Plateau complex terrain vegetation growth remote sensing monitoring method

Abstract

The application discloses a kind of highland complex terrain vegetation growth remote sensing monitoring methods, including for obtaining target area vertical observation optical remote sensing image data and carrying out image extraction processing to output vertical observation vegetation index, while obtaining real-time meteorological wind direction vector data;Extract target pixel point whose vegetation index drop amplitude is in preset interval;Obtain windward and leeward view angle image parallel to wind direction vector and extract corresponding vegetation index;The aforementioned index and wind direction vector are input into multi-view asymmetric classification model to output vegetation state category;When judging as wind pressure directional lodging category, stop using vertical observation vegetation index, extract specified band reflectivity data for leeward view angle image and input into calculation module to output canopy vigor index.The application solves the problem that vegetation physical lodging is easily misjudged as physiological degradation under highland strong wind environment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a remote sensing monitoring method for vegetation growth in complex plateau terrain. Background Technology

[0002] The growth status of vegetation, such as grasslands and forests, is an important indicator reflecting the health of ecosystems and global climate change. In the complex terrain of plateau regions, traditional ground-based manual observation is often difficult to conduct due to the harsh natural environment and vast spatial span. Therefore, satellite remote sensing technology, with its advantages of large-scale and periodic observation, has become the core means of monitoring vegetation phenology and growth in high-altitude areas. Existing technologies typically calculate parameters such as vegetation indices by extracting reflectance in the visible and near-infrared bands, and then construct time series to reflect the fluctuations in vegetation growth. For example, existing patent CN117689959A discloses a remote sensing classification method that integrates vegetation life cycle characteristics. This method uses multi-temporal remote sensing images to extract vegetation life cycle characteristics, and achieves dynamic classification and monitoring of vegetation coverage over a large area by fitting and smoothing the time series curves.

[0003] However, existing remote sensing monitoring methods suffer from serious misjudgments when faced with the extreme weather, physiological habits of unique species, and special animal micro-topography characteristic of plateaus. The main problems are as follows: The plateau's terrain is highly undulating and often accompanied by strong canyon winds or ridge winds. During the peak vegetation growth period, continuous unidirectional strong winds can physically flatten meadows or shrubs. This directional lodging flattens the originally dense canopy, exposing large areas of the underlying soil background. Consequently, when remote sensing satellites observe vertically downwards, the reflected signals from green vegetation received are sharply reduced. Current technology cannot distinguish between this physical tilting caused by wind and genuine physiological withering of vegetation, making it highly susceptible to erroneous assessments of large-scale vegetation degradation after strong winds.

[0004] The plateau is home to a large number of alpine plants, such as wolfsbane, which bloom explosively during the short summer, with large areas of bright flowers covering the original green leaves. Because conventional monitoring relies heavily on the absorption characteristics of chlorophyll in the visible light band, the drastic change in reflectance during the flowering period can cause a cliff-like drop in vegetation index, making vegetation that is in its most vigorous reproductive growth stage misjudged as extremely deteriorated.

[0005] The plateau is home to burrowing animals such as pikas, whose frequent digging creates tiny mounds that expose large amounts of deep, loose, bare soil to the surface. After rain, these micro-landforms, lacking root systems to retain water, experience rapid evaporation, exhibiting high-frequency short-wave infrared oscillations in the optical characteristics of mixed pixels. This severely interferes with the moisture signal of normal meadows, leading existing systems to mistakenly believe that the overall vegetation canopy is undergoing severe drought stress. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a remote sensing monitoring method for vegetation growth in complex plateau terrain, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A remote sensing monitoring method for vegetation growth in complex plateau terrain includes the following steps: The system acquires vertical observation optical remote sensing image data of the target area in the current observation period, performs image extraction processing on the vertical observation optical remote sensing image data, and outputs the vertical observation vegetation index of each pixel; acquires real-time meteorological wind direction vector data of the target area; acquires the historical vertical observation vegetation index of each pixel in the historical observation period, calculates the decrease of the vertical observation vegetation index compared with the historical vertical observation vegetation index, and extracts the target pixels whose decrease is within a first preset range. For the target pixel, acquire windward optical remote sensing image data and leeward optical remote sensing image data that are parallel to the direction of the real-time meteorological wind direction vector data, and perform the image extraction processing on the windward optical remote sensing image data and the leeward optical remote sensing image data respectively, and output the windward vegetation index and the leeward vegetation index. The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the vegetation index from the windward view, and the vegetation index from the leeward view are input into a pre-trained machine learning-based multi-view asymmetric classification model to output the vegetation state category of the target pixel.

[0008] Specifically, when the vegetation status category is wind pressure directional lodging, the vertical observation vegetation index of the target pixel is stopped, the image extraction processing is performed on the leeward view optical remote sensing image data of the target pixel, the reflectance data of the specified band is output, and the reflectance data of the specified band is input to the vegetation index calculation module, and the vegetation index calculation module outputs the canopy growth index of the target pixel.

[0009] Specifically, the machine learning-based multi-view asymmetric classification model is a multilayer perceptron network model. The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the windward view vegetation index, and the leeward view vegetation index are input into the pre-trained machine learning-based multi-view asymmetric classification model, and the output vegetation state category of the target pixel includes: The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the windward view vegetation index, and the leeward view vegetation index are concatenated into a first input feature vector. The first input feature vector is input to the input layer of the multilayer perceptron network model; The first input feature vector is nonlinearly mapped through the hidden layer of the multilayer perceptron network model to obtain a multidimensional feature vector. The multidimensional feature vector is input to the output layer of the multilayer perceptron network model, and the output layer outputs a probability distribution vector. The category corresponding to the maximum value in the probability distribution vector is extracted as the vegetation state category.

[0010] Specifically, the training steps of the multilayer perceptron network model include: Acquire historical wind pressure lodging sample data and historical physiological degradation sample data; Extract the historical vertical observation vegetation index, historical meteorological wind direction vector data, historical windward view vegetation index and historical leeward view vegetation index corresponding to the historical wind pressure lodging sample data and the historical physiological degradation sample data, and concatenate them into a training input feature vector. Assign the corresponding true class label to the training input feature vector; The training input feature vector is input into the multilayer perceptron network model to be trained, and the predicted class distribution vector is output. The predicted category distribution vector and the true category label are input into the cross-entropy loss function calculation module, and the cross-entropy loss value is output. The backpropagation algorithm is applied and the network weight parameters of the multilayer perceptron network model to be trained are updated using the cross-entropy loss value until the cross-entropy loss value is less than a preset loss threshold, thereby obtaining the pre-trained machine learning-based multi-view asymmetric classification model.

[0011] Specifically, the image extraction processing performed on the leeward-view optical remote sensing image data of the target pixel, outputting reflectance data of a specified band, and inputting the reflectance data of the specified band into the vegetation index calculation module, with the vegetation index calculation module outputting the canopy growth index of the target pixel, includes: The image extraction process is performed on the leeward-view optical remote sensing image data to output near-infrared band reflectance data and red band reflectance data as the specified band reflectance data. The near-infrared band reflectance data and the red band reflectance data are input into the vegetation index calculation module, and the vegetation index calculation module outputs the canopy growth index of the target pixel.

