Forest grass health monitoring and predicting method based on image recognition

Abstract

The invention discloses a forest grass health monitoring and prediction method based on image recognition. The method comprises the following steps: S1, constructing a regional ecological parameter set; s2, acquiring multi-source image data of the target forest and grass region by using the regional ecological parameter set; s3, preprocessing the multi-source image data; s4, performing forest and grass health state recognition by using the image recognition model, and outputting a health level label; s5, constructing a two-channel prediction model, and outputting a trend prediction value and a risk probability value based on the health level label and the regional ecological parameter set; s6, when the risk probability value presents a rising acceleration trend or the fluctuation amplitude exceeds a threshold value in a plurality of continuous time periods, triggering a structure reconstruction mechanism; and S7, carrying out forest grass health monitoring and prediction again by using the reconstructed image recognition model. The method is based on VOLO network fusion ecological modeling, intelligently predicts forest grass health and dynamically reconstructs the structure, and has the advantages of high adaptability, high precision and high response.

Description

Technical Field

[0001] The present invention relates to the field of ecological monitoring technology, and in particular to a forest and grass health monitoring and prediction method based on image recognition. Background Art

[0002] The health status of forestry and grassland ecosystems is an important supporting factor for national ecological security and sustainable development. Especially in many key areas such as forest fire prevention, desertification control, grassland degradation control, and pest and disease control, accurate and efficient forest and grassland health monitoring methods have become an important technical foundation for promoting the modernization of ecological governance. Against this background, regional ecological perception systems built based on remote sensing images, satellite monitoring, aerial photography, geographic information systems, etc. have gradually become mainstream. However, existing forest and grassland health monitoring methods still have certain limitations and cannot meet the systematic needs of the complete monitoring goal of "wide-area coverage, dynamic response, health diagnosis, and trend prediction."

[0003] Current forest and grassland ecological monitoring systems are primarily based on two approaches: one is traditional manual inspections and ground sampling, which rely on manual experience to determine vegetation status and collect a certain number of ecological samples for analysis. Although this approach offers high accuracy, it suffers from low coverage efficiency, long response cycles, and high labor costs, making it difficult to apply to large-scale, cross-regional, and long-term continuous monitoring of forest and grassland health. The other is data analysis methods based on remote sensing satellite or drone imagery. These methods, such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDVI), and Leaf Area Index (LAI), serve as key ecological indicators, combined with meteorological data for time-series monitoring and status assessment. While remote sensing methods offer advantages in spatial coverage and monitoring cycles, their health assessments are often based on a single threshold or rule set, lacking the ability to comprehensively model multi-source heterogeneous data. This is particularly problematic for responses to sudden ecological disturbances, such as disease, drought, and insect pests, which exhibit low recognition accuracy and insufficient sensitivity.

[0004] In the field of image processing, deep learning, especially convolutional neural networks, has been widely used in tasks such as image classification, target detection, and semantic segmentation, and has achieved remarkable results in segmented scenarios such as agricultural remote sensing, urban monitoring, and biomedical images. In recent years, the Transformer structure and its extensions in the visual field (such as ViT, Swin, VOLO, etc.) have provided new global modeling capabilities for image recognition, and have also promoted the improvement of feature extraction capabilities of large-scale remote sensing images and ecological images. However, the direct application of these models to forest and grassland ecological monitoring still faces significant challenges. On the one hand, forest and grassland images have visual characteristics such as high similarity, high redundancy, and low contrast, and traditional classification networks are prone to overfitting or misclassification; on the other hand, the model structure is mostly static and lacks the ability to adaptively adjust to dynamic changes in the environment. It is impossible to flexibly adjust the recognition structure according to changes in the ecological status of different regions, and it is difficult to achieve early warning of sudden ecological events.

[0005] Furthermore, while some studies have attempted to use time series analysis methods (such as ARIMA and LSTM) to model trends in ecological indicators, these approaches generally model only health level sequences, ignoring the complex effects of ecological and environmental variables on health status and lacking a predictive response mechanism based on risk fluctuations. More importantly, existing methods often lack a linkage mechanism between prediction results and identification models, failing to form a systematic closed loop of monitoring, identification, prediction, and reconstruction.

[0006] In summary, existing forest and grassland health monitoring technologies have the following major defects in terms of response timeliness, model adaptability, multi-source data fusion, and prediction feedback mechanism: First, health status identification mostly relies on static network structures and cannot dynamically reconstruct models according to trend changes; second, ecological characteristic modeling is mostly based on single variables or simple weighting rules, and lacks conditional modeling mechanisms; third, prediction results are only used for state estimation and lack the ability to reverse drive identification structures, making it impossible to achieve structural enhancement and resource scheduling in high-risk areas; fourth, there is a lack of spatiotemporal binding mechanisms for image units, making it impossible to systematically archive and trend analyze the health evolution process. Summary of the Invention

[0007] One purpose of the present invention is to propose a forest and grassland health monitoring and prediction method based on image recognition. The present invention constructs an image recognition model based on the VOLO network, integrates the Outlooker and Transformer modules to realize local structure perception and global semantic modeling, and introduces regional adaptability in combination with the ecological label conditional convolution mechanism to construct a health prediction and structural self-reconstruction closed-loop mechanism, thereby realizing accurate identification and trend prediction of the ecological health status of forests and grasslands, and having the advantages of high recognition accuracy, adaptive model structure and strong response mechanism.

