Prediction model and application of carbon emission of dairy cattle

Abstract

This invention belongs to the field of dairy cow performance measurement and dairy cow greenhouse gas emission monitoring, specifically disclosing a predictive model and application for carbon emissions from the nose of dairy cows. Regarding the selection of characteristic bands, it breaks away from the commonly used method of selecting characteristic bands using algorithms, instead employing a method of manual selection and multiple iterations. In particular, it confirms that incorporating water absorption and near-infrared regions, which are discarded in conventional milk component analysis modeling, can improve model performance. Ultimately, the characteristic bands for predicting Holstein dairy cow nose carbon emissions (carbon dioxide emission equivalent) were selected. The optimal preprocessing and algorithm combination corresponding to the selection of the optimal model for predicting Holstein dairy cow nose carbon emissions (carbon dioxide emission equivalent) was determined, the optimal parameters were identified, and the accuracy of the model was improved.

Description

Technical Field

[0001] This invention belongs to the field of dairy cow performance testing and dairy cow greenhouse gas emission monitoring, specifically involving the prediction model and application of carbon emission equivalent dairy cow oral and nasal carbon emissions (carbon dioxide emission equivalent). Background Technology

[0002] To more objectively assess the contribution of dairy cows to climate change, methane emissions are typically quantified in carbon dioxide equivalents (CO2-eq). While numerous methods and technologies exist for measuring individual greenhouse gas emissions from dairy cows, such as respiratory chambers (RC), gas tracer techniques, sniffers, and GreenFeed emission monitoring (GEM) systems, they generally suffer from limitations such as high cost, low throughput, and a limited range of gases measured, hindering large-scale application. Modeling to predict greenhouse gas emissions from dairy cows is a viable alternative, but indicators such as body weight, energy intake, and caloric output—essential for CO2 prediction models—and dry matter intake (DMI), gross energy intake (GEI), and dietary fiber content—essential for methane prediction models, are difficult to collect on commercial farms, limiting the large-scale usability of such models. Therefore, using low-cost and routinely recorded traits as predictor variables in model building may be a more practical choice.

[0003] When a substance absorbs infrared light in the 2.5μm-25μm range, the chemical bonds or functional groups in the molecule undergo specific vibrations or rotations, resulting in characteristic absorption peaks. Mid-infrared spectroscopy (MIRS) analysis technology utilizes the interaction characteristics between mid-infrared light and matter to analyze and determine the structure and composition of substances, thereby obtaining rich information on the chemical components in the sample. It features high sensitivity, high throughput, simplicity, and non-destructive operation. MIRS of dairy cow milk can be obtained in large quantities, at low cost, and efficiently from routine DHI measurements in dairy cows. The technique of predictive analysis using milk MIRS has been routinely used for the quantitative or qualitative analysis of milk components, and in recent years it has been increasingly used to predict phenotypic information such as health diseases and physiological conditions in dairy cows.

[0004] Currently, many international teams are exploring the feasibility of predicting greenhouse gas emissions from dairy cows using mid-infrared spectroscopy (MIRS) of milk. For example, it has been confirmed that predicting methane emissions from dairy cows based on mid-infrared spectroscopy of milk is feasible, with strong biological rationality and moderate predictive accuracy. However, current research focuses more on establishing predictive models for methane emissions from the mouth and nose of dairy cows, with relatively few studies on predicting the carbon emission equivalent from the mouth and nose. Furthermore, the accuracy of these models is not high, and the selection of wavelengths used in their development is limited to those used in conventional milk component analysis, often neglecting the near-infrared and water absorption regions of milk MIRS. There is still much room for improvement. Therefore, there is an urgent need to establish a predictive model for carbon dioxide emission equivalent from the mouth and nose of Holstein dairy cows to quickly understand the greenhouse gas emissions of large numbers of individual dairy cows, improve the breeding efficiency of low-carbon dairy cows, and thus not only mitigate greenhouse gas emissions from dairy farming but also ensure milk production and reduce feed waste.

[0005] Therefore, the purpose of this invention is to establish a rapid batch detection technology for the carbon emissions (carbon dioxide emission equivalent) of Holstein dairy cows based on milk MIRS, improve the efficiency of low-carbon dairy cow breeding, and provide technical support for the low-carbon, healthy, and sustainable development of the dairy industry. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid batch detection method for carbon emissions from the mouth and nose of dairy cows. The method is simple, fast, and has high accuracy compared with the true value.