[0012] Specifically, when the vegetation status category is non-wind pressure directional lodging, the following steps are performed: The visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data of the target pixel are obtained from the vertical observation optical remote sensing image data. The visible light band reflectance data, the near-infrared band reflectance data, and the short-wave infrared band reflectance data are input into a pre-trained corolla spectral feature classification model, and the phenological stage category of the target pixel is output. When the phenological stage category is the concentrated flowering period category, the red-edge band reflectance data of the target pixel is obtained from the vertical observation optical remote sensing image data; The red-edge band reflectance data and the near-infrared band reflectance data are input into the structural vegetation index calculation module, and the structural vegetation index calculation module outputs the pure structural vegetation growth value of the target pixel. The pure structure vegetation growth value is used as the vegetation growth status characterization parameter of the target pixel.

[0013] Specifically, the corolla spectral feature classification model adopts a support vector machine model. The visible light band reflectance data, the near-infrared band reflectance data, and the short-wave infrared band reflectance data are input into the pre-trained corolla spectral feature classification model, and the phenological stage category of the target pixel is output, including: The visible light band reflectance data, the near-infrared band reflectance data, and the short-wave infrared band reflectance data are concatenated into a second input feature vector; The second input feature vector is input into the support vector machine model, and the support vector machine model calculates the spatial distance between the second input feature vector and the preset hyperplane. Extract the classification spatial labels corresponding to the spatial distance values, and map the classification spatial labels to the phenological stage categories.

[0014] Specifically, the training steps of the support vector machine model include: Obtain actual sample data of plots during the concentrated flowering period and sample data of plots with real physiological degradation; Extract the visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data of the sample data of the actual measured plots during the concentrated flowering period and the sample data of the actual physiologically degraded plots, and then concatenate the extracted data into a training feature matrix; The training feature matrix is ​​labeled with the corresponding true classification labels of the samples. The training feature matrix with the data annotations is input into the support vector machine model to be trained, and the maximum margin hyperplane parameters are solved by applying the sequence minimum optimization algorithm to obtain the pre-trained corolla spectral feature classification model.

[0015] Specifically, before acquiring vertically observed optical remote sensing image data of the target area during the current observation period, the following steps are also included: Acquire the first bidirectional reflectance distribution function data of the target area within the first solar altitude angle interval and the second bidirectional reflectance distribution function data within the second solar altitude angle interval, wherein all angle values ​​within the first solar altitude angle interval are less than all angle values ​​within the second solar altitude angle interval; The first bidirectional reflectance distribution function data and the second bidirectional reflectance distribution function data are input into a pre-trained micro-topographic feature classification model, and the micro-topographic category of each pixel in the target area is output. Pixels belonging to the mixed category of mouse-rabbit hills in the micro-topography are extracted, and the extracted pixels are marked as micro-topography interference pixels.

[0016] Specifically, for the aforementioned micro-topographical interference pixels, the following steps are performed: Acquire data on the occurrence of precipitation meteorological events; Shortwave infrared time series data of the micro-topographic interference pixels within a first preset time span after extracting the time node corresponding to the precipitation meteorological event occurrence record data; The shortwave infrared band time series data is input into the empirical mode decomposition algorithm module, and the empirical mode decomposition algorithm module outputs the first eigenmode function component set and residual components. Extract the intrinsic mode function components whose frequency values ​​are in the first preset frequency range from the first intrinsic mode function component set, and mark the extracted intrinsic mode function components as high-frequency drying signal components; The shortwave infrared band time series data and the high-frequency drying signal component are input to the signal subtraction operation module, which performs a subtraction operation on the input data and outputs a low-frequency stable signal component. The low-frequency stable signal component is input to the vegetation moisture content assessment module, and the vegetation moisture content assessment module outputs the vegetation canopy moisture state index of the micro-topographic interference pixels.

[0017] The advantage of this invention over existing technologies lies in its innovative introduction of wind-driven canopy directional tilt and multi-view reflectance asymmetry, cleverly solving the technical challenge of misjudging physical vegetation lodging as physiological degradation under strong winds in high-altitude areas. In strong winds, vegetation that has undergone directional lodging often exposes its bottom stems and soil on the windward side, while the leeward side displays densely packed green leaves flattened by the wind. This invention utilizes this highly concrete optical spatial geometric relationship, combined with real-time meteorological wind direction vectors, to extract remote sensing reflectance characteristics of the target area from both windward and leeward perspectives. By identifying this significant directional asymmetry in reflectance through a machine learning model, the system can accurately determine the wind-induced lodging state. Subsequently, it discards distorted vertical observation data and directly uses the reflectance of a specified band from the leeward perspective to characterize the true canopy growth. This mechanism not only overcomes the limitations of a single vertical perspective but also completely avoids complex algebraic models, significantly improving the accuracy and robustness of monitoring results under severe weather conditions.

[0018] Building upon this foundation, this invention further integrates the principles of corolla spectral inversion and physical structure conservation, effectively resolving the misjudgment of pseudo-degradation caused by the concentrated flowering period of unique plateau vegetation. Addressing the characteristic that corolla color only strongly interferes with the visible light band, while the near-infrared and short-wave infrared bands are more sensitive to and stable in terms of physical three-dimensional structure and water content, this invention constructs a corolla spectral feature classification model. Upon identifying a target area entering its concentrated flowering period, it automatically stops using conventional band data interfered with by flower color and instead extracts red-edge and near-infrared band data to calculate pure structural vegetation growth values. This unique technical feature ensures the objectivity of vegetation evaluation during its peak growth period, preventing peak growth from being mistaken for withering. Furthermore, this invention delves into the asymmetric laws governing micro-topographical shadows and subsoil desorption, completely eliminating the interference from subsoil wet-dry oscillations caused by the micro-topographical features of plateau pika mounds. By analyzing the bidirectional reflectance distribution function under different solar altitude angles, the mixed pixel points of the mouse and rabbit mounds were accurately located. Then, using an empirical mode decomposition algorithm, the high-frequency rapid drying signal of bare soil was precisely extracted from the shortwave infrared time series after precipitation events, allowing only the low-frequency stable signal components to be used for moisture status assessment. This innovative signal separation technique significantly improves the accuracy of vegetation moisture content monitoring in high-altitude animal activity areas. Attached Figure Description

[0019] Figure 1 This is a geometric and physical schematic diagram of the multi-view remote sensing observation of the present invention; Figure 2 This is a schematic diagram illustrating the phenomenon of wind pressure-induced directional collapse and the asymmetry of the observation surface in this invention; Figure 3 This is a structural diagram of the multi-view asymmetric classification model of the present invention; Figure 4This is a flowchart of the growth index extraction process under the lodging state of the present invention; Figure 5 This is a schematic diagram illustrating the principle of bidirectional reflectance distribution observation of micro-topography in the mouse-rabbit hill of this invention. Detailed Implementation

[0020] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0021] It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make several equivalent substitutions or conventional modifications, all of which should fall within the scope of protection of the present invention.

