Digging shovel for detecting content of organic matters in soil in real time

Abstract

The digging shovel capable of detecting the content of the organic matter in the soil in real time comprises a detachable digging shovel body, the front end of the digging shovel body is a digging blade, and the rear end of the digging shovel body is provided with a sensor integrated cavity; the data acquisition and communication module is integrated in the digging shovel, comprises a near infrared spectrum sensor and acquires the soil reflectivity by adopting a V-shaped optical fiber diffuse reflection structure; the high-resolution camera is used for collecting a soil surface texture image; the GPS module is used for recording longitude, latitude and elevation data of sampling points; the main control unit is used as a core controller for coordinating data acquisition, processing and communication; the data processing unit realizes accurate soil organic matter content data based on spectrum and image information fusion, helps workers to grasp soil conditions in real time, and improves excavation production efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural machinery, and provides a digging shovel for real-time detection of soil organic matter content. Background Art

[0002] The content of soil organic matter is the core index to measure soil fertility and directly affects the growth efficiency of crops. Traditional detection methods (such as high-temperature combustion method, wet acid digestion method) rely on laboratory chemical analysis (refer to national standards GB 9834-1988, NY / T 1121.6-2006, etc.), and have the following significant defects: First, the samples need to be pretreated, chemically decomposed and instrumentally detected, and the single detection cycle is as long as several hours to several days, which cannot meet the requirements of real-time field operations. Second, professional technicians are required to operate precision instruments (such as elemental analyzers), the equipment maintenance cost is high, and it is not suitable for non-laboratory environments. Third, laboratory detection only supports discrete point sampling, and it is difficult to generate a continuous soil organic matter distribution map, resulting in a lack of global data support for fertilization decisions.

[0003] In recent years, near-infrared spectroscopy (NIRS) technology and image recognition technology have been introduced into the field of soil detection, but there are still the following technical bottlenecks: First, low equipment integration: existing portable spectrometers only support single-spectrum data collection, and are not integrated with high-resolution image sensors, GPS modules and agricultural implement adaptation structures, resulting in low field operation efficiency. For example, the operator needs to manually collect samples and conduct separate detections after agricultural machinery operations, and cannot achieve an integrated process of "operating while analyzing"; Second, single data model: existing technologies mostly use linear regression models (such as partial least squares regression PLSR) to process spectral data, and do not fuse multi-source information such as soil surface texture (such as gray-level co-occurrence matrix features) and color distribution (RGB histogram statistics), and the prediction error rate is generally ≥8%, which is difficult to meet the requirements of precision agriculture; Third, lack of real-time performance: the end-to-end response time of the solution that relies on the cloud server for data processing exceeds 10 seconds, and it cannot support real-time agricultural machinery operation decisions (such as dynamically adjusting the fertilization amount); Fourth, poor environmental adaptability: the high-dust and high-humidity environment in the field easily causes sensor contamination. The existing equipment protection design (such as bag filters) needs to be frequently cleaned (once every 2 hours), and is not optimized for working conditions such as agricultural machinery vibration and impact, and the equipment reliability is insufficient; In addition, existing agricultural implement integrated sensors (such as soil humidity and temperature sensors) are mostly independent modules, lack a multi-modal data fusion architecture, and the dust-proof design efficiency is low (the dust barrier rate of traditional atomization dust removal solutions is ≤92%). The above problems seriously restrict the large-scale application of soil organic matter detection technology in agricultural mechanization operations.

[0004] In summary, the existing technology has not solved the core problems such as low real-time detection accuracy in the field, weak multi-source data fusion ability, and poor environmental adaptability of the embedded system of agricultural implements. There is an urgent need for an innovative solution that combines laboratory-level accuracy and field operation efficiency. Summary of the Invention

[0005] The object of the present invention is to provide a digging shovel for real-time detection of soil organic matter content, which realizes accurate soil organic matter content data based on the fusion of spectral and image information, grasps the soil condition in real time during actual operation, and then makes more scientific and reasonable fertilization decisions, improves production efficiency, avoids waste of resources, and realizes more accurate soil nutrient management.

[0006] The object of the present invention is achieved by the following technical solutions: A digging shovel for real-time detection of soil organic matter content, comprising: A digging shovel, with a digging edge at the front end and a sensor integration cavity at the rear end, and is connected to agricultural machinery through a fixing bolt and a connecting rod; A dust-proof component, including a dust-proof cover and a pneumatic balance valve built in the dust-proof cover; A main control unit, coordinating data acquisition, communication and processing; A Global Positioning System (GPS) module, recording the longitude, latitude and elevation data of the sampling point; A data acquisition and communication unit, integrated in the sensor integration cavity, including: a near-infrared spectral sensor for collecting soil reflectance, covering the wavelength range of 400 - 2500 nm; a camera for collecting soil surface texture images, equipped with a light source and a retractable light-shielding plate; A data processing unit, including: A spectral data processing module: using the de-envelope method to eliminate spectral baseline drift, and then using the three-dimensional correlation coefficient method to screen sensitive bands strongly correlated with organic matter. The correlation between soil reflectance and organic matter content is strongly correlated when it is between 0.6 and 1; An image data processing module: extracting the contrast, energy, entropy and color histogram statistics of the gray-level co-occurrence matrix (GLCM); A multi-source data fusion model: using a three-branch convolutional neural network (3D CNN) to fuse spectral, texture and color features, and outputting the predicted value of soil organic matter content; A result output module: generating a soil organic matter distribution map and displaying it in real time through a display screen, and synchronously uploading it to the management platform.