[0007] Another objective of this invention is to provide an application of a rapid batch detection method for carbon emissions from the mouth and nose of dairy cows.

[0008] To achieve the above objectives, the present invention adopts the following technical measures:

[0009] A rapid batch detection method for carbon emissions from the mouth and nose of dairy cows includes the following steps:

[0010] 1. The characteristic wavelength band of the infrared spectrum collected from the milk sample is 1103.39 cm⁻¹. -1 -1180.55cm -1 1562.49cm -1 -1604.93cm -1 1689.80cm -1 -1716.81cm -1 2044.74cm -1 -2133.47cm -1 2195.20cm -1 -2426.68cm -1 2434.40cm -1 -2573.29cm -12604.15cm -1 -2808.62cm -1 3109.55cm -1 -3225.29cm -1 3279.30cm -1 -3422.05cm -1 3672.82cm -1 -3830.99cm -1 3966.02cm -1 -4008.46cm -1 4016.18cm -1 -4074.05cm -1 4413.55cm -1 -4594.88cm -1 4641.17cm -1 -4776.20cm -1 4810.93cm -1 -5007.68cm -1 MIR data in the middle;

[0011] 2. Substitute the measured MIR data into the constructed model of second derivative + multivariate scattering correction + partial least squares regression to output the predicted carbon emissions from the mouth and nose of dairy cows.

[0012] The principal component of the partial least squares regression is 9.

[0013] Preferably, in the method described above, each segment of the band is allowed to have a difference of two wave points before and after.

[0014] Preferably, the milk described in the above method is milk produced by Holstein cows.

[0015] Application of a rapid batch detection method for carbon emissions from the mouth and nose of dairy cows in dairy cow performance testing.

[0016] Compared with the prior art, the advantages of the present invention are as follows:

[0017] 1. Regarding the selection of feature bands, we broke away from the commonly used method of using algorithms to select feature bands. Instead, we used a method of manual selection and multiple iterations. In particular, we confirmed that adding water absorption regions and near-infrared regions that were discarded in conventional milk component analysis modeling can improve the performance of the model. Finally, we selected the feature band of Holstein carbon emission equivalent dairy cow mucus carbon emission (carbon dioxide emission equivalent).

[0018] 2. The optimal preprocessing and algorithm combination for establishing the optimal model for predicting carbon emission equivalent (carbon dioxide emission equivalent) of dairy cows was selected, the optimal parameters were determined, and the accuracy of the model was improved.

[0019] 3. It enables rapid, accurate, low-cost, and high-throughput monitoring of dairy cow carbon emission equivalents. It only takes 10-15 seconds to obtain milk MIRS and milk composition information. After importing the data into the model, the carbon emission equivalent (carbon dioxide emission equivalent) of dairy cows can be directly obtained, which greatly improves the monitoring efficiency and has strong practicality. It can be widely used in dairy cow performance testing and dairy cow greenhouse gas emission monitoring. Attached Figure Description

[0020] Figure 1 This is the original mid-infrared spectrum of the milk sample.

[0021] Figure 2 This is a general spectral diagram of the 15 selected characteristic bands.

[0022] Figure 3 A magnified set of images for each characteristic band.

[0023] Figure 4 The correlation between the true and predicted values ​​and the fitted straight line plots for the model training and test sets are shown. Detailed Implementation

[0024] Unless otherwise specified, the technical solutions described in this invention are all conventional solutions in the field; unless otherwise specified, the reagents or materials described are all from commercial sources.

[0025] 1. Experimental Materials

[0026] Greenhouse gas monitoring of dairy cows is costly, difficult, and time-consuming, making it challenging to obtain large sample sizes and data. This study used the GreenFeed system (C-Lock Inc. GF384, USA) to conduct greenhouse gas monitoring on four batches of dairy cows at a dairy farm in Hohhot, Inner Mongolia, my country. A total of 76 healthy Holstein cows participated in the monitoring, representing cows of different parities and lactation stages to ensure data representativeness and diversity. Milk produced by the cows was collected during the monitoring period and sent to the local DHI center for mid-infrared spectral data collection.

[0027] Sample Information Statistics Table

[0028] number of cows lactation Measurement cycle 20 Mid-term 3 20 Early stage 3 18 Later 3 18 Mid-term 1

[0029] 2. Method for determining the true (reference) level of greenhouse gas emissions from the mouth and nose of individual dairy cows.

