Urban cinnamomum camphora health monitoring method based on hyperspectrum of unmanned aerial vehicle

Abstract

The invention discloses an urban cinnamomum camphora health monitoring method based on unmanned aerial vehicle hyperspectrum, and relates to the technical field of ecological environment monitoring, and the method comprises the following steps: 1, obtaining a cinnamomum camphora forest canopy image of a monitored area through a hyperspectral camera carried by an unmanned aerial vehicle; 2, collecting a leaf sample of a cinnamomum camphora forest in a monitoring area; 3, extracting a spectral band, a vegetation index and texture features from the hyperspectral image, and matching the spectral band, the vegetation index and the texture features with an on-site acquired data space; 4, generating synthetic samples by using a generative adversarial network, and amplifying the number of the samples; 5, screening key feature combinations, and constructing a gradient lifting regression health parameter estimation model; and 6, outputting a health grading map of the cinnamomum camphora forest in the monitored area by applying the gradient lifting regression health parameter estimation model, monitoring the health condition of the cinnamomum camphora forest in the whole area, and positioning abnormal plants. According to the urban cinnamomum camphora health monitoring method based on the hyperspectrum of the unmanned aerial vehicle, the health condition of a cinnamomum camphora forest is accurately, rapidly and nondestructively monitored.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring technology, and in particular to a method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging. Background Technology

[0002] Current methods for monitoring the health of camphor trees mainly consist of two types: manual ground surveys and satellite data monitoring. Manual ground surveys, as a traditional monitoring method, rely on portable instruments for individual tree measurements. Taking the SPAD-502 chlorophyll meter as an example, it obtains the relative chlorophyll content (SPAD value) through a dual-wavelength optical measurement principle of 650nm and 940nm, and calculates the leaf area index based on the canopy gap fraction obtained by the LAI-2200 canopy analyzer using a fisheye lens. While this method can obtain accurate data for individual trees (leaf-level precision), it has significant limitations: First, the sampling efficiency is extremely low. According to actual measurement data from *Forest Ecology and Management* (2019), a four-person working group could only sample a maximum of 15-20 trees per day. Second, destructive sampling requires cutting leaves from the canopy, and long-term monitoring can lead to physiological damage to the sample trees. A 2021 report from the U.S. Forest Service showed that repeated sampling increases the annual variability rate of camphor tree chlorophyll content by 23%. More importantly, manual measurements are susceptible to subjective factors. Wu et al. (2020) confirmed that the SPAD value difference between leaves on the sunny and shaded sides can reach 12.3%, and standard sampling procedures often fail to cover this microenvironmental variation. Satellite remote sensing technology, represented by Landsat-9 (30m resolution) and Sentinel-2 (10m resolution), acquires vegetation indices (such as NDVI and EVI) through multispectral sensors. While this technology can achieve large-scale coverage (single scene coverage of 185km), it is not universally applicable. 2 However, it suffers from a fundamental technical flaw: insufficient spatial resolution leads to low single-tree recognition rates.

[0003] The study *Remote Sensing of Environment* (2021) showed that for camphor saplings with a crown width <4m, the false negative rate of Sentinel-2 data exceeded 40%. Furthermore, atmospheric interference severely impacts data quality; Zhang et al. (2022) verified through MODIS data that aerosols cause NDVI value fluctuations of ±0.15. While the "Landsat 8OLI band 5-4 difference method" used in patent CN112730132A can assess the overall health of a forest stand, it cannot distinguish individual differences between adjacent trees (Example 1 in the patent specification). It is noteworthy that the temporal resolution of satellite data (revisit period 5-16 days) also makes it difficult to capture the rapid changes in pest and disease outbreaks. Summary of the Invention

[0004] The purpose of this invention is to provide a method for monitoring the health of urban camphor trees based on hyperspectral data from unmanned aerial vehicles (UAVs). This method uses UAVs equipped with hyperspectral equipment to acquire canopy spectral data, analyzes the spectral response characteristics and patterns of health status, and constructs a health parameter estimation model for camphor forests, thereby achieving rapid, efficient, and accurate monitoring of the health status of camphor forests.

[0005] To achieve the above objectives, this invention provides a method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging, comprising the following steps:

[0006] Step 1: Under clear and windless conditions, use a drone equipped with a hyperspectral camera to acquire camphor forest canopy images with a spatial resolution of 0.1m. Based on laboratory calibration cloths with 20% and 40% reflectance, use the empirical linear method to perform atmospheric correction on the original images.

