Forest ecology monitoring system and method with unmanned aerial vehicle carrying multispectral sensor

Abstract

The invention relates to the technical field of data processing, and further relates to a forest ecology monitoring system and method with an unmanned aerial vehicle carrying a multispectral sensor. The system comprises an unmanned aerial vehicle, a hyperspectral data processing device and an ecological health detection platform, the unmanned aerial vehicle carries a multispectral sensor so as to acquire multispectral data of a forest area and perform atmospheric correction on the surface reflectance; the hyperspectral data processing device is used for reconstructing hyperspectral data from the multispectral data by using a compressed sensing theory and carrying out noise reduction by using wavelet transform to obtain the hyperspectral data after noise reduction; and the ecological health detection platform is used for calculating forest health indexes and forest carbon reserves according to the hyperspectral data after noise reduction, and presenting the forest health indexes and the forest carbon reserves to users. According to the method, comprehensive and accurate evaluation of the health condition and the carbon reserve of the forest ecosystem is realized, and the defects in the aspects of data precision, processing efficiency and evaluation comprehensiveness in the prior art are overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a forest ecological monitoring system and method using an unmanned aerial vehicle equipped with a multispectral sensor. Background Art

[0002] As global climate change intensifies and the ecological environment deteriorates, the health status and carbon sink capacity of forest ecosystems, as a key component of terrestrial ecosystems, have an important impact on the global carbon cycle and climate change. Forests absorb carbon dioxide from the atmosphere through photosynthesis and fix it in biomass, thus playing an important role as a carbon sink. In recent years, forest ecological monitoring methods based on remote sensing technology have gradually become a research hotspot in academia and industry, especially multispectral and hyperspectral remote sensing technology, which provides a new way to quickly and accurately obtain forest ecological information. However, the existing forest ecological monitoring technology still has many problems in terms of accuracy, real-time performance, cost, etc., and more perfect technical solutions are urgently needed to improve the accuracy of forest health assessment and carbon storage calculation.

[0003] Traditional forest ecological monitoring methods mainly rely on ground observations, such as sample plot surveys, biomass determination, and tree counting. Although these methods can provide detailed vegetation characteristics and biomass information, their limitations are also very obvious. First, ground observation methods usually require a lot of manpower and time, with limited coverage, making it difficult to quickly monitor large areas of forest. In addition, due to the complexity and heterogeneity of forest ecosystems, ground observations often cannot fully reflect the growth status of the entire forest area and are easily affected by sample bias, resulting in errors in the evaluation results. With the development of remote sensing technology, remote sensing observations based on satellites and aircraft have gradually become an important means of forest ecological monitoring. Among the current remote sensing technologies, multispectral remote sensing is the most widely used one. Multispectral remote sensing uses multispectral sensors to observe the reflectance characteristics of the forest surface in different bands and extract ecological parameters related to forest biomass, vegetation cover, canopy moisture content, etc. Traditional multispectral remote sensing usually contains several to more than a dozen bands, such as visible light (red, green, blue), near infrared, and short-wave infrared. Multispectral remote sensing has the advantages of fast data acquisition and wide coverage, and has been widely used in forest health monitoring, pest detection, fire assessment, etc. However, multispectral data has low spectral resolution and a limited number of bands, which cannot capture more detailed spectral feature information and make it difficult to fully reflect the physiological state, carbon sequestration capacity and health status of vegetation. Summary of the invention

[0004] The main purpose of the present invention is to provide a forest ecological monitoring system and method using an unmanned aerial vehicle equipped with a multispectral sensor, which realizes a comprehensive and accurate assessment of the health status and carbon reserves of the forest ecosystem, and makes up for the shortcomings of the existing technology in terms of data accuracy, processing efficiency and comprehensiveness of assessment.

[0005] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0006] A forest ecological monitoring system with a multispectral sensor mounted on an unmanned aerial vehicle, the system comprising: an unmanned aerial vehicle, a hyperspectral data processing device and an ecological health detection platform; the unmanned aerial vehicle is equipped with a multispectral sensor to collect multispectral data of a forest area and perform atmospheric correction on the surface reflectance; the hyperspectral data processing device is used to reconstruct hyperspectral data from the multispectral data using compressed sensing theory, and use wavelet transform to perform noise reduction to obtain the noise-reduced hyperspectral data; the ecological health detection platform is used to calculate the forest health index and the forest carbon stock based on the noise-reduced hyperspectral data; the forest carbon stock and the forest health index are presented to users as monitoring results.

[0007] Furthermore, the ecological health detection platform calculates the forest health index and the forest carbon stock based on the denoised hyperspectral data, which specifically includes: calculating the normalized vegetation index and leaf area index based on the near-infrared band reflectance and the red light band reflectance in the denoised hyperspectral data; estimating the canopy moisture content using the short-wave infrared band reflectance; calculating the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculating the net primary productivity using the light utilization efficiency model; calculating the forest health index based on the net primary productivity, photochemical vegetation index, canopy moisture content and leaf area index; calculating the stand biomass based on the leaf area index, canopy moisture content and photochemical vegetation index; and calculating the forest carbon stock based on the stand biomass.

[0008] Furthermore, suppose that the multispectral data of the forest area collected by the multispectral sensor is X; the multispectral data X is a multidimensional matrix containing the spectral information of each band, and its dimension is expressed as [number of pixel rows, number of pixel columns, number of bands]; the reflectance information of each pixel in the multiband is stored as a vector; the atmospheric correction converts the original radiation brightness data L with band λ into λ Converted into reflectivity data R that reflects the true surface characteristics λ ; X is a set of atmospherically corrected multispectral data, including the surface reflectance of each band; each third-dimensional slice in X corresponds to the corrected surface reflectance data R of a band λ λ .

[0009] Further, through the following formula, the atmospheric correction converts the original radiance data L of the band λ into the reflectance data R that reflects the true surface characteristics λ :

[0010]

[0011] where L Rλ is the Rayleigh scattering radiation component; L Mλ is the Mie scattering radiation component; L Mλ is the residue of the radiation after atmospheric absorption and scattering; τ λ↑ is the optical thickness from the surface to the multispectral sensor direction, τ λ↓ is the optical thickness from the sun to the surface direction; E 0λ is the solar top irradiance; θ s is the solar zenith angle; e is the natural constant.

[0012] Further, the specific process of the hyperspectral data processing device to obtain the denoised hyperspectral data includes: through the following formula, using the compressed sensing theory, the multispectral data and the hyperspectral data are connected:

[0013] X = ΦS + ∈;

[0014] where X is the multispectral data, with a shape of [m×p], where m is the total number of spatial pixels and p is the number of multispectral bands; Φ is the measurement matrix, with a shape of [p×n], which compresses the hyperspectral data S into the multispectral data X; Φ is a sparse matrix, p << n, where n is the number of hyperspectral bands; S is the hyperspectral data matrix, with a shape of [m×n], where n is the number of hyperspectral bands; ∈ is the noise term, which follows a Gaussian distribution: σ is the standard deviation of the Gaussian distribution; in order to reconstruct the hyperspectral data S from the multispectral data X, an optimization problem needs to be solved, and the optimization problem is transformed into the minimization of the sparse signal, and the formula is as follows:

[0015]

[0016] where ||·|| F is the Frobenius norm, which is used to measure the two-norm distance of the matrix; W is the wavelet transform matrix, with a shape of [n×n], which sparsely represents the hyperspectral data S. By performing wavelet transform on the hyperspectral data, it is transformed into the sparse domain; λ is the regularization parameter, which is used to balance the reconstruction error and sparsity.

