Ground hyperspectral field greenhouse gas CO2 monitoring method
By processing ground-based hyperspectral data and using deep learning models, the uncertainty problem in monitoring CO2 emissions from farmland has been solved, achieving high-precision and low-cost CO2 gas monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies are insufficient for efficiently and cost-effectively monitoring CO2 emissions from farmland, especially the dynamic changes in soil respiration and root respiration. Furthermore, traditional methods suffer from uncertainties in time and space, leading to inaccurate carbon flux estimations.
By preprocessing and mathematical transformation of ground-based hyperspectral data, combined with significance testing and screening using multiple algorithms, characteristic spectral bands sensitive to CO2 are extracted, and a deep learning model is established for CO2 gas monitoring.
It enables precise monitoring of CO2 gas at the farmland scale, reduces equipment costs, and improves the continuity and accuracy of monitoring over time.
Smart Images
Figure CN121783868A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of hyperspectral remote sensing technology, and in particular relates to a ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields. Background Technology
[0002] Currently, atmospheric greenhouse gas monitoring methods mainly include spectroscopic methods, satellite remote sensing, and gas chromatography. However, these methods largely rely on imported equipment, which is expensive and requires high maintenance costs. Greenhouse gas satellites, on the other hand, have insufficient detection capabilities, short observation durations, and inherent uncertainties, posing challenges to the long-term stable capture of greenhouse gas concentrations. Furthermore, CO2 emissions in farmland ecosystems (primarily including soil respiration and root respiration) are complexly influenced by various environmental and anthropogenic factors such as soil temperature, humidity, vegetation type, farming practices, and organic fertilizer application. Their emission fluxes exhibit significant dynamic variations across diurnal and seasonal scales. Traditional chamber methods, through intermittent manual sampling and low measurement frequency (typically every few hours or even days), struggle to capture key peaks in diurnal variations (such as emissions during peak photosynthesis and soil respiration at midday) and cannot cover drastic changes caused by different weather events (such as before and after rainfall). This leads to significant uncertainties in the estimation of regional carbon fluxes, potentially resulting in a systematic underestimation or overestimation of carbon emissions. Therefore, research on high-precision, low-cost CO2 gas monitoring has become particularly important. Summary of the Invention
[0003] Technical Solution: To address the aforementioned technical problems, this invention preprocesses and performs spectral mathematical transformation on the collected ground hyperspectral data. Combined with indoor measured CO2 emission flux, correlation analysis with significance testing and spectral feature screening are conducted to obtain characteristic spectral bands sensitive to CO2. Specifically, a ground hyperspectral method for monitoring greenhouse gas CO2 in farmland is proposed. The method comprises the following steps: Step 1: Data Acquisition and Preprocessing Hyperspectral data and corresponding gas samples were collected from field sampling points. The hyperspectral data were preprocessed and noise was reduced to obtain the measured spectral data. The gas samples were analyzed and calculated in the laboratory to obtain the measured CO2 gas emission flux data. Step 2: Mathematical Transformation of the Spectrum Perform mathematical transformations on the spectral data collected and preprocessed in step one; Step 3: Extraction of Sensitive Bands By calculating the Pearson correlation coefficient rThe correlation between CO2 emission flux and different mathematical transformations in the 400-2400 nm band was obtained and significance tests were performed. Two statistical methods were combined to extract the sensitive band for CO2 emission flux, where the correlation coefficient... r The formula for calculating heat generation is: (1) In the formula, n For the sample size, x i The reflectivity of the band within the spectral range. y i This represents a sample value of CO2 gas emission flux. The mean reflectance of the sample. This represents the average CO2 gas emission flux for the sample.
[0004] Step 4: Screening of characteristic bands Elimination of uninformed variables UVE Continuous projection SPA Least angular regression LARS Genetic Algorithm GA Competitive adaptive reweighted sampling CARS Five algorithms were used to perform in-depth screening of the sensitive bands extracted in step three to obtain the feature bands with the lowest threshold for modeling effect. The feature bands selected by different algorithms were then used to model and analyze CO2 emission flux, and the goodness of fit R of the modeled predicted values was measured. 2 The closer the value is to 1, the better the modeling effect of the characteristic bands and the more effective the screening algorithm.