[0007] Step 2: Collect leaf samples from the sunny side of the canopy of camphor trees in the field, distinguish between new leaves and old leaves, record the location of each sampled tree, and measure the chlorophyll content, leaf nitrogen content, and leaf area index of new and old leaves.

[0008] Step 3: Based on the location of each sampled tree, extract the 3m×3m canopy area corresponding to the location of each sample tree from the atmospherically corrected canopy image, extract the band features, principal component features, texture features and vegetation index features of the hyperspectral data, and perform spatial matching with the ground data;

[0009] Step 4: Construct a generative adversarial network model to generate synthetic samples and increase the number of samples for chlorophyll, leaf nitrogen content, and leaf area index;

[0010] Step 5: Design a dynamic feature optimization algorithm to optimize feature combinations. Starting from the first feature, add features one by one. After traversing all features, determine the optimal feature combination. After determination, use a gradient boosting regressor to establish a health parameter estimation model.

[0011] Step 6: Based on the measured value range, the chlorophyll content, leaf nitrogen content and leaf area index are classified using the natural breakpoint method. The canopy hyperspectral characteristics of all camphor forests in the monitoring area are input into the health parameter estimation model. The model simultaneously outputs the predicted values ​​of chlorophyll content, leaf nitrogen content and leaf area index, and generates a thematic map of camphor forest health classification based on the three parameters.

[0012] Preferably, the hyperspectral camera carried by the UAV in step one has a hyperspectral module covering a spectral range of 400-1000nm, a spectral resolution of 2.68nm, and a total of 224 bands.

[0013] Preferably, in step two, the new and old leaves are distinguished by the color of the leaves and petioles. During the process of measuring the leaf area index, each camphor tree sample is measured three times, and the average value is used for subsequent experiments.

[0014] Preferably, the band features extracted from the hyperspectral data in step three include 14 hyperspectral bands: 450nm, 550nm, 571nm, 615nm, 641nm, 650nm, 690nm, 706nm, 720nm, 728nm, 790nm, 800nm, 397nm, and 405nm.

[0015] Preferably, the principal component features in step three include the first 10 principal components obtained from principal component analysis of the original 224 bands.

[0016] Preferably, the texture features in step three include six texture features: contrast, dissimilarity, homogeneity, autocorrelation matrix, energy, and correlation.

[0017] Preferably, the vegetation index features in step three include 14 vegetation indices: normalized vegetation index, corrected chlorophyll absorption and reflectance index, converted chlorophyll absorption and reflectance index, chlorophyll absorption and reflectance index, chlorophyll index, red-edge normalized difference index, proportional vegetation index, difference vegetation index, soil regulation index, green index, normalized difference water body index, terrestrial chlorophyll index, triangular vegetation index, and photochemical reflectance index.

[0018] Preferably, the generative adversarial network model in step four consists of a generator and a discriminator. During training, the generator and the discriminator engage in a game until their losses reach equilibrium, at which point training is complete. The training effect of the model is judged by examining the error between the generated data and the real data.

[0019] Preferably, in step five, the gradient boosting regressor minimizes the loss function by iteratively training the decision tree model. Each tree is trained on the residual of the previous tree. The parameter combination used in the model construction is n_estimators=500, learning_rate=0.05, max_depth=4, subsample=0.9, and an dom_state=42.

[0020] Preferably, in step six, the natural breakpoint method is used to divide the model predicted values ​​of the three parameters—chlorophyll content, leaf nitrogen content, and leaf area index—into five grade ranges.

[0021] Therefore, the above-mentioned method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging has the following beneficial effects:

[0022] (1) A drone with a flight altitude of 100m was used to carry a 0.1m resolution hyperspectral camera to achieve fine analysis of a single camphor tree. Through laboratory calibration and atmospheric correction using empirical linear method, the data accuracy error was controlled within ±5%.

[0023] (2) By adopting the generative adversarial network framework and constructing a high-fidelity synthetic sample generation system through the adversarial training mechanism of generator-discriminator, the bottleneck problem of insufficient sample size in traditional monitoring methods is effectively solved, the effective training samples are expanded, and the sample diversity is improved.

