Soil health comprehensive evaluation system and method based on artificial intelligence

Abstract

The invention relates to the technical field of soil evaluation, in particular to a soil health comprehensive evaluation system and method based on artificial intelligence. The method and the system comprise a segmented sensing probe which is integrated with a stress release structure at a signal cable outlet; the main control module is used for sensing the connection stability in real time and generating a credibility label; and the data processing and analyzing unit is used for performing artificial intelligence evaluation by using the soil data with the credibility label. The invention provides a soil health comprehensive assessment system and method based on artificial intelligence, and solves or at least alleviates the problem that the robustness and accuracy of AI assessment are affected due to sensor connection stability reduction, data distortion or drifting caused by the environment during multi-depth monitoring in the existing soil health assessment. The reliability and precision of soil health assessment can be improved, and the accuracy, reliability and long-term effectiveness of an assessment result are ensured.

Description

Technical Field

[0001] This invention relates to the field of soil assessment technology, and in particular to a comprehensive soil health assessment system and method based on artificial intelligence. Background Technology

[0002] As global agricultural production increasingly demands efficiency and sustainability, the penetration and application of artificial intelligence (AI) technology in agriculture are becoming more widespread. Among these applications, AI-based soil health assessment systems have become a key tool for promoting precision agriculture, realizing smart farmland management, and optimizing environmental protection strategies. These systems, through intelligent analysis of multi-dimensional soil parameters, aim to provide scientific basis for crop growth, water and fertilizer management, and pest and disease early warning, with the goal of increasing yields, reducing resource consumption, effectively mitigating soil degradation, and maintaining ecological balance.

[0003] Some existing methods and systems for assessing soil pollution employ hyperspectral remote sensing and artificial intelligence technologies. For example, patent CN120102476B discloses a method and system for assessing soil pollution. Its technical solution sets sensitive bands and contribution values ​​to determine a correlation matrix, correct hyperspectral data, and then assess soil pollution. This technical solution can determine its effectiveness by measuring pollutant concentrations and judging a comprehensive pollution index, thereby identifying soil pollution. This technical solution is suitable for monitoring large-scale, non-contact soil pollution. Its beneficial effect lies in utilizing the rich spectral information of hyperspectral data and using intelligent algorithms to separate, judge, and identify specific pollutants from numerous mixed information, providing a basis for soil pollution prevention and control. However, the aforementioned technical solution does not address how to conduct long-term, reliable, multi-depth vertical profile monitoring for soil health status, nor does it discuss the physical issues of sensor connections in an artificial intelligence soil health assessment system under long-term field and multi-depth vertical profile monitoring conditions.

[0004] Soil is a complex medium of physical motion. Sensor probes placed at soil observation points are continuously affected by physical environmental factors such as soil creep (slow displacement of soil particles and reshaping of soil structure), periodic expansion and contraction due to wet and dry conditions (causing volume changes within the probe or at the mechanical connection points between the probe and the connecting cable), and diurnal or seasonal thermal expansion and contraction, rather than simply being affected by the slow embedding or creep of the static soil. These physical movements cause the mechanical connection points within the probe or between the probe and the connecting cable to be continuously subjected to unpredictable micro-displacement forces, thus generating physical stress. This physical stress does not simply damage the sensor, but rather more subtly and dangerously gradually induces minute deformations at the connecting cables or solder joints within the sensor, thereby altering the physical properties of the signal transmission circuit, such as contact resistance drift, impedance fluctuations, and even intermittent micro-disruptions or signal attenuation.

[0005] This instability in the probe's internal connectivity, caused by the movement of the soil's physical environment, is harmful not only because it constantly poses a direct cause of probe failure, but more importantly, because it damages the sensor's acquired data signals in a hidden way, without manifesting as a malfunction. This leads to increased data distortion or drift, and such distortion or drift may only occur in one or a few depth layers of a multi-depth probe, exhibiting high locality and uncertainty. When some data signals connected to the AI ​​evaluation system are unknowingly and silently contaminated by physical connectivity instability, and the AI ​​evaluation system's machine learning model cannot determine the physical reliability of the data signal source, it becomes difficult to distinguish whether the signal fluctuations originate from changes in soil health parameters or from false signals caused by hardware connectivity instability.