[0017] Further, the noise in the hyperspectral data is distributed in the high-frequency part. The wavelet transform is used to sparsely represent S and denoise it in the sparse domain; the wavelet transform matrix W is decomposed into:

[0018]

[0019] Among them, W spatial is the spatial wavelet transform matrix. Considering the spatial information of the matrix, discrete wavelet transform is used to implement it. spectral is the spectral wavelet transform matrix, which considers the spectral information between bands and is implemented using one-dimensional wavelet transform; is the tensor product; after adding wavelet denoising and sparsity constraints, the following complete optimization problem is obtained:

[0020]

[0021] in, is the sparse regularization term after wavelet transformation, indicating the sparsity of hyperspectral data after wavelet transformation; ||S|| TV is the total variation regularization term, which is used to maintain the edge characteristics of the matrix and reduce noise, and is defined as:

[0022]

[0023] λ1 and λ2 are regularization parameters, which are used to weigh the impact of sparsity and total variation on reconstruction respectively; i and j are subscript indices; S i，j Represents the element in the i-th row and j-th column in the hyperspectral data matrix; The hyperspectral data after denoising.

[0024] Furthermore, the complete optimization problem is a non-convex optimization problem; for the complete optimization problem, the alternating direction multiplier method is used to decompose it into multiple sub-problems for solution; wherein the soft threshold algorithm is used to solve Solved by gradient descent method

[0025] Furthermore, the forest health index is calculated using the following formula:

[0026]

[0027] Among them, FHI is the forest health index, ranging from 0 to 1; LAI is the leaf area index; CWC is the canopy water content; PRI is the vegetation photosynthetic efficiency; NPP is the net primary productivity; ω1, ω2, ω3 and ω4 are all weight coefficients, satisfying LAI max is the maximum value of the theoretical leaf area index;

[0028] CWC max is the theoretical maximum canopy moisture content; NPP max is the theoretical maximum value of net primary productivity.

[0029] Furthermore, forest carbon storage is calculated using the following formula:

[0030] C = B·CF·(1+R);

[0031] Among them, C is the forest carbon stock; CF is the coefficient of biomass conversion to carbon, ranging from 0.4 to 0.6; R is the ratio of underground biomass to aboveground biomass, ranging from 0.2 to 0.3.

[0032] A forest ecological monitoring method using a multispectral sensor mounted on an unmanned aerial vehicle, the method comprising:

[0033] Step 1: Collect multispectral data of the forest area using a drone equipped with a multispectral sensor and perform atmospheric correction on the surface reflectance;

[0034] Step 2: Using compressed sensing theory, reconstruct hyperspectral data from multispectral data, and use wavelet transform to reduce noise to obtain denoised hyperspectral data;

[0035] Step 3: Calculate the normalized vegetation index and leaf area index based on the near-infrared band reflectance and red light band reflectance in the denoised hyperspectral data; estimate the canopy moisture content using the short-wave infrared band reflectance; calculate the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculate the net primary productivity using the light use efficiency model; calculate the forest health index based on the net primary productivity, photochemical vegetation index, canopy moisture content and leaf area index; calculate the stand biomass based on the leaf area index, canopy moisture content and photochemical vegetation index; calculate the forest carbon stock based on the stand biomass; present the forest carbon stock and forest health index as monitoring results to the user.

[0036] The forest ecological monitoring system and method using a multispectral sensor mounted on an unmanned aerial vehicle of the present invention have the following beneficial effects:

[0037] The present invention fully utilizes the high resolution and multi-band advantages of hyperspectral data through the organic combination of multispectral and hyperspectral data, and provides more detailed spectral information than traditional multispectral remote sensing. Hyperspectral data contains detailed reflection characteristics of forest vegetation in different bands, so that the present invention can accurately extract key ecological parameters reflecting the health of the forest, such as leaf area index (LAI), canopy water content (CWC), photosynthetic efficiency index (PRI) and net primary productivity (NPP). These parameters are important indicators for evaluating the growth status and carbon sequestration capacity of forests, and are helpful for comprehensively judging the health status of forests. However, due to the limited bands of traditional multispectral data, it is impossible to capture such detailed spectral characteristics, which limits its accuracy in health monitoring and carbon storage assessment. Therefore, the present invention has obvious advantages in data quality and evaluation accuracy.

[0038] Secondly, the present invention uses compressed sensing theory to process the conversion and reconstruction between multispectral and hyperspectral data. In the process of acquiring and processing hyperspectral data, problems such as large data volume and noise interference usually affect the accuracy and efficiency of analysis. By using compressed sensing technology, the present invention compresses high-dimensional hyperspectral data into low-dimensional multispectral data, and then performs sparse representation and noise reduction on the hyperspectral data through an optimization algorithm. This not only greatly reduces the amount of data and improves processing efficiency, but also ensures the integrity and accuracy of hyperspectral data. In particular, the present invention uses wavelet transform to perform sparse processing on the data, reduces noise on the data through two dimensions of space and spectrum, and retains the main characteristics of the signal. The alternating direction multiplier method (ADMM) is introduced into the optimization algorithm to decompose the complex non-convex optimization problem into multiple sub-problems for solution, thereby improving the solution efficiency. The application of these technical means enables the present invention to accurately reconstruct hyperspectral data while reducing the computational complexity, providing a high-quality data basis for subsequent ecological parameter extraction and analysis. In terms of forest health assessment, the present invention proposes a comprehensive forest health index (FHI), and through a weighted combination of multiple ecological parameters, provides a quantitative indicator that comprehensively reflects the health of the forest. The calculation of FHI takes into account key ecological parameters such as leaf area index, canopy water content, photosynthetic efficiency and net primary productivity, and introduces weight coefficients for weighted processing, so that the proportion of each parameter in the health assessment can be flexibly adjusted. This indicator design overcomes the defects of the single assessment dimension of traditional vegetation indices (such as NDVI) and can more comprehensively reflect the health level of forest ecosystems. In addition, the FHI indicator ranges from 0 to 1, which has the advantages of being easy to interpret and compare, which helps forest managers to evaluate and compare the health status of forests in different regions and at different times, detect abnormal situations in a timely manner and make corresponding management decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the structure of a forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.

[0041] Example 1, reference Figure 1: A forest ecological monitoring system with a multispectral sensor mounted on an unmanned aerial vehicle, the system comprising: an unmanned aerial vehicle, a hyperspectral data processing device and an ecological health detection platform; the unmanned aerial vehicle is equipped with a multispectral sensor to collect multispectral data of a forest area and perform atmospheric correction on the surface reflectance; the hyperspectral data processing device is used to reconstruct hyperspectral data from multispectral data using compressed sensing theory, and use wavelet transform for noise reduction to obtain noise-reduced hyperspectral data; the ecological health detection platform is used to calculate the forest health index and the forest carbon stock based on the noise-reduced hyperspectral data; the forest carbon stock and the forest health index are presented to users as monitoring results.

[0042] Specifically, the UAV is based on the design and structural principles of multi-rotor or fixed-wing, and achieves stable flight, hovering and cruising through components such as power system, flight control system, navigation system and power battery. Unlike traditional ground monitoring equipment, UAVs can fly flexibly at different heights and angles in forest areas. This advantage not only breaks through the limitations of ground viewing angles, but also can more comprehensively cover the vertical structural characteristics of forests. During the actual flight process, the UAV uses its flight control system and multiple sensors such as inertial measurement unit, barometer, GPS, etc. to adjust its posture and position in real time to ensure stable flight in a changing forest environment. Especially when the UAV flies in forest areas with complex terrain and dense vegetation, the accuracy of the navigation and positioning system is crucial to ensure that the UAV can fly according to the preset route and collect the required multispectral data at an appropriate altitude.