[0005] Step 5: Construction of the CO2 gas monitoring model Using the spectral data obtained from the mathematical transformation in step two and the characteristic bands obtained in step four as independent variables, and the CO2 gas emission flux corresponding to each sample as the dependent variable, the training set and validation set are divided according to the proportion of 25%-45% of the data. A CO2 gas monitoring model is established through a deep learning algorithm, and the results of CO2 gas monitoring obtained by different processing methods for the characteristic bands are output.
[0006] As an improvement, the specific steps in step one are as follows: Acquiring hyperspectral data: Spectral data acquisition and usage ASD FieldSpec 4. On-site spectrometer, wherein the fiber optic probe of the analyzer is placed vertically 40-55cm above the sampling point, and 4-10 spectral curves are collected at each sampling point, with a spectral range of 350nm to 2500nm and an interval of 1-1.5nm; CO2 gas data collection: After the spectral data collection is completed, a transparent collection device is placed above each sampling point. The syringe collects air gas samples from 0 to 30 minutes, with a total of 4 air gas samples at each time point, which are then placed in an opaque gas bag. The collected gas samples are then analyzed and measured in a gas chromatograph in the laboratory, and the CO2 gas emission flux is calculated.
[0007] As an improvement, the mathematical transformation of the spectrum in step two includes three steps, specifically: Savitzky-Golay Polynomial smoothing, first-order differential D1, second-order differential D2.
[0008] As an improvement, in step three, the correlation processing is performed by setting a significance level of P=0.01 to extract the sensitive bands that pass the 99% significance test in the correlation analysis.
[0009] As an improvement, the deep learning algorithm in step five is at least one of Random Forest (RF) or Generalized Additive Model (GAM).
[0010] As an improvement, the accuracy evaluation of the estimation model is also included. The specific steps of the evaluation process are as follows: when constructing the crop phenotypic parameter estimation model, the commonly used coefficient of determination R is calculated. 2 The root mean square error (RMSE) is used to evaluate model performance. 2 The closer the value is to 1 and the smaller the RMSE, the better the model performs. (2) (3) in n It is the number of samples. y i It is the first i The true value of each sample It is the first i Estimates for each sample It is the average of the actual observed values.
[0011] Beneficial effects: This invention proposes a ground-based hyperspectral method for monitoring CO2 greenhouse gases in open fields, which involves preprocessing and... SG, D1, D2 Three spectral mathematical transformations, combined with indoor measured CO2 emission fluxes, were analyzed using correlation analysis with significance tests, and... UVE , SPA , LARS , GA , CARS Five screening algorithms were used to select characteristic spectral bands sensitive to CO2, which were then used to establish a CO2 monitoring model, ultimately achieving accurate monitoring of CO2 gas at the farmland scale.
[0012] In addition, compared with conventional monitoring methods, this invention predicts CO2 emissions by modeling the CO2 emission flux through outdoor hyperspectral data collection and indoor measurement. This can reduce the problems of high cost, discontinuous monitoring time, and low accuracy of previous CO2 monitoring equipment, and achieve accurate results through simple operation. Attached Figure Description
[0013] Figure 1 This is a flowchart of the present invention.
[0014] Figure 2 Figure (a) shows the effect of spectral mathematical transformation in Embodiment 1 of the present invention; Figure (b) shows the original spectral curve; Figure (c) shows the transformation of the original spectral curve. Savitzky-Golay The curve after polynomial smoothing smooths certain values; Figures (c) and (d) show the D1 and D2 transformations performed on the basis of the smoothing in Figure (b).
[0015] Figure 3 This is a visualization of the Pearson correlation with a significance test set to P=0.01 in Embodiment 1 of the present invention.
[0016] Figure 4 In Embodiment 1 of the present invention UVE , SPA , LARS , GA , CARS Screening results of five algorithms.