[0022] This invention addresses vegetation monitoring scenarios in complex plateau terrain areas, including grasslands, shrublands, and sparse forest understory, focusing on resolving three types of remote sensing misjudgments that are highly prone to occur in plateau environments. One type is the misjudgment of vegetation lodging caused by strong winds as physiological degradation; another is the false degradation caused by flowers obscuring green leaves during peak bloom; and the third type is pika mounds (a collective term for small, raised mounds and surrounding loosened bare soil areas formed by burrowing mammals such as plateau pikas during digging, expanding, and excavating burrows). These surface units are typically small in scale, densely distributed, have high surface roughness, and are a mixture of bare soil and vegetation, with the surface soil often being recently disturbed and loose. Fine soil has a weak water retention capacity and is more prone to rapid water loss than the surrounding intact meadow canopy under solar radiation and wind. Therefore, pika mounds not only alter the area composition of vegetation and bare soil within a pixel, but also cause significant differences in directional reflection and shortwave infrared fluctuations due to local shading, soil exposure, and rapid drying after precipitation, thus interfering with the vegetation growth and moisture status retrieved by remote sensing. High-frequency drying disturbances in shortwave infrared caused by burrowing animal micro-topography are misinterpreted as a sharp decline in canopy moisture. Therefore, this invention does not rely solely on single vertical observation results, but combines wind direction, observation angle, spectral characteristics of flowering period, and the changing patterns of shortwave infrared time series after precipitation to construct a remote sensing monitoring link that more accurately reflects the actual ecological processes on the plateau.

[0023] In one embodiment, the target area can be alpine meadow, alpine shrub meadow, mountain grassland, or other plateau vegetation-covered areas dominated by low-lying vegetation. The current observation period can be set to any one of 1 day, 5 days, 8 days, 10 days, or 16 days, based on the satellite revisit cycle or operational monitoring cycle. The pixel spatial resolution can be uniformly resampled to 2 m, 10 m, 20 m, or 30 m according to the data source. This invention does not strictly limit the resolution, as long as multi-source image registration, pixel-to-pixel comparison, and time-series analysis can be completed under the same coordinate reference.

[0024] The image extraction processing described in this invention preferably includes steps such as radiometric calibration, atmospheric correction, geometric correction, topographic correction, cloud and cloud shadow removal, snow or strongly reflective features removal, pixel-level registration, target area cropping, and band reflectance extraction. Because the terrain in plateau regions is highly undulating and the illuminance varies significantly between sunny and shady slopes, without topographic correction, simple spectral differences can be mixed with the influence of slope aspect and solar incidence angle, easily misinterpreting terrain brightness as vegetation changes. Therefore, in one embodiment, a digital elevation model can be introduced to correct the slope and aspect of the image, making the subsequently extracted reflectance closer to the intrinsic spectral information of vegetation.

[0025] The various vegetation indices, structural indices, and water state indices involved in this invention can all be directly adopted from mature indicators that are clearly defined in existing remote sensing technologies. Their definitions are not the core innovation of this invention. The core of this invention lies in choosing which type of index in which scenario, from which observation perspective, and how to avoid misuse of distorted data through classification models. In other words, the mathematical definitions of the indices can directly utilize existing remote sensing algorithm libraries, industry software platforms, or existing public standards. For example, vertically observed vegetation indices and canopy growth indices after wind-induced lodging can use normalized vegetation indices constructed based on red and near-infrared reflectance, or enhanced vegetation indices or soil-adjusted vegetation indices with stronger soil background suppression capabilities; pure structural vegetation growth values ​​under concentrated flowering periods can use red-edge normalized difference indices constructed based on red-edge and near-infrared bands; canopy water state indices can use water content indices constructed based on near-infrared and short-wave infrared, or water stress indices sensitive to short-wave infrared. Those skilled in the art can simply make equivalent substitutions based on whether the data source has conditions such as red-edge or short-wave infrared bands.

[0026] In a preferred embodiment, the present invention first executes the wind pressure-oriented lodging identification branch. This branch is executed first as the core branch because strong wind lodging directly destroys the ability of a single vertical observation to characterize the canopy geometry. If this type of error is not corrected first, all subsequent growth judgments based on the vertical perspective may be based on distorted input.

[0027] As shown in Figure 1, specifically, vertical observation optical remote sensing image data of the target area in the current observation period is first acquired. Here, vertical observation refers to an observation method where the sensor's line of sight is basically consistent with or has a small angle with the corresponding local surface normal direction, used to minimize shadow expansion and lateral occlusion caused by oblique viewing. In one embodiment, the observation zenith angle of the vertical observation viewpoint can be controlled between 0° and 15°, preferably not exceeding 10°. Subsequently, the aforementioned image extraction processing is performed on the vertical observation optical remote sensing image data, outputting the vertical observation vegetation index for each pixel.

[0028] While extracting the vertically observed vegetation index, real-time meteorological wind direction vector data for the target area is acquired. This real-time meteorological wind direction vector data can come from automatic weather stations, regional meteorological reanalysis data, numerical weather prediction grid products, or be obtained by fusing multi-source wind field data. To adapt to pixel-level applications, wind direction is preferably represented as an east-west component and a north-south component, with wind speed as an additional meteorological feature input to the model. This design aims to avoid the angular discontinuity problem caused by using only a single wind direction angle; for example, a change in wind direction from 359° to 1° is physically minimal, but numerically it represents a significant jump in angle. Using a fractional representation makes it easier for the model to learn the continuous correspondence between wind field and reflectivity differences.

[0029] Subsequently, the historical vertical vegetation index of each pixel within the historical observation period is obtained, and the decrease in the current vertical vegetation index compared to the historical vertical vegetation index is calculated. The historical observation period here is preferably not an arbitrary past moment, but a historical window with a comparable phenological stage to the current period. For example, images from the same period within the past 3 to 5 years, differing from the current date by no more than 10 or 15 days, can be selected. Alternatively, a pixel baseline can be established after smoothing the previous 3 to 6 monitoring periods. This is because the seasonal changes in plateau vegetation are very drastic; directly comparing spring data with midsummer data would misinterpret normal phenological fluctuations as abnormal degradation.

[0030] In one embodiment, the decrease can be expressed as a relative decrease ratio or an absolute exponential difference. When using a relative decrease ratio, the first preset range is preferably set to 10% to 45%, more preferably 15% to 35%. When using an absolute difference, if the exponent range is 0 to 1, the first preset range can be set to 0.08 to 0.30. This range should not be too small, otherwise a large number of normal fluctuations will be included in suspected abnormal pixels; nor should it be too large, otherwise a large number of moderately abnormal pixels caused by wind-blown landslides that have not yet been completely flattened will be missed. By extracting target pixels whose decrease range falls within this first preset range, subsequent multi-view analysis can focus on the most problematic pixels that need to be identified, reducing computational load and minimizing misjudgments caused by irrelevant pixels.