[0006] Soil health often exhibits complex vertical gradient variations; for example, the distribution of nutrients, moisture, and microbial activity may differ significantly at different depths. AI models, when performing vertical gradient analysis and calculating stratified health indices, require extremely high accuracy and overall consistency between data from different depth layers. If data from any layer is distorted or drifts due to unstable physical connections, it will directly lead to biased assessment results for that layer and further distort the gradient relationships between adjacent layers, severely impacting the accuracy and reliability of the entire vertical profile health assessment, and potentially even triggering erroneous agricultural management decisions. Existing data correction algorithms, such as those mentioned above for hyperspectral data correction, typically focus on addressing environmental interference during data acquisition or calibration errors of the sensors themselves. However, they rarely address or fail to provide solutions for ensuring the stability and status perception of the physical connection links between sensors and the system under long-term dynamic stress in the field. Summary of the Invention

[0007] The purpose of this invention is to provide an artificial intelligence-based comprehensive soil health assessment system and method, which can solve, or at least in the prior art, the problem that under long-term field deployment and multi-depth vertical profile monitoring conditions, environmental factors lead to a decrease in the stability of sensor hardware connections, which in turn causes data distortion or drift, and ultimately affects the robustness and accuracy of artificial intelligence assessment results.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a comprehensive soil health assessment system based on artificial intelligence, comprising:

[0009] One or more segmented sensing probes are used to deploy in the soil to acquire multi-depth vertical profile data;

[0010] A main control module is used to control the operation of the segmented sensing probe, and to perform data preprocessing and connection stability sensing.

[0011] A data processing and analysis unit is used to receive and process soil data with confidence labels from the main control module, and perform artificial intelligence assessment based on the confidence labels; and

[0012] A human-computer interaction interface is used to display evaluation results and system status.

[0013] To further realize the present invention, the following technical solutions may be preferred:

[0014] Preferably, the main body of the segmented sensing probe is a cylindrical shell, with multiple independent sensing units integrated inside along its axial direction. The number of sensing units is not less than three, and they are uniformly or non-uniformly distributed along the axial direction to correspond to different soil depth layers. Each sensing unit independently contains at least one soil parameter sensing electrode, which includes at least one of a conductivity sensing electrode, a humidity sensing electrode, and a temperature sensing electrode. The signal cable of each sensing unit is led out through an independent wiring channel and connected to a multi-channel electronic switch array disposed inside the probe base. The multi-channel electronic switch array is controlled by the main control module to realize the on-demand conduction or disconnection of the power supply circuit and signal path of any sensing unit.

[0015] Preferably, the probe housing integrates an annular groove at the signal cable outlet corresponding to each sensing unit to form a local stress relief structure. The design of the groove allows the signal cable to have greater bending flexibility in the local area. When the soil generates micro-displacement stress on the probe body due to creep, wet-dry cycles, or thermal expansion and contraction, the signal cable preferentially undergoes elastic bending deformation at the groove, thereby absorbing and dispersing mechanical stress and ensuring the stability of the signal connection.

[0016] Preferably, when the target sensing unit is activated, the main control module injects a weak test current into the signal circuit of the unit and simultaneously monitors the equivalent impedance of the signal circuit and its fluctuation amplitude and trend. The main control module compares the monitored impedance fluctuation with a preset threshold. If the fluctuation exceeds the threshold, the signal connection is determined to be unstable, and a "low confidence state" mark is generated. Otherwise, it is marked as a "high confidence state". The main control module binds this confidence state information with the original data of the corresponding sensing unit.

[0017] Preferably, the data processing and analysis unit assigns a dynamic credibility weight factor to the raw sensing data of each depth layer based on the connection stability perception result; the artificial intelligence evaluation engine receives the weighted profile data and calculates the vertical gradient of specific parameters between adjacent depth layers, and then multiplies the gradient value with the combined value of the credibility weights of the corresponding two depth layers to obtain the credibility gradient feature; if the credibility weight corresponding to a certain depth layer is too low, the contribution of the data of that depth layer in participating in the vertical gradient calculation is suppressed or compensated by interpolation.