[0030] The GreenFeed system (C-Lock Inc. GF384, USA) was selected to monitor the carbon dioxide production (CDP, g / d) and methane production (MEP, g / d) of dairy cows. The air filter of the device was replaced weekly, and a CO2 gas recovery test was conducted monthly. Each measurement was conducted for at least 2 minutes per cow. The raw data was calibrated and processed by the C-Lock data team and then exported from https: / / greenfeed.c-lockinc.com.

[0031] Considering the variations in methane and carbon dioxide emissions from dairy cows throughout the day, a special design was employed to ensure that greenhouse gas emission levels at different times were obtained from the participating cows as comprehensively as possible. Measurements were conducted in 8-day cycles, with each batch of cows measured for 3 cycles (except for the fourth batch due to weather conditions). Typically, the first day of a cycle began measurement at 0:00 AM, 8:00 AM, and 4:00 PM; the second day at 1:00 AM, 9:00 AM, and 5:00 PM, and so on, ultimately covering the 8-day measurement period with a 24-hour measurement cycle. Ideally, each cow should have 24 measurement records at different times per cycle. At least 20 valid measurement records per cow per cycle were guaranteed, and the arithmetic mean was used as the final data; data not meeting this condition were discarded.

[0032] 3. Milk collection and mid-infrared spectroscopy determination

[0033] The dairy cows were milked three times daily, approximately at 3:00 AM, 11:00 AM, and 7:00 PM, using automated milking equipment from the Swedish company DeLaval, and milk yield was recorded. Milk samples were collected throughout the milking process during methane emission monitoring, with 40 mL collected from each cow per milking cycle. The samples were dispensed into cylindrical sampling tubes (3.5 cm in diameter and 9 cm in height), numbered sequentially, and bromonitrile glycol preservative was immediately added to each tube and gently shaken to dissolve completely. The milk samples were stored at 4°C to prevent spoilage before being transported to the DHI center. Spectroscopic acquisition was performed immediately upon arrival at the DHI center.

[0034] Upon arrival at the DHI testing center, the sample tube containing the milk sample was placed in a 42°C water bath for 15-20 minutes. The sample was then analyzed using a high-throughput milk somatic cell analyzer (COMBIFOSS 7DC) from FOSS. After shaking the sample well, it was sent into the instrument for scanning to obtain milk composition information and MIRS. Figure 1 This is the original mid-infrared spectrum of the milk sample.

[0035] 4. Data Processing

[0036] All milk samples collected during the monitoring period were discarded, and invalid data caused by sample deterioration, empty tubes (no samples collected), abnormal milk composition measurements (only records of measurements with 1.5% < milk fat percentage < 9%, 1% < milk protein percentage < 7%, and somatic cell count < 1000K were retained), and abnormal mid-infrared spectroscopy measurements were removed. Subsequently, the retained milk composition records and MIRS of the single cycle were weighted and averaged according to the milk yield of each milking, and then the average MIRS of the participating cows for that cycle was calculated.

[0037] According to the experimental design, ideally, 76 participating dairy cows should have generated 192 daily average carbon dioxide emission equivalent records for the measurement period. However, after removing outliers, only 129 valid data points from 66 participating cows were retained (the test set and training set of Example 1 totaled 124, with five data points reserved for external validation) for model building, optimization, and validation. First, cows that withdrew due to illness during the monitoring period were removed. Then, only cows with at least 20 measurement records per period were retained to ensure the accuracy of the measurement results.

[0038] Finally, the average of all measurement records within a single period is used as the daily average carbon dioxide emission production (CDP) and daily average methane emission production (MEP) of the tested cattle for that period. Then, the daily average carbon emission equivalent of the tested cattle for the measurement period is calculated using the formula: TotalCO2 equivalent (g / d) = CDP (g / d) + MEP (g / d) * 27 (where "27" represents the global warming potential of methane gas from the IPCC Sixth Assessment Report on Climate Change). The oral and nasal carbon dioxide emission equivalents of dairy cows monitored by the GreenFeed system described in this invention are all calculated using this formula.

[0039] Example 1:

[0040] Establishment of a predictive model for carbon emissions (carbon dioxide emission equivalent) from the mouth and nose of dairy cows:

[0041] 1. The GreenFeed system monitors the oral and nasal carbon dioxide emission equivalents of dairy cows and divides the modeling dataset as follows:

[0042]

[0043] In this embodiment, the modeling dataset is divided into a training set (75%) and a test set (25%). The ratio of the training set to the test set is 3:1.