[0031] For the aforementioned target pixels, further acquire windward and leeward optical remote sensing image data parallel to the real-time meteorological wind direction vector data. Here, "parallel to the wind direction" preferably means that the observation azimuth angle is substantially consistent with the wind direction within the horizontal projection plane, rather than requiring the three-dimensional observation direction to completely coincide with the wind direction vector. The windward angle can be defined as the angle between the observation azimuth and the wind direction not exceeding 30°, and the leeward angle can be defined as the angle between the observation azimuth and the opposite wind direction not exceeding 30°. Preferably, near-temporal multi-angle remote sensing data is used to ensure that the reflectance comparison between the front and rear views is completed within a time window where wind field changes are minimal. In one embodiment, the observation time interval between the windward and leeward views is controlled within 0.5 h to 24 h, preferably not exceeding 6 h, to avoid additional errors introduced by diurnal variations in vegetation or weather changes.

[0032] After acquiring the aforementioned multi-view images, image extraction processing was performed on the windward and leeward optical remote sensing image data, respectively, outputting the windward and leeward vegetation indices. The reason for not directly comparing the original band reflectance, but instead further extracting the vegetation index, is that the vegetation index has a certain robustness to changes in soil background and illumination, making it more suitable for quickly distinguishing between lodged and non-lodged states. As shown in Figure 2, the essence of wind-directed lodging is that the canopy geometry is compressed along the wind direction. Therefore, more stems and subsoil are often exposed on the windward side, while the compacted green leaf layer is more easily observed on the leeward side. This directional asymmetry is the basis for identifying wind-lodging in this invention.

[0033] In one embodiment, the vertical observation vegetation index, real-time meteorological wind direction vector data, windward view vegetation index, and leeward view vegetation index of the target pixel are input into a pre-trained machine learning-based multi-view asymmetric classification model, which outputs the vegetation state category of the target pixel. Multi-view asymmetry refers to the systematic, non-random difference in reflectance of the same pixel across different wind-related observation directions. This difference is not caused by ordinary terrain undulations or sensor noise, but by changes in canopy attitude caused by wind pressure. Explicitly encoding this physical phenomenon into a classification problem avoids hastily concluding degradation based solely on a decrease in the index at a single point in time.

[0034] In a preferred embodiment, as shown in Figure 3, the multi-view asymmetric classification model is a multilayer perceptron network model. Specifically, the vertical observation vegetation index, windward view vegetation index, and leeward view vegetation index of the target pixel can be used as vegetation observation features. Real-time meteorological wind direction vector data can be expanded into east-west and north-south wind components, and wind speed can be used as an additional meteorological feature. These components are then concatenated into the first input feature vector. To improve training stability, it is preferable to first standardize each feature dimension to make its mean close to 0 and its scale within a similar range.

[0035] The first input feature vector is then fed into the input layer of the multilayer perceptron network model. Two to four hidden layers are then added after the input layer, with each layer containing 32, 64, 128, or 256 neurons. Two or three hidden layers are preferred to balance nonlinear fitting capability and generalization performance. The activation function is preferably a modified linear unit (MRU), but a leaky modified linear unit (LCU) can also be used. To suppress overfitting, deactivation layers can be added between the hidden layers, with a deactivation ratio of 0.1 to 0.5, preferably 0.2 to 0.3. The multilayer perceptron performs nonlinear mapping on the first input feature vector through the hidden layers to obtain a multidimensional feature vector. This multidimensional feature vector is then input into the output layer, which outputs a probability distribution vector and extracts the category corresponding to the largest value as the vegetation state category.

[0036] In one embodiment, the vegetation status categories include at least wind-pressure directional lodging and non-wind-pressure directional lodging categories, and can be further subdivided into obvious wind-pressure lodging, mild wind-pressure lodging, physiological degradation, and uncertain categories. The purpose of adding the mild wind-pressure lodging and uncertain categories is to reserve more granular risk classification space for business applications, enabling the system to perform both rapid binary classification and more fine-grained expert-assisted interpretation.

[0037] The training steps of the multilayer perceptron network model can be implemented as follows: First, acquire historical wind pressure lodging sample data and historical physiological degradation sample data. Historical wind pressure lodging samples can come from ground quadrat surveys after strong winds, confirmation by UAV side-view images, manual visual interpretation results, or labeled data formed by combining on-site vegetation lodging records; historical physiological degradation samples can come from naturally yellowing plots at the end of the season, actual degradation plots caused by drought stress, and wilting plots after freeze-thaw damage, etc. By simultaneously introducing these two types of samples, the model can learn to distinguish between situations that superficially exhibit exponential decline but have different mechanisms of decline.

[0038] Then, historical vertical vegetation indices, historical meteorological wind direction vectors, historical windward vegetation indices, and historical leeward vegetation indices are extracted from historical wind pressure lodging sample data and historical physiological degradation sample data, and concatenated into a training input feature vector. After assigning corresponding ground truth class labels to the training input feature vector, it is input into the multilayer perceptron network model to be trained, outputting a predicted class distribution vector. The predicted class distribution vector and the ground truth class labels are then input into the cross-entropy loss function calculation module, outputting the cross-entropy loss value. Subsequently, the backpropagation algorithm is applied, and the network weight parameters are updated based on the cross-entropy loss value until the loss value is less than a preset loss threshold, resulting in a pre-trained machine learning-based multi-view asymmetric classification model.

[0039] The preset loss threshold can be set according to the sample size and class complexity, generally ranging from 0.05 to 0.30, preferably from 0.10 to 0.20. The learning rate can be set from 0.0001 to 0.01, preferably around 0.001; the batch size can be from 32 to 512, preferably from 64 to 256; the number of training epochs can be set from 50 to 500, and an early stopping mechanism is used to stop training when the validation set loss no longer decreases for several consecutive epochs. The reason for this setting is that plateau data is greatly affected by clouds, snow, and terrain occlusion, and sample noise is unavoidable. If we blindly pursue extremely low training loss, it is easy to overfit local regional features.

[0040] As shown in Figure 4, when the output vegetation status category is wind-induced directional lodging, the vertical observation vegetation index of the target pixel is discontinued. This discontinuation means not only removing this value from the current period's growth evaluation, but also marking it as a distorted observation value in subsequent time series fitting, growth trend regression, and degradation early warning processes, thus preventing this outlier from contaminating the entire time series curve. Subsequently, image extraction processing is performed on the leeward-view optical remote sensing image data of the target pixel, outputting the reflectance data of a specified band. This reflectance data is then input into the vegetation index calculation module, which outputs the canopy growth index of the target pixel.

