An unmanned aerial multi-modal vegetation observation system

Abstract

The application discloses a kind of unmanned aerial vehicle multi-modal vegetation observation systems, it is related to agricultural remote sensing and intelligent monitoring technical field.The unmanned aerial vehicle platform is included;Multi-modal data acquisition module is used to collect the multi-modal data of target area and generate current observation state vector;Mirror state modeling module is used to construct the vegetation mirror state vector of target area under health state based on historical flight data;Deviation calculation module is used to calculate the state deviation of current observation state vector and vegetation mirror state vector;Path optimization module is used to iteratively update the path control point of unmanned aerial vehicle according to the spatial distribution of state deviation in target area, and generate optimized path for subsequent flight.The application introduces vegetation mirror state as health reference, realizes the priority of unmanned aerial vehicle to abnormal area by path optimization algorithm driven by state deviation.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and intelligent monitoring technology, and more specifically, to an unmanned aerial vehicle (UAV)-borne multimodal vegetation observation system. Background Technology

[0002] With the development of precision agriculture and smart agriculture, using drones equipped with sensors for high-frequency, high-precision remote sensing observation of farmland has become an important means of crop health monitoring, growth assessment, and pest and disease early warning. Vegetation observation systems typically rely on vegetation indices (such as NDVI and NDRE) extracted from remote sensing images to characterize crop physiological states. These indices can reflect growth characteristics such as chlorophyll content, vegetation density, and water status to a certain extent. Drones have the advantages of maneuverability, low-altitude high resolution, and can quickly cover large areas of farmland, becoming an important supplement to traditional ground surveys and satellite remote sensing.

[0003] Currently, multimodal remote sensing data acquisition is becoming increasingly widespread, with typical modalities including visible light images, infrared images, and multispectral reflectance data. Different modalities are complementary: visible light reflects apparent color changes, infrared is used to sense thermal stress, and multispectral data can be used to calculate various vegetation indices. However, fusing these modal data into a consistent representation remains a challenge. Furthermore, vegetation status exhibits strong spatiotemporal dynamics, influenced by multiple factors such as water, nutrients, pests and diseases, and climate fluctuations. Therefore, observation systems not only need to identify anomalous areas but also need dynamic response capabilities to support subsequent intervention decisions.

[0004] In current research and applications, the following key issues still urgently need to be addressed:

[0005] 1. Vegetation observations are mostly based on remote sensing data at a single moment, lacking a clear reference for "health status", making it difficult to quantitatively identify abnormal areas that deviate from the ideal growth state;

[0006] 2. Multimodal remote sensing data are often processed independently and are not fully integrated into a unified representation of crop status, which affects the ability to conduct comprehensive analysis;

[0007] 3. Drone flight paths usually rely on manual settings or static rules, lacking a mechanism for dynamically optimizing paths based on the actual vegetation conditions, resulting in insufficient observation efficiency and coverage quality;

[0008] 4. Although abnormal areas can be initially identified, the judgment of the specific stress type behind the abnormality (such as drought, pests and diseases, nutrient deficiency) still relies on human experience or subsequent ground verification, making it difficult to achieve automated response.

[0009] The aforementioned problems have not yet been effectively solved in current agricultural remote sensing monitoring practices. There is an urgent need to build an intelligent vegetation observation system with state reference capabilities, multimodal fusion expression capabilities, dynamic path scheduling capabilities, and stress classification capabilities to better support efficient and precise crop health monitoring and agricultural decision-making. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide an unmanned aerial vehicle (UAV)-borne multimodal vegetation observation system to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] An unmanned aerial vehicle (UAV)-borne multimodal vegetation observation system includes:

[0013] Unmanned aerial vehicle (UAV) platforms are used to carry multimodal sensors and perform observation flights in target areas;

[0014] A multimodal data acquisition module, installed on the UAV platform, is used to simultaneously acquire visible light images, infrared images, and multispectral reflectance data of the target area, and generate a current observation state vector;

[0015] The mirror state modeling module is used to construct a vegetation mirror state vector of the target area in a healthy state based on historical flight data. The vegetation mirror state vector is a multi-dimensional feature representation of the ideal vegetation growth state.

[0016] The deviation calculation module is used to compare the current observed state vector with the vegetation mirror state vector and calculate the state deviation between the two.

[0017] The path optimization module is used to iteratively update the path control points of the UAV based on the spatial distribution of the state deviation within the target area using a gradient descent optimization algorithm with the state deviation as the objective function, thereby generating an optimized path for subsequent flight and enabling the UAV to prioritize observation of areas with significant deviations.