[0024] (3) Develop an intelligent feature selection algorithm based on recursive feature elimination and introduce an early stopping mechanism to optimize computational efficiency. The system can automatically identify 67 key feature bands from 224 original hyperspectral bands, reducing the feature dimension by 70% while ensuring that the modeling accuracy loss is controlled within 5%, and improving the overall modeling efficiency by 3.2 times.

[0025] (4) An innovative multi-task learning framework based on gradient boosting regression tree was constructed. By sharing the feature extraction layer and designing a joint loss function, the chlorophyll content, leaf nitrogen content and leaf area index were estimated simultaneously and accurately.

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0027] Figure 1 This is a spatial distribution map of chlorophyll content, leaf nitrogen content, and leaf area index collected in the field in a camphor forest according to an embodiment of the present invention.

[0028] Figure 2 The importance ranking of features after the amplification of the chlorophyll estimation model of this invention;

[0029] Figure 3 The importance ranking of features after the amplification of the leaf nitrogen content estimation model of this invention;

[0030] Figure 4 The importance ranking of features after the amplification of the leaf area index estimation model of this invention;

[0031] Figure 5 This is a schematic diagram of the gradient boosting regressor architecture of the present invention;

[0032] Figure 6 This is a thematic map of the health classification of camphor forests in a certain monitoring area, generated by the present invention based on a health parameter estimation model. Detailed Implementation

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0035] Example

[0036] Please see Figures 1-6 This invention provides a method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging, comprising the following steps:

[0037] Step 1: Using a drone equipped with a hyperspectral camera, under clear and windless weather conditions, acquire camphor forest canopy images with a spatial resolution of 0.1m at a flight altitude of 100m. Based on laboratory calibration cloths with 20% and 40% reflectance, perform atmospheric correction on the original images using the empirical linear method to eliminate interference from illumination and atmospheric scattering.

[0038] The drone is equipped with a hyperspectral camera whose hyperspectral module covers a spectral range of 400-1000nm, with a spectral resolution of 2.68nm and a total of 224 bands, including near-infrared and red-edge bands.

[0039] Step 2: Use high-branch shears to collect leaf samples from the sunny side of the camphor tree canopy. During the collection, distinguish between new leaves and old leaves, record the location of each sampled tree, and measure the chlorophyll content, leaf nitrogen content, and leaf area index of new and old leaves.

[0040] like Figure 1 As shown, researchers collected chlorophyll content, leaf nitrogen content, and leaf area index from camphor forest sample trees in the monitoring area and recorded the location of each sample tree. During the collection, they distinguished between new leaves and old leaves. The significant feature for the distinction was the difference in the color of the leaves and petioles between the new and old leaves. In the process of measuring the leaf area index, each camphor sample tree was measured three times, and the average value was used for subsequent experiments.

[0041] Step 3: Based on the location of each sampled tree, extract the 3m×3m canopy area corresponding to the location of each sample tree from the atmospherically corrected canopy image, extract the band features, principal component features, texture features and vegetation index features of the hyperspectral data, and perform spatial matching with the ground data.

[0042] The hyperspectral data were extracted to include 14 bands: 450nm, 550nm, 571nm, 615nm, 641nm, 650nm, 690nm, 706nm, 720nm, 728nm, 790nm, 800nm, 397nm, and 405nm. Principal component features included the first 10 principal components obtained from principal component analysis of the original 224 bands. Texture features included six texture features: contrast, dissimilarity, homogeneity, autocorrelation matrix, energy, and correlation. Vegetation index features included 14 vegetation indices: normalized vegetation index, corrected chlorophyll absorption and reflectance index, converted chlorophyll absorption and reflectance index, chlorophyll absorption and reflectance index, chlorophyll index, red-edge normalized difference index, proportional vegetation index, differential vegetation index, soil regulation index, green index, normalized difference water index, terrestrial chlorophyll index, triangular vegetation index, and photochemical reflectance index.

[0043] An innovative 34-dimensional composite feature pool was constructed. By carefully selecting 14 sensitive physical bands such as 450nm and 550nm, six types of GLCM texture indicators such as contrast and energy were extracted. Integrating 14 vegetation indices such as NDVI and MCARI, after dimensionality reduction by principal component analysis (PCA), 90% of the information was retained, reducing the feature dimension from the original 224 to 34 and improving the computational efficiency by 40%.

[0044] Step 4: To address the issue of insufficient sample size, a generative adversarial network model is constructed to generate synthetic samples, thereby increasing the sample size for chlorophyll, leaf nitrogen content, and leaf area index.