[0041] The leeward perspective was chosen because, for directional lodging vegetation, the leeward side more easily reveals the green leaf layer that maintains a high coverage even after being compacted along the windward direction. Compared to vertical observation and windward side views, this is closer to the actual canopy biomass and leaf area. This invention does not simply recalculate the original indices from a different angle; rather, after confirming the lodging, it actively abandons distorted perspectives and uses observational geometry that more closely reflects the true canopy state for characterization. This is the key reason why this invention is more robust than traditional methods.

[0042] In a preferred embodiment, when performing image extraction processing on leeward-view optical remote sensing image data, near-infrared band reflectance data and red band reflectance data are output as reflectance data for a specified band. These near-infrared and red band reflectance data are then input into a vegetation index calculation module, which outputs a canopy growth index. This canopy growth index can directly use an existing normalized vegetation index or an enhanced vegetation index. When there is significant exposure of meadow soil background or a high proportion of bare soil, a vegetation index more robust to soil background is preferred. If the data source lacks stable red light quality, a conventional green vegetation index constructed using near-infrared and other alternative red bands can also be used. Those skilled in the art can make substitutions based on the characteristics of the data source without affecting the essence of the invention.

[0043] In another embodiment, a strong wind trigger condition can be added to the wind pressure collapse recognition branch. That is, multi-view asymmetric classification is only performed on the target pixel when the real-time wind speed reaches a second preset wind speed threshold. This wind speed threshold can be set to 5 m / s to 15 m / s, preferably 7 m / s to 10 m / s. The purpose of this is to reduce the need for meaningless calls to the multi-view model under low wind speed conditions, improve computational efficiency, and prevent ordinary swaying from being misjudged as collapse.

[0044] When the vegetation state category is classified as non-wind-pressure directional lodging, it indicates that the current pixel's index decrease is not primarily caused by wind-pressure lodging. In this case, the branch for identifying corolla spectral features is initiated. This branch addresses reflectance anomalies caused by short-term concentrated flowering at high altitudes. Many alpine plants simultaneously enter their reproductive growth phase during their short growing season, and large areas of corolla significantly alter visible light reflectance, but this does not necessarily indicate a decline in vegetation growth. Therefore, if the scene of abundant flowers and still vigorous foliage is not separated from the scene of withered leaves and decaying canopy, the vertical vegetation index will be systematically underestimated.

[0045] Specifically, visible light reflectance data, near-infrared reflectance data, and short-wave infrared reflectance data for target pixels are obtained from vertically observed optical remote sensing image data. The visible light band can include one or more of blue, green, and red light; the near-infrared band is used to sense canopy internal scattering and structural continuity; the short-wave infrared band is more sensitive to differences in water content, petal and leaf tissue, and can provide supplementary information for distinguishing between flowering and degeneration stages. Subsequently, the reflectance data is input into a pre-trained corolla spectral feature classification model, which outputs the phenological stage category of the target pixel.

[0046] In a preferred embodiment, the corolla spectral feature classification model employs a support vector machine (SVM) model. Specifically, visible light reflectance data, near-infrared reflectance data, and short-wave infrared reflectance data can be concatenated into a second input feature vector, which is then input into the SVM model. The SVM model calculates the decision distance of the second input feature vector relative to a preset classification hyperplane, and determines the corresponding classification spatial label based on the side of the decision distance or the output of the decision function. This classification spatial label is then mapped to a phenological stage category. The phenological stage categories include at least concentrated flowering period categories and non-concentrated flowering period categories, and can also be extended to vegetative growth period categories, budding period categories, full bloom period categories, post-flowering color-changing period categories, and true degeneration period categories. When there are two or more phenological stage categories, a one-to-one or one-to-many multi-class SVM implementation can be used.

[0047] Support vector machines (SVMs) are suitable for this branch because the flowering and degeneration periods typically have clear boundaries in several key band combinations, and SVMs demonstrate stable performance in learning classification boundaries under small to medium sample conditions. High-altitude field sample collection is costly, and it is often difficult to obtain large-scale labeled data as in plains areas. Therefore, using SVMs can maintain good discriminative ability even with limited sample sizes.

[0048] In one embodiment, the support vector machine model preferably uses a radial basis function kernel. The penalty parameter can be set to 0.1 to 100, preferably 1 to 20; the kernel width parameter can be set to 0.001 to 10, preferably 0.01 to 1. The parameters can be automatically determined through grid search or cross-validation. During training, sample data of actual plots during the concentrated flowering period and sample data of actual physiologically degraded plots are first acquired. Samples during the concentrated flowering period can be obtained through ground transect surveys, high-resolution color images from UAVs, plant phenological survey records, etc.; samples of actual physiologically degraded plots can be obtained through naturally withered plots after the end of the growing season, degraded plots caused by prolonged drought, and plots with significant decreases in measured leaf water content. Then, the visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data of these samples are extracted, and the extracted data are concatenated into a training feature matrix. The training feature matrix is ​​then labeled with the corresponding sample real classification labels. Then, the training feature matrix with data annotations is input into the support vector machine model to be trained, and the maximum margin hyperplane parameters are solved by applying the sequence minimum optimization algorithm to obtain the pre-trained corolla spectral feature classification model.

[0049] When the phenological stage category is concentrated flowering period category, the red edge band reflectance data of the target pixel is obtained from the vertical observation optical remote sensing image data, and the red edge band reflectance data and near-infrared band reflectance data are input into the structural vegetation index calculation module. The module outputs the pure structural vegetation growth value of the target pixel, and then uses the pure structural vegetation growth value as the vegetation growth status characterization parameter of the target pixel.

[0050] The red-edge and near-infrared bands were chosen because corolla color has the most significant impact on visible light, while red-edge and near-infrared bands are more sensitive to changes in leaf internal structure, canopy density, and leaf area, and have a relatively weaker direct response to flower color itself. By switching from color-dominated visible light evaluation to structure-dominated red-edge near-infrared evaluation, the vigorous flowering period can be separated from pseudo-decline. The pure structural vegetation growth value can use existing red-edge normalized difference indices, or other structural vegetation indices based on red-edge and near-infrared. The index definition is existing technology, and those skilled in the art can directly call existing algorithms to implement it. For example, when the data source has both the red-edge center band and the near-infrared band, the structural growth index can be calculated using the normalized difference method of the two; when multiple narrow red-edge bands are available, more stable red-edge structural characterization values ​​can also be obtained through interpolation or combination.

[0051] When the phenological stage category is non-concentrated flowering period, it indicates that the current pixel is neither affected by wind-induced lodging nor belongs to the pseudo-degradation of concentrated flowering period. In one embodiment, the aforementioned vertically observed vegetation index can be directly used as a parameter representing the vegetation growth status of the pixel in the current observation period, or it can be input into a conventional time-series smoothing module for subsequent growth trend output. This ensures that the system maintains an efficient processing method compatible with existing technologies for ordinary pixels, while only activating a dedicated correction mechanism for abnormal situations.