[0018] In some embodiments, the path optimization module includes:

[0019] A deviation heatmap generation unit is used to visualize the spatial distribution of the state deviation within the target area as a deviation heatmap;

[0020] The path control point setting unit is used to set multiple adjustable path control points on the initial path and use the position of each path control point as an optimization variable.

[0021] The gradient calculation unit is used to calculate the deviation gradient information based on the position of each path control point in the deviation heatmap.

[0022] The control point update unit is used to iteratively update the position of each path control point according to the deviation gradient information using a gradient descent optimization algorithm until the optimized path converges within a preset error range or reaches the maximum number of iterations.

[0023] In some embodiments, the multimodal data acquisition module includes a time synchronization device and a spatial calibration unit. The time synchronization device is used to align the acquisition clocks of visible light images, infrared images, and multispectral reflectance data. The spatial calibration unit is used to unify different modal data into an observation state vector under the same geographic coordinate system.

[0024] In some embodiments, the mirror state modeling module selects observation data from multimodal samples collected in the target area over a historical time period that have NDVI and NDRE vegetation indices in the optimal range, and constructs a vegetation mirror state vector that includes statistical mean and principal component features.

[0025] In some embodiments, the system further includes a stress classification module for classifying and judging the crop stress type in the significantly biased region based on the state deviation, wherein the crop stress type includes any one or more of the following: drought stress, pest and disease stress, or nutrient deficiency.

[0026] In some embodiments, the stress classification module automatically determines the crop stress type in areas with significant deviations using a machine learning classification model trained on it. The machine learning classification model is trained under supervision using historical labeled state deviation sample data, where the labels are crop stress types.

[0027] In some embodiments, the path optimization module further includes a stress type awareness submodule, which is used to introduce stress type matching constraints during the path optimization process based on the target stress type specified by the user.

[0028] In some embodiments, the process of introducing stress type matching constraints specifically includes:

[0029] Based on the stress type labels and their confidence levels at each spatial location output by the stress classification module, a spatial distribution map of each stress type within the target area is generated.

[0030] During the path optimization process, based on the spatial distribution map, deviation regions that match the user's interest type are selected as priority observation targets, and the weight coefficients of the corresponding regions in the deviation heatmap are adjusted to give them higher attention in gradient descent optimization.

[0031] In some embodiments, the process of calculating the state deviation specifically includes:

[0032] Align the current observation state vector with the vegetation mirror state vector at the corresponding location according to dimensions to construct a difference vector, where each dimension of the difference vector represents the degree of deviation of a certain modal feature;

[0033] The difference vector is normalized using a weighted Euclidean distance to obtain the state deviation that reflects the intensity of the overall deviation.

[0034] In some embodiments, the deviation calculation module further includes a deviation significance determination unit, used to determine whether there is a significant deviation region in the target region based on the state deviation, specifically including:

[0035] The target area is divided into several spatial grid cells, and the state deviation of all observation points in each grid cell is calculated.

[0036] Based on the distribution characteristics of the state deviation values ​​across the entire region, a significance threshold is set. The threshold can be the global mean plus a preset multiple of the standard deviation, or it can be set as the lower limit of the values ​​of the top several percentiles of the deviation values.

[0037] Grid cells whose state deviation values ​​exceed the significance threshold are marked as areas of significant deviation, and their spatial coordinates are output.

[0038] The advantages of this invention compared to existing technologies lie in its ability to construct a mirror-image state vector of vegetation in a healthy target area and compare it with the currently observed state vector to obtain a state deviation reflecting the degree of vegetation degradation. Based on this, the system uses the state deviation as the objective function and dynamically updates the flight path using a gradient descent algorithm, prioritizing the observation of areas with significant deviations by the UAV. This mechanism breaks through the limitations of traditional path planning that relies on static preset routes or rule-based judgments, achieving dynamic flight scheduling driven by "crop state perception," thus improving observation accuracy and efficiency. Building upon this core mechanism, this invention also introduces a crop stress type classification module. By training and modeling the state deviation using machine learning methods, it can automatically determine the stress type (such as drought, pests and diseases, nutrient deficiency) of areas with significant deviations, providing fine-grained response information for agricultural management. Furthermore, the system allows users to set the stress types they are interested in and dynamically adjusts the priority of target areas during path optimization, achieving "on-demand scheduling" directional observation capabilities. Attached Figure Description