[0008] A forest and grass health monitoring and prediction method based on image recognition according to an embodiment of the present invention includes the following steps:

[0009] S1. Construct regional ecological parameter set;

[0010] S2, adopting a regionally differentiated adaptive sampling strategy and using regional ecological parameter sets to obtain multi-source image data of the target forest and grassland areas;

[0011] S3, preprocessing the multi-source image data to generate a standardized image dataset;

[0012] S4. Inputting the standardized image dataset into an image recognition model to identify the health status of forests and grasslands and output a health grade label;

[0013] S5. Construct a dual-channel prediction model to output trend prediction values and risk probability values based on health level labels and regional ecological parameter sets, and generate a health prediction result set;

[0014] S6. When the risk probability value corresponding to any image unit in the health prediction result set shows an upward acceleration trend or the fluctuation amplitude exceeds a threshold in multiple consecutive time periods, a structural reconstruction mechanism is triggered;

[0015] S7. Use the reconstructed image recognition model to re-monitor and predict forest and grassland health.

[0016] Optionally, the regional ecological parameter set includes historical environmental data, meteorological data, geographic information and vegetation type labels of the target forest and grassland area.

[0017] Optionally, the multi-source image data includes visible light images, multispectral images and drone aerial images.

[0018] Optionally, the S2 specifically includes:

[0019] S21, spatially dividing the target forest and grassland area based on the regional ecological parameter set, dividing the entire monitoring area into image units according to a set spatial resolution, and assigning a unique spatial identification number to each image unit;

[0020] S22. Analyze ecological indicators for each image unit and generate corresponding sampling priority scores based on vegetation index changes, meteorological fluctuation characteristics, historical health status stability, and human disturbance frequency;

[0021] S23. Combining the sampling priority score of the image unit with the geographical distribution density to form a sampling density control strategy with spatial weights, and dynamically adjusting the sampling frequency of the image unit according to the strategy;

[0022] S24. Select a target sampling area for the current cycle according to the control strategy, and collect images.

[0023] Optionally, the preprocessing includes image registration, illumination normalization, geometric correction and image enhancement.

[0024] Optionally, the image recognition model includes a backbone feature extraction network constructed based on the VOLO network and an eco-label conditional convolution module. The eco-label conditional convolution module selects the corresponding convolution channel path according to the vegetation type label in the regional ecological parameter set. The VOLO network obtains local structural features through the Outlooker module, and models global semantic relationships through the Transformer module to output health recognition results. The Outlooker module is a local attention mechanism module.

[0025] Optionally, the S4 specifically includes:

[0026] S41. Construct an input image tensor based on a standardized image dataset;

[0027] S42: Input the input image tensor into the backbone feature extraction network built based on the VOLO network, perform local structure perception through the Outlooker module, and extract the local context feature map of the image:

[0028]

[0029] Among them, F L (i, j) represents the local context feature map, describing the local feature value at position (i, j). Represents the local receptive field neighborhood centered at position (i, j), A i,j,m,n represents the attention weight of position (m,n) to position (i,j) in the input image tensor, and X(m,n) represents the pixel value at position (m,n) in the input image tensor;

[0030] S43, pass the local context feature map into the multi-layer Transformer module, use the multi-head attention mechanism to calculate the global dependency feature map, and obtain the fused feature map:

[0031] F O =Transformer(F L )=Concat(head1,...,head k )·W O ;

[0032] Among them, F O Represents the fusion feature map, head k represents the output of the kth attention head, W O Represents the output linear transformation weight matrix, Concat(...) represents the concatenation operation of the outputs of multiple attention heads;

[0033]

[0034] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, d k Represents the dimension of each attention head, Softmax represents the row normalization operation, and T represents the transposition operation;

[0035] S44. Input the fused feature map into the ecological label conditional convolution module, generate a one-hot encoding vector based on the vegetation type label in the regional ecological parameter set, and control the activation state of multiple predefined convolution kernel paths based on the one-hot encoding vector to output the ecological modulation feature map:

[0036]

[0037] Among them, F E Represents the ecological modulation characteristic map, e k Indicates the kth bit in the one-hot encoding vector. When the value is 1, it means the current vegetation type label is l k , activate the corresponding convolution kernel W k , otherwise it will not participate in the calculation. Indicates the use of convolution kernel W k Perform convolution operation on the fused feature map, W k represents the convolution kernel;

[0038] S45. Input the ecological modulation feature map into the classification head, perform image unit health status recognition, and output a health recognition result, which includes a health level label corresponding to the image unit:

[0039] Y h =arg max(Softmax(W c ·F E +b c ));

[0040] Among them, Y h Indicates the health level label, with values of 0, 1, or 2, representing healthy, sub-healthy, and abnormal, respectively. c represents the weight matrix of the classification linear layer, b c Represents the bias term of the classification linear layer, Softmax represents the normalization of each category score, and arg max represents the variable value that maximizes the function value.