[0018] An artificial intelligence-based comprehensive soil health assessment method, applicable to the aforementioned system, includes the following steps:

[0019] Step S1: Deploy a soil probe with segmented sensing and stress relief structure. The probe has multiple independent sensing units integrated along the axial direction, and each sensing unit has a stress relief interface formed by an annular groove integrated at the signal cable outlet.

[0020] Step S2: Perform segmented selective activation and synchronous sensing of connection status. Select the target depth layer according to the evaluation requirements and activate the corresponding sensing unit, and synchronously collect the connection stability status information of the activated segment.

[0021] Step S3: Construct a hierarchical confidence-weighted soil profile dataset, and assign dynamic confidence weight factors to the original sensor data of each activated depth layer according to the connection stability state information;

[0022] Step S4: Perform a hierarchical fusion-based artificial intelligence comprehensive assessment of soil health. The artificial intelligence model analyzes the data based on weighted profile data and constructed reliable gradient features, and outputs a hierarchical health index and comprehensive assessment results.

[0023] Step S5: Feedback optimization and long-term monitoring adaptation. The data processing and analysis unit records the historical connectivity stability trend of each depth layer and dynamically adjusts the subsequent sampling strategy.

[0024] Preferably, the connection status synchronous sensing in step S2 includes: while activating the target sensing unit, injecting a weak test current into the signal circuit of the unit, and monitoring the equivalent impedance of the signal circuit and its fluctuation amplitude and trend; if the impedance fluctuation amplitude continues to exceed a preset threshold, the connection status of the signal segment is determined to be unstable, and a "low confidence status" mark is generated.

[0025] Preferably, step S3 includes: for sensing units determined to be in a "high confidence state", directly reading their output parameter values; for sensing units determined to be in a "low confidence state", starting a data verification process, repeatedly activating and collecting data, and if the difference between multiple readings is still large, marking it as "data to be corrected"; and assigning a dynamic confidence weight factor to each depth layer according to the determination result, and binding the original parameter values ​​with the weight factor.

[0026] Preferably, step S4 includes: calculating the parameter change rate between adjacent depth layers, and multiplying the gradient value by the combined value of the confidence weights of the corresponding two layers to obtain the confidence gradient feature; if the weight of a certain layer is too low, its contribution to the gradient calculation is suppressed; the artificial intelligence model performs deep learning analysis based on the weighted profile data and the confidence gradient feature.

[0027] Preferably, step S4 further includes: outputting soil health sub-indices for each depth layer, wherein the calculation of the sub-indices has embedded the confidence weight of the data for that layer; fusing all sub-indices to generate an overall comprehensive soil health assessment value and outputting a confidence status label; step S5 includes: continuously storing connection stability data for each depth layer to form an interface health profile; if a certain depth layer repeatedly exhibits a low confidence status, a maintenance warning is triggered, and subsequent sampling strategies are dynamically adjusted, including reducing the sampling frequency of that layer or prioritizing the activation of adjacent high-stability layers for data compensation.

[0028] The beneficial effects of this invention are:

[0029] This invention overcomes the challenges of data distortion and AI model blind spots caused by hardware connection instability in existing technologies for multi-depth vertical profile soil health assessment. It ensures data source reliability at the physical level through segmented probes combined with stress-relief structures; segmented selective activation and synchronous connection state perception mechanisms transform the physical connection state into digital credibility labels; credibility-weighted profile datasets enable AI models to distinguish between real soil changes and secondary noise; hierarchical fusion AI assessment utilizes these weighted reliable data for more accurate and robust analysis; and finally, feedback optimization and adaptive sampling strategies construct a long-term monitoring closed loop capable of self-learning and evolution. This invention ensures the accuracy, reliability, and long-term effectiveness of soil health assessment results through stress-relief ring grooves, selective activation and impedance fluctuation monitoring, dynamic weight allocation and credibility gradient construction, and gradient suppression and fusion. Attached Figure Description

[0030] Figure 1 A block diagram of the system architecture of this invention;

[0031] Figure 2A structural block diagram of the segmented sensing probe of the present invention;