[0007] As a more preferable technical solution of the present invention, the sensitive bands strongly correlated with organic matter include a combination of 918 nm, 580.8 nm and 831.8 nm.

[0008] As a more preferable technical solution of the present invention, the color histogram statistics include the mean, standard deviation, skewness and kurtosis of the RGB three channels.

[0009] As a more preferable technical solution of the present invention, the three-branch convolutional neural network includes: a Spectral feature branch, where the spectral data filtered by the three-dimensional correlation coefficient method is processed by a 1D convolution layer; b Texture feature branch, the texture feature parameters of the soil image extracted by processing the gray-level co-occurrence matrix by the 1D convolution layer; c Color feature branch, where the color features of the soil image extracted by the color histogram are processed by a 1D convolutional layer; The above features are subjected to trilinear pooling and multi-source data fusion to establish a 3D CNN model.

[0010] As a more optimal technical solution of the present invention, the soil organic matter distribution map generated by the result output module satisfies: Spatial interpolation: Generate 10m×10m grid data based on Kriging interpolation method; Visual mapping: Using the HSL (hue, saturation, brightness) color space, the purple to red gradient represents 0% to 5% organic matter content.

[0011] As a better technical solution of the present invention, the main control unit is a Raspberry Pi, equipped with a TensorRT inference engine, with an end-to-end response time of ≤2 seconds, and manages continuous spectral data streams through a circular buffer.

[0012] As a more optimal technical solution of the present invention, the adjustable angle connecting rod has an adjustment range of 0°-45°, is suitable for the interface of a rotary tiller and a seed drill, and the carbon steel material of the fixing bolt has a load-bearing capacity of ≥50 kg.

[0013] As a more optimal technical solution of the present invention, the near infrared spectrum sensor is detachably connected to the camera and the sensor integrated cavity, including: (a) The near-infrared spectral sensor is connected to the sensor integrated cavity through a base with a threaded interface; (b) The camera is connected to the sensor integrated cavity through a slide rail.

[0014] As a more optimal technical solution of the present invention, the power supply module is a replaceable lithium-ion battery pack or a solar thin-film battery, and its battery management system BMS has: aDynamic power management, giving priority to powering the spectrum sensor and GPS module; b Predict the remaining power based on the operation trajectory of agricultural machinery and provide an early warning 30 minutes in advance.

[0015] As a more optimal technical solution of the present invention, the annular dust cover has a built-in air pressure balancing valve, which automatically opens the balancing airflow when the pressure difference with the external environment is ≤10 Pa.

[0016] The beneficial effects are as follows: Through the collaborative breakthrough of the "Multi-modal Data Fusion Architecture (Spectrum + Image + GPS)" and the "Real-time Embedded Detection System for Agricultural Machinery", the present invention first introduces the three-dimensional dynamic weight allocation of spectrum, image texture, and spatial position into the field of soil detection; through the detachable structure and modular design, the unity of laboratory-level accuracy and field operation efficiency is achieved; the combination of the three-branch CNN model and the TensorRT acceleration engine realizes the algorithm-hardware co-optimization and defines a new standard for real-time detection. Description of the Drawings

[0017] Figure 1 It is a schematic structural diagram of the digging shovel.

[0018] Figure 2 It is a structural diagram of the spectrum module inside the digging shovel.

[0019] Figure 3 It is an example of three-dimensional correlation coefficient screening.

[0020] Figure 4 It is the result predicted by the soil organic matter prediction model, which can be used to draw the distribution image of the soil organic matter content subsequently.

[0021] Wherein: 1. Display screen; 2. Power supply module; 3. Near-infrared spectrum sensor; 4. Light shield; 5. GPS module; 6. Connecting rod; 7. Main control unit; 8. Transmission line; 9. Data acquisition and communication module; 10. Camera; 11. Digging shovel; 12. Fixed bolt; 13. Dust cover. Detailed Description of the Invention

[0022] The present invention will be further described in detail below in conjunction with specific embodiments.

[0023] By designing a subsoiling digging shovel integrated with a spectrum sensor, a camera, a control unit, and a GPS positioning module, the present invention can collect the spectrum data and image information of the soil while digging the soil during actual operation, and inversely calculate the soil organic matter content in real time through a data fusion algorithm. In addition, the digging shovel provided by the present invention can also record the geographical location of each sampling point while collecting data, generate a distribution map of the soil organic matter content, provide accurate soil quality data for agricultural managers, and assist in agricultural decisions such as fertilization, irrigation, and soil improvement. It solves the problems of poor timeliness, cumbersome operation, and heavy equipment existing in the existing soil organic matter monitoring technology. Especially during the agricultural production process, it can quickly and accurately obtain the soil organic matter content data and be efficiently applied in the field environment.

[0024] Embodiment 1: Structure and hardware configuration of the digging shovel.