[0039] Figure 1 This is an overall workflow diagram of the present invention;

[0040] Figure 2 This is a flowchart of the observation state vector generation process of this invention;

[0041] Figure 3 This is a flowchart of the state deviation calculation and significant region identification process of the present invention;

[0042] Figure 4 This is a flowchart of the path optimization and stress type priority response mechanism of the present invention. Detailed Implementation

[0043] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0044] This invention proposes an unmanned aerial vehicle (UAV)-borne multimodal vegetation observation system, which aims to achieve accurate monitoring of vegetation status in target areas and priority observation of abnormal areas through multimodal data acquisition, state modeling, deviation analysis, and dynamic path optimization.

[0045] like Figure 1 As shown, the system of the present invention mainly consists of an unmanned aerial vehicle platform, a multimodal data acquisition module, an image state modeling module, a deviation calculation module, a path optimization module, and a stress classification module.

[0046] The drone platform serves as the hardware carrier, carrying multimodal sensors and performing flight missions.

[0047] The data acquisition module is responsible for synchronously acquiring observation data from multiple modalities and generating the current observation state vector.

[0048] The mirror state modeling module uses historical data to construct a reference state for healthy vegetation;

[0049] The deviation calculation module quantifies the degree of vegetation deviation by comparing the current state with the reference state.

[0050] The path optimization module dynamically adjusts the flight path based on the spatial distribution of the deviation, prioritizing coverage of abnormal areas.

[0051] The stress classification module further analyzes the stress types in abnormal regions and supports priority observation of user-defined stress types.

[0052] To ensure the stable operation of the system, the drone platform is selected from multi-rotor drones with high load capacity and long endurance, such as hexa-rotor or octo-rotor models, with a maximum takeoff weight of not less than 10 kg and an endurance of not less than 30 minutes.

[0053] The drone is equipped with sensors including a high-resolution visible light camera, an infrared thermal imager, and a multispectral sensor covering the 400-1000nm wavelength band, including at least the blue, green, red, red-edge, and near-infrared bands. These sensors ensure data alignment in time and space through a unified time synchronization device and spatial calibration unit.

[0054] Among them, such as Figure 2As shown, the multimodal data acquisition module is responsible for acquiring visible light images, infrared images, and multispectral reflectance data, and integrating them into a unified current observation state vector. To achieve data synchronization, the time synchronization device uses a high-precision GPS module to uniformly calibrate the acquisition clocks of all sensors through a trigger signal. For example, during a flight, the UAV acquires data at a frequency of 1Hz, generating a set of samples containing visible light, infrared, and multispectral data per second.

[0055] The spatial calibration unit calibrates the data from different sensors to the WGS84 coordinate system using geographic coordinate system calibration. In practice, the UAV's inertial navigation system is first used to record the geographic location of each frame of data, including longitude, latitude, and altitude. Then, the image pixels are mapped to geographic coordinates to generate spatially aligned multimodal data.

[0056] For multispectral data, band registration techniques are employed to ensure a one-to-one correspondence between reflectance data from different bands and pixels in the visible light image. Ultimately, the current observation state vector consists of multidimensional features, including, but not limited to, the following:

[0057] Visible light characteristics, the average intensity of the RGB color channels, ranging from [0,255], reflect the color and morphology of vegetation.

[0058] Infrared features, the average temperature value of infrared images, in degrees Celsius, typically in the range [0, 50], are used to detect heat stress in vegetation.

[0059] Multispectral characteristics, Normalized Difference Vegetation Index (NDVI) and Normalized Difference Red Edge Index (NDRE), are calculated using the following formulas:

[0060]

[0061] NIR represents near-infrared reflectance, Red represents red reflectance, and RedEdge represents red-edge reflectance. These indices range from -1 to 1 and reflect the health status and photosynthetic activity of vegetation.

[0062] The process of generating state vectors includes preprocessing of the original data, such as noise removal, correction for illumination effects, and normalization.

[0063] Suppose the original data of an observation point within the target area is as follows:

[0064] (R,G,B,T,ρ Blue ,ρ Green ,ρ Red ,ρ RedEdge ,ρ NIR );

[0065] R, G, B represent RGB intensity, T is infrared temperature, and ρ represents reflectivity of each band. The current observation state vector can then be expressed as: S t =[R n G n B n ,T n ,NDVI,NDRE];

[0066] Among them, (R) n G n B n ,T n These are the normalized RGB intensity and temperature values, respectively. The normalization formula is:

[0067]

[0068] Where, x min and x max These are the minimum and maximum values ​​of the corresponding features, respectively.