[0052] like Figure 5 As shown, in addition to the two branches directly related to canopy growth mentioned above, this invention further includes a micro-topographic interference identification and moisture signal stripping branch. This branch mainly targets areas frequently inhabited by burrowing animals such as pika mounds. Since pika mounds often form small raised mounds, bare soil patches, and uneven shadows, these micro-topographic features exhibit significant rapid drying characteristics after rainfall. Short-wave infrared radiation is highly sensitive to moisture; therefore, the rapid evaporation of bare soil in mixed pixels can create high-frequency, strong fluctuations in a short time, causing the system to mistakenly believe that the entire vegetation is in a state of severe water loss. To avoid misjudging rapid drying of the subsoil as canopy water loss, this invention pre-identifies and specifically processes micro-topographic interference pixels before formally conducting monitoring for the current cycle.

[0053] Specifically, before acquiring vertical observation optical remote sensing image data of the target area in the current observation period, the first bidirectional reflectance distribution function data of the target area within the first solar altitude angle interval and the second bidirectional reflectance distribution function data within the second solar altitude angle interval are acquired. All angle values ​​within the first solar altitude angle interval are smaller than all angle values ​​within the second solar altitude angle interval. The bidirectional reflectance distribution function data can be directly obtained from existing multi-angle remote sensing products or retrieved from reflectance data observed at different time phases or angles. The first solar altitude angle interval can be set to 10° to 30°, and the second solar altitude angle interval can be set to 40° to 70°, preferably with an interval of at least 10° between them, to differentiate the response differences of shadows and surface roughness at different solar altitude angles.

[0054] The reason for using bidirectional reflectance distribution function data under different solar altitude angles is that micro-topography such as pika hills is extremely sensitive to illumination geometry. At lower solar altitude angles, local shadows are more easily formed on the shaded side of the hills and in the low-lying areas between the hills, making the differences in reflectance direction more obvious. As the solar altitude angle increases, the shadow length shortens, and the differences in scattering between bare soil and vegetation become apparent in another way. Combining the bidirectional reflectance distribution functions under low and high solar altitude angle conditions allows for a more robust distinction between true micro-topography mixed pixels and ordinary vegetation pixels.

[0055] In one embodiment, the first bidirectional reflectance distribution function data and the second bidirectional reflectance distribution function data are input into a pre-trained micro-topographic feature classification model. The model outputs the micro-topographic category of each pixel within the target area and extracts pixels belonging to the pika-hill mixed category, marking them as micro-topographic interference pixels. The so-called pika-hill mixed category refers to a single pixel containing both vegetation canopy and loose bare soil, hill shadows, or small-scale raised structures formed by pika digging. Therefore, its optical response does not represent a pure canopy state.

[0056] In a preferred embodiment, the micro-topographic feature classification model can employ a random forest model. Input features may include forward scattering response, backscattering response, and azimuth anisotropy difference under a first solar altitude angle interval, as well as corresponding similar features under a second solar altitude angle interval. It may also include response difference features between the two solar altitude angle intervals. The number of trees in the random forest can be set to 100 to 500, preferably 150 to 300; the maximum tree depth can be set to 8 to 20; and the minimum number of leaf node samples can be set to 1 to 5. Random forests are preferred because this model is highly adaptable to multidimensional heterogeneous features, robust to a small number of outliers and nonlinear boundaries, and facilitates the assessment of the importance of different bidirectional reflectance features.

[0057] The training of the micro-topographic feature classification model can be implemented as follows: First, a mixed pixel sample library of mouse-rabbit hills and a pixel sample library of ordinary meadows are established through methods such as UAV low-altitude photogrammetry, ground quadrat surveys, and manual fine interpretation. Then, the bidirectional reflectance distribution function features of each sample under different solar altitude angle intervals are extracted to construct a training sample matrix, which is then labeled using the true class labels. The training samples are then input into the random forest model to be trained, and the model parameters are adjusted based on classification accuracy, recall, and validation set stability to obtain the pre-trained micro-topographic feature classification model. For this invention, the specific machine learning implementation of the micro-topographic classification model is not unique; it can also be replaced with support vector machines, gradient boosting trees, or lightweight neural networks, as long as it can stably identify mixed pixels of mouse-rabbit hills using the bidirectional reflectance direction features under different solar altitude angles.

[0058] For the aforementioned micro-topographic interference pixels, further post-precipitation rapid drying signal stripping processing is performed. First, precipitation meteorological event records are acquired. These precipitation events can be defined based on rainfall thresholds; for example, a period with accumulated precipitation of 0.5 mm, 1 mm, 2 mm, or 5 mm can be defined as a valid precipitation event. Then, shortwave infrared time series data of the micro-topographic interference pixels are extracted within a first preset time span following the corresponding time node in the precipitation meteorological event records. The first preset time span is preferably set to 6 h to 72 h, more preferably 12 h to 48 h. This post-precipitation time window is focused on because the rapid drying effect of bare soil primarily occurs within a short period after rainfall, and this stage best reflects the high-frequency disturbance of shortwave infrared signals by the micro-topography of the pika hill.

[0059] Subsequently, the shortwave infrared time series data is input into the Empirical Mode Decomposition (EMD) algorithm module, which outputs the first set of intrinsic mode function (EMF) components and the residual components. Empirical Mode Decomposition is an adaptive decomposition method suitable for non-stationary and nonlinear time series. It does not require pre-setting fixed basis functions and is particularly suitable for handling complex shortwave infrared fluctuations in plateau regions influenced by weather, topography, and surface conditions. By decomposing the original shortwave infrared sequence into several EMF components ranging from high to low frequencies, the fast-changing components corresponding to rapid drying of bare soil and the slow-changing components corresponding to slow water loss or slow recovery of vegetation canopy can be separated.

[0060] In one embodiment, intrinsic mode function (IMF) components whose frequency values ​​fall within a first preset frequency range are extracted from the first IMF component set. These extracted IMF components are then superimposed to form a high-frequency drying signal component. The first preset frequency range can be dynamically set according to the sampling interval. If the time series is sampled hourly, it is preferable to extract high-frequency components that exhibit significant fluctuations and attenuation within 1 to 6 sampling intervals; if sampled daily, it is preferable to extract the first 1 to 3 high-frequency IMF components that complete the main changes within a 1 to 3-day timescale. The purpose of this setting is to remove short-term spikes caused by rapid evaporation of bare soil, while retaining the main trend of slow changes in canopy moisture.

[0061] Then, the shortwave infrared time series data and the high-frequency drying signal component are input to the signal subtraction module. This module performs a subtraction operation on the input data and outputs a low-frequency stable signal component. Here, the low-frequency stable signal component does not mean absolutely stable, but rather refers to effective information that better represents the true moisture changes in the vegetation canopy after removing high-frequency disturbances caused by the rapid drying of bare soil on pika mounds. This low-frequency stable signal component is then input to the vegetation moisture content assessment module, which outputs the vegetation canopy moisture state index for micro-topographic interference pixels.