[0041] Optionally, the S5 specifically includes:

[0042] S51. Constructing a health status sequence based on the health level labels and corresponding timestamps, and constructing an ecological status sequence based on the regional ecological parameter set;

[0043] S52. Calculate the slope of the health change trend based on the health status sequence:

[0044]

[0045] Among them, Δ h represents the slope of health change trend, Indicates the health level label at the current moment, represents the health level label of the first q time steps, t n and t n-q are the timestamps of the current and previous q time steps respectively;

[0046] S53. Introduce a trend-driven evolutionary regression mechanism, which dynamically adjusts the prediction model parameters based on the slope of the health change trend to generate evolutionary parameters:

[0047] ω′ k =ω k ·(1+α·Δ h ),θ′ j =θ j ·(1+β·|Δ h |);

[0048] Among them, ω′ k represents the updated regression weight, ω k represents the regression weight, α represents the trend sensitivity coefficient, Δ h represents the slope of health change trend; θ′ j represents the updated ecological variable weight, θ j represents the ecological variable weight, β represents the risk sensitivity coefficient, |Δ h | represents the absolute value of the slope of the health change trend;

[0049] S54. Input the health status sequence and the ecological status sequence into the dual-channel prediction model, perform parallel prediction using the evolutionary parameters, and obtain trend prediction values and risk probability values, respectively. Use a dynamic weighted regression model to obtain the trend prediction value. The dual-channel prediction model includes a dynamic weighted regression model and a logistic regression model:

[0050]

[0051] in, represents the health grade label at the t+1 time step, ω0 represents the regression bias term, ω′ k represents the updated regression weight, represents the health level label of the t+1-kth time step, and n represents the total number of time steps;

[0052] Use the logistic regression model to output the risk probability value:

[0053]

[0054] in, represents the risk probability value at time step t+1, θ0 represents the logistic regression bias term, and θ′ j represents the updated ecological variable weight, represents the value of the jth ecological variable at the current moment, m represents the total number of ecological variables, and exp represents the natural exponential function;

[0055] S55 , repeating the prediction process in the sliding time window to generate a health prediction result set for future time steps, wherein the health prediction result set includes a trend prediction value, a risk probability value, and a corresponding timestamp.

[0056] Optionally, the S6 specifically includes:

[0057] S61. Calculate the risk growth slope and fluctuation range of the target image unit within the predicted time period based on the predicted risk probability value:

[0058]

[0059] Among them, Δ r represents the risk growth slope, and Represent the risk probability values at the last and first time points of the forecast period, t n+M and t n+1 Indicates the corresponding timestamp; σ r Indicates the fluctuation range, M indicates the total length of the forecast period, represents the risk probability value of n+k time steps, Indicates the average risk probability value within a time period;

[0060] S62. Set the structural reconstruction triggering condition. When any inequality is met, the structural reconstruction mechanism is triggered:

[0061] Δ r >λ1,σ r >λ2;

[0062] Among them, λ1 represents the risk slope trigger threshold, λ2 represents the risk fluctuation trigger threshold, Δ r represents the risk growth slope, σ r Indicates the amplitude of fluctuation;

[0063] S63. When the structural reconstruction mechanism is triggered, the reconstruction level factor is determined based on the risk growth slope and volatility:

[0064] Γ=γ1·Δ r +γ2·σ r ;

[0065] Where Γ represents the reconstruction level factor in the current forecast period, γ1 and γ2 represent the risk growth and volatility response weights preset by the system;

[0066] S64. Performing dynamic structural adjustment of the image recognition model according to the structural reconstruction level factor, including:

[0067] When Γ<τ1, the original structure remains unchanged;

[0068] When τ1≤Γ<τ2, the deep Outlooker branch is activated and the preset deep local attention path is automatically loaded. The deep local attention path is composed of multiple Outlooker modules with extended receptive fields connected in series, and the window size and receptive field range are adjusted;

[0069] When Γ≥τ2, switch to the high-resolution image processing path, expand the width of the backbone feature extraction network, and increase the number of output channels of the Transformer module;

[0070] where τ1 and τ2 represent the structural reconstruction response thresholds.

[0071] Optionally, the structural reconstruction mechanism dynamically adjusts the backbone feature extraction network of the image recognition model according to the risk growth slope and fluctuation amplitude in the health prediction result set.

[0072] The beneficial effects of the present invention are:

[0073] First of all, the present invention provides a forest and grassland health monitoring and prediction method based on image recognition, which breaks through the technical bottlenecks of existing forest and grassland ecological monitoring means that rely on static model structure, single recognition ability, disconnection between prediction and recognition, and lagging response mechanism, and constructs an intelligent health monitoring system integrating image perception, ecological modeling, state recognition, trend prediction and structural reconstruction. This method integrates image recognition model and regional ecological parameter set, and realizes deep coupling of multi-source images and ecological labels at the input layer, effectively enhancing the model's adaptability to heterogeneous image features of forest and grassland areas, especially in the identification of atypical vegetation areas such as drought patches, diseased textures, and degraded edges, showing higher accuracy and stability.

[0074] Secondly, by introducing the image recognition backbone structure built based on the VOLO network and integrating the Outlooker module and the Transformer module, the model not only has the ability to perceive the local structure of the image, but also can establish dependencies between global semantics, significantly improving the ability to capture fine-grained differences in large-scale forest and grassland images. At the same time, the designed ecological label conditional convolution mechanism uses the vegetation type label in the regional ecological parameter set to control the convolution kernel channel path, realizing the conditional adjustment of the recognition model structure to the vegetation type, so that the same model can automatically adjust the structural parameters in different ecological regions, reflecting significant domain adaptability.