[0069] The mirror state modeling module aims to construct a vegetation reference state for a target area under healthy conditions, generating a vegetation mirror state vector. To this end, the system utilizes historical flight data to select sample data where the vegetation is in its optimal growth state. Specifically, the selection criteria are based on NDVI and NDRE indices, with the optimal interval set as follows:

[0070] NDVI∈[0.6,0.9] indicates that the vegetation has high photosynthetic activity.

[0071] NDRE∈[0.4,0.8] indicates that the reflectance of the red edge band of vegetation is normal.

[0072] The filtered sample data is used to construct a statistical model with multidimensional features. Assume the historical data contains N observation points, and the feature vector for each observation point is S. i =[R n,i G n,i B n,i ,T n,i NDVI i ,NDRE i ];

[0073] The steps for calculating the mirror state vector are as follows:

[0074] Calculate the mean for each feature dimension to obtain:

[0075]

[0076] Where, μ = [μ R ,μ G ,μ B ,μ T ,μNDVI ,μ NDRE ] represents the average feature of each component of the feature vector under healthy conditions.

[0077] To capture the correlation between features, PCA can be used to extract the main feature components and generate principal component feature vectors.

[0078] Combining the mean vector and principal component features yields the vegetation mirror state vector:

[0079] S m =[μ R ,μ G ,μ B ,μ T ,μ NDVI ,μ NDRE PC1, PC2, ..., PC k ];

[0080] Among them, PC i Principal component components enhance the expressive power of the state vector, and can also be calculated for the current observed state vector.

[0081] For example, in an observation of a cornfield, historical data might show that healthy vegetation has an average NDVI of 0.75, an average NDRE of 0.55, a greenish RGB color, and an average infrared temperature of 25°C. Using the method described above, the generated mirrored state vector can comprehensively characterize the multimodal features of healthy vegetation.

[0082] like Figure 3 As shown, the deviation calculation module quantifies the degree of vegetation degradation by comparing the current observed state vector with the mirror state vector. For any observation point within the target area, let the current state vector be S. t The mirror state vector is S m The deviation calculation process is as follows:

[0083] Vector alignment, ensuring S t and S m The dimensions are consistent. If the dimensions are different (e.g., due to the addition of a new sensor), they are made up by interpolation or dimensionality reduction.

[0084] Difference vector calculation: Calculate the difference vector.

[0085]

[0086] Where, d i This represents the deviation of the i-th dimension feature.

[0087] To integrate multidimensional deviations, weighted Euclidean distance is used to calculate state deviations:

[0088] Among them, w i The weight coefficients for the i-th feature reflect its importance to vegetation health. For example, the weights for NDVI and NDRE can be set to 0.3, RGB features to 0.15, infrared temperature to 0.2, principal component to 0.1, and the sum to 1. The weights are chosen based on the experience of agricultural experts and the feature's ability to distinguish stress types. NDVI and NDRE have higher weights because they directly reflect the photosynthetic capacity and health status of the vegetation.

[0089] To identify anomalous regions, the target area is divided into a regular spatial grid (e.g., 10m × 10m per grid). For each observation point within the grid, the average state deviation is calculated:

[0090] Where M is the number of observation points within the grid. Based on the distribution of deviation values ​​across the entire region, a significance threshold is set: τ = μ Δ +k·σ Δ ;

[0091] Where, μ Δ and σ Δ These represent the mean and standard deviation of the deviation across the entire region, respectively, with k being a multiplier, typically between 1.5 and 2. Alternatively, a threshold can be set as the lower limit of the top 10% of deviation values. Grids with deviation values ​​exceeding the threshold are marked as areas of significant deviation, and their center coordinates are recorded.

[0092] For example, in monitoring an orchard, the average deviation of a certain grid is 0.8, while the mean for the entire area is 0.3 and the standard deviation is 0.2. If k = 2, then the threshold is 0.3 + 2 * 0.2 = 0.7. This grid is marked as an abnormal area because its deviation exceeds the threshold.

[0093] like Figure 4 As shown, the path optimization module dynamically adjusts the UAV's flight path based on the spatial distribution of the deviation, prioritizing coverage of areas with significant deviations. The specific implementation includes the following steps:

[0094] The state deviation values ​​are mapped to a two-dimensional space of the target area to generate a deviation heatmap. The heatmap uses interpolation methods, such as inverse distance weighting, to transform discrete deviation values ​​into a continuous distribution map. Areas with higher deviation values ​​are highlighted in the heatmap.