[0043] The multispectral sensor carried by the drone can sense electromagnetic wave information in different bands, including visible light, near infrared, short-wave infrared, etc., reflecting various physiological and biochemical characteristics of forest vegetation. The sensor receives the spectral information reflected by the forest surface and converts the light signals of different bands into digital signals, thereby obtaining the spectral reflectance information of different vegetation in the forest area. This information is an important basis for evaluating the health status of the forest, plant growth status, and soil and water conditions. In the actual collection process, the working principle of the multispectral sensor is similar to that of a highly sensitive camera, except that it has multi-channel spectroscopic capabilities and can refine the spectral information into several bands. In order to ensure the accuracy and consistency of the collected multispectral data, the drone needs to perform atmospheric correction on the data. The principle of atmospheric correction is to correct the spectral data according to the lighting conditions of the real-time environment, the height of the sensor, the angle of the sensor, and other factors, remove the influence of scattering and absorption in the atmosphere on the spectral reflectance, and obtain a more realistic reflectance of the ground objects. After collecting the multispectral data, the drone needs to transmit the data to the hyperspectral data processing device on the ground in real time for subsequent processing and analysis. In the process of data transmission, wireless communication technology is usually used, such as Wi-Fi, 4G / 5G or dedicated long-distance communication modules. This real-time transmission capability ensures the timeliness of data processing, allowing the forest ecological monitoring system to complete ecological data collection and analysis of large areas in a short period of time, improving the efficiency and accuracy of monitoring. The stability and efficiency of data transmission are directly related to the performance of the entire system. Therefore, drones are usually equipped with high-gain antennas, data compression modules and data redundancy check mechanisms to ensure the quality of data transmission in complex forest environments.

[0044] The hyperspectral data processing device can be divided into two core steps: reconstructing hyperspectral data from multispectral data using compressed sensing theory, and performing noise reduction on the reconstructed hyperspectral data through wavelet transform. The hyperspectral data processing device plays a key bridge role in the entire forest ecological monitoring system, converting a large amount of multidimensional spectral data collected from drone multispectral sensors into high-resolution, high-signal-to-noise ratio hyperspectral data, thereby providing accurate data support for the ecological health detection platform.

[0045] First of all, the theory of compressed sensing is the core of the hyperspectral data processing device to realize the reconstruction of multispectral data into hyperspectral data. Generally speaking, hyperspectral data contains more spectral dimensions and higher resolution than multispectral data. However, directly collecting hyperspectral data not only requires high-cost sensor equipment, but also generates a huge amount of data, which increases the burden of data processing and transmission. The theory of compressed sensing compresses the signal at the sampling stage based on the principles of signal sparsity and random sampling, thereby realizing the reconstruction of the original hyperspectral data with less sampling data. In the multispectral data collected by the multispectral sensor carried by the drone, each pixel has its specific spectral characteristics in different bands, and these spectral characteristics are sparse, that is, they can be represented by a small number of basis vectors in a certain transform domain. Based on this, the theory of compressed sensing can reconstruct hyperspectral data by sparsely encoding and decoding multispectral data using a small amount of sampling information, so that the hyperspectral information that could only be obtained through high-cost collection can be restored under low-cost conditions. During this process, the device uses a series of complex matrix operations and optimization algorithms. By selecting appropriate sampling matrices and sparse bases, it converts multispectral data into a sparse representation of hyperspectral data, and then uses optimization algorithms to solve the complete expression of the hyperspectral data.

[0046] After completing the reconstruction of the hyperspectral data, the noise in the data is a problem that needs to be further processed. Due to the influence of factors such as acquisition conditions, sensor characteristics, and atmospheric environment, hyperspectral data usually contains a lot of noise, which may affect the accurate judgment of forest ecological characteristics. In order to improve the signal-to-noise ratio of the data and ensure the reliability of the data, the hyperspectral data processing device uses wavelet transform for noise reduction. Wavelet transform is an analytical tool that decomposes signals into multi-scale and multi-frequency domains, which can effectively separate high-frequency noise and low-frequency useful information in the signal. In this device, the wavelet transform decomposes the hyperspectral signal into different scales through multi-level decomposition, so that the performance of noise and useful signals at different scales is clearly distinguished. Specifically, the noise in hyperspectral data usually manifests as large fluctuations on the high-frequency scale, while the useful signal manifests as a smooth change on the lower-frequency scale. The wavelet transform uses this feature to filter out fluctuations exceeding a certain amplitude as noise by applying a threshold on the high-frequency scale, thereby achieving noise reduction of the hyperspectral data.

[0047] The denoising process of the device is to threshold the wavelet coefficients of the hyperspectral data, retain the main signal components, discard the noise components, and then restore the processed signal to the time and space domain through wavelet reconstruction. In this way, the device can effectively reduce the noise in the data, improve the quality of the signal, and provide more accurate spectral data for the subsequent forest health index and carbon storage calculation. In the process of denoising, the multi-scale characteristics of the wavelet transform ensure the analysis and processing of the signal at different resolutions, so that the processed hyperspectral data can retain global and local ecological characteristic information at the same time. The hyperspectral data processing device plays an irreplaceable role in the entire forest ecological monitoring system. Its principle and performance are directly related to the system's monitoring accuracy and data processing efficiency of the forest ecological status. Compared with the traditional hyperspectral data acquisition method, the reconstruction using compressed sensing theory makes the system unnecessary to be equipped with expensive hyperspectral sensors, reduces hardware costs, and reduces data redundancy and storage pressure. Through the denoising processing of wavelet transform, the system effectively improves the signal-to-noise ratio of hyperspectral data, ensuring that the forest health index and carbon storage calculation based on hyperspectral data are more accurate and reliable. Therefore, the principle of the hyperspectral data processing device not only realizes the efficient processing of multispectral data, but also ensures the overall performance and practicality of the forest ecological monitoring system, and provides technical support for accurately assessing the health status of forest ecology.

[0048] The ecological health detection platform first performs feature extraction and spectral analysis based on hyperspectral data that has been processed by compressed sensing and wavelet noise reduction. Since hyperspectral data contains the spectral reflectance characteristics of vegetation in multiple bands, it provides more detailed vegetation information than ordinary multispectral data, including the biochemical composition, chlorophyll content, and water content of vegetation leaves. These characteristics can reflect the physiological state, growth status, and response to environmental stresses (such as drought, pests and diseases, and nutrient deficiencies) of vegetation. The ecological health detection platform uses this feature to construct index models related to forest health, such as the normalized difference vegetation index (NDVI), by analyzing the reflectance changes of hyperspectral data in specific bands or combined bands. These index models can reflect different ecological parameters such as vegetation coverage, health, and chlorophyll content, and thus reflect the ecological health status of the entire forest area. Next, the platform will synthesize these calculated index values ​​into a forest health index to quantify the ecological health status of the forest. The forest health index is an indicator that can comprehensively reflect the overall condition of the forest. The platform will set different weights according to different ecological parameters to achieve a comprehensive assessment of forest health. For example, if the chlorophyll index of a forest area is low and the moisture index is also on a downward trend, this may indicate that the forest in the area is in a state of water shortage or poor growth. The ecological health detection platform will use the weight model to obtain a lower health index, thereby warning the health status of the area. This process involves multivariate statistics and data fusion technology to ensure that the platform can uniformly process hyperspectral data from different sources and dimensions, so as to obtain a more comprehensive and accurate assessment of forest health. In addition to calculating the forest health index, the ecological health detection platform also evaluates forest carbon storage. Carbon storage is a key indicator to measure the carbon sink function of forest ecosystems in the global carbon cycle. The platform calculates the biomass and carbon content of the forest by analyzing the spectral features in the hyperspectral data and combining the existing empirical models. The vegetation feature information provided by the hyperspectral data, such as the photosynthetic capacity of the leaves, the leaf area index, and the biomass distribution, is closely related to the forest carbon storage. The ecological health detection platform uses these feature information, combined with parameters such as vegetation coverage area and canopy height in the multispectral data, to estimate the carbon storage of the entire forest area through the carbon storage model. This model is usually based on the correlation between vegetation index and carbon storage, and obtains the empirical relationship between hyperspectral data and carbon storage through fitting and optimization algorithms. When calculating carbon storage, the platform not only focuses on the biomass of the aboveground parts (such as trunks and leaves), but also estimates the underground parts (such as roots and soil carbon pools) to provide a more comprehensive carbon storage assessment.