[0075] In addition, after the health status identification results are generated, the present invention innovatively introduces a dual-channel trend prediction model, jointly models the health level label and the ecological variable sequence, and outputs the trend prediction value and the health risk probability value respectively. This prediction structure not only takes into account the time dependence of the health status, but also integrates the dynamic influence of ecological factors, and uses the trend-driven evolutionary regression mechanism to dynamically adjust the model parameters, making the prediction results more time-sensitive and risk-aware. Compared with traditional static prediction models, this method can effectively identify areas where risks are accelerating or fluctuating violently in advance, providing timely intervention basis for ecological management departments.

[0076] Finally, the present invention constructs a closed-loop response mechanism that reversely drives the "model structure" from the "prediction results". When the system detects that the risk probability of image units in a certain area shows an upward trend or fluctuates beyond a set threshold over multiple consecutive time periods, it will automatically trigger a structural reconstruction mechanism. This mechanism jointly calculates the reconstruction level factor based on the risk growth slope and fluctuation amplitude, and then drives the image recognition model to make structural adjustments, including activating deep Outlooker branches, expanding the receptive field, adjusting the resolution path, and expanding the Transformer trunk channel width. This dynamic structural expansion strategy not only improves the model's recognition accuracy in high-risk areas, but also realizes the differentiated allocation of computing resources, taking into account both performance and efficiency.

[0077] In summary, the present invention realizes the multi-link linkage of image recognition, ecological status modeling, trend prediction, and structural response adjustment in forest and grassland health monitoring. It has the advantages of strong structural adaptability, predictable health trends, and high degree of automation of early warning response. It is suitable for intelligent monitoring and dynamic management of large-scale and multi-time-scale forest and grassland ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 This is a flow chart of a forest and grass health monitoring and prediction method based on image recognition proposed by the present invention;

[0080] Figure 2 This is a structural block diagram of an image recognition model for a forest and grassland health monitoring and prediction method based on image recognition proposed in the present invention;

[0081] Figure 3 This is the structure of a dual-channel prediction model for a forest and grassland health monitoring and prediction method based on image recognition proposed by the present invention. DETAILED DESCRIPTION

[0082] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0083] refer to Figure 1-3 , a forest and grassland health monitoring and prediction method based on image recognition, comprising the following steps:

[0084] S1. Construct regional ecological parameter set;

[0085] S2, adopting a regionally differentiated adaptive sampling strategy and using regional ecological parameter sets to obtain multi-source image data of the target forest and grassland areas;

[0086] S3, preprocessing the multi-source image data to generate a standardized image dataset;

[0087] S4. Inputting the standardized image dataset into an image recognition model to identify the health status of forests and grasslands and output a health grade label;

[0088] S5. Construct a dual-channel prediction model to output trend prediction values and risk probability values based on health level labels and regional ecological parameter sets, and generate a health prediction result set;

[0089] S6. When the risk probability value corresponding to any image unit in the health prediction result set shows an upward acceleration trend or the fluctuation amplitude exceeds a threshold in multiple consecutive time periods, a structural reconstruction mechanism is triggered;

[0090] S7. Use the reconstructed image recognition model to re-monitor and predict forest and grassland health.

[0091] The present invention realizes full-process monitoring and intelligent prediction of the health status of forests and grasslands by constructing a complete method flow that integrates image recognition, ecological modeling, trend prediction and structural reconstruction. It has the advantages of strong closed-loop, fast response mechanism and high system linkage capability.

[0092] In this embodiment, the regional ecological parameter set includes historical environmental data, meteorological data, geographic information and vegetation type labels of the target forest and grassland area.

[0093] The present invention integrates historical environment, meteorological, geographical and vegetation type information in the construction of regional ecological parameter sets, providing a multi-dimensional ecological foundation for subsequent image sampling, recognition and prediction, and improving the regional adaptability and recognition accuracy of the model.

[0094] In this embodiment, the multi-source image data includes visible light images, multispectral images and drone aerial images.

[0095] The multi-source image data collected by the present invention includes visible light, multispectral and drone aerial images, which significantly enhances the data dimension and image information integrity of health monitoring, and provides data guarantee for health identification at different scales and states.

[0096] In this embodiment, S2 specifically includes:

[0097] S21, spatially dividing the target forest and grassland area based on the regional ecological parameter set, dividing the entire monitoring area into image units according to a set spatial resolution, and assigning a unique spatial identification number to each image unit;

[0098] S22. Analyze ecological indicators for each image unit and generate corresponding sampling priority scores based on vegetation index changes, meteorological fluctuation characteristics, historical health status stability, and human disturbance frequency;

[0099] S23. Combining the sampling priority score of the image unit with the geographical distribution density to form a sampling density control strategy with spatial weights, and dynamically adjusting the sampling frequency of the image unit according to the strategy;

[0100] S24. Select a target sampling area for the current cycle according to the control strategy, and collect images.

[0101] The present invention constructs a sampling priority score and dynamically adjusts the sampling frequency through regional differentiated ecological factor analysis, thereby realizing intelligent sampling control of image units and improving data acquisition efficiency and representativeness.