[0095] On the initial path, such as a grid path based on full area coverage, several adjustable control points are evenly distributed, for example, one control point every 100 meters. The position of each control point is determined by two-dimensional coordinates (x, y, z). i ,y i ) indicates that it is used as an optimization variable.

[0096] Based on the deviation heatmap, calculate the deviation gradient at each control point:

[0097] The gradient is approximated by numerical methods, such as the central difference method. Where δ is the minute step size, usually taken as 1 meter.

[0098] The control point positions are updated using the gradient descent algorithm: Where η is the learning rate, which ranges from [0.01, 0.1] and is dynamically adjusted according to the size of the region and the distribution of the bias.

[0099] Iterate and update until the path converges, i.e., the deviation change is less than the preset error, such as 0.01, or the maximum number of iterations is reached, such as 1000.

[0100] For example, in monitoring a field, the initial path is a straight scan. It is found that the deviation value in a certain area is significantly higher than that in other areas. Through gradient descent optimization, the path control points are shifted towards the high deviation area, eventually generating a winding path that covers all high deviation areas.

[0101] The stress classification module of this invention uses a machine learning model to automatically determine the stress type in areas with significant deviations. Supported stress types include drought stress, pest and disease stress, and nutrient deficiency. The classification model employs a convolutional neural network (CNN) architecture, with the specific structure as follows:

[0102] The input layer receives multimodal data (visible light image, infrared image, multispectral reflectance) and state deviation values. The input dimension is H×W×C, where H and W are the image resolution and C is the number of channels, for example, 3 for RGB images and 5 for multispectral data.

[0103] The convolutional layer contains 3 convolutional layers, each using 32 3×3 convolutional kernels with a stride of 1 and the activation function being ReLU.

[0104] A max pooling layer is added after each convolutional layer, with a pooling window of 2×2, to reduce the dimension of the feature map.

[0105] After flattening the convolutional features, they are input into two fully connected network layers with 128 and 64 neurons respectively, and the activation function is ReLU.

[0106] Use the Softmax function to output the probability of each stress type, with 3 categories (drought, pests and diseases, and nutrient deficiency).

[0107] Model training utilizes historical labeled data, with labels provided by agricultural experts based on field surveys. The training process employs a cross-entropy loss function, uses the Adam optimizer, a learning rate of 0.001, a batch size of 32, and 50 training epochs. Data augmentation techniques, such as random pruning and flipping, are used to improve the model's generalization ability. The model typically achieves accuracy exceeding 85% on the validation set.

[0108] The stress type-aware submodule further enhances the flexibility of path optimization. Users can specify the stress types of interest, such as focusing only on drought stress.

[0109] The system generates spatial distribution maps for each stress type based on the stress type labels and their confidence scores output by the classification model. For example, the confidence map for drought stress regions consists of the drought probability for each grid cell. During path optimization, the system prioritizes user-specified stress type regions by adjusting the weighting coefficients of the bias heatmap: w' g =w g ·(1+α·P g ), where w g For the original weights, P g The confidence level is the target coercion type, and α is the adjustment coefficient, which ranges from [0.5, 2] and is adjusted according to user priority.

[0110] For example, in monitoring a rice paddy, the user specified that they should focus on pest and disease stress. The system identified the pest and disease confidence level in the northeast corner area as 0.9, increased the weight coefficient of the corresponding grid by 50%, and after path optimization, the drone prioritized flying to this area.

[0111] Through the aforementioned technical solutions, the system can achieve precise monitoring of vegetation status in complex agricultural environments. For example, in a 100-hectare cornfield, the system identified significant deviations in 10% of the area during a single flight, including 5% due to drought stress, 3% due to pest and disease stress, and 2% due to nutrient deficiency. The optimized path allows the drone to focus 80% of its observation time on these abnormal areas during subsequent flights, significantly improving monitoring efficiency.

[0112] This invention's system is applicable to various agricultural scenarios, including but not limited to field crops (corn, rice), cash crops (orchards, tea gardens), and forestry monitoring. By adjusting weighting coefficients and stress type priorities, the system can flexibly adapt to different user needs, such as prioritizing the monitoring of specific diseases or optimizing irrigation management.