[0017] Figure 5 This is a graph showing the CO2 prediction results in Embodiment 1 of the present invention. Detailed Implementation
[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0019] The ground-based hyperspectral method for monitoring greenhouse gas CO2 in the field in this invention Step 1: Data Acquisition and Preprocessing use ASD FieldSpec 4. The on-site spectrometer performs hyperspectral acquisition at the sampling point. The fiber optic probe is placed vertically 40-50cm above the sampling point. Five spectral curves are acquired at each sampling point, with a spectral band range of 350nm to 2500nm and an interval of 1-1.5nm.
[0020] Preferably, in ViewSpecPro The software performs mean processing on the five spectral curves for each sampling point and deletes spectral data in the bands of 300–399 nm and 2401 nm–2500 nm to avoid the influence of the spectral bands of the absorption peaks of the spectral curves caused by atmospheric moisture absorption on the results.
[0021] After the spectral data acquisition was completed, a transparent acquisition device was placed above each sampling point. Four air gas samples were collected using a syringe from 0 to 30 minutes, with each time point containing four samples, and these samples were placed in an opaque gas bag. The collected gas samples were analyzed and measured using a gas chromatograph in the laboratory, and the CO2 emission flux was calculated using the following formula.
[0022] (4) In the formula: F CO2 emission flux (mg / m³) 2 / h); ρ This represents the density of CO2 gas under standard conditions. dc / dt The slope of the regression curve of gas concentration versus time in the static chamber; P 0 The air pressure under standard conditions ( kPa T represents the average temperature (°C) of the air inside the static chamber during sampling. P atmospheric pressure ( kPa ); H The height of the air chamber inside the box ( cm ).
[0023] Step 2: Mathematical Transformation of the Spectrum exist Python In this process, hyperspectral data is imported, and spectral calculations are performed. Savitzky-Golay The program performs three mathematical transformations: polynomial smoothing, first-order differential D1, and second-order differential D2. The results are then output and plotted.
[0024] Step 3: Extraction of Sensitive Bands The mathematically transformed data obtained in step two and the CO2 gas emission flux data calculated in step one are combined, based on... Python China Pearson The correlation index is calculated using a method that extracts the relevant correlation index. P (Significance level) <= 0.01, i.e., bands that pass the 99% significance test, are considered bands sensitive to CO2 emission flux, and the extraction results are visualized.
[0025] Step 4: Feature Band Screening exist Python In the process of constructing the elimination of uninformed variables UVE Continuous projectionSPA Least angular regression LARS Genetic Algorithm GA Competitive adaptive reweighted sampling CARS The formulas for five algorithms are used to perform deep algorithm screening on the sensitive bands extracted in step three. To achieve stable screening results, each algorithm is run repeatedly, preferably 4-10 times. The different screening effects of each preprocessing method are also presented.
[0026] Step 5: Construction of the CO2 gas monitoring model Using various spectral data transformations and special screening algorithms to obtain various characteristic bands as independent variables, and the CO2 gas emission flux corresponding to each sample as the dependent variable, the training set and validation set are divided according to a data ratio of 25%-45%. A deep learning algorithm, preferably random forest, is then applied. RF and generalized additive model GAM Two machine learning algorithms were used to establish a CO2 gas monitoring model. Five-fold cross-validation was added during the modeling process to find the optimal modeling parameters. Finally, the characteristic bands obtained by different processing methods were output as CO2 gas monitoring results, which are inversion combinations with a certain accuracy. Example
[0027] use ASD FieldSpec 4. The on-site spectrometer performs hyperspectral acquisition at the sampling points. The fiber optic probe is placed vertically 50 cm above the sampling point. Five spectral curves are acquired at each sampling point, with a spectral range of 350 nm to 2500 nm and an interval of 1 nm. ViewSpecPro The software performs mean processing on the five spectral curves for each sampling point and deletes spectral data in the band ranges of 300–399 nm and 2401 nm–2500 nm.
[0028] After spectral data acquisition, a transparent acquisition device was placed above each sampling point. Four air gas samples were collected using a syringe over a period of 0–30 minutes, with each time point containing four samples, and these samples were placed in an opaque gas bag. The collected gas samples were analyzed and measured using a gas chromatograph in the laboratory, and the CO2 emission flux was calculated.
[0029] The calculation of CO2 emission flux is performed using Excel, and the process is shown in Table 1 below.