[0032] Figure 3 A structural block diagram of the sensing unit of the present invention;

[0033] Figure 4 A structural block diagram of the signal cable outlet stress relief structure of the present invention;

[0034] Figure 5 A flowchart illustrating the method of the present invention;

[0035] Figure 6 A flowchart illustrating step S2 of the present invention;

[0036] Figure 7 A schematic diagram of the processing logic of the data processing and evaluation module of this invention;

[0037] Figure 8 A schematic diagram of the segmented sensing probe of the present invention;

[0038] Figure 9 A cross-sectional view of the segmented sensing probe of the present invention;

[0039] Figure 10 The present invention Figure 9 Sectional view at point AA;

[0040] Figure 11 The present invention Figure 9 Enlarged view of point B in the middle.

[0041] The attached figures are labeled as follows:

[0042] 1-Housing; 2-Sensing unit; 201-Conductivity sensing electrode; 202-Humidity sensing electrode; 203-Temperature sensing electrode; 3-Probe base; 4-Signal cable; 5-Annular groove. Detailed Implementation

[0043] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] This embodiment discloses a comprehensive soil health assessment system based on artificial intelligence. (Refer to...) Figure 1 The system block diagram shown primarily comprises one or more segmented sensing probes, a main control module, a data processing and analysis unit, and a human-machine interface. The segmented sensing probes are designed for multi-depth vertical profile data acquisition in soil. The main control module controls the operation of the segmented sensing probes and performs data preprocessing and critical connectivity stability sensing functions. The data processing and analysis unit receives and processes soil data with confidence labels from the main control module, and then performs complex artificial intelligence assessments. Finally, the human-machine interface intuitively presents the assessment results and system operating status to the user.

[0047] like Figure 2 and Figures 8-11 As shown, the main structure of the probe is a cylindrical outer shell 1, made of high-strength, corrosion-resistant 316L stainless steel to ensure long-term stable operation in complex soil environments, and has a uniform cross-section with a diameter of 38mm. The length of the probe can be customized according to specific monitoring depth requirements; in this embodiment, it is set to 1500mm to cover a soil profile from the surface to a depth of 1.5 meters. Five independent sensing units 2 are precisely integrated along the probe's axial direction, i.e., N=5. These five sensing units are non-uniformly distributed along the axial direction within the probe to optimize monitoring for different soil layers: the first sensing unit is located in the 0-100mm depth range, the second in the 100-300mm range, the third in the 300-600mm range, the fourth in the 600-1000mm range, and the fifth in the 1000-1500mm range. This non-uniform distribution strategy is designed to focus on the root-active layer and the important soil and water transport layer.

[0048] like Figure 3As shown, each sensing unit 2 independently includes multiple soil parameter sensing electrodes, specifically a conductivity sensing electrode 201, a humidity sensing electrode 202, and a temperature sensing electrode 203. The conductivity sensing electrode consists of a pair of platinum parallel plate electrodes with a spacing of 10 mm. This electrode uses an applied AC signal with a frequency of 1 kHz and an amplitude of 100 mV to measure the conductivity of the soil solution via AC impedance method. The humidity sensing electrode employs dielectric constant sensor technology, accurately reflecting the soil volumetric water content by measuring the capacitance change caused by the change in the soil dielectric constant. The temperature sensing electrode is a high-precision NTC thermistor encapsulated in corrosion-resistant epoxy resin, with a measurement range of -20℃ to 60℃.

[0049] Each sensing unit's signal cable 4 is made of polytetrafluoroethylene (PTFE) insulated multi-strand copper wire. These cables are led upwards inside the probe through independent wiring channels and connected to a multi-channel electronic switch array located inside the probe base 3. The multi-channel electronic switch array consists of five groups of low on-resistance (typically 100mΩ) solid-state relays, each group of relays responsible for the power supply circuit and signal path of one sensing unit. This array is controlled by the main control module via an SPI bus interface, enabling on-demand switching of the power supply circuit and signal path of any sensing unit, thereby effectively reducing system power consumption and eliminating signal crosstalk that may be generated by inactive sensing units.