[0045] A Generative Adversarial Network (GAN) model consists of two parts: a generator and a discriminator. The generator produces data as close to real-world data as possible. It typically receives a random noise vector as input and transforms it into data with the same dimensions and distribution as real data. The discriminator distinguishes real data from the data generated by the generator, outputting the probability that the generated data is real. During training, the generator and discriminator engage in a game until their losses reach equilibrium, at which point training is considered complete. The training effectiveness can be assessed by examining the error between the generated and real data.

[0046] Generative adversarial networks ensure the synergy of physiological parameters in the synthesized samples, with a correlation coefficient r > 0.92 between the generated and measured data; dynamic feature optimization introduces an early stopping mechanism, with three consecutive R... 2The modeling process terminates upon no improvement, reducing modeling time from 25 minutes to 8 minutes; the multi-task learning framework achieves chlorophyll (R) enhancement by sharing the underlying feature extraction network and joint loss function. 2 =0.87), leaf nitrogen (R 2 =0.86), LAI(R) 2 Synchronous high-precision inversion (=0.89).

[0047] Step 5: Design a dynamic feature optimization algorithm to optimize feature combinations. The gradient boosting regressor is used as the basic model and the coefficient of determination is used as the evaluation index. Starting from the first feature, features are added one by one. If the coefficient of determination increases, the feature is retained; otherwise, it is discarded. After traversing all features, the optimal feature combination is determined. After determination, the gradient boosting regressor is used to establish a health parameter estimation model.

[0048] like Figures 2-4 As shown, using the gradient boosting regressor as the basic model and the coefficient of determination as the evaluation index, the importance ranking of the features of the optimized estimation models for chlorophyll, leaf nitrogen content, and leaf area index (after expansion) was obtained.

[0049] like Figure 5 As shown, the gradient boosting regressor is an ensemble learning method based on gradient boosting decision trees. It minimizes the loss function by iteratively training the decision tree model. Each tree is trained on the residuals of the previous tree, thereby gradually improving the model's prediction accuracy. The parameter combinations used in model construction are as follows:

[0050] n_estimators=500, learning_rate=0.05, max_depth=4, subsample=0.9, andom_state=42.

[0051] Step 6: Based on the measured value range, the three parameters of chlorophyll content, leaf nitrogen content and leaf area index are divided into 5 levels using the natural breakpoint method. The canopy hyperspectral characteristics of all camphor forests in the monitoring area are input into the health parameter estimation model. The model simultaneously outputs the predicted values ​​of chlorophyll content, leaf nitrogen content and leaf area index, and generates thematic maps of camphor forest health classification for the three parameters to monitor the health status of camphor forests in the whole area and locate abnormal plants.

[0052] like Figure 6 As shown, it is a thematic map of the health classification of camphor forests in the monitoring area, generated by the model output, based on three types of parameters;

[0053] The natural breakpoint method was used to divide the model prediction values ​​of three parameters—chlorophyll content, leaf nitrogen content, and leaf area index—into five levels: very low, relatively low, average, relatively high, and very high. Based on this, a thematic map of the health classification of camphor forests in the monitoring area was generated.