[0025] As Figure 1As shown in the figure, the digging shovel for real-time detection of soil organic matter content of the present invention includes the following components: Removable digging shovel 11: The front end is a digging edge made of tungsten carbide alloy, and a sensor integration cavity is provided at the rear end. The cavity size is 120 mm × 80 mm × 50 mm (length × width × height), and it adopts an IP67 sealing design and is internally provided with a shock-absorbing bracket with a spring stiffness coefficient of 5 N / mm. It is connected to the agricultural implement through a quick-release fixing bolt 12 (M8 bolt, pre-tightening torque 15 - 20 N·m) and a connecting rod 6 with adjustable angle (adjustment range 0° - 45°, locking force ≥ 200 N). Data acquisition and communication module 9: Near-infrared spectroscopy sensor 3: Adopts a V-type optical fiber diffuse reflection structure, covering the wavelength band of 400 - 2500 nm, the light source is a 10W halogen lamp, and it is calibrated every 30 minutes through a Labsphere Spectralon standard whiteboard. The optical fiber is configured with a core diameter of 400 μm (light source optical fiber) and 600 μm (detection optical fiber), with a 60° included angle, and the effective acquisition depth is 2 - 5 mm soil layer.

[0026] High-resolution camera 10: Equipped with a 2-million-pixel CMOS sensor and a ring-shaped LED light source, and a retractable light shield 4, supporting ±5 cm axial slide rail adjustment, and collecting soil surface texture images (resolution 1920×1080).

[0027] GPS module 5: Adopts a Ublox NEO-M8N module, with a positioning accuracy of ≤2 cm, and real-time records the longitude, latitude and elevation data of the sampling point.

[0028] The main control unit 7 is implemented based on a Raspberry Pi, equipped with a TensorRT inference engine, with an end-to-end response time of ≤2 seconds, manages continuous data streams through a circular buffer, and the data acquisition and communication module 9 transmits data to the main control unit 7 through a transmission line 8.

[0029] Dust-proof component: Includes a ring-shaped dust-proof cover 13 with a hydrophobic nano-coating and a pneumatic balance valve (nylon 66 valve body, silicone diaphragm thickness 0.5 mm), which automatically opens when the internal and external air pressure difference ≥8 Pa, and the dust barrier rate ≥99.2%.

[0030] Example 2: Data processing flow.

[0031] Spectral data preprocessing includes: Reflectance calculation: Through the formula CR ( λ )=( I ( λ )-min( I )) / (max( I )- min( I )) to eliminate ambient light interference, where I ( λ) is the value of the spectral signal at the wavelength λ , min( I ) and max( I ) are the minimum and maximum values of the signal respectively, and CR ( λ ) is the envelope value at the wavelength λ ; Sensitive band screening: Based on the Pearson correlation coefficient (r > 0.85), 918 nm (C-H bond stretching vibration), 580.8 nm (N-H bond deformation), and 831.8 nm (O-H bond bending) are selected as characteristic bands.

[0032] Image feature extraction: Texture features: Use GLCM to calculate contrast, energy, entropy, homogeneity, correlation, and angular second moment; Color features: Extract the mean, standard deviation, skewness, and kurtosis of the RGB three channels.

[0033] Example 3: System performance verification.

[0034] Experimental conditions: Mounted on a tractor in the Northeast black soil area (organic matter content 2.1% - 4.8%), operating speed 5 km / h, and continuously collecting 1200 groups of samples.

[0035] Result analysis: Prediction accuracy: R 2 = 0.94, RMSE = 0.31 g / kg (the error is reduced by 42% compared with the traditional method).

[0036] Real-time performance: End-to-end response time is 1.6 seconds (95% confidence interval), supporting the processing of 100 data points per second.

[0037] Anti-interference performance: Under the vibration condition of 5 - 8G, the prediction error volatility ≤ 3% (the traditional model ≥ 12%).

[0038] Example 4: Generation of soil organic matter distribution map.

[0039] Spatial interpolation: Use the Kriging interpolation method (spherical model, range 35 m) to generate 10 m × 10 m grid data.

[0040] Visualization mapping: Represent the organic matter content gradient in the HSL color space (purple to red corresponding to 0% - 5%), and export the report in PDF or CSV format.

[0041] Example 5: Modular extended application.

[0042] Power supply module 2: Support replaceable lithium-ion battery packs (8-hour battery life) or thin-film solar cells. The BMS system predicts the battery power according to the agricultural machinery trajectory and gives an early warning 30 minutes in advance.

[0043] Algorithm adaptation: For the red soil area, dynamically adjust the sensor acquisition frequency to 2 times (when the soil humidity > 60%), and optimize the model parameters through transfer learning.

[0044] Example 6: Dust prevention and durability test.

[0045] After 8 hours of continuous operation, the dust content in the sensor cavity ≤ 0.1 mg / m³, and the contact angle of the dust-proof cover surface ≥ 150°, indicating that the hydrophobic performance has not deteriorated. The polyurethane damping pad (Shore hardness 70A) of the connecting rod effectively absorbs 83% of the high-frequency vibration, and the component life is extended to 2000 hours.

[0046] The above examples show that the present invention realizes the efficient and accurate detection of soil organic matter content through multi-sensor integration, advanced algorithm fusion and modular design, providing reliable technical support for precision agriculture.

[0047] By combining a near-infrared spectroscopy sensor and an image acquisition system, the present invention can obtain the spectral and image information of the soil in real time on site, thereby realizing the instant monitoring of the soil organic matter content. Avoiding the time delay and data feedback lag problems in traditional methods, it helps agricultural workers obtain soil information in a timely manner and quickly take appropriate agricultural measures.