[0050] As a preferred embodiment of this example, Figure 4As shown, at the exit of the signal cable 4 corresponding to each sensing unit 2, i.e., where the signal cable exits from the probe's interior into the more spacious internal cavity of the probe base, the probe housing integrates an annular groove 5 to form a local stress relief structure. The annular groove is 0.5 mm deep and 0.5 mm wide, and its shape is a smooth U-shape, surrounding the signal cable at its entry or exit point from the probe wall. This groove design allows the signal cable to have a smaller cross-sectional area and greater bending flexibility in this local region compared to other parts. Specifically, the effective bending radius in this region is significantly increased, reducing the force required per unit deformation. When the soil exerts micro-displacement stress on the probe body due to creep, periodic wet-dry cycles causing volume expansion and contraction (e.g., clay layers shrink when dry and expand when wet, resulting in approximately 5% volume change), or diurnal / seasonal thermal expansion and contraction (a 20°C change in soil temperature can cause micron-level displacement of the probe body), the signal cable will preferentially undergo elastic bending deformation at the annular groove. This preferential deformation mechanism effectively transforms the shear, tensile, or torsional stresses that would normally act directly on the cable solder joints or connector pins into elastic bending strain in the local area, thereby absorbing and dispersing the mechanical stresses imposed by external environmental factors. This stress relief mechanism ensures the stability of the physical connection between the signal cable and its internal conductors, micro-soldering points, or connector pins of the multi-channel electronic switch array, significantly avoiding solder joint cracking, contact resistance drift (e.g., from 0.01Ω to 0.5Ω), or intermittent micro-open circuits caused by direct mechanical stress, thus maintaining the integrity of the electrical connection and the reliability of signal transmission of the sensing unit over the long term. The external encapsulation material of the stress relief interface is medical-grade silicone rubber, which is tightly bonded to the probe shell and cable through injection molding to form a watertight encapsulation, preventing the intrusion of moisture and contaminants, and further enhancing the stress buffering effect, ensuring that it can maintain its physical properties under extreme temperature (-40℃ to 85℃) and humidity conditions.

[0051] Example 2

[0052] This embodiment discloses an artificial intelligence-based comprehensive soil health assessment method, which is implemented on the aforementioned system. The specific process is as follows. Figure 5 As shown.

[0053] Step S1: Deploy a soil probe with segmented sensing and stress relief structure.

[0054] Step S101: Construct a multi-segment sensing probe body. N independent sensing units are integrated axially within a single cylindrical probe; N is set to 5 in this embodiment. Each sensing unit independently corresponds to a specific soil depth layer. As mentioned earlier, each sensing unit includes at least a conductivity sensing electrode, a humidity sensing electrode, and a temperature sensing electrode. The conductivity sensing electrode is implemented as a pair of parallel platinum electrodes spaced 10 mm apart. It obtains conductivity data by measuring the AC impedance of the soil medium. Its front end is equipped with a high-precision constant current source and synchronous demodulation circuit to ensure stable measurement even in complex electromagnetic environments. The humidity sensing electrode is implemented as a dielectric constant sensor; its capacitance changes with soil moisture content, and it incorporates a high-linearity capacitance-to-digital converter (CDC). The temperature sensing electrode is implemented as an NTC thermistor; its resistance changes with soil temperature and is converted into a voltage signal by a high-precision Wheatstone bridge circuit. Each sensing unit's data acquisition circuit includes a low-noise preamplifier with an input noise voltage density of less than 10nV / √Hz, and an analog-to-digital converter (ADC) with a resolution of at least 16 bits and a sampling rate of at least 200Hz, ensuring accurate capture and rapid response to weak signals. The multi-channel electronic switch array consists of N independent low-on-resistance solid-state relays with an on-resistance of less than 50mΩ. Controlled by SPI protocol commands from the main control module, it independently switches the power supply (3.3V DC) and signal paths (analog signal lines and digital I / O lines) of any sensing unit, ensuring complete isolation in non-operating states, minimizing system power consumption to the microampere level, and completely eliminating crosstalk.