[0054] Therefore, this invention employs a method for urban camphor tree health monitoring based on UAV hyperspectral imaging. A UAV flying at a height of 100m is equipped with a 0.1m resolution hyperspectral camera to achieve detailed analysis of individual camphor trees. Atmospheric correction using laboratory calibration and empirical linear methods ensures data accuracy errors are controlled within ±5%. A generative adversarial network (GAN) framework is used, employing a generator-discriminator adversarial training mechanism to construct a high-fidelity synthetic sample generation system. This effectively solves the bottleneck problem of insufficient sample size in traditional monitoring methods, expanding the effective training samples and improving sample diversity. An intelligent feature selection algorithm based on recursive feature elimination is developed, introducing an early stopping mechanism to optimize computational efficiency. This system can automatically identify 67 key feature bands from 224 original hyperspectral bands, reducing feature dimensionality by 70% while ensuring modeling accuracy loss is controlled within 5%, resulting in an overall modeling efficiency improvement of 3.2 times. An innovative multi-task learning framework based on gradient boosting regression trees is constructed. By sharing feature extraction layers and designing joint loss functions, simultaneous and accurate estimation of chlorophyll content, leaf nitrogen content, and leaf area index is achieved.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging, characterized in that, Includes the following steps: Step 1: Under clear and windless conditions, use a drone equipped with a hyperspectral camera to acquire camphor forest canopy images with a spatial resolution of 0.1m. Based on laboratory calibration cloths with 20% and 40% reflectance, use the empirical linear method to perform atmospheric correction on the original images. Step 2: Collect leaf samples from the sunny side of the canopy of camphor trees in the field, distinguish between new leaves and old leaves, record the location of each sampled tree, and measure the chlorophyll content, leaf nitrogen content, and leaf area index of new and old leaves. Step 3: Based on the location of each sampled tree, extract the 3m×3m canopy area corresponding to the location of each sample tree from the atmospherically corrected canopy image, extract the band features, principal component features, texture features and vegetation index features of the hyperspectral data, and perform spatial matching with the ground data; Step 4: Construct a generative adversarial network model to generate synthetic samples and increase the number of samples for chlorophyll, leaf nitrogen content, and leaf area index; Step 5: Design a dynamic feature optimization algorithm to optimize feature combinations. Starting from the first feature, add features one by one. After traversing all features, determine the optimal feature combination. After determination, use a gradient boosting regressor to establish a health parameter estimation model. Step 6: Based on the measured value range, the chlorophyll content, leaf nitrogen content and leaf area index are classified using the natural breakpoint method. The canopy hyperspectral characteristics of all camphor forests in the monitoring area are input into the health parameter estimation model. The model simultaneously outputs the predicted values ​​of chlorophyll content, leaf nitrogen content and leaf area index, and generates a thematic map of camphor forest health classification based on the three parameters.

2. The method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 1, characterized in that: The hyperspectral camera carried by the drone in step one has a hyperspectral module covering a spectral range of 400-1000nm, a spectral resolution of 2.68nm, and a total of 224 bands.

3. The method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 2, characterized in that: In step two, the new and old leaves were distinguished by the color of the leaves and petioles. During the determination of the leaf area index, each camphor tree sample was measured three times, and the average value was used for subsequent experiments.

4. The method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 3, characterized in that: Step 3 extracts the band features of the hyperspectral data, which include 14 hyperspectral bands: 450nm, 550nm, 571nm, 615nm, 641nm, 650nm, 690nm, 706nm, 720nm, 728nm, 790nm, 800nm, 397nm, and 405nm.

5. The method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 4, characterized in that: Step 3 includes the first 10 principal components obtained from principal component analysis of the original 224 bands.

6. The method for urban camphor tree health monitoring based on UAV hyperspectral imaging according to claim 5, characterized in that: In step three, the texture features include six texture features: contrast, dissimilarity, homogeneity, autocorrelation matrix, energy, and correlation.

7. The method for urban camphor tree health monitoring based on UAV hyperspectral imaging according to claim 6, characterized in that: Step 3 includes 14 vegetation index features, including Normalized Difference Vegetation Index (NDVI), Corrected Chlorophyll Absorption and Reflectance Index (CDRI), Converted Chlorophyll Absorption and Reflectance Index (CDR), Chlorophyll Absorption and Reflectance Index (CDR), Chlorophyll Index, Red-edged Normalized Difference Index (NDDI), Proportional Vegetation Index (PPI), Differential Vegetation Index (DDI), Soil Regulation Index (SRI), Green Index, Normalized Difference Water Index (NDDI), Terrestrial Chlorophyll Index (TRI), Triangular Vegetation Index (TDI), and Photochemical Reflectance Index (PRRI).

8. The method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 7, characterized in that: In step four, the generative adversarial network model consists of a generator and a discriminator. During training, the generator and the discriminator engage in a game until their losses reach equilibrium, at which point training is complete. The training effect of the model is judged by examining the error between the generated data and the real data.

9. A method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 8, characterized in that: In step five, the gradient boosting regressor minimizes the loss function by iteratively training the decision tree model. Each tree is trained on the residual of the previous tree. The parameter combination used in the model construction is n_estimators=500, learning_rate=0.05, max_depth=4, subsample=0.9, an dom_state=42.

10. A method for monitoring the health of urban camphor trees based on UAV hyperspectral imaging according to claim 9, characterized in that: In step six, the natural breakpoint method is used to divide the model prediction values ​​of the three parameters, namely chlorophyll content, leaf nitrogen content and leaf area index, into five levels.

Images