[0048] The present invention integrates a spectral sensor, an image acquisition device and a control unit on an excavation shovel that can be connected to agricultural machinery. Agricultural workers can directly conduct soil detection in the fields without carrying heavy laboratory equipment, greatly improving agricultural production efficiency.

[0049] The present invention adopts a multi-band spectral sensor and a high-resolution camera, combined with advanced data processing algorithms (three-dimensional correlation coefficient screening method, three-branch convolutional neural network integrating multi-source information), to achieve high-precision prediction of soil organic matter content. Through data fusion technology, the spectral data and image information complement each other, which can more comprehensively reflect the real situation of the soil and reduce measurement errors.

[0050] The present invention combines a GPS positioning module, which can record the geographical location of each sampling point in real time and generate a soil organic matter distribution map. These charts not only provide intuitive soil quality information for agricultural managers, but also can be used for the monitoring and analysis of soil quality change trends, supporting precision fertilization and soil improvement decisions.

[0051] The excavation shovel design of the present invention has a high degree of modularity and customization, and can be configured according to different soil types, crop requirements and agricultural environments, and is applicable to various agricultural operation scenarios, especially having a wide application prospect in the field of precision agriculture.

[0052] The digging shovel of the present invention can be mechanically connected to the agricultural implement in operation. The digging shovel head obtains samples by digging the soil, and the spectral sensor and the camera simultaneously collect the spectral information and the surface image of the soil. The control unit transmits these data to the processing unit, and analyzes the spectral and image features through algorithms to extract the predicted data of the organic matter content. The GPS module records the sampling location, and the system combines the soil data and the location information to generate a soil organic matter distribution map. Finally, the control unit stores these data and exports reports or distribution maps.

[0053] The present invention combines spectral analysis, image processing, data fusion, and GPS positioning technologies to provide a portable soil organic matter detection tool that is convenient for agricultural workers to use. Through the optimized design of the subsoiling digging shovel, this tool can collect the spectral data and image information of the soil in real time while digging the soil, and accurately predict the soil organic matter content through the built-in control unit and advanced algorithms.

[0054] The digging shovel for real-time detection of soil organic matter content provided by the present invention includes: a detachable digging shovel 11 with a digging edge at the front end and a sensor integration cavity at the rear end, which is connected to the agricultural implement through a quick-release fixing bolt 12 and an adjustable-angle connecting rod 6; a data acquisition and communication module integrated inside the digging shovel, including: a near-infrared spectral sensor 3 that uses a V-type optical fiber diffuse reflection structure to collect the soil reflectance, covering the wavelength range of 400 - 2500 nm; a high-resolution camera 10 equipped with a ring-shaped LED light source and a telescopic light shield 4 for collecting the soil surface texture image; a GPS module 5 that records the longitude, latitude, and elevation data of the sampling point, with a positioning accuracy of ≤ 2 cm; a Raspberry Pi as the core controller to coordinate data acquisition, processing, and communication; a data processing unit, including: a spectral data processing module: preprocessing and feature extraction of spectral data using the de-envelope method and three-dimensional correlation coefficient screening; an image data processing module: extracting the contrast, energy, entropy of the gray-level co-occurrence matrix, and color histogram statistics; a multi-source data fusion model: fusing spectral, texture, and color features using a three-branch convolutional neural network to output the predicted value of the soil organic matter content; a result output module: generating a soil organic matter distribution map and displaying it in real time through a display screen, and uploading it to the management platform synchronously; a dust-proof component, including a hydrophobic nano-coated ring-shaped dust-proof cover 13 and a pneumatic balance valve to reduce field dust pollution.

[0055] The digging shovel of the present invention is designed based on a detachable subsoiling digging shovel, combined with a spectral sensor, an image camera, and a control unit, and has the following characteristics: Excavation shovel head part: The excavation shovel head is a normal excavation shovel, and multiple sensors are integrated behind it, including a near-infrared spectroscopy sensor and a visible light image camera, etc. The design of the excavation shovel head ensures that when excavating the soil, it can simultaneously collect the spectral data and surface image information of the soil. Spectral module: This module consists of a light source, an optical fiber transmission system, a photoelectric sensor, etc. It contains multiple V-shaped optical fibers for transmitting light. The light is diffusely reflected by the soil and then transmitted to the photoelectric sensor. The sensor converts the optical signal into an electrical signal, and the signal processing unit performs amplification, filtering, and A / D conversion to ensure high-quality and low-noise data. This module can provide spectral information for inverting soil organic matter. Camera: The excavation shovel head is also equipped with a high-resolution visible light camera, which is responsible for collecting the image information of the soil surface. The camera extracts the features of the soil surface through image processing algorithms, uses the gray-level co-occurrence matrix to extract the texture features of the soil surface, and uses the color histogram to extract the color features of the soil image, etc., to assist in improving the accuracy of predicting the soil organic matter content. The control unit is the core part of the excavator, responsible for collecting, processing, and calculating sensor data. It specifically includes the following parts: Hardware part: The control unit uses Raspberry Pi as the main control unit to process the signals collected from the spectral sensor, camera, and other devices. Raspberry Pi has strong computing power, can process a large amount of data in real time, and can be effectively connected to other modules. Software part: The software part uses the Python programming language to implement data collection and processing, image processing, data fusion, and soil organic matter prediction. This part includes the following functions: Data collection: Collect soil spectral data and image data from the spectral sensor and camera, and the collected data is transmitted to the control unit. Spectral data processing: First, use the detrending method to preprocess the spectral data, and use the three-dimensional correlation coefficient screening method to select high-correlation bands; the calculation formula and screening results of some indices are as follows: CR ( λ ) = ( I ( λ ) - min( I )) / (max( I ) - min( I )) where I ( λ ) is the value of the spectral signal at wavelength λ , min( I ) and max( I ) are the minimum and maximum values of the signal respectively, and CR( λ ) is the envelope value at wavelength λ .