[0102] In this embodiment, the preprocessing includes image registration, illumination normalization, geometric correction and image enhancement.

[0103] The present invention introduces registration, illumination normalization, geometric correction and image enhancement steps in the image preprocessing stage, which significantly improves the spatial consistency and quality stability of multi-source image data and lays a good foundation for subsequent model input.

[0104] In this embodiment, the image recognition model includes a backbone feature extraction network constructed based on the VOLO network and an eco-label conditional convolution module. The eco-label conditional convolution module selects the corresponding convolution channel path according to the vegetation type label in the regional ecological parameter set. The VOLO network obtains local structural features through the Outlooker module, and models global semantic relationships through the Transformer module to output health recognition results. The Outlooker module is a local attention mechanism module.

[0105] The present invention realizes adaptive health recognition of different vegetation types through an image recognition model built based on the VOLO network, integrating the Outlooker local modeling and Transformer global modeling capabilities, and combining the eco-label conditional convolution module.

[0106] In this embodiment, the S4 specifically includes:

[0107] S41. Construct an input image tensor based on a standardized image dataset;

[0108] S42: Input the input image tensor into the backbone feature extraction network built based on the VOLO network, perform local structure perception through the Outlooker module, and extract the local context feature map of the image:

[0109]

[0110] Among them, F L (i, j) represents the local context feature map, describing the local feature value at position (i, j). Represents the local receptive field neighborhood centered at position (i, j), A i,j,m,n represents the attention weight of position (m,n) to position (i,j) in the input image tensor, and X(m,n) represents the pixel value at position (m,n) in the input image tensor;

[0111] S43, pass the local context feature map into the multi-layer Transformer module, use the multi-head attention mechanism to calculate the global dependency feature map, and obtain the fused feature map:

[0112] F O =Transformer(F L )=Concat(head1,...,head k )·W O ;

[0113] Among them, F O Represents the fusion feature map, head k represents the output of the kth attention head, W O Represents the output linear transformation weight matrix, Concat(...) represents the concatenation operation of the outputs of multiple attention heads;

[0114]

[0115] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, d k Represents the dimension of each attention head, Softmax represents the row normalization operation, and T represents the transposition operation;

[0116] S44. Input the fused feature map into the ecological label conditional convolution module, generate a one-hot encoding vector based on the vegetation type label in the regional ecological parameter set, and control the activation state of multiple predefined convolution kernel paths based on the one-hot encoding vector to output the ecological modulation feature map:

[0117]

[0118] Among them, F E Represents the ecological modulation characteristic map, e k Indicates the kth bit in the one-hot encoding vector. When the value is 1, it means the current vegetation type label is l k , activate the corresponding convolution kernel W k , otherwise it will not participate in the calculation. Indicates the use of convolution kernel W k Perform convolution operation on the fused feature map, W k represents the convolution kernel;

[0119] S45. Input the ecological modulation feature map into the classification head, perform image unit health status recognition, and output a health recognition result, which includes a health level label corresponding to the image unit:

[0120] Y h =arg max(Softmax(W c ·F E +b c ));

[0121] Among them, Y h Indicates the health level label, with values of 0, 1, or 2, representing healthy, sub-healthy, and abnormal, respectively. c represents the weight matrix of the classification linear layer, b c Represents the bias term of the classification linear layer, Softmax represents the normalization of each category score, and arg max represents the variable value that maximizes the function value.

[0122] The present invention refines the processing flow of standardized images and constructs a complete recognition path from local feature extraction to ecological modulation to health level classification, thereby improving the classification accuracy and recognition ability of the model for heterogeneous image regions.

[0123] In this embodiment, the S5 specifically includes:

[0124] S51. Constructing a health status sequence based on the health level labels and corresponding timestamps, and constructing an ecological status sequence based on the regional ecological parameter set;

[0125] S52. Calculate the slope of the health change trend based on the health status sequence:

[0126]

[0127] Among them, Δ h represents the slope of health change trend, Indicates the health level label at the current moment, represents the health level label of the first q time steps, t n and t n-q are the timestamps of the current and previous q time steps respectively;

[0128] S53. Introduce a trend-driven evolutionary regression mechanism, which dynamically adjusts the prediction model parameters based on the slope of the health change trend to generate evolutionary parameters:

[0129] ω′ k =ω k ·(1+α·Δ h ),θ′ j =θ j ·(1+β·|Δ h |);

[0130] Among them, ω′ k represents the updated regression weight, ω k represents the regression weight, α represents the trend sensitivity coefficient, Δ h represents the slope of health change trend; θ′ j represents the updated ecological variable weight, θ j represents the ecological variable weight, β represents the risk sensitivity coefficient, |Δ h | represents the absolute value of the slope of the health change trend;

[0131] S54. Input the health status sequence and the ecological status sequence into the dual-channel prediction model, perform parallel prediction using the evolutionary parameters, and obtain trend prediction values and risk probability values, respectively. Use a dynamic weighted regression model to obtain the trend prediction value. The dual-channel prediction model includes a dynamic weighted regression model and a logistic regression model:

[0132]

[0133] in, represents the health grade label at the t+1 time step, ω0 represents the regression bias term, ω′ k represents the updated regression weight, represents the health level label of the t+1-kth time step, and n represents the total number of time steps;

[0134] Use the logistic regression model to output the risk probability value:

[0135]

[0136] in, represents the risk probability value at time step t+1, θ0 represents the logistic regression bias term, and θ′ j represents the updated ecological variable weight, represents the value of the jth ecological variable at the current moment, m represents the total number of ecological variables, and exp represents the natural exponential function;

[0137] S55 , repeating the prediction process in the sliding time window to generate a health prediction result set for future time steps, wherein the health prediction result set includes a trend prediction value, a risk probability value, and a corresponding timestamp.