[0030] Table 1 Calculated values of CO2 gas emission flux
[0031] Note: The slope of the regression curve of gas concentration versus time in the static chamber in Table 1. dc / dt -0.1604 。
[0032] Tablex The time points are 0-30 minutes. y The values represent the CO2 gas concentrations measured at each time point. ρ, H, T These are the density of CO2 gas under standard conditions, the average sampling temperature, and the height of the sampling device, obtained through actual measurements. F for Excel The CO2 emission flux for the corresponding sample is calculated by importing the emission flux formula.
[0033] exist Python In this process, hyperspectral data is imported, and spectral calculations are performed. Savitzky-Golay The program performs three mathematical transformations: polynomial smoothing, first-order differential D1, and second-order differential D2. The results are then output and plotted.
[0034] See Figure 2 The results are shown below. From the figure, (a) represents the spectral curve of the original sample, and (b) represents the spectral curve of the processed sample. SG The smoothed spectral curve shows that the original spectral curve is more stable. SG The smoothing effect is not very significant. (c) represents the spectral curve after D1 transformation based on (b), showing obvious peaks and troughs representing spectral variations. (d) represents the spectral curve after D2 transformation based on (b), showing more pronounced wavelet troughs compared to (c), indicating more prominent spectral features. The purpose of these transformations is to extract the variation characteristics of the original spectral curve, specifically the changes in reflectance across different wavebands.
[0035] like Figure 3 As shown, the broken line represents the magnitude of the correlation coefficient, the black dashed line represents the confidence level of p=0.01, and the scatter plot below the graph shows the bands that passed the significance test. It can be seen that: SG The reflectance of the transformed spectral curve bands is negatively correlated with CO2 emission flux (a), with a low correlation coefficient, mainly concentrated between -0.2 and -0.4. Even if some bands pass the significance analysis, they are difficult to be statistically significant.
[0036] Furthermore, the correlation between the band reflectance of the spectral curve after D1 transformation and CO2 emission flux shows alternating positive and negative values (b), with a relatively high correlation coefficient, mainly concentrated between -0.5 and 0.5. Several sensitive bands with varying degrees of positive and negative correlation were extracted, demonstrating the initial effectiveness of the spectral mathematical transformation. The correlation between the band reflectance of the spectral curve after D2 transformation and CO2 emission flux shows even more frequent alternations of positive and negative values (c), with a better correlation coefficient than (b), mainly concentrated between -0.55 and 0.6. A considerable number of sensitive bands were extracted, indicating a significant effect of the spectral mathematical transformation.
[0037] like Figure 4As shown, the scatter plot represents the selected bands, and the ridge pattern behind it represents the density of selected bands within a band interval. This is achieved by constructing... UVE (Elimination of uninformed variables) SPA (Continuous projection) LARS (Least Angle Regression) GA (Genetic Algorithm) CARS The formulas for five algorithms (competitive adaptive reweighted sampling) are used to further filter the extracted sensitive bands. It can be seen that, from the perspective of different mathematical transformations of the spectrum, based on correlation analysis with significance testing, if only... SG Even after smoothing the spectral data and performing feature band filtering (a), a large number of spectral bands are still retained, making it difficult to effectively reduce the dimensionality of the spectral data; SG +D1 and SG The +D2 processed spectral data underwent feature band filtering ((b) and (c)), effectively reducing the data dimensionality and filtering out spectral bands that are mostly distributed within the visible light range (400nm-760nm). From the perspective of different filtering algorithms, UVE It exhibits significant instability due to its applicability to linear relationships. SPA and LARS A certain number of characteristic bands can be stably selected through algorithm settings. GA and CARS Compared to other algorithms, this algorithm can filter out more feature bands.
[0038] Using various characteristic bands obtained from different spectral data transformations and special screening algorithms as independent variables, and the CO2 gas emission flux corresponding to each sample as the dependent variable, the training and validation sets were divided according to a 30% data ratio. Random forest was then used to... RF and generalized additive model GAM Two machine learning algorithms were used to build a CO2 gas monitoring model. Five-fold cross-validation was incorporated into the modeling process to find the optimal modeling parameters. The final output showed the CO2 gas monitoring results obtained from different processing methods for characteristic bands, and a relatively accurate inversion combination was achieved. Figure 5 As shown, we can conclude that: SG +D2+ SPA The combination of modeling methods can achieve the most effective inversion results, with the highest accuracy on the test set reaching 0.614.