[0049] Embodiment 2: The ecological health detection platform calculates the forest health index and the forest carbon stock based on the denoised hyperspectral data, specifically including: calculating the normalized vegetation index and leaf area index based on the near-infrared band reflectance and the red light band reflectance in the denoised hyperspectral data; estimating the canopy moisture content using the short-wave infrared band reflectance; calculating the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculating the net primary productivity using the light utilization efficiency model; calculating the forest health index based on the net primary productivity, the photochemical vegetation index, the canopy moisture content and the leaf area index; calculating the stand biomass based on the leaf area index, the canopy moisture content and the photochemical vegetation index; and calculating the forest carbon stock based on the stand biomass.

[0050] Specifically, by calculating the improved normalized difference vegetation index (NDVI) and estimating the leaf area index (LAI):

[0051]

[0052] Among them, R NIR is the reflectivity in the near-infrared band; R RED is the reflectivity of red light band; f APAR is the ratio of photosynthetically active radiation absorbed, estimated from NDVI: APAR =1.24·NDVI-0.168; κ: extinction coefficient, determined by the solar zenith angle θ s calculate: Estimate canopy moisture content using shortwave infrared (SWIR) reflectance:

[0053] CWC=α·(1-R SWIR1 )·(1-R SWIR2 )·LAI;

[0054] Where CWC is the canopy water content; R SWIR1 is the short-wave infrared 1 band reflectivity; R SWIR2 is the shortwave infrared 2 band reflectance; α is an empirical coefficient, with a value of 0.5-0.7. Calculate the photochemical vegetation index (PRI) to evaluate vegetation photosynthetic efficiency:

[0055]

[0056] Among them, R 531 is the reflectivity in the 531nm band; R 570 It is the reflectance in the 570nm band; combined with LAI, CWC and PRI, the stand biomass is estimated.

[0057]

[0058] Where B is the stand biomass; H is the average stand height, which can be obtained through LiDAR or stereo photogrammetry; β0, β1, β2, β3, β4 are regression coefficients, which are calibrated by field sampling data. Net primary productivity is estimated using the light use efficiency model:

[0059] NPP=ε·APAR·f(T)·f(W)·f(N);

[0060] Among them, NPP is the net primary productivity; ε is the maximum light energy utilization efficiency, which can be estimated by PRI as follows:

[0061] ε=ε max (1+PRI);

[0062] APAR is the absorbed photosynthetically active radiation, which is calculated from LAI as:

[0063] APAR=PAR·(1-e -κ·LAI );

[0064] f(T), f(W), f(N) are the limiting factors of temperature, water and nutrients respectively, and their values ​​range from 0 to 1.

[0065] Example 3: Let the multispectral data of the forest area collected by the multispectral sensor be X; the multispectral data X is a multidimensional matrix, containing the spectral information of each band, and its dimension is expressed as [number of pixel rows, number of pixel columns, number of bands]; the reflectance information of each pixel in the multiband is stored as a vector; the atmospheric correction converts the original radiation brightness data L of the band λ into λ Converted into reflectivity data R that reflects the true surface characteristics λ ; X is a set of atmospherically corrected multispectral data, including the surface reflectance of each band; each third-dimensional slice in X corresponds to the corrected surface reflectance data R of a band λ λ .

[0066] Specifically, during the flight of the drone, the multispectral sensor captures not only the spectral reflectance of the surface, but also the effects of the atmosphere on light scattering and absorption. Therefore, the raw data is expressed as radiance data L λ , which contains information about the influence of the atmosphere after the solar radiation reaches the surface and is reflected back. In order to accurately extract the surface characteristics, these raw radiation brightness data must be atmospherically corrected and converted into surface reflectance data R λ. The principle of atmospheric correction is to adjust the data of different bands according to various environmental parameters such as sensor height, observation angle, atmospheric composition and solar radiation to remove the interference of the atmosphere on the spectral signal. The result of this process is to obtain reflectance data that is closer to the actual surface reflectance characteristics, thereby eliminating the error influence of atmospheric scattering and absorption on the spectral data. In the multispectral data set X after atmospheric correction, the data of each band reflects the surface characteristics corresponding to the band. Specifically, each slice in the third dimension of the matrix X is the reflectance data R of a separate band λ. λ , these data represent the true reflectance characteristics of the forest floor in this band. In this way, the entire multispectral data cube X completely covers the two-dimensional range of the forest area in space, and covers multiple bands in the spectrum, providing a detailed information basis for in-depth analysis of the health of the forest ecosystem.

[0067] Each pixel in the multispectral data stores its reflectance in different bands in the form of a vector. These vectors contain the spectral characteristics of the surface reflectance and can be used to further calculate various spectral indices, such as the Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), etc. These indices are closely related to the forest ecological status. The spectral information of different bands provides multi-level ecological characteristics of the forest surface, such as vegetation density, chlorophyll content, canopy moisture, etc. Therefore, the multispectral data after atmospheric correction not only has the characteristics of high spatial resolution and wide coverage, but also has extremely strong spectral information integrity, which lays a good data foundation for subsequent ecological feature extraction and health assessment. Atmospheric correction is a key step in multispectral data processing. Since the UAV flies at different times, climates and altitudes, the original radiance L λ The data will fluctuate significantly due to changes in environmental conditions. If these data are not corrected for the atmosphere, using them directly for ecological analysis may lead to inaccurate and inconsistent results. Atmospheric correction can ensure that multispectral data collected under different conditions can be compared and analyzed comprehensively, thereby accurately reflecting the true ecological characteristics of the forest area. This accurate data reflection is of great significance for forest health monitoring, carbon storage assessment, ecological change detection, etc.

[0068] Embodiment 4: Atmospheric correction converts the original radiance data L of the band λ into reflectance data R reflecting the real surface characteristics through the following formula: λ :

[0069]

[0070] Among them, L Rλ is the Rayleigh scattered radiation component; L Mλ is the Mie scattered radiation component; L Hλis the residual amount of radiation after atmospheric absorption and scattering; τ λ↑ is the optical thickness from the surface to the multispectral sensor, τ λ↓ is the optical thickness from the sun to the earth's surface; E 0λ is the solar top irradiance; θ s is the solar zenith angle; e is a natural constant.

[0071] Specifically, starting from the multispectral data collected by the drone flying in the forest area, the radiation brightness data L captured by the sensor λ It includes the solar radiation reflected directly from the surface, as well as the components scattered and absorbed in the atmosphere. The numerator of the formula focuses on these atmospheric scattering effects. Rayleigh scattered radiation component L Rλ and the Mie scattered radiation component L Mλ There are two main scattering components that need to be removed in atmospheric correction. Rayleigh scattering is mainly caused by the scattering of short-wavelength light by small molecules in the atmosphere (such as oxygen and nitrogen), resulting in more obvious scattering of the blue light. Mie scattering is caused by larger atmospheric particles (such as dust and water droplets), which usually affects longer wavelength light. The formula is obtained by multiplying these components by the transmission factor T λ↓ The subtraction corrects the scattering effect of the atmosphere on the surface radiation. In addition, other molecules in the atmosphere absorb part of the solar radiation, which is then scattered to the multispectral sensor, forming the residual radiation L after atmospheric absorption and scattering. Hλ This component is also derived from the total radiance L λ to ensure that the corrected data more accurately reflects the reflection from the surface.