[0138] The present invention constructs a dual-channel prediction model, combines health status and ecological variable sequences, dynamically adjusts prediction parameters, and accurately outputs trend and risk information, thereby realizing intelligent prediction and trend perception of the healthy evolution process of forests and grasslands.

[0139] In this embodiment, S6 specifically includes:

[0140] S61. Calculate the risk growth slope and fluctuation range of the target image unit within the predicted time period based on the predicted risk probability value:

[0141]

[0142] Among them, Δ r represents the risk growth slope, and Represent the risk probability values at the last and first time points of the forecast period, t n+M and t n+1 Indicates the corresponding timestamp; σ r Indicates the fluctuation range, M indicates the total length of the forecast period, represents the risk probability value of n+k time steps, Indicates the average risk probability value within a time period;

[0143] S62. Set the structural reconstruction triggering condition. When any inequality is met, the structural reconstruction mechanism is triggered:

[0144] Δ r >λ1,σ r >λ2;

[0145] Among them, λ1 represents the risk slope trigger threshold, λ2 represents the risk fluctuation trigger threshold, Δ r represents the risk growth slope, σ r Indicates the amplitude of fluctuation;

[0146] S63. When the structural reconstruction mechanism is triggered, the reconstruction level factor is determined based on the risk growth slope and volatility:

[0147] Γ=γ1·Δ r +γ2·σ r ;

[0148] Where Γ represents the reconstruction level factor in the current forecast period, γ1 and γ2 represent the risk growth and volatility response weights preset by the system;

[0149] S64. Performing dynamic structural adjustment of the image recognition model according to the structural reconstruction level factor, including:

[0150] When Γ<τ1, the original structure remains unchanged;

[0151] When τ1≤Γ<τ2, the deep Outlooker branch is activated and the preset deep local attention path is automatically loaded. The deep local attention path is composed of multiple Outlooker modules with extended receptive fields connected in series, and the window size and receptive field range are adjusted;

[0152] When Γ≥τ2, switch to the high-resolution image processing path, expand the width of the backbone feature extraction network, and increase the number of output channels of the Transformer module;

[0153] where τ1 and τ2 represent the structural reconstruction response thresholds.

[0154] The present invention constructs a structural reconstruction trigger mechanism based on the growth trend and fluctuation amplitude of risk probability, and introduces a reconstruction level factor to dynamically adjust the model structure, thereby improving the identification flexibility and responsiveness of high-risk areas.

[0155] In this embodiment, the structural reconstruction mechanism dynamically adjusts the backbone feature extraction network of the image recognition model according to the risk growth slope and fluctuation range in the health prediction result set.

[0156] The present invention utilizes the risk trend characteristics in the prediction results to dynamically control the backbone structure of the image recognition model, achieves local structure accuracy enhancement and differential configuration of computing resources, and optimizes the overall recognition performance and processing efficiency of the model.

[0157] Example 1:

[0158] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a forest-grassland transition zone within the jurisdiction of the Ewenki Autonomous Banner, Hulunbuir City, Inner Mongolia Autonomous Region for actual deployment and testing from June 2023 to October 2023. The ecological environment in this area is complex. It contains typical temperate coniferous and broad-leaved mixed forest areas, large areas of degraded grasslands and partial desertification expansion zones. It has obvious coexistence characteristics of multiple ecological types and is an ideal area for forest and grassland health monitoring and trend forecasting research.

[0159] The system deployed in this implementation plan uses fixed ground camera poles combined with multi-rotor drones to collect image data, and at the same time retrieves environmental data from local ecological monitoring stations in the area to construct a complete multi-source image input and ecological parameter set. The image resolution is set to 512×512, and a multi-scale stitching method is used to cover large area image units. The ecological parameter sampling interval is once every 6 hours, and the image sampling frequency is twice a day.

[0160] Based on data collection, the system first performs image preprocessing on the collected images, including image registration, illumination normalization and geometric correction, to ensure the spatial consistency and stability of the image. In the image recognition stage, the image recognition backbone is constructed based on the VOLO-D2 network, the Outlooker module is used to extract local structural features, and the Transformer module completes the global modeling. Then, the ecological label conditional convolution module is used to realize structural path selection and ecological feature modulation in combination with the vegetation type label. The recognition module outputs the health level label of each image unit at each moment, including three states: healthy (0), sub-healthy (1) and abnormal (2).