[0062] Regarding the specific implementation of the vegetation canopy moisture state index, in one embodiment, near-infrared time series data can be acquired synchronously at the time node corresponding to the shortwave infrared time series, and the low-frequency stable signal component and the synchronous near-infrared data can be input together into an existing moisture index calculation module to output a normalized canopy moisture index or a difference-type moisture content index. In another embodiment, the low-frequency stable signal value corresponding to the current assessment time point can also be selected and directly mapped to a moisture state level value between 0 and 1 through a pre-established empirical regression model or lookup table. The reason for allowing these two implementations is that different data sources have different availability of shortwave infrared and near-infrared, and what this invention truly emphasizes is the removal of high-frequency soil drying disturbances from shortwave infrared, not the formula definition of a specific moisture index.

[0063] To facilitate engineering implementation, in one overall embodiment, the present invention can operate according to the following process. First, micro-topographic interference pixels are identified and an interference marker layer is established using the bidirectional reflectance distribution function under different solar altitude angles. Then, vertical observation optical remote sensing image data for the current observation period is acquired, and the vertical observation vegetation index of each pixel is output. For pixels whose current index decreases relative to the historical baseline within a first preset range, a multilayer perceptron network model is used to identify whether they belong to wind-pressure directional lodging, combining the real-time wind direction vector and the vegetation index from the windward and leeward perspectives. If they belong to wind-pressure directional lodging, the canopy growth index is directly calculated using the reflectance of the specified band from the leeward perspective, replacing the vertical observation vegetation index as the growth characterization value for the current period of that pixel. If they do not belong to wind-pressure directional lodging, a support vector machine corolla spectral feature classification model is further used to identify whether they are in the concentrated flowering period. If they are in the concentrated flowering period, the red edge and near-infrared bands are extracted to calculate the pure structural vegetation growth value, which is used as the growth status characterization parameter for the current period of that pixel; if they are not in the concentrated flowering period, the vertical observation vegetation index is maintained as the regular growth characterization value for that pixel. For areas marked as micro-topographic interference pixels, empirical mode decomposition is additionally called to strip high-frequency dry signals when outputting moisture status, and low-frequency stable signal components are used to calculate the vegetation canopy moisture status index.

[0064] In a more specific application scenario, if a pixel in an alpine meadow shows a 22% decrease in its vertically observed vegetation index relative to the same historical period, and the wind speed in the area reaches 8 m / s during the same period, while the vegetation index at the windward view is significantly lower than that at the leeward view, the multilayer perceptron model will tend to output a wind-pressure-induced lodging category. In this case, the system will not directly classify the pixel as a degraded or withered area, but instead uses the red light and near-infrared reflectance at the leeward view to calculate the canopy growth index, thus preserving the true growth information of the meadow, which is still vigorous. If another pixel does not have obvious wind field asymmetry characteristics, but the combination of visible light and short-wave infrared shows typical corolla shading characteristics, the support vector machine model will classify it as a concentrated flowering period category. In this case, the system further uses red edge and near-infrared to extract pure structural vegetation growth values, avoiding mistaking full bloom for decay. If a certain area is identified as a mixed pixel of pika and rabbit hills, when performing canopy moisture analysis, the system will remove the fast-drying high-frequency component from the shortwave infrared time series after rainfall and retain only the stable low-frequency component to estimate the canopy water content, thereby avoiding mistaking the rapid drying of the subsoil for a rapid loss of water in the vegetation canopy.

[0065] Furthermore, to enhance the migration capability between different regions, in one embodiment, the multilayer perceptron model, support vector machine model, and micro-topographic feature classification model can all adopt a strategy of regional pre-training followed by target region fine-tuning. That is, a general sample library is first established in multiple plateau regions for initial training, and then a small number of measured samples from the target region are used for parameter correction. This can reduce the impact of differences in species composition, soil color, and surface roughness in a single region on the model's generalization ability. For regions with a small sample size, it is preferable to retain the existing remote sensing index calculation method unchanged, and only make minor updates to the classification threshold and model parameters.

[0066] In this invention, if multiple indices are output in parallel, growth and moisture layers can be created separately according to business needs. The growth layer prioritizes the use of the canopy growth index after wind pressure correction, the pure structural vegetation growth value after flowering correction, or the conventional vertical observation vegetation index. The moisture layer prioritizes the use of the vegetation canopy moisture status index after high-frequency stripping for pixels with micro-topographic interference, and the conventional moisture index for non-interference pixels. The advantage of this approach is that different ecological parameters are generated by the observation link most suitable for their physical meaning, rather than compressing all ecological issues into a single vegetation index for processing.

[0067] It should also be noted that the various index acquisition methods mentioned in this invention can all be implemented using existing technologies. Taking the vertical observation vegetation index and the canopy growth index after wind-induced lodging as examples, they can be calculated from red band reflectance and near-infrared band reflectance using existing normalized difference methods; taking the pure structural vegetation growth value as an example, it can be calculated from red edge band reflectance and near-infrared band reflectance according to the existing red edge index definition; taking the canopy moisture state index as an example, it can be obtained from the existing moisture index definitions of near-infrared and shortwave infrared, or from the corrected shortwave infrared stable component through existing empirical models. Since these indices all have publicly available, mature, and widely used definitions in the field of remote sensing, those skilled in the art can make equivalent selections based on the satellite platform, band configuration, and operational practices used, without changing the core implementation idea of ​​this invention.

[0068] In summary, the specific implementation of this invention does not simply superimpose multiple conventional remote sensing processing steps. Instead, it establishes correction paths corresponding to the mechanisms of several special ecological processes most prone to misjudgment in plateau regions. For canopy geometric deflection caused by strong winds, wind-driven multi-view asymmetric classification combined with a leeward perspective is used; for visible light distortion caused by concentrated flowering periods, corolla spectral feature classification combined with red-edge near-infrared structural characterization is employed; and for short-wave infrared high-frequency dryness disturbances caused by pika hills, bidirectional reflectance distribution function identification combined with empirical mode decomposition signal stripping is used. Through these technical links, the accuracy of remote sensing monitoring of vegetation growth in complex plateau terrain can be significantly improved.

Claims

1. A remote sensing monitoring method for vegetation growth in complex plateau terrain, characterized in that, Includes the following steps: The system acquires vertical observation optical remote sensing image data of the target area in the current observation period, performs image extraction processing on the vertical observation optical remote sensing image data, and outputs the vertical observation vegetation index of each pixel; acquires real-time meteorological wind direction vector data of the target area; acquires the historical vertical observation vegetation index of each pixel in the historical observation period, calculates the decrease of the vertical observation vegetation index compared with the historical vertical observation vegetation index, and extracts the target pixels whose decrease is within a first preset range. For the target pixel, acquire windward optical remote sensing image data and leeward optical remote sensing image data that are parallel to the direction of the real-time meteorological wind direction vector data, and perform the image extraction processing on the windward optical remote sensing image data and the leeward optical remote sensing image data respectively, and output the windward vegetation index and the leeward vegetation index. The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the vegetation index from the windward view, and the vegetation index from the leeward view are input into a pre-trained machine learning-based multi-view asymmetric classification model to output the vegetation state category of the target pixel.