[0072] In the denominator, this formula describes the change in energy that light undergoes as it propagates through the atmosphere. 0λ is a key factor, which represents the energy distribution of sunlight before entering the atmosphere. Since sunlight is affected by many factors when passing through the atmosphere, the transmission path of these energies also needs to pass through the optical thickness parameter τ λ↑ and τ λ↓ to correct. Optical thickness is a parameter that describes the attenuation of light due to scattering and absorption in the propagation path. The introduction of the optical thickness parameter in the denominator ensures that the optical characteristics are fully considered in the entire light propagation path from the sun to the surface and from the surface to the sensor. In addition, the exponential attenuation term e used in the formula -τ It reflects the energy attenuation caused by the absorption and scattering of light by the atmosphere. This attenuation follows the Belland-Bouget law, which describes that when light passes through an absorbing substance (such as the atmosphere), its intensity decreases exponentially. sThe overall design of the denominator ensures that the radiation intensity is corrected to accurately reflect the energy changes that light undergoes on its path from the sun to the ground and then to the sensor.

[0073] During the atmospheric correction process, the solar zenith angle θ s The solar zenith angle refers to the angle between the sun's rays and the vertical direction of the earth's surface. The size of this angle determines the angle of incidence of the light when it enters the atmosphere. Because the sun's rays do not always shine vertically on the earth's surface, but change with time, geographical location, and season, the intensity of solar radiation received at the same location at different times is different. Therefore, the correction of the zenith angle is crucial for the accurate calculation of the surface reflectivity. The cosθ involved in the denominator s The term is used to describe the irradiance change caused by different solar incidence angles. The role of the atmospheric correction formula is to remove the effects of scattering, absorption and secondary scattering in the original sensor data by comprehensively considering various atmospheric interference factors in the radiance data, so that the final reflectivity R λ It can more realistically represent the reflectance characteristics of the surface. This is especially critical when drones are equipped with multispectral sensors for forest ecological monitoring. Because the altitude, angle and atmospheric conditions of the drone's flight will cause the collected radiation data to vary greatly at different locations and times, if these changes are not corrected, it may lead to incomparability of data and inaccurate analysis results. The atmospheric correction achieved by this formula quantifies these complex environmental influencing factors, allowing the surface reflectance characteristics to be accurately extracted, thereby providing a reliable data basis for subsequent ecological monitoring and analysis.

[0074] Embodiment 5: The specific process of obtaining the denoised hyperspectral data by the hyperspectral data processing device includes: connecting the multispectral data with the hyperspectral data by using the compressed sensing theory through the following formula:

[0075] X = ΦS + ∈;

[0076] Where X is multispectral data with a shape of [m×p], where m is the total number of spatial pixels and p is the number of multispectral bands; Φ is the measurement matrix with a shape of [p×n], which compresses the hyperspectral data S to the multispectral data X; Φ is a sparse matrix with p<<n, where n is the number of hyperspectral bands; S is the hyperspectral data matrix with a shape of [m×n], where n is the number of hyperspectral bands; ∈ is the noise term, which obeys the Gaussian distribution: σ is the standard deviation of the Gaussian distribution. In order to reconstruct the hyperspectral data S from the multispectral data X, it is necessary to solve the optimization problem and transform the optimization problem into the minimization of the sparse signal. The formula is as follows:

[0077]

[0078] Among them, ||·|| F is the Frobenius norm, which is used to measure the binary norm distance of the matrix; W is the wavelet transform matrix with a shape of [n×n], which sparsely represents the hyperspectral data S. By performing wavelet transform on the hyperspectral data, it is converted to the sparse domain; λ is the regularization parameter, which is used to balance the reconstruction error and sparsity.

[0079] Specifically, in the multispectral data X collected by the drone, the shape of the matrix is ​​[m×p], where m represents the total number of spatial pixels in the forest area and p is the number of bands of the multispectral data. The dimension of the hyperspectral data S is higher, with a shape of [m×n], where n is the number of bands of the hyperspectral data. Usually n is much larger than p, which means that the hyperspectral data contains more spectral information at each pixel. In practical applications, directly obtaining such high-dimensional spectral data requires a lot of storage space and computing resources. Therefore, in order to efficiently obtain hyperspectral information, the compressed sensing theory provides a method to compress the high-dimensional data S into low-dimensional data X through the measurement matrix Φ. Here, Φ is a sparse measurement matrix of shape [p×n], which compresses the hyperspectral data S into a low-dimensional space through linear projection to form the multispectral data X. Since p<<n, this means that in the compression process, the hyperspectral data is effectively reduced in dimension, thereby greatly reducing the amount of data. In this process, noise ∈ is inevitably introduced, which obeys the Gaussian distribution and describes the random error in the multispectral data acquisition and compression process. The existence of this noise interference requires noise reduction in the reconstruction process of hyperspectral data. The goal of the hyperspectral data processing device is to accurately reconstruct the hyperspectral data S from the multispectral data X while minimizing the influence of noise. To achieve this goal, the device needs to solve an optimization problem whose objective function comprehensively considers the reconstruction error and sparsity of the data.

[0080] The objective function of the optimization problem consists of two parts: data reconstruction error and sparsity constraint. The reconstruction error is calculated by the Frobenius norm. It describes the difference between the multispectral data X and the hyperspectral data S after the measurement matrix Φ is projected into a low-dimensional space. The Frobenius norm is a two-norm that represents the sum of the squares of the matrix elements. By minimizing the reconstruction error, the optimization process ensures that the hyperspectral data reconstructed from the multispectral data X is As close to the true value as possible, so as to restore the fine features of the forest floor in different spectral dimensions. The sparsity constraint is reflected by ||WS||1, where W is the wavelet transform matrix, which is used to transform the hyperspectral data S into a sparse domain. The sparsity of hyperspectral data means that after wavelet transform, it can be represented by a few non-zero coefficients. Through wavelet transform, the hyperspectral signal is mapped from the spatiotemporal domain to the frequency domain, where most of the energy is concentrated in a few coefficients, thus showing sparse characteristics. The optimization process ensures that the representation of hyperspectral data in the sparse domain is as simple as possible by minimizing ||WS||1, reducing unnecessary information redundancy. Sparsity not only helps to compress data, but also effectively suppresses noise, because the distribution of noise in the sparse domain is usually uneven, and most of the noise components can be filtered out through sparse representation. The regularization parameter λ plays a role in balancing the reconstruction error and sparsity in the optimization problem. A larger λ value means that sparsity is given a greater weight, and the optimization process tends to obtain a sparse hyperspectral signal; on the contrary, a smaller λ value pays more attention to minimizing the reconstruction error. Therefore, choosing an appropriate λ value is crucial to ensure accurate reconstruction and noise suppression of hyperspectral data.