[0056] .

[0057] R1, R2, and R3 are reflectance data selected from different spectral bands, and the selection of these bands is based on the correlation analysis of soil organic carbon (SOC) content. The purpose of selecting the above bands is to find the wavelength combination with the strongest correlation with SOC content. After the data is transformed by TDI, the Pearson correlation coefficient method is used to calculate the correlation coefficient between the data column and the soil organic matter content column.

[0058]

[0059] r is the Pearson correlation coefficient. x i and y i are the observed values of the two variables respectively. and are respectively x and y the average values of n is the number of samples.

[0060] Figure 3 are the three-dimensional coefficient correlation diagrams obtained by screening through different TDI formulas. In each diagram, three sensitive bands can be seen, such as Figure 3 the three sensitive bands in (a) are 918mm, 580.8nm, and 831.8nm. After obtaining the spectral data each time, the algorithm model in the control unit will calculate multiple TDIs and select the TDI model with the highest correlation as one of the input parts of the subsequent multi-source data model.

[0061] Image data processing: Use image recognition algorithms to extract soil surface features. Use the gray-level co-occurrence matrix method to extract the texture feature indices (contrast, energy, entropy, homogeneity, correlation, angular second moment) of the soil surface, and use the color histogram method to extract the color feature indices (red mean, green mean, blue mean, red standard deviation, green standard deviation, blue standard deviation, red skewness, green skewness, blue skewness, red kurtosis, green kurtosis, blue kurtosis) of the soil image.

[0062] Three-branch convolutional neural network model for multi-source data fusion: Perform multi-source information (spectral features, texture features, color features) fusion on spectral data and image features, and substitute the fused data into the support vector machine model (SVM), random forest model (RF), partial least squares model (PLS), and 3D CNN model. The model results are as Figure 4 shown.

[0063]

[0064]

[0065] RMSE represents the root mean square error of prediction, and R² represents the correlation coefficient. In regression analysis, the closer R² is to 1 and the lower the RMSE value, the stronger the model fitting ability, the smaller the prediction error, and the better the effectiveness and reliability of the model.

[0066] Soil prediction: Based on the results predicted by the soil organic matter prediction model and the recorded geographical coordinates, the distribution image of the soil organic matter content is drawn.

[0067] To ensure the accurate association between soil data and geographical locations, the present invention designs an integrated GPS module responsible for real-time recording of the geographical locations of each sampling point. Assuming the accuracy of the GPS module is ±5 meters, through real-time recording of location data, the geographical coordinates of each sampling point can be accurately associated with the collected soil data, thereby improving the accuracy of the soil organic matter content distribution map. For example, if 100 soil samples are collected within a 10-hectare farmland, through precise GPS positioning, agricultural managers can track the soil quality distribution in real time, and then carry out precise fertilization and soil management. Combined with high-precision GPS positioning (such as the Ublox NEO-M8N module with an accuracy of ±2.5 meters), this technology can provide efficient spatial data support.

[0068] All collected soil data (including spectral data, image data, GPS positioning data, etc.) will be stored in the control unit and can be displayed in real time through the display screen 1. Assuming the system collects 100 data points per second, after 1 hour of soil detection, a total of 360,000 data points are collected, and these data will be stored and processed in real time. Agricultural workers can view the prediction results of the soil organic matter content in real time through the control unit, and agricultural workers can make more precise adjustments based on the data. The system also supports exporting the data in CSV or PDF formats, which is convenient for generating the soil organic matter distribution map and prediction report, further assisting agricultural management decisions.

[0069] The system of the present invention has good user interactivity. Users can intuitively view the real-time collected data and the predicted values of the soil organic matter content through the display screen. The system design adopts a simple and intuitive user interface, and users can select different soil samples through the touch screen to view or modify the collection parameters. Assuming the system can complete data display and update within 1 second, when the predicted value of the soil organic matter content deviates from the predetermined threshold (such as exceeding ±10%), the system will remind agricultural workers to adjust the fertilization plan in real time. Through this intuitive user interaction method, agricultural workers can respond more quickly, improving the efficiency and precision of agricultural production.

[0070] The excavation shovel design of the present invention features a high degree of modularity and customizability. Users can select appropriate sensor and algorithm configurations according to soil types, crop requirements, and the needs of specific agricultural environments. For example, the system can dynamically adjust the sensitivity of sensors and the data acquisition frequency based on information such as the pH value, humidity, and temperature of the soil. When the soil humidity exceeds 60%, the acquisition frequency of the spectral sensor can be doubled to ensure data accuracy. This modular design enables the device to adapt to different agricultural operation scenarios, especially in areas such as precision agriculture (e.g., crop yield prediction), smart agriculture (e.g., automated irrigation), and environmental monitoring (e.g., detection of polluted soil), with broad applicability. By adopting standardized interfaces, the device can expand new sensor modules or algorithms according to different needs, further enhancing the flexibility and adaptability of the system.