[0055] Step S102: Integrate a stress relief interface at the exit of each signal cable segment. In the transition area between the probe housing and the internal flexible signal cable, an annular groove is formed for each signal cable corresponding to each sensing unit. The depth of the annular groove is controlled between 0.5mm and 0.7mm, and the width is 0.5mm. Its edge is designed as a smooth arc to avoid any form of damage to the signal cable. This groove design enhances the elastic deformation capability of the signal cable in this local area, improves the tensile strength at local stress concentration points, and extends fatigue life. The structural characteristics of this annular groove allow it to preferentially undergo elastic bending at the groove when subjected to micro-displacement stress caused by environmental factors such as soil creep (e.g., an average annual soil horizontal displacement of 1-3mm), periodic wet-dry cycles, or thermal expansion and contraction, effectively buffering and dispersing the mechanical stress acting on the signal cable connection point. For example, when the probe experiences a local displacement of 0.2mm, the cable solder joint without a groove design may withstand shear stress up to 50MPa, while the groove design of this invention can reduce this stress to below 15MPa. This elastic deformation mechanism ensures that the internal conductors, solder joints, and electrical contact points of the signal cable are protected from the direct effects of tensile, shear, or torsional stresses, thereby maintaining the long-term stability of the electrical connection of the sensing unit, avoiding unpredictable drift of contact resistance or intermittent interruption of the signal transmission path, and keeping the signal noise level below 0.5mV during long-term operation.

[0056] Step S2: Perform segmented selective activation and connection state synchronization sensing.

[0057] Reference Figure 6The detailed flowchart shown first proceeds to step S201: Selecting the target depth layer and activating the corresponding sensing unit based on the evaluation requirements. The main control module receives instructions from the artificial intelligence evaluation engine 31 in the data processing and analysis unit. These instructions contain information about the current soil depth range to be collected, such as discrete depth intervals expressed as 0-100mm, 100-300mm, 300-600mm, 600-1000mm, and 1000-1500mm. The main control module parses the instructions to determine the sensing unit number corresponding to the specific depth layer to be activated. For example, if the instructions require data collection from the 100-300mm and 600-1000mm depth layers, the main control module will identify the second and fourth sensing units to be activated. Based on the identified sensing unit numbers, the main control module sends control commands to the multi-channel electronic switch array via the SPI bus. These commands precisely connect the power supply circuit and signal path of the sensing unit corresponding to the target depth layer. Meanwhile, the remaining non-target depth layer sensing units will remain powered off, and their power and signal connections will be completely isolated by the switch array. This effectively reduces the overall power consumption of the system (for example, when only one of the five sensing units is activated, the power consumption can be reduced by 80%), and completely eliminates the interference of electromagnetic crosstalk that may be generated by the inactive sensing units on the data acquisition of the target sensing unit, ensuring that the signal-to-noise ratio of the target signal is higher than 60dB.

[0058] Step S202 is performed synchronously: The connection stability status information of the activated segment is collected simultaneously. While activating the target sensing unit, the main control module injects a weak AC test current with a preset frequency of 8kHz and an amplitude precisely controlled between 200 and 300 microamps into the signal loop of the unit, i.e., the specific line connected to the sensor. This current amplitude is far lower than the normal operating current of the sensor to avoid affecting normal sensor measurements. The main control module synchronously monitors the equivalent loop impedance of the signal loop, specifically by measuring the voltage drop generated by the injected current in the loop and calculating the voltage-to-current ratio to obtain the impedance value in real time. For example, if a 250 microamp current is injected and a voltage drop of 50 microvolts is measured on a certain section of cable, then its impedance is 0.2 ohms. Furthermore, the main control module continuously monitors the instantaneous fluctuation amplitude and long-term trend of this impedance value. The impedance value is sampled 2000 times within a 0.1-second time window, and its standard deviation is calculated. For example, if the impedance value fluctuates between 0.20Ω ± 0.05Ω within 0.1 seconds, its standard deviation is 0.025Ω. The main control module compares the monitored loop impedance fluctuation amplitude with multiple preset thresholds in real time. The preset thresholds are calibrated in the laboratory; for example, if the standard deviation exceeds 0.08 ohms, it is judged as "slightly unstable"; if it exceeds 0.4 ohms, it is judged as "significantly unstable". If the impedance fluctuation amplitude continuously exceeds a certain threshold (e.g., 0.15 ohms) within a set monitoring period (e.g., 10 consecutive seconds), the main control module determines that the signal cable segment, including its internal connection points and external leads, has an unstable physical connection state due to soil disturbance, and generates a "low confidence state" mark. Conversely, if the impedance fluctuation amplitude is continuously below the threshold, it is marked as a "high confidence state". The main control module binds this confidence status information (e.g., binary bit 0 represents low confidence and bit 1 represents high confidence) with the original data of the corresponding sensing unit to form an initial data tuple containing a physical connection confidence label, for example: {Depth: 0-100mm, Temperature: 25.3℃, Humidity: 35.1%, Conductivity: 1.2mS / cm, Confidence: 1}.