[0081] By solving this optimization problem, we get It is the optimal hyperspectral reconstruction result of multispectral data X under the projection of measurement matrix Φ. Since the optimization problem combines the sparsity of data with the reconstruction error, and considering the sparse representation characteristics of wavelet transform, this method can reconstruct hyperspectral data with minimum error and high accuracy under limited sampling conditions. The final result is a set of hyperspectral data that has been denoised and retains rich spectral features, providing a detailed and reliable basis for the health monitoring and analysis of forest ecosystems. In the forest ecological monitoring system equipped with multispectral sensors by unmanned aerial vehicles, the application of compressed sensing theory has greatly improved the efficiency of data processing. Multispectral data compresses hyperspectral information through the measurement matrix, reducing the cost of acquisition and storage. At the same time, sparsity constraints and wavelet transform are introduced in the reconstruction of hyperspectral data to achieve high-precision noise reduction and signal restoration. This method makes full use of the sparse characteristics of hyperspectral data, ensuring that the system can conduct comprehensive and detailed monitoring of forest ecological characteristics under limited computing resources and data acquisition conditions, providing reliable technical support for the assessment of forest health and the calculation of carbon reserves.

[0082] Example 6: The noise in the hyperspectral data is distributed in the high-frequency part. Wavelet transform is used to sparsely represent S and perform noise reduction in the sparse domain. The wavelet transform matrix W is decomposed into:

[0083]

[0084] Among them, W spatialis the spatial wavelet transform matrix. Considering the spatial information of the matrix, discrete wavelet transform is used to implement it. spectral is the spectral wavelet transform matrix, which considers the spectral information between bands and is implemented using one-dimensional wavelet transform; is the tensor product; after adding wavelet denoising and sparsity constraints, the following complete optimization problem is obtained:

[0085]

[0086] in, is the sparse regularization term after wavelet transformation, indicating the sparsity of hyperspectral data after wavelet transformation; ||S|| TV is the total variation regularization term, which is used to maintain the edge characteristics of the matrix and reduce noise, and is defined as:

[0087]

[0088] λ1 and λ2 are regularization parameters, which are used to weigh the impact of sparsity and total variation on reconstruction respectively; i and j are subscript indices; S i，j Represents the element in the i-th row and j-th column in the hyperspectral data matrix; The hyperspectral data after denoising.

[0089] Specifically, wavelet transform is used to capture the different features of hyperspectral data in both spatial and spectral dimensions. Hyperspectral data S is essentially a three-dimensional data cube, in which each pixel has its position in the spatial dimension and has information of multiple bands in the spectral dimension. In order to make full use of this information, two wavelet transform matrices are introduced in this embodiment: spatial wavelet transform matrix W spatial and the spectral wavelet transform matrix W spectral .W spatial It is mainly used to analyze the spatial distribution of data between pixels and capture the association and changes between different pixels; while W spectral It is used to analyze the information of hyperspectral data in the spectral dimension and describes the spectral characteristics between different bands. These two matrices are tensor-producted Combined together, it forms a complete wavelet transform matrix It is possible to analyze the spatial and spectral characteristics of hyperspectral data simultaneously. After wavelet transform, the data is decomposed into different scales and frequency components, so that the signal and noise can be separated in the frequency domain. Since the noise is mainly distributed in the high-frequency part, and the main features of the hyperspectral signal are concentrated in the low-frequency part, the high-frequency noise can be effectively suppressed and the main features of the signal can be retained through sparse processing in the wavelet domain. To achieve this, the objective function of the optimization problem is designed as a combination of reconstruction error, wavelet sparsity and total variation regularization terms.

[0090] In this optimization objective, the first term is the reconstruction error Where X is the multispectral data collected by the drone sensor, and φ is the measurement matrix used to compress the hyperspectral data S to multispectral data. By minimizing this error term, the reconstructed hyperspectral data is ensured to be consistent with the collected multispectral data, thereby accurately restoring the original signal. The second term introduces the wavelet sparse regularization term This term emphasizes the sparsity of hyperspectral data in the wavelet domain. The wavelet transformed signal has a small number of non-zero coefficients in the sparse domain, while most of the coefficients are close to zero. This sparse representation property helps to separate the signal from the noise, and a sparse hyperspectral signal can be obtained through optimization. Here, λ1 is a regularization parameter used to balance the weight between reconstruction error and sparsity. The optimization process hopes to find a hyperspectral data that is sparse in the wavelet domain so that most of the noise components are filtered out and the main features of the signal are retained. The third term is the total variation (TV) regularization term ||S|| TV , which is used to preserve the edge characteristics and structural information in hyperspectral data. Total variation is a denoising technique used in signal and image processing, which emphasizes the continuity of adjacent pixels in the signal while retaining mutation features such as edges and textures. In the formula, the TV regularization term measures the degree of change of the signal by the sum of the squares of the differences between adjacent pixels. Minimizing this term can smooth the high-frequency noise of the signal while retaining the edge characteristics of the signal. For hyperspectral data, edge characteristics usually correspond to the boundaries between different vegetation or surface features, so maintaining these edge information is crucial for subsequent forest ecological monitoring and analysis. λ2 is introduced as a regularization parameter to balance the contribution of sparsity and total variation to signal reconstruction. By combining these three items in the optimization objective, the optimization problem strikes a balance between minimizing the reconstruction error, sparse representation, and edge preservation, ensuring the denoising effect and reconstruction quality of hyperspectral data. When solving the optimization problem, the final result is It is the hyperspectral data that has been processed by noise reduction. It not only eliminates the high-frequency noise caused by sensor noise and environmental interference in the original data, but also retains the important characteristics of the signal in spatial and spectral dimensions, thus providing high-quality data support for the accurate monitoring of forest ecosystems.

[0091] Embodiment 7: The complete optimization problem is a non-convex optimization problem; for the complete optimization problem, the alternating direction multiplier method is used to decompose it into multiple sub-problems for solution; wherein the soft threshold algorithm is used to solve Solved by gradient descent method

[0092] Specifically, in Example 7, the complete optimization problem is a non-convex optimization problem, so direct solution will face great computational challenges. In order to solve this problem, the system adopts the alternating direction multiplier method (ADMM) to decompose the original optimization problem into several solvable sub-problems, and gradually approach the optimal solution of the original problem by alternately solving each sub-problem. ADMM shows high efficiency in dealing with non-convex optimization problems, and is particularly suitable for optimization objectives with multiple constraints and regularization terms. Specifically, the different terms in the objective function reflect the different characteristics and constraints of hyperspectral data. The goal of ADMM is to optimize these terms separately, thereby achieving the overall optimization goal.

[0093] The first of the three parts of the optimization problem is It measures the reconstruction error between the multispectral data X and the reconstructed hyperspectral data φS. Since this part is a convex function, it can be solved by the gradient descent method. The gradient descent method is an iterative optimization algorithm that calculates the gradient of the objective function and updates the parameter S along the negative direction of the gradient to minimize the reconstruction error. In each iteration, the gradient is calculated and the hyperspectral data S is updated, gradually approaching the global optimal solution. The advantage of this process is that it is simple and effective, and can converge quickly, especially when dealing with convex functions, it has high computational efficiency.

[0094] The second part is It is a sparse regularization term in the optimization objective. This term involves the sparse representation of hyperspectral data after wavelet transformation, and the goal is to remove noise through sparsification in the wavelet domain. Since this term contains the L1 norm, direct solution may lead to instability. For this reason, the system uses a soft-thresholding algorithm to solve it. The soft threshold algorithm is a commonly used sparse signal processing method. By thresholding the coefficients of the signal in each iteration, the coefficients less than the threshold are set to zero or reduced, and the larger signal coefficients are retained. In this way, the effect of sparse data can be achieved while retaining the main features of the signal. Specifically, the soft threshold algorithm can effectively reduce the interference of noise on hyperspectral data, make the data after wavelet transformation sparse, and retain useful feature information in the hyperspectral signal.

