula, R NIR is the reflectivity in the near-infrared band, R Red is the reflectivity in the red light band, R Blue is the reflectivity in the blue light band, R 531 is the reflectivity in the 531 nm band, R 570 is the reflectivity in the 570 nm band.

4. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The mathematical expression of the decision-level fusion function is: F(x) = σ(w s • S(x) + w t • T(x) + w p • P(x)) In the formula, F(x) is a current degradation probability, S(x) is a spectral feature vector, T(x) is a texture feature vector, P(x) is a phenology feature vector, sigma is an activation function, and w s , w t , and w p are feature weight coefficients of S(x), T(x), and P(x) respectively, and F(x) is a forest degradation probability. In the formula, F(x) is a current degradation probability, S(x) is a spectral feature vector, T(x) is a texture feature vector, P(x) is a phenology feature vector, sigma is an activation function, and w s , w t , and w p are feature weight coefficients of S(x), T(x), and P(x) respectively, and F(x) is a forest degradation probability.

5. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 4, characterized in that, S(x) = [VI fused , red edge slope, Moisture stress index T .

6. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The mathematical expression of the degradation trend prediction model is: s t+1 = f θ (s t , a t )+ ∈ t where s t+1 is the state vector at time t + 1, s t is the state vector at time t, a t is the action space, f θ is the state transition function, θ is the network parameter of the state transition function, ∈ t is the Gaussian noise.

7. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The mathematical expression of the degradation probability gradient is: wherein is the degradation probability gradient for the time period t→k, F(x) t+k is the degradation probability at time t+k, F(x) t is the degradation probability at time t.

8. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The spectral decomposition features include but are not limited to red edge slope and water stress index.

9. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The mathematical expression of the loss function is: where, is the loss function, θ is the network parameter of the state transition function, φ is the model parameter of the policy network, E (st,at,st+1) denotes the expectation over the transition samples sampled from the experience replay pool D, D is the experience replay pool, r(s t ,a t ) is the immediate reward function, γ is the discount factor, V ψ is the state value function, ψ is the parameter of the state value function, KL is the KL divergence, π φ is the policy network, π old is the policy network before update, α * is the regularization coefficient.

10. The forest degradation early warning platform fusing vegetation spectral features with reinforcement learning according to claim 1, characterized in that, The early warning trigger logic is:

Images