[0059] Step S3: Construct a hierarchical confidence-weighted soil profile dataset.

[0060] This step processes the collected raw data and assigns it a credibility weight.

[0061] First, step S301 is performed: acquiring the raw sensing data for each activation depth layer. For sensing units determined to be in a "high confidence state," the main control module directly reads and records their output parameters such as conductivity, humidity, and temperature. After sampling by a 16-bit ADC, the data undergoes noise reduction processing using a digital filter (e.g., a second-order Kalman filter with a cutoff frequency of 0.5Hz connected in series with a 10-point moving average filter), which can reduce the noise variance by more than 70%. For sensing units determined to be in a "low confidence state," the main control module immediately initiates a data verification process. This verification process includes repeatedly activating the sensing unit 2 to 3 times, and re-acquiring its sensing data and connection stability information after each activation. For example, a low confidence unit is sampled three times, with a 5-second interval between each sampling. If, during multiple repeated data acquisitions, the difference between the obtained sensor parameter values ​​(such as conductivity readings), measured by the relative standard deviation (RSD), still exceeds a preset threshold (e.g., 5%), the system marks the acquired data as "data to be corrected," along with the labels "low confidence" and "needs correction," for subsequent special processing by the artificial intelligence model. For example, if three conductivity readings are 1.0, 1.05, and 1.15 mS / cm, with an average of 1.1 mS / cm, a standard deviation of approximately 0.06 mS / cm, and an RSD of approximately 5.45%, exceeding the 5% threshold, then it is marked as data to be corrected.

[0062] Then, step S302 is performed: a dynamic reliability weight is assigned to each layer of data. Based on the determination result of step S202, i.e., based on the loop impedance fluctuation amplitude, the main control module assigns a dynamic reliability weight factor to the raw sensing data of each depth layer. If the determination is "high reliability state" (impedance fluctuation standard deviation less than 0.08Ω), the corresponding weight factor is set to 1.0. If the determination is "low reliability state" (impedance fluctuation standard deviation greater than 0.08Ω), the weight factor will be linearly decayed according to the degree of impedance fluctuation. For example, when the impedance fluctuation amplitude (standard deviation) is in the lowest unstable range (0.08Ω to 0.15Ω), the weight factor is 0.7; when the fluctuation amplitude reaches the highest unstable range (greater than 0.4Ω), the weight factor can decay to 0.3. The linear attenuation function can be expressed as: W = 1.0 - k * (ΔZ - ΔZ_min) / (ΔZ_max - ΔZ_min), where W is the weighting factor, ΔZ is the standard deviation of impedance fluctuation, ΔZ_min is the lower limit of the low confidence threshold (0.08Ω), ΔZ_max is the upper limit of the high instability threshold (0.4Ω), and k is the attenuation coefficient, set to 0.7 here. Using this function, when ΔZ = 0.08Ω, W = 1.0; when ΔZ = 0.4Ω, W = 0.3. The main control module binds the original sensing parameter values ​​with the calculated dynamic confidence weighting factor. This binding operation forms a structured weighted profile data tuple, whose format can be {depth layer D_i, parameter P_1, weight W_1; parameter P_2, weight W_2; ...}. The weighted profile data tuples are then encapsulated and transmitted to the data processing and analysis unit via a wireless communication module (e.g., a module based on the LoRaWAN protocol) as input features for the artificial intelligence evaluation model.