[0095] The working principle of ADMM is to decompose the original optimization problem into independent sub-problems, optimize each sub-problem separately, and update the solutions of all sub-problems after each iteration step. In this case, the sub-problems of reconstruction error term and sparse regularization term are solved separately by combining gradient descent method and soft threshold algorithm. First, the gradient descent method is used to optimize the reconstruction error to ensure that the projection error between multispectral data X and hyperspectral data S is minimized. Then, the soft threshold algorithm processes the sparse regularization term to ensure that the hyperspectral data has good sparsity in the wavelet domain. By repeatedly optimizing these two sub-problems alternately, ADMM gradually approaches the optimal solution of the original optimization problem. This alternating solution method helps to overcome the difficulty of non-convex optimization problems, because the direct solution of the original problem involves complex multivariable coupling relationships, while ADMM decomposes these coupling relationships into more manageable independent parts by introducing auxiliary variables. In each iteration step, the optimization of each sub-problem can converge quickly, so that the overall solution gradually converges to the optimal solution of the original problem.

[0096] Example 8: The forest health index is calculated using the following formula:

[0097]

[0098] Among them, FHI is the forest health index, ranging from 0 to 1; LAI is the leaf area index; CWC is the canopy water content; PRI is the vegetation photosynthetic efficiency; NPP is the net primary productivity; ω1, ω2, ω3 and ω4 are all weight coefficients, satisfying LAI max is the maximum value of the theoretical leaf area index;

[0099] CWC max is the theoretical maximum canopy moisture content; NPP max is the theoretical maximum value of net primary productivity.

[0100] Specifically, the calculation of the forest health index FHI involves four core indicators: leaf area index (LAI), canopy water content (CWC), photosynthetic efficiency index (PRI) and net primary productivity (NPP). These indicators reflect the health of forest ecosystems from different perspectives. UAV multispectral sensors can capture spectral information of forest areas in different bands through hyperspectral imaging, so as to accurately calculate parameters such as LAI, CWC, PRI and NPP for each pixel. These parameters are key representations of vegetation health: LAI reflects the coverage and growth density of forest leaves. A higher LAI usually indicates that the forest has a good growth state; CWC is used to measure the water content in the canopy, reflecting the forest's ability to absorb and retain water resources. Forests with sufficient water are generally healthier and more resistant to stress; PRI is mainly used to evaluate the photosynthetic efficiency of vegetation, because plant photosynthesis directly affects the energy balance and carbon fixation of forests; NPP represents net primary productivity, which is the core indicator for measuring the carbon sink capacity of forest ecosystems, reflecting the total amount of carbon dioxide absorbed by forest vegetation in photosynthesis. These parameters are normalized to values ​​between 0 and 1 in the FHI formula to facilitate comprehensive analysis and comparison. The normalization process takes into account the theoretical maximum value of each indicator, for example, LAI max Indicates the theoretical maximum leaf area per unit area, CWC max and NPP max It represents the theoretical maximum value of canopy water content and net primary productivity. After normalization, each parameter reflects its health level relative to the most ideal state under specific time and space conditions. In addition, the original value of PRI may be negative, so it is used in the calculation. to normalize it to a positive value.

[0101] In order to comprehensively evaluate forest health, four indicators are given different weights ω1, ω2, ω3, and ω4 in the FHI formula, and the sum of the weights is equal to 1. This means that the impact of different indicators on the overall health index can be adjusted according to research focus or actual application needs. For example, in arid areas, the impact of canopy water content CWC on forest health may be more significant, so the weight of ω2 can be increased accordingly. In the assessment of ecosystem productivity, more attention may be paid to the contribution of NPP to the carbon cycle, so the weight of ω4 may be greater. This flexibility in weight setting makes FHI highly adaptable and can be optimized for different forest types, ecological conditions and monitoring objectives. The hyperspectral data obtained by the multispectral sensor of the drone is extracted through a specific algorithm to extract indicators such as LAI, CWC, PRI, and NPP. When calculating FHI, the normalized values ​​and weights of these indicators are used for weighted averaging, and finally a comprehensive index FHI between 0 and 1 is obtained. The closer the index is to 1, the healthier the forest is and the better the ecosystem is. On the contrary, the closer the index is to 0, the more likely the forest may have health problems, such as water stress, pest infestation, or limited photosynthesis. By continuously monitoring changes in FHI, abnormalities in forest health can be detected in a timely manner, assisting forestry departments in making corresponding management decisions. This forest health index assessment method based on hyperspectral data provides a scientific and efficient means for forest ecological monitoring systems. It can comprehensively reflect the health characteristics of forest ecosystems in terms of structure, function, and physiological state, and provides strong data support for the protection and management of forest resources.

[0102] Example 9: Forest carbon storage is calculated using the following formula:

[0103] C = B·CF·(1+R);

[0104] Among them, C is the forest carbon stock; CF is the coefficient of biomass conversion to carbon, ranging from 0.4 to 0.6; R is the ratio of underground biomass to aboveground biomass, ranging from 0.2 to 0.3.

[0105] Specifically, the formula C = B·CF·(1+R) is the key to the entire calculation, where C represents forest carbon storage, B is biomass, CF is the coefficient for converting biomass to carbon content, and R reflects the proportional relationship between underground biomass and aboveground biomass. Biomass B is an indicator of the total volume, density and growth status of forest vegetation, including all organic matter such as trunks, branches, leaves and underground roots. The hyperspectral data collected by the multispectral sensor of the drone contains the reflection characteristics of vegetation in different bands. By analyzing and processing these reflection characteristics, the biomass of the forest area can be estimated. Hyperspectral data has high resolution and multi-dimensional characteristics, which can capture the subtle differences of vegetation in multiple bands such as near-infrared and visible light, making the estimation of biomass more accurate. When calculating carbon storage, biomass B needs to be converted into carbon content, which is achieved through CF. CF is the conversion coefficient from biomass to carbon content, usually between 0.4 and 0.6, and the specific value depends on the vegetation type, environmental conditions and carbon content characteristics of the forest biomass. For example, the carbon conversion coefficients of coniferous forests and broad-leaved forests will be different because they differ in cell structure, wood fiber content, and carbon content. The hyperspectral data collected by the multispectral sensor of the drone can reflect these subtle spectral differences and help select the appropriate CF to ensure the accuracy of carbon storage estimation. In addition, the carbon storage of the forest is not only contributed by the aboveground biomass, but the carbon storage of the underground biomass is also an important part. The R in the formula describes the proportional relationship between the underground biomass and the aboveground biomass, which is usually between 0.2 and 0.3. The underground biomass mainly includes roots, soil microorganisms, etc., which together with the aboveground biomass constitute the overall carbon storage capacity of the forest ecosystem. By introducing R, the calculation of carbon storage is no longer limited to the aboveground part, but covers the biomass of the entire ecosystem, thereby more comprehensively evaluating the ability of the forest to serve as a carbon sink.

[0106] Embodiment 10: A method for monitoring forest ecology using a multispectral sensor mounted on an unmanned aerial vehicle, the method comprising:

[0107] Step 1: Collect multispectral data of the forest area using a drone equipped with a multispectral sensor and perform atmospheric correction on the surface reflectance;

[0108] Step 2: Using compressed sensing theory, reconstruct hyperspectral data from multispectral data, and use wavelet transform to reduce noise to obtain denoised hyperspectral data;

[0109] Step 3: Calculate the normalized vegetation index and leaf area index based on the near-infrared band reflectance and red light band reflectance in the denoised hyperspectral data; estimate the canopy moisture content using the short-wave infrared band reflectance; calculate the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculate the net primary productivity using the light use efficiency model; calculate the forest health index based on the net primary productivity, photochemical vegetation index, canopy moisture content and leaf area index; calculate the stand biomass based on the leaf area index, canopy moisture content and photochemical vegetation index; calculate the forest carbon stock based on the stand biomass; present the forest carbon stock and forest health index as monitoring results to the user.