[0071] In some embodiments, the texture features extracted by the image data processing module include the contrast, energy, entropy, homogeneity, correlation, and angular second moment of the gray-level co-occurrence matrix; the color features include the mean, standard deviation, skewness, and kurtosis of the RGB three channels.

[0072] In some embodiments, the 3D CNN includes: a spectral feature branch, which processes the spectral data screened by the three-dimensional correlation coefficient method through a 1D convolutional layer; b texture feature branch, which processes the texture feature parameters of the soil image extracted from the gray-level co-occurrence matrix through a 1D convolutional layer; c color feature branch, which processes the color features of the soil image extracted through the color histogram through a 1D convolutional layer. The above features are fused through trilinear pooling to establish a 3D CNN model, and the model evaluation satisfies the determination coefficient R 2 ≥0.92 and the root mean square error RMSE ≤ 0.35 g / kg.

[0073] Verification of sensitive bands: The Pearson correlation coefficient matrix of 918 nm (C-H bond stretching vibration), 580.8 nm (N-H bond deformation), and 831.8 nm (O-H bond bending) is shown in Table 1.

[0074] Table 1

[0075] Comparative test with the method of NY / T 1121.6 - 2006 (n = 500).

[0076] Achieve 4.3 times acceleration through INT8 quantization of TensorRT: 200 groups of representative spectral images.

[0077] Quantization error: ≤ 0.02 g / kg (based on RMSE ≤ 0.35 g / kg).

[0078] Compared with traditional dual-branch models (such as 1D+2D CNN), this structure realizes dynamic weight allocation through the multi-head attention mechanism, and the comparison on the ImageNet test set is shown in Table 2 below.

[0079] Table 2

[0080] The channel attention mechanism (SE module) is adopted to enhance the robustness to vibration noise. Under the working condition of 5-8G acceleration, the prediction error volatility ≤ 3% (traditional model ≥ 12%). The dynamic weight allocation of feature channels is realized through the following three-stage operations.

[0081] Squeeze stage (global information compression): Input feature map dimension U : (H=224, W=224, C=512 corresponding to the output of the final residual block of ResNet-18).

[0082] Global average pooling operation: Generate a channel statistical vector , where z c represents the global spatial feature of the c th channel.

[0083] Excitation stage (channel correlation modeling): Two fully connected layer structure:

[0084] Dimensionality reduction fully connected layer ( r =16 is the compression ratio, which is verified to be the optimal balance point through experiments).

[0085] : ReLU activation function (to prevent gradient disappearance): : Dimensionality increase fully connected layer (restore the original number of channels); : Sigmoid function (output weight value range [0,1]); Reweight stage (feature channel calibration): Channel-level multiplication: where s c represents the importance weight of the c th channel, realizing noise suppression and key feature enhancement.

[0086] The determination of the compression ratio r is shown in Table 3 below.

[0087] Table 3

[0088] When r = 16 is selected, the model reaches Pareto optimality in the three-dimensional metrics of the number of parameters - accuracy - speed.

[0089] Vibration and noise suppression: Through channel weights.

[0090] Dynamically reduce the weights of the channels related to high-frequency vibration and noise (experiments show that the weight attenuation of the noise channels under vibration conditions is ≥ 70%).

[0091] Enhanced feature robustness: Assign a weight gain of 1.2 - 1.5 times to the channels related to the soil surface texture (GLCM features) and color distribution (RGB histograms).

[0092] Convert the weights of the fully connected layer of the SE module from FP32 to INT8, achieving a 4.3-fold acceleration on the Raspberry Pi 4B (the latency is reduced from 85 ms to 20 ms).

[0093] Adopt a circular buffer to manage the SE weight vector, reducing the memory occupancy by 62% (from 8.2 MB to 3.1 MB).

[0094] The verification of the effectiveness of channel attention is shown in Table 4.

[0095] Table 4

[0096] The calculation overhead analysis is shown in Table 5.

[0097] Table 5

[0098] The circular buffer management supports breakpoint resumption for up to 30 seconds, adapting to the intermittent loss of the agricultural machinery GPS signal (the average packet loss rate ≤ 2%).

[0099] In some embodiments, the soil organic matter distribution map generated by the result output module satisfies: Spatial interpolation: Generate 10 m × 10 m grid data based on the Kriging interpolation method; Visualization mapping: Adopt the HSL color space, and the gradient from purple to red represents the organic matter content from 0% to 5%.

[0100] In some embodiments, the main control unit of the data processing unit is a Raspberry Pi, equipped with a TensorRT inference engine, the end-to-end response time ≤ 2 seconds, and the continuous spectral data stream is managed through a circular buffer.

[0101] In some embodiments, the adjustable-angle connecting rod 6 has an adjustment range of 0° - 45°, is adapted to the interfaces of rotary tillers and seeders, and the carbon steel material of the fixing bolt (12) has a load-bearing capacity of ≥ 50 kg.

[0102] In some embodiments, the near-infrared spectrum sensor 3 and the camera 10 adopt a modular quick-release design, including: (a) A spectrum sensor base provided with an M12 threaded interface and a waterproof sealing ring; (b) A camera slide rail bracket that supports axial position adjustment of ±5 cm.