[0044] 2. Selection of Preprocessing Methods for Modeling MIRS Data

[0045] This invention uses the partial least squares regression (PLSR) algorithm to build the model.

[0046] Effective feature selection is a fundamental operation in processing spectral data, aiming to reduce noise and lay the foundation for feature extraction. Effective feature selection mainly includes three types: feature extraction, feature preprocessing, and feature dimensionality reduction. This embodiment primarily focuses on feature preprocessing of the spectral data. During modeling, sometimes a single preprocessing step can limit the model's generalization ability and optimization potential; ensemble preprocessing has been proven to produce better results during model building. Therefore, before partitioning the dataset, this embodiment selects two preprocessing methods from 12 options—None (no preprocessing), MMS (Maximum-Minimum Normalization), SS (Standardization), CT (Mean Centering), MA (Moving Average Smoothing), SG (Convolutional Smoothing), MSC (Multivariate Scatter Correction), DT (Trend Correction), SNV (Standard Normal Transform), wave (Wavelet Transform), D1 (First Derivative), and D2 (Second Derivative)—for ensemble preprocessing of the spectral data. Comparative analysis of single preprocessing revealed a significant performance difference between the training and test sets when using the original spectra for modeling, indicating overfitting. While the D2 preprocessed spectra showed the best prediction performance on the training set, and overfitting was not improved, this study will employ an ensemble preprocessing approach. Combining the D2 preprocessing method with other preprocessing methods offers greater potential for improvement and optimization of the test set results. Therefore, this embodiment chooses to use D2 preprocessing as the basis for ensemble preprocessing.

[0047] Comparative analysis of integrated preprocessing methods based on D2 preprocessing revealed that integrating D2 preprocessing with SNV, MSC, and D2 methods all improved the prediction performance on the test set and reduced model overfitting. Among these, the spectral modeling using the D2+MSC integrated preprocessing showed the best improvement in overfitting, and due to the subsequent selection of characteristic bands, this model has greater potential for performance improvement and optimization. Therefore, this embodiment selects the D2+MSC integrated preprocessing.

[0048] 3. Manual selection and determination of modeling characteristic bands

[0049] There are many methods for selecting feature bands, mainly including algorithmic and manual selection. Algorithmic feature selection is based on the correlation between each wave point and a reference value. Its advantages are speed and efficiency, but it neglects the synergistic effect between adjacent wave points and has a relatively simplistic approach. Manual feature selection, on the other hand, strengthens the role of wave bands (i.e., adjacent wave points) during the selection process, and retains more of the original spectral information during model improvement. It has stronger inclusiveness and generalization ability, and the selected wave bands are more accurate. However, its disadvantages are slow selection speed and low efficiency.

[0050] In this embodiment, the characteristic bands are selected manually, and the steps are as follows:

[0051] (1) Determine the modeling algorithm. In this embodiment, the classic partial least squares regression algorithm is selected as the algorithm for predicting the equivalent carbon dioxide emissions from the mouth and nose of dairy cows.

[0052] (2) Determine the optimal preprocessing. In this embodiment, the optimal preprocessing combination selected before splitting the dataset is the D2+MSC preprocessing combination.

[0053] (3) Manually select the characteristic band. The specific process is as follows:

[0054] 1) The obtained 925.92cm -1 -5011.54cm -1 The spectral region is divided into 11 segments, including 925.92 cm⁻¹. -1 -4780.06cm -1 The average length is divided into 10 segments, with 100 wave points per segment, and the length is greater than 4780.06 cm. -1 This is the last segment, containing 60 polka dots.

[0055] 2) Using the partial least squares regression algorithm, firstly, add or subtract a set of wave points (0 to 100 wave points) at both ends of the critical point of the first wave segment. Select the optimal number of wave points to add or subtract based on the model effect, and then perform similar operations on the second wave segment based on this. Finally, after completing one round of operations on all eleven wave segments, the first traversal is considered complete.

[0056] 3) After the first traversal is completed, perform a second, third or more manual traversals until all the wave points no longer change, which is the optimal characteristic wave band.

[0057] 4) If adding or subtracting polka dots at both ends of a band does not improve the model's performance and the band is long, consider dividing the band in half and adding or subtracting polka dots again from the two ends of the dividing point.