[0110] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The forest ecological monitoring system equipped with a multispectral sensor by a drone is characterized by: The system includes: an unmanned aerial vehicle, a hyperspectral data processing device and an ecological health detection platform; the unmanned aerial vehicle is equipped with a multispectral sensor to collect multispectral data of a forest area and perform atmospheric correction on the surface reflectance; the hyperspectral data processing device is used to reconstruct hyperspectral data from multispectral data using compressed sensing theory, and use wavelet transform to perform noise reduction to obtain noise-reduced hyperspectral data; the ecological health detection platform is used to calculate a forest health index and forest carbon reserves based on the noise-reduced hyperspectral data; and the forest carbon reserves and forest health index are presented to users as monitoring results.

2. The forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in claim 1, characterized in that: The ecological health detection platform calculates the forest health index and the forest carbon stock based on the denoised hyperspectral data, and the process specifically includes: calculating the normalized vegetation index and the leaf area index based on the near-infrared band reflectance and the red light band reflectance in the denoised hyperspectral data; estimating the canopy moisture content using the short-wave infrared band reflectance; calculating the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculating the net primary productivity using the light utilization efficiency model; calculating the forest health index based on the net primary productivity, the photochemical vegetation index, the canopy moisture content and the leaf area index; calculating the stand biomass based on the leaf area index, the canopy moisture content and the photochemical vegetation index; and calculating the forest carbon stock based on the stand biomass.

3. The forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in claim 2, characterized in that: Suppose the multispectral data of the forest area collected by the multispectral sensor is X; the multispectral data X is a multidimensional matrix containing the spectral information of each band, and its dimension is expressed as [number of pixel rows, number of pixel columns, number of bands]; the reflectance information of each pixel in the multiband is stored as a vector; the atmospheric correction converts the original radiance data L with band λ into λ Converted into reflectivity data R that reflects the true surface characteristics λ ; X is a set of atmospherically corrected multispectral data, including the surface reflectance of each band; each third-dimensional slice in X corresponds to the corrected surface reflectance data R of a band λ λ .

4. The forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in claim 3, characterized in that: The atmospheric correction converts the original radiation brightness data L in the band λ into λ Converted into reflectivity data R that reflects the true surface characteristics λ : Among them, L Rλ is the Rayleigh scattered radiation component; L Mλ is the Mie scattered radiation component; L Hλ is the residual amount of radiation after atmospheric absorption and scattering; τ λ↑ is the optical thickness from the surface to the multispectral sensor, τ λ↓ is the optical thickness from the sun to the earth's surface; E 0λ is the solar top irradiance; θ s is the solar zenith angle; e is a natural constant.

5. The forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in claim 3, characterized in that: The specific process of obtaining the denoised hyperspectral data by the hyperspectral data processing device includes: connecting the multispectral data with the hyperspectral data by using the compressed sensing theory through the following formula: X = ΦS + ∈; Where X is multispectral data with a shape of [m×p], where m is the total number of spatial pixels and p is the number of multispectral bands; Φ is the measurement matrix with a shape of [p×n], which compresses the hyperspectral data S to the multispectral data X; Φ is a sparse matrix with p<<n, where n is the number of hyperspectral bands; S is the hyperspectral data matrix with a shape of [m×n], where n is the number of hyperspectral bands; ∈ is the noise term, which obeys the Gaussian distribution: σ is the standard deviation of the Gaussian distribution. In order to reconstruct the hyperspectral data S from the multispectral data X, it is necessary to solve the optimization problem and transform the optimization problem into the minimization of the sparse signal. The formula is as follows: Among them, ||·|| F is the Frobenius norm, which is used to measure the binary norm distance of the matrix; W is the wavelet transform matrix with a shape of [n×n], which sparsely represents the hyperspectral data S. By performing wavelet transform on the hyperspectral data, it is converted to the sparse domain; λ is the regularization parameter, which is used to balance the reconstruction error and sparsity.

6. The forest ecological monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in claim 5, characterized in that: The noise in the hyperspectral data is distributed in the high-frequency part. Wavelet transform is used to sparsely represent S and perform noise reduction in the sparse domain. The wavelet transform matrix W is decomposed into: Among them, W spatial is the spatial wavelet transform matrix. Considering the spatial information of the matrix, discrete wavelet transform is used to implement it. spectral is the spectral wavelet transform matrix, which considers the spectral information between bands and is implemented using one-dimensional wavelet transform; is the tensor product; after adding wavelet denoising and sparsity constraints, the following complete optimization problem is obtained: in, is the sparse regularization term after wavelet transform, indicating the sparsity of hyperspectral data after wavelet transform; ||S|| TV is the total variation regularization term, which is used to maintain the edge characteristics of the matrix and reduce noise, and is defined as: λ1 and λ2 are regularization parameters, which are used to weigh the impact of sparsity and total variation on reconstruction respectively; i and j are subscript indices; S i，j Represents the element in the i-th row and j-th column in the hyperspectral data matrix; The hyperspectral data after denoising.

7. The forest ecological monitoring system using a multi-spectral sensor mounted on an unmanned aerial vehicle as claimed in claim 6, characterized in that: The complete optimization problem is a non-convex optimization problem. For the complete optimization problem, the alternating direction multiplier method is used to decompose it into multiple sub-problems for solution. Among them, the soft threshold algorithm is used to solve Solved by gradient descent method 8. The forest ecological monitoring system using a multi-spectral sensor mounted on an unmanned aerial vehicle as claimed in claim 7, characterized in that: The forest health index is calculated using the following formula: Among them, FHI is the forest health index, ranging from 0 to 1; LAI is the leaf area index; CWC is the canopy water content; PRI is the vegetation photosynthetic efficiency; NPP is the net primary productivity; ω1, ω2, ω3 and ω4 are all weight coefficients, satisfying LAI max is the maximum value of the theoretical leaf area index; CWC max is the theoretical maximum canopy moisture content; NPP max is the theoretical maximum value of net primary productivity.

9. The forest ecological monitoring system using a multi-spectral sensor mounted on an unmanned aerial vehicle as claimed in claim 8, characterized in that: Forest carbon storage is calculated using the following formula: C = B·CF·(1+R); Among them, C is the forest carbon stock; CF is the coefficient of biomass conversion to carbon, ranging from 0.4 to 0.6; R is the ratio of underground biomass to aboveground biomass, ranging from 0.2 to 0.

3.

10. A method for monitoring forest ecology using a multispectral sensor mounted on an unmanned aerial vehicle, for implementing a forest ecology monitoring system using a multispectral sensor mounted on an unmanned aerial vehicle as claimed in any one of claims 1 to 9, characterized in that: The method comprises: Step 1: Collect multispectral data of the forest area using a drone equipped with a multispectral sensor and perform atmospheric correction on the surface reflectance; Step 2: Using compressed sensing theory, reconstruct hyperspectral data from multispectral data, and use wavelet transform to reduce noise to obtain denoised hyperspectral data; Step 3: Calculate the normalized vegetation index and leaf area index based on the near-infrared band reflectance and red light band reflectance in the denoised hyperspectral data; estimate the canopy moisture content using the short-wave infrared band reflectance; calculate the photochemical vegetation index using the 531nm band reflectance and the 570nm band reflectance; calculate the net primary productivity using the light use efficiency model; calculate the forest health index based on the net primary productivity, photochemical vegetation index, canopy moisture content and leaf area index; calculate the stand biomass based on the leaf area index, canopy moisture content and photochemical vegetation index; calculate the forest carbon stock based on the stand biomass; present the forest carbon stock and forest health index as monitoring results to the user.

Images