[0103] In some embodiments, the power supply module 2 is a replaceable lithium-ion battery pack or a thin-film solar battery, and its battery management system (BMS) has: (a) Dynamic power management to prioritize power supply to the spectrum sensor and the GPS module; (b) Predict the remaining power based on the operation trajectory of the agricultural machinery and give an early warning 30 minutes in advance.

[0104] In some embodiments, the annular dust-proof cover 13 of the dust-proof component is internally provided with a pneumatic balance valve, which automatically opens to balance the air flow when the air pressure difference from the external environment is ≤ 10 Pa.

[0105] In some embodiments, the quick-release fixing bolt 12 uses an M8 bolt with a pre-tightening torque of 15 - 20 N·m, and is adapted to the ISO2320 standard agricultural machinery interface.

[0106] In some embodiments, a polyurethane damping pad (Shore hardness 70A) is added between the connecting rod and the fixing bolt, which can absorb more than 80% of the high-frequency vibration (frequency > 50 Hz) during agricultural machinery operation.

[0107] In some embodiments, the end of the adjustable-angle connecting rod 6 is configured with an ISO 5675 standard quick hitch, which supports the common interfaces of rotary tillers (such as John Deere 5E series) and seeders (such as Case IH 2150), with an adjustment angle accuracy of ±1° and an adjustment locking force of ≥ 200 N.

[0108] In some embodiments, the cavity size of the sensor integration cavity is 120 mm × 80 mm × 50 mm (length × width × height), and it adopts an IP67-level sealing design, with an internal pre-set damping bracket (spring stiffness coefficient 5 N / mm).

[0109] In some embodiments, the light source optical fiber (core diameter 400 μm, NA 0.22) and the detection optical fiber (core diameter 600 μm, NA 0.24) form a 60° angle, with a spacing of 0.5 mm, and the effective acquisition depth is 2 - 5 mm of the soil layer.

[0110] In some embodiments, the halogen light source (wavelength range 400 - 2500 nm, power 10 W, color temperature 2856 K) is automatically calibrated every 30 minutes (using a Labsphere Spectralon standard whiteboard).

[0111] In some embodiments, the base is fixed through an M12 threaded interface (compliant with IEC 61076 - 2 - 101 standard), and the waterproof sealing ring is made of fluororubber (temperature resistance - 20°C to 120°C).

[0112] In some embodiments, the spectral signal is normalized to reflectance: R(λ) = ((λ)−D(λ)) / (W(λ)−D(λ)).

[0113] In some embodiments, the rule for selecting the envelope connection points is to use the minimum values of the absorption peaks (such as at 1410 nm and 1910 nm) as nodes for piecewise linear fitting.

[0114] In some embodiments, the wavelength band (400 - 2500 nm), space (average value in the area of 1 m² around the sampling point), and time (sliding window of 10 consecutive samplings).

[0115] In some embodiments, based on the screening of the Pearson correlation coefficient (r > 0.85), experimental verification shows that the combination of 918 nm (C - H bond stretching vibration), 580.8 nm (NO3⁻ absorption peak), and 831.8 nm (characteristic of aromatic C = C in organic matter) has the lowest prediction error (RMSE = 0.28 g / kg).

[0116] In some embodiments, the spectral branch: 3 - layer 1D convolution (kernel size [3,5,3], stride 1, number of channels [32,64,128]), followed by a global average pooling layer.

[0117] In some embodiments, the fusion branch: multi - head attention mechanism (4 heads, embedding dimension 256), and the fusion weights are dynamically assigned through learnable parameters.

[0118] In some embodiments, the model training process is as follows: Dataset: 5000 groups of soil samples (organic matter content 0.5% - 5.2%), 70% training set, 15% validation set, 15% test set;

[0119] Hyperparameters: initial learning rate 0.001 (Cosine decay), batch size 32, loss function is Huber Loss (δ = 0.5), optimizer AdamW;

[0120] Lightweight measure: FP16 quantization is adopted during TensorRT deployment, and the measured model inference time is 1.3 - 1.8 seconds (Raspberry Pi 4B, ambient temperature 25°C).

[0121] In some embodiments, the interpolation parameters are as follows: Semivariogram: Spherical model (nugget value 0.12, sill value 1.05, range 35 m), anisotropy ratio 1.5 (azimuth angle 30°). Search strategy: Maximum radius 50 m, at least 8 neighborhood points, and the interpolation error is controlled by cross-validation (mean absolute error MAE ≤ 0.2 g / kg).

[0122] In some embodiments, the body material of the air pressure balance valve is nylon 66 (temperature resistance -40°C to 120°C), with a built-in silicone diaphragm (thickness 0.5 mm, elastic modulus 1.5 MPa), and spring stiffness 0.8 N / mm. The differential pressure sensor (model Honeywell HSC series) monitors the internal and external air pressures in real time, and triggers the valve to open when ΔP ≥ 8 Pa (response time ≤ 0.5 seconds).

[0123] The embodiment of the present invention is in the Northeast Black Soil Region (organic matter content 2.1% - 4.8%), mounted on a Lovol M904 tractor, with an operating speed of 5 km / h. Prediction accuracy: R² = 0.94, RMSE = 0.31 g / kg (n = 1200). Response time: 1.6 seconds (95% confidence interval), dust prevention effect: no dust pollution during 8 hours of continuous operation (dust barrier rate 99.2%).