[0058] After multiple rounds of selection, the optimal results were obtained, as shown in the table below:

[0059]

[0060] A total of 15 characteristic bands were selected, and the result was: 1103.39cm. -1 -1180.55cm -1 1562.49cm -1 -1604.93cm -1 1689.80cm -1 -1716.81cm -1 2044.74cm -1 -2133.47cm -12195.20cm -1 -2426.68cm -1 2434.40cm -1 -2573.29cm -1 2604.15cm -1 -2808.62cm -1 3109.55cm -1 -3225.29cm -1 3279.30cm -1 -3422.05cm -1 3672.82cm -1 -3830.99cm -1 3966.02cm -1 -4008.46cm -1 4016.18cm -1 -4074.05cm -1 4413.55cm -1 -4594.88cm -1 4641.17cm -1 -4776.20cm -1 4810.93cm -1 -5007.68cm -1 Each segment is allowed a gap of two polka dots. Figure 2 This is the overall spectral plot of the 15 selected characteristic bands. Figure 3 A magnified set of images for each characteristic band.

[0061] 4. Selection and determination of model parameters

[0062] Model parameters typically include parameters from the preprocessing method and parameters from the modeling algorithm. Since this model uses the D2+MSC ensemble preprocessing method and the PLSR modeling algorithm, only the algorithm parameters need to be selected. The parameters for the partial least squares regression algorithm are the principal components (n_components). The parameter selection results are compared below:

[0063]

[0064]

[0065] It can be observed that when the principal component (n_component) is 9, the model performs best on the test set and does not overfit. Therefore, the principal component (n_component) is ultimately chosen to be 9.

[0066] In summary, after comparative analysis, the optimal regression model for cow snout carbon emissions (carbon dioxide emission equivalent) is the D2+MSC+PLSR (n_component=9) model. The correlation coefficients between the training and test sets are 0.85 and 0.84, respectively; the root mean square errors between the training and test sets are 1453.45 and 1489.55, respectively. Figure 4 The correlation between the true and predicted values ​​and the fitted straight line plots for the model training and test sets are shown.

[0067] Example 2:

[0068] Application of predictive models for carbon emissions (CO2 equivalent) from the mouth and nose of dairy cows:

[0069] The optimal regression model (D2+MSC+PLSR(n_component=9)) of the established prediction model for carbon emissions from the mouth and nose of dairy cows was used to predict five randomly selected samples (not one of the 124 modeling samples), and the prediction results were compared with the actual values.

[0070] How to use the model:

[0071] 1. Collect MIRS data of 15 characteristic bands (selected in Example 1) from the infrared spectrum of milk samples.

[0072] 2. Substitute the obtained MIRS data into the D2+MSC+PLSR (n_component=9) model constructed in Example 2 to output the predicted results of carbon emissions from the mouth and nose of dairy cows.

[0073] The model's predictions are very close to the actual results of monitoring dairy cow nasal carbon emissions (CO2 equivalent) using the GreenFeed system (C-Lock Inc. GF384, USA) (see table below). Therefore, the model has high accuracy and can be used to predict dairy cow nasal carbon emissions.

[0074]

Claims

1. A rapid batch detection method for carbon emissions from the mouth and nose of dairy cows, comprising the following steps: 1) The characteristic wavelength of the infrared spectrum collected from the milk sample is 1103.39 cm⁻¹. -1 -1180.55 cm -1 1562.49cm -1 -1604.93 cm -1 1689.80 cm -1 -1716.81 cm -1 2044.74 cm -1 -2133.47 cm -1 2195.20 cm -1 -2426.68 cm -1 2434.40 cm -1 -2573.29 cm -1 2604.15 cm -1 -2808.62 cm -1 3109.55 cm -1 -3225.29 cm -1 3279.30 cm -1 -3422.05 cm -1 3672.82 cm -1 -3830.99 cm -1 3966.02 cm -1 -4008.46 cm -1 4016.18 cm -1 -4074.05 cm -1 4413.55 cm -1 -4594.88 cm -1 4641.17 cm -1 -4776.20 cm -1 4810.93 cm -1 -5007.68 cm -1 MIR data in the middle; 2) Substitute the measured MIR data into the constructed model of second derivative + multivariate scattering correction + partial least squares regression to output the predicted carbon emissions from the mouth and nose of dairy cows; The principal component of the partial least squares regression is 9.

2. The method according to claim 1, characterized in that: Each segment of the aforementioned band is allowed to have a gap of two wave points before and after it.

3. The method according to claim 1, characterized in that: The milk mentioned is from Holstein cows.

4. The application of the rapid batch detection method for carbon emissions from the mouth and nose of dairy cows as described in claim 1 in the performance evaluation of dairy cows.

Images