[0039] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields, characterized in that, The steps of the method are as follows: Step 1: Data Acquisition and Preprocessing Hyperspectral data and corresponding gas samples were collected from field sampling points. The hyperspectral data were preprocessed and noise was reduced to obtain the measured spectral data. The gas samples were analyzed and calculated in the laboratory to obtain the measured CO2 gas emission flux data. Step 2: Mathematical Transformation of the Spectrum Perform mathematical transformations on the spectral data collected and preprocessed in step one; Step 3: Extraction of Sensitive Bands By calculating the Pearson correlation coefficient r The correlation between CO2 emission flux and different mathematical transformations in the 400-2400 nm band was obtained and significance tests were performed. Two statistical methods were combined to extract the sensitive band for CO2 emission flux, where the correlation coefficient... r The calculation formula is: (1) In the formula, n For the sample size, x i The reflectivity of the band within the spectral range. y i This represents a sample value of CO2 gas emission flux. The mean reflectance of the sample. This represents the average CO2 gas emission flux for the sample. Step 4: Screening of characteristic bands Elimination of uninformed variables UVE Continuous projection SPA Least angular regression LARS Genetic Algorithm GA Competitive adaptive reweighted sampling CARS Five algorithms were used to perform in-depth screening of the sensitive bands extracted in step three to obtain the feature bands with the lowest modeling effect threshold. The feature bands selected by different algorithms were then used to model and analyze CO2 emission flux, and the goodness of fit R of the modeled predicted values was measured. 2 The closer it is to 1, the better the modeling effect of the characteristic bands and the more effective the screening algorithm. Step 5: Construction of the CO2 gas monitoring model Using the spectral data obtained from the mathematical transformation in step two and the characteristic bands obtained in step four as independent variables, and the CO2 gas emission flux corresponding to each sample as the dependent variable, the training set and validation set are divided according to the proportion of 25%-45% of the data. A CO2 gas monitoring model is established through a deep learning algorithm, and the results of CO2 gas monitoring obtained by different processing methods for the characteristic bands are output.
2. The ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields according to claim 1, characterized in that, The specific steps in step one are as follows: Hyperspectral data acquisition: Spectral data acquisition was performed using an ASD FieldSpec 4 field spectrometer. The fiber optic probe of the analyzer was placed vertically 40-55 cm above the sampling point. 4-10 spectral curves were acquired at each sampling point, with a spectral band range of 350 nm to 2500 nm and an interval of 1-1.5 nm. CO2 gas data collection: After the spectral data collection is completed, a transparent collection device is placed above each sampling point. The syringe collects air gas samples from 0 to 30 minutes, with a total of 4 air gas samples at each time point, which are then placed in an opaque gas bag. The collected gas samples are then analyzed and measured in a gas chromatograph in the laboratory, and the CO2 gas emission flux is calculated.
3. The ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields according to claim 1, characterized in that, Step two involves three mathematical transformations of the spectrum, specifically: Savitzky-Golay Polynomial smoothing, first-order differential D1, second-order differential D2.
4. The ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields according to claim 1, characterized in that, In step three, correlation processing is performed by setting a significance level of P=0.01 to extract sensitive bands that pass the 99% significance test in the correlation analysis.
5. The ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields according to claim 1, characterized in that, In step five, the deep learning algorithm is at least one of Random Forest (RF) or Generalized Additive Model (GAM).
6. The ground-based hyperspectral method for monitoring greenhouse gas CO2 in open fields according to claim 1, characterized in that, It also includes the accuracy evaluation of the estimation model. The specific steps of the evaluation process are as follows: when constructing the crop phenotypic parameter estimation model, the commonly used coefficient of determination R is calculated. 2 The root mean square error (RMSE) is used to evaluate model performance. 2 The closer the value is to 1 and the smaller the RMSE, the better the model performs; （2） （3） in n It is the number of samples. y i It is the first i The true value of each sample It is the first i Estimates for each sample It is the average of the actual observed values.