2. The method according to claim 1, characterized in that, When the vegetation status category is wind pressure directional lodging, the vertical observation vegetation index of the target pixel is stopped, the image extraction processing is performed on the leeward view optical remote sensing image data of the target pixel, the reflectance data of the specified band is output, and the reflectance data of the specified band is input into the vegetation index calculation module, and the canopy growth index of the target pixel is output by the vegetation index calculation module.

3. The method as described in claim 1, characterized in that, The machine learning-based multi-view asymmetric classification model is a multilayer perceptron network model. The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the windward view vegetation index, and the leeward view vegetation index are input into the pre-trained machine learning-based multi-view asymmetric classification model. The output vegetation state category of the target pixel includes: The vertical observation vegetation index of the target pixel, the real-time meteorological wind direction vector data, the windward view vegetation index, and the leeward view vegetation index are concatenated into a first input feature vector. The first input feature vector is input to the input layer of the multilayer perceptron network model; The first input feature vector is nonlinearly mapped through the hidden layer of the multilayer perceptron network model to obtain a multidimensional feature vector. The multidimensional feature vector is input to the output layer of the multilayer perceptron network model, and the output layer outputs a probability distribution vector. The category corresponding to the maximum value in the probability distribution vector is extracted as the vegetation state category.

4. The method as described in claim 3, characterized in that, The training steps of the multilayer perceptron network model include: Acquire historical wind pressure lodging sample data and historical physiological degradation sample data; Extract the historical vertical observation vegetation index, historical meteorological wind direction vector data, historical windward view vegetation index and historical leeward view vegetation index corresponding to the historical wind pressure lodging sample data and the historical physiological degradation sample data, and concatenate them into a training input feature vector. Assign the corresponding true class label to the training input feature vector; The training input feature vector is input into the multilayer perceptron network model to be trained, and the predicted class distribution vector is output. The predicted category distribution vector and the true category label are input into the cross-entropy loss function calculation module, and the cross-entropy loss value is output. The backpropagation algorithm is applied and the network weight parameters of the multilayer perceptron network model to be trained are updated using the cross-entropy loss value until the cross-entropy loss value is less than a preset loss threshold, thereby obtaining the pre-trained machine learning-based multi-view asymmetric classification model.

5. The method as described in claim 2, characterized in that, The image extraction processing is performed on the leeward-view optical remote sensing image data of the target pixel, outputting reflectance data of a specified band, and inputting the reflectance data of the specified band into the vegetation index calculation module. The vegetation index calculation module outputs the canopy growth index of the target pixel, including: The image extraction process is performed on the leeward-view optical remote sensing image data to output near-infrared band reflectance data and red band reflectance data as the specified band reflectance data. The near-infrared band reflectance data and the red band reflectance data are input into the vegetation index calculation module, and the vegetation index calculation module outputs the canopy growth index of the target pixel.

6. The method as described in claim 1, characterized in that, When the vegetation status category is non-wind pressure directional lodging, the following steps are performed: The visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data of the target pixel are obtained from the vertical observation optical remote sensing image data. The visible light band reflectance data, the near-infrared band reflectance data, and the short-wave infrared band reflectance data are input into a pre-trained corolla spectral feature classification model, and the phenological stage category of the target pixel is output. When the phenological stage category is the concentrated flowering period category, the red-edge band reflectance data of the target pixel is obtained from the vertical observation optical remote sensing image data; The red-edge band reflectance data and the near-infrared band reflectance data are input into the structural vegetation index calculation module, and the structural vegetation index calculation module outputs the pure structural vegetation growth value of the target pixel. The pure structure vegetation growth value is used as the vegetation growth status characterization parameter of the target pixel.

7. The method as described in claim 6, characterized in that, The corolla spectral feature classification model employs a support vector machine model. The visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data are input into the pre-trained corolla spectral feature classification model, and the phenological stage category of the target pixel is output, including: The visible light band reflectance data, the near-infrared band reflectance data, and the short-wave infrared band reflectance data are concatenated into a second input feature vector; The second input feature vector is input into the support vector machine model, and the support vector machine model calculates the spatial distance between the second input feature vector and the preset hyperplane. Extract the classification spatial labels corresponding to the spatial distance values, and map the classification spatial labels to the phenological stage categories.

8. The method as described in claim 7, characterized in that, The training steps of the support vector machine model include: Obtain actual sample data of plots during the concentrated flowering period and sample data of plots with real physiological degradation; Extract the visible light band reflectance data, near-infrared band reflectance data, and short-wave infrared band reflectance data of the sample data of the actual measured plots during the concentrated flowering period and the sample data of the actual physiologically degraded plots, and then concatenate the extracted data into a training feature matrix; The training feature matrix is ​​labeled with the corresponding true classification labels of the samples. The training feature matrix with the data annotations is input into the support vector machine model to be trained, and the maximum margin hyperplane parameters are solved by applying the sequence minimum optimization algorithm to obtain the pre-trained corolla spectral feature classification model.

9. The method as described in claim 1, characterized in that, Before acquiring vertically observed optical remote sensing image data of the target area in the current observation period, the following steps are also included: Acquire the first bidirectional reflectance distribution function data of the target area within the first solar altitude angle interval and the second bidirectional reflectance distribution function data within the second solar altitude angle interval, wherein all angle values ​​within the first solar altitude angle interval are less than all angle values ​​within the second solar altitude angle interval; The first bidirectional reflectance distribution function data and the second bidirectional reflectance distribution function data are input into a pre-trained micro-topographic feature classification model, and the micro-topographic category of each pixel in the target area is output. Pixels belonging to the mixed category of mouse-rabbit hills in the micro-topography are extracted, and the extracted pixels are marked as micro-topography interference pixels.

10. The method as described in claim 9, characterized in that, For the aforementioned micro-topographical interference pixels, perform the following steps: Acquire data on the occurrence of precipitation meteorological events; Shortwave infrared time series data of the micro-topographic interference pixels within a first preset time span after extracting the time node corresponding to the precipitation meteorological event occurrence record data; The shortwave infrared band time series data is input into the empirical mode decomposition algorithm module, and the empirical mode decomposition algorithm module outputs the first eigenmode function component set and residual components. Extract the intrinsic mode function components whose frequency values ​​are in the first preset frequency range from the first intrinsic mode function component set, and mark the extracted intrinsic mode function components as high-frequency drying signal components; The shortwave infrared band time series data and the high-frequency drying signal component are input to the signal subtraction operation module, which performs a subtraction operation on the input data and outputs a low-frequency stable signal component. The low-frequency stable signal component is input to the vegetation moisture content assessment module, and the vegetation moisture content assessment module outputs the vegetation canopy moisture state index of the micro-topographic interference pixels.

Images