[0113] Through the detailed implementation methods described above, this system invention can achieve efficient and accurate vegetation observation in various agricultural scenarios, providing strong support for agricultural production management and ecological protection.

[0114] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A UAV-borne multimodal vegetation observation system, characterized in that, include: Unmanned aerial vehicle (UAV) platforms are used to carry multimodal sensors and perform observation flights in target areas; A multimodal data acquisition module, installed on the UAV platform, is used to simultaneously acquire visible light images, infrared images, and multispectral reflectance data of the target area, and generate a current observation state vector; The mirror state modeling module is used to construct a vegetation mirror state vector of the target area in a healthy state based on historical flight data. The vegetation mirror state vector is a multi-dimensional feature representation of the ideal vegetation growth state. The deviation calculation module is used to compare the current observed state vector with the vegetation mirror state vector and calculate the state deviation between the two. The path optimization module is used to iteratively update the path control points of the UAV based on the spatial distribution of the state deviation within the target area using a gradient descent optimization algorithm with the state deviation as the objective function, thereby generating an optimized path for subsequent flight and enabling the UAV to prioritize the observation of areas with significant deviations. The system also includes a stress classification module, which is used to classify and judge the crop stress type in the deviation area based on the state deviation. The crop stress type includes any one or more of the following: drought stress, pest and disease stress, or nutrient deficiency. The stress classification module automatically determines the crop stress type in areas with significant deviations using a trained machine learning classification model. The machine learning classification model is trained under supervision using historical labeled state deviation sample data, where the labels represent crop stress types. The path optimization module further includes a stress type awareness submodule, which is used to introduce stress type matching constraints during the path optimization process based on the target stress type specified by the user. The process of introducing stress type matching constraints specifically includes: Based on the stress type labels and their confidence levels at each spatial location output by the stress classification module, a spatial distribution map of each stress type within the target area is generated. During the path optimization process, based on the spatial distribution map, deviation regions that match the user's interest type are selected as priority observation targets, and the weight coefficients of the corresponding regions in the deviation heatmap are adjusted to give them higher attention in gradient descent optimization.

2. The UAV-borne multimodal vegetation observation system according to claim 1, characterized in that, The path optimization module includes: A deviation heatmap generation unit is used to visualize the spatial distribution of the state deviation within the target area as a deviation heatmap; The path control point setting unit is used to set multiple adjustable path control points on the initial path and use the position of each path control point as an optimization variable. The gradient calculation unit is used to calculate the deviation gradient information based on the position of each path control point in the deviation heatmap. The control point update unit is used to iteratively update the position of each path control point according to the deviation gradient information using a gradient descent optimization algorithm until the optimized path converges within a preset error range or reaches the maximum number of iterations.

3. The UAV-borne multimodal vegetation observation system according to claim 1, characterized in that, The multimodal data acquisition module includes a time synchronization device and a spatial calibration unit. The time synchronization device is used to align the acquisition clocks of visible light images, infrared images, and multispectral reflectance data. The spatial calibration unit is used to unify different modal data into observation state vectors under the same geographic coordinate system.

4. The UAV-borne multimodal vegetation observation system according to claim 1, characterized in that, The mirror state modeling module selects observation data in which the NDVI and NDRE vegetation indices are in the optimal range from multimodal samples collected in the target area over a historical period, and constructs a vegetation mirror state vector containing statistical mean and principal component features.

5. The UAV-borne multimodal vegetation observation system according to claim 1, characterized in that, The process of calculating the state deviation specifically includes: Align the current observation state vector with the vegetation mirror state vector at the corresponding location according to dimensions to construct a difference vector, where each dimension of the difference vector represents the degree of deviation of a certain modal feature; The difference vector is normalized using a weighted Euclidean distance to obtain the state deviation that reflects the intensity of the overall deviation.

6. The UAV-borne multimodal vegetation observation system according to claim 1, characterized in that, The deviation calculation module further includes a deviation significance determination unit, used to determine whether there is a significant deviation region in the target region based on the state deviation, specifically including: The target area is divided into several spatial grid cells, and the state deviation of all observation points in each grid cell is statistically analyzed. Based on the distribution characteristics of the state deviation values ​​across the entire region, a significance threshold is set. The threshold can be the global mean plus a preset multiple of the standard deviation, or it can be set as the lower limit of the values ​​of the top several percentiles of the deviation values. Grid cells whose state deviation values ​​exceed the significance threshold are marked as areas of significant deviation, and their spatial coordinates are output.

Images