[0161] Subsequently, a health status time series is constructed based on the recognition results. Ecological variables such as regional temperature, humidity, NDVI, and soil moisture are introduced to form an ecological status sequence. A dual-channel prediction module uses a dynamic weighted regression model and a logistic regression model to output health trend predictions and risk probability values, respectively. A trend-driven evolutionary mechanism adaptively adjusts the regression parameters, enabling the model to respond to changes in slope. If the risk probability value shows an upward trend or fluctuates significantly over multiple consecutive time periods, the system triggers a structural reconstruction mechanism, activating a deep Outlooker branch or switching to a high-resolution image channel. The system also dynamically adjusts the width of the Transformer module, thereby improving recognition accuracy in complex areas.

[0162] Table 1 Comparison of experimental performance between the forest and grass health monitoring method of the present invention and the prior art

[0163]

[0164] From the perspective of overall recognition capability, the present invention achieved an image recognition accuracy of 91.6%, which is a significant improvement compared to the traditional benchmark method (77.8%) based on NDVI threshold judgment, with an accuracy increase of 13.8 percentage points. This result fully verifies the effectiveness of the present invention in complex forest and grassland image recognition through the VOLO backbone network structure, eco-label conditional convolution module and local-global feature fusion mechanism, especially showing stronger discrimination ability in areas with strong heterogeneity and complex textures.

[0165] In terms of recognition performance in each category, the F1-score of the healthy category increased from 84.1% of the traditional method to 92.3%, indicating that the model has stable and accurate recognition effects in normal areas. In the more challenging sub-healthy and abnormal categories, the F1-score of the present invention reached 89.1% and 91.5%, respectively, which are 15.3% and 18.7% higher than the traditional method, fully demonstrating the adaptability of the conditional convolution structure and ecological modulation mechanism to difficult-to-identify targets such as abnormal plaques and lesions.

[0166] In terms of trend prediction performance, the trend-driven evolution prediction model proposed in this paper demonstrates significant superiority. In a three-day health trend forecast, the root mean square error (RMSE) of the prediction was only 0.248, and the mean absolute error (MAE) was 0.194. These performances significantly outperformed the baseline ARIMA model's RMSE (0.347) and MAE (0.286), representing reductions of 28.4% and 32.2%, respectively. This demonstrates that the present invention not only accurately identifies the current state but also possesses strong trend tracking capabilities, enabling earlier judgments for ecological interventions.

[0167] The structural reconstruction mechanism, a key innovation of this invention, has demonstrated practical application in high-risk areas. Upon identifying high-risk evolutionary trends, the system activates the structural reconstruction logic. This logic improves recognition capabilities by invoking deep Outlooker modules and switching to high-resolution paths, resulting in an average accuracy improvement of approximately 9.8%. Furthermore, the system's computational resource overhead is kept within a 3.1% time increase, achieving a good balance between performance and efficiency, making it suitable for engineering deployment.

[0168] In addition, the present invention has for the first time achieved "trend-guided" prediction in terms of early warning capabilities in high-risk areas, and can issue abnormal status warnings on average 2.6 days earlier than traditional methods, significantly improving the response time and intervention efficiency of ecological management. This capability is due to the fusion modeling mechanism of risk probability paths and ecological variables, which enables the system to not only identify the current status, but also have insight into potential evolution directions.

[0169] In terms of image data input compatibility, the present invention supports multi-source fusion processing of visible light, multispectral and drone aerial images, which is far superior to most traditional methods that only support NDVI single-channel input. It exhibits strong adaptability to ecological heterogeneity and can meet monitoring needs in different regions and seasonal changes.

[0170] Overall, the present invention performs superiorly in terms of recognition accuracy, prediction capability, intelligent structural response, and engineering practicality in forest and grassland ecological monitoring, significantly outperforming traditional static methods, and has value for promotion and practical application in a wide range of ecological health management systems.

[0171] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A forest and grass health monitoring and prediction method based on image recognition, characterized in that: The steps include: S1. Construct regional ecological parameter set; S2, adopting a regionally differentiated adaptive sampling strategy and using regional ecological parameter sets to obtain multi-source image data of the target forest and grassland areas; S3, preprocessing the multi-source image data to generate a standardized image dataset; S4. Inputting the standardized image dataset into an image recognition model to identify the health status of forests and grasslands and output a health grade label; S5. Construct a dual-channel prediction model to output trend prediction values and risk probability values based on health level labels and regional ecological parameter sets, and generate a health prediction result set; S6. When the risk probability value corresponding to any image unit in the health prediction result set shows an upward acceleration trend or the fluctuation amplitude exceeds a threshold in multiple consecutive time periods, a structural reconstruction mechanism is triggered; S7. Use the reconstructed image recognition model to re-monitor and predict forest and grassland health.

2. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The regional ecological parameter set includes historical environmental data, meteorological data, geographic information and vegetation type labels of the target forest and grassland area.

3. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The multi-source image data includes visible light images, multispectral images and drone aerial images.

4. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The S2 specifically includes: S21, spatially dividing the target forest and grassland area based on the regional ecological parameter set, dividing the entire monitoring area into image units according to a set spatial resolution, and assigning a unique spatial identification number to each image unit; S22. Analyze ecological indicators for each image unit and generate corresponding sampling priority scores based on vegetation index changes, meteorological fluctuation characteristics, historical health status stability, and human disturbance frequency; S23. Combining the sampling priority score of the image unit with the geographical distribution density to form a sampling density control strategy with spatial weights, and dynamically adjusting the sampling frequency of the image unit according to the strategy; S24. Select a target sampling area for the current cycle according to the control strategy, and collect images.

5. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The preprocessing includes image registration, illumination normalization, geometric correction and image enhancement.

6. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The image recognition model includes a backbone feature extraction network built based on the VOLO network and an eco-label conditional convolution module. The eco-label conditional convolution module selects the corresponding convolution channel path based on the vegetation type label in the regional ecological parameter set. The VOLO network obtains local structural features through the Outlooker module and models global semantic relationships through the Transformer module to output health recognition results. The Outlooker module is a local attention mechanism module.

7. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The S4 specifically includes: S41. Construct an input image tensor based on a standardized image dataset; S42, inputting the input image tensor into the backbone feature extraction network built based on the VOLO network, performing local structure perception through the Outlooker module, and extracting the image local context feature map; S43, passing the local context feature map into the multi-layer Transformer module, using the multi-head attention mechanism to calculate the global dependency feature map to obtain the fused feature map; S44. Input the fused feature map into the ecological label conditional convolution module, generate a one-hot encoding vector based on the vegetation type label in the regional ecological parameter set, and control the activation state of multiple predefined convolution kernel paths based on the one-hot encoding vector to output the ecological modulation feature map: Among them, F E Represents the ecological modulation characteristic map, e k Indicates the kth bit in the one-hot encoding vector. When the value is 1, it means the current vegetation type label is l k , activate the corresponding convolution kernel W k , otherwise it will not participate in the calculation. Indicates the use of convolution kernel W k Perform convolution operation on the fused feature map, W k represents the convolution kernel; S45 , input the ecological modulation feature map into the classification head, perform image unit health status recognition, and output a health recognition result, wherein the health recognition result includes a health level label corresponding to the image unit.

8. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The S5 specifically includes: S51. Constructing a health status sequence based on the health level labels and corresponding timestamps, and constructing an ecological status sequence based on the regional ecological parameter set; S52. Calculating the slope of the health change trend based on the health status sequence; S53. Introduce a trend-driven evolutionary regression mechanism, which dynamically adjusts the prediction model parameters based on the slope of the health change trend to generate evolutionary parameters: oh' k =ω k ·(1+a·D h ),θ′ j =θ j ·(1+β·|D h |); Among them, ω′ k represents the updated regression weight, ω k represents the regression weight, α represents the trend sensitivity coefficient, Δ h represents the slope of health change trend; θ′ j represents the updated ecological variable weight, θ j represents the ecological variable weight, β represents the risk sensitivity coefficient, |Δ h | represents the absolute value of the slope of the health change trend; S54. Input the health status sequence and the ecological status sequence into the dual-channel prediction model, perform parallel prediction using the evolutionary parameters, and obtain trend prediction values and risk probability values, respectively. Use a dynamic weighted regression model to obtain the trend prediction value. The dual-channel prediction model includes a dynamic weighted regression model and a logistic regression model: in, represents the health grade label at the t+1 time step, ω0 represents the regression bias term, ω′ k represents the updated regression weight, represents the health level label of the t+1-kth time step, and n represents the total number of time steps; Use the logistic regression model to output the risk probability value: in, represents the risk probability value at time step t+1, θ0 represents the logistic regression bias term, and θ′ j represents the updated ecological variable weight, represents the value of the jth ecological variable at the current moment, m represents the total number of ecological variables, and exp represents the natural exponential function; S55 , repeating the prediction process in the sliding time window to generate a health prediction result set for future time steps, wherein the health prediction result set includes a trend prediction value, a risk probability value, and a corresponding timestamp.

9. The forest and grass health monitoring and prediction method based on image recognition according to claim 1 is characterized in that: The S6 specifically includes: S61. Calculate the risk growth slope and fluctuation range of the target image unit within the predicted time period based on the predicted risk probability value; S62. Set the structural reconstruction triggering conditions. When any inequality is met, the structural reconstruction mechanism is triggered: D r >λ1,σ r >λ2; Among them, λ1 represents the risk slope trigger threshold, λ2 represents the risk fluctuation trigger threshold, Δ r represents the risk growth slope, σ r Indicates the amplitude of fluctuation; S63. When the structural reconstruction mechanism is triggered, the reconstruction level factor is determined based on the risk growth slope and volatility: C=c1·D r +γ2·σ r ; Where Γ represents the reconstruction level factor in the current forecast period, γ1 and γ2 represent the risk growth and volatility response weights preset by the system; S64. Performing dynamic structural adjustment of the image recognition model according to the structural reconstruction level factor, including: When Γ<τ1, the original structure remains unchanged; When τ1≤Γ<τ2, the deep Outlooker branch is activated and the preset deep local attention path is automatically loaded. The deep local attention path is composed of multiple Outlooker modules with extended receptive fields connected in series, and the window size and receptive field range are adjusted; When Γ≥τ2, switch to the high-resolution image processing path, expand the width of the backbone feature extraction network, and increase the number of output channels of the Transformer module; where τ1 and τ2 represent the structural reconstruction response thresholds.

10. The forest and grass health monitoring and prediction method based on image recognition according to claim 9 is characterized in that: The structural reconstruction mechanism dynamically adjusts the backbone feature extraction network of the image recognition model according to the risk growth slope and fluctuation amplitude in the health prediction result set.

Images