[0124] In summary, the present invention aims to provide a soil organic matter detection tool with high real-time performance, convenience, and accuracy. By obtaining soil data in real time, it helps agricultural producers make scientific and reasonable decisions, and promotes the development of agricultural production towards intelligent and precise directions.

[0125] The shovel head part of the excavating shovel of the present invention integrates multiple sensors (such as near-infrared spectroscopy sensors, image cameras, etc.) and has normal excavating functions at the same time. By optimizing the structural design of the excavating shovel head, it can synchronously collect the spectral data and surface image information of the soil when performing soil excavation tasks. This design enables agricultural workers to directly obtain soil information in the field without carrying heavy equipment, thus improving the convenience and efficiency of soil detection. The main aspects of three-dimensional sensitive band screening and 3D CNN are as follows: First, the three-dimensional correlation coefficient screening method is used to select the bands highly correlated with the soil organic matter content from multiple spectral bands to ensure the extraction of the most predictive spectral data and improve the prediction accuracy of the model. Then, a three-branch convolutional neural network is used to process the fused spectral data and image features. The spectral features, texture features, and color features are independently processed using three branches respectively, and the results are fused to achieve a more comprehensive prediction of the soil organic matter content. This method effectively combines the advantages of multi-source data, reduces errors, and improves the accuracy and reliability of the prediction. The system can record the geographical location of each data sampling point, combine the collected soil data with the geographical location, and generate a soil organic matter distribution map to help agricultural managers make decisions on precise fertilization, irrigation, and soil improvement.

[0126] The spectral sensor and image sensor of the excavating shovel of the present invention can be configured according to different soil types and environmental conditions. The spectral sensor can be automatically adjusted according to the soil humidity and other environmental factors to ensure the accuracy of the data. At the same time, the algorithm will be optimized according to the different characteristics of the soil to provide accurate organic matter prediction.

Claims

1. An excavation shovel for real-time detection of soil organic matter content, characterized in that, Comprising: A digging shovel (11), with a digging edge at the front end and a sensor integration cavity at the rear end, and is connected to the agricultural implement through a fixing bolt (12) and a connecting rod (6); A dust-proof component, including a dust-proof cover (13) and a pneumatic balance valve built therein; A main control unit (7) that coordinates data acquisition, communication, and processing; A GPS module (5) that records the longitude, latitude, and elevation data of the sampling points; A data acquisition and communication unit (9), integrated in the sensor integration cavity, including: a near-infrared spectroscopy sensor (3) for collecting soil reflectance, covering the wavelength range of 400 - 2500 nm; a camera (10) for collecting soil surface texture images, equipped with a light source and a retractable light-shielding plate (4); A data processing unit, including: A spectral data processing module: using the baseline drift elimination method of removing the envelope line, and then using the three-dimensional correlation coefficient method to screen the sensitive bands strongly correlated with organic matter. The correlation between soil reflectance and organic matter content is strongly correlated when it is between 0.6 and 1; An image data processing module: extracting the contrast, energy, entropy, and color histogram statistics of the gray-level co-occurrence matrix; A multi-source data fusion model: using a three-branch convolutional neural network to fuse spectral, texture, and color features, and outputting the predicted value of soil organic matter content; A result output module: generating a soil organic matter distribution map and displaying it in real time through a display screen (1), and synchronously uploading it to the management platform.

2. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The sensitive bands strongly correlated with organic matter include the combination of 918 nm, 580.8 nm, and 831.8 nm.

3. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The color histogram statistics include the mean, standard deviation, skewness, and kurtosis of the RGB three channels.

4. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The three-branch convolutional neural network includes: (a) A spectral feature branch, where the spectral data screened by the three-dimensional correlation coefficient method is processed by a 1D convolutional layer; (b) A texture feature branch, where the texture feature parameters of the soil image extracted from the gray-level co-occurrence matrix are processed by a 1D convolutional layer; (c) A color feature branch, where the color features of the soil image extracted through the color histogram are processed by a 1D convolutional layer; The above features are fused through trilinear pooling for multi-source data fusion to establish a 3D CNN model.

5. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The soil organic matter distribution map satisfies: Spatial interpolation, generating 10 m × 10 m grid data based on the Kriging interpolation method; Visualization mapping, using the HSL color space, with a purple to red gradient representing an organic matter content of 0% to 5%.

6. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The main control unit (7) is a Raspberry Pi, equipped with a TensorRT inference engine, with an end-to-end response time ≤ 2 seconds, and manages the continuous spectral data stream through a circular buffer.

7. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The angle adjustment range of the connecting rod (6) is 0° to 45°, adapting to the interfaces of rotary tillers and seeders.

8. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The near-infrared spectroscopy sensor (3) and the camera (10) are both detachably connected to the sensor integration cavity, including: (a) The near-infrared spectroscopy sensor (3) is connected to the sensor integration cavity through a base with a threaded interface; (b) The camera (10) is connected to the sensor integration cavity through a slide rail.

9. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, It further includes a power supply module (2), which is a replaceable lithium-ion battery pack or a thin-film solar cell, and is used to supply power to the display screen (1), the main control unit (7), the GPS module (5), and the data acquisition and communication unit (9).

10. The excavation shovel for real-time detection of soil organic matter content according to claim 1, characterized in that, The dust cover (13) is internally provided with a pneumatic balance valve, which automatically opens to balance the air flow when the air pressure difference from the external environment is ≤ 10 Pa.

Images