[0063] Step S4: Perform a hierarchical and integrated artificial intelligence assessment of soil health.

[0064] Reference Figure 7The diagram illustrates the internal processing logic of the data processing and evaluation module. Step S401 involves constructing vertical gradient features based on weighted profile data. The artificial intelligence evaluation engine in the data processing and analysis unit receives the weighted profile data tuples. The artificial intelligence model first calculates the rate of change of a specific parameter (e.g., conductivity or humidity) between adjacent depth layers, i.e., the vertical gradient. The gradient calculation formula is G_i,i+1=(P_{i+1}-P_i) / (D_{i+1}-D_i), where P is the parameter value and D is the depth. For example, if the humidity of D1 (0-100mm) is 35.1% (P1) and the humidity of D2 (100-300mm) is 32.5% (P2), then G1,2=(32.5-35.1) / (300-100)=-0.013% / mm. Subsequently, the AI ​​model multiplies the gradient value by the geometric mean of the credibility weights of the two corresponding depth layers to obtain a "credible gradient feature". The formula for calculating the geometric mean weight is W_avg=sqrt(W_i*W_{i+1}). For example, if the weight of D1 is 1.0 and the weight of D2 is 0.7, then W_avg=sqrt(1.0*0.7)≈0.837. The final credible gradient feature is G1,2_credible=G1,2*W_avg=-0.013*0.837≈-0.0109% / mm. This processing method ensures that the gradient calculation fully considers the reliability of the data source, avoiding false gradient signals caused by unreliable data. If the credibility weight corresponding to a certain depth layer is too low (e.g., below 0.2) due to its extremely unstable connection state, the contribution of the data of that depth layer in the vertical gradient calculation will be significantly suppressed or even ignored. This suppression mechanism ensures that extreme outliers do not distort the true trend of the entire soil vertical profile due to their high uncertainty, thereby effectively improving the robustness of gradient features. This suppression can be achieved by replacing the parameter values ​​of low-weight layers with interpolation results obtained from adjacent high-weight layers using Bayesian linear interpolation, or by directly introducing a weight penalty term into the gradient calculation. For example, if the weight of D2 is reduced to 0.15, its data will be considered low-reliability when calculating the gradient with D1 and D3. The system will then use the data from D1 and D3 to interpolate D2 to construct a more reliable profile.

[0065] Further, step S402 is performed: outputting hierarchical health indices and comprehensive evaluation results. The artificial intelligence model is based on a Long Short-Term Memory (LSTM) network architecture, containing three layers of LSTM units, each with 128 neurons, and supplemented by a self-attention mechanism. During the training phase, the model uses historical soil health datasets, meteorological data, and crop growth data as training data. The Adam optimizer is used for optimization during training, with the mean squared error loss function, and an early stopping strategy is employed to prevent overfitting. Specific hyperparameters of the model include a batch size of 32 and a training epoch count of 100. In the prediction phase, the model receives the weighted profile data and constructed reliable gradient features, and performs deep learning analysis based on the aforementioned training parameters. The model outputs soil health sub-indices for each depth layer, including but not limited to nutrient activity index, aeration index, and salinization risk index. In the calculation of each sub-index, a reliability weight for the data at that depth layer is embedded. For example, the reliability level is reflected by nonlinearly multiplying or fusing the original calculation result (such as the initial value of the nutrient activity index N_raw) with the corresponding weight factor (W) (e.g., N_final=N_raw*(1-exp(-k*W))). For example, if the initial value of the nutrient activity index at a certain depth layer is 0.8, but its data weight is only 0.7, after fusion, the final nutrient activity index may be adjusted to 0.75 to reflect its low data reliability. The artificial intelligence model then performs a weighted average of the sub-indices at all depth layers or uses an independent decision fusion layer (based on a multilayer perceptron) for nonlinear fusion. The fusion weight can be dynamically adjusted based on the historical prediction accuracy of each sub-index or its impact on crop growth. Finally, the model generates an overall comprehensive soil health assessment value (e.g., range 0-100). Simultaneously, based on the average confidence weight of all input data layers and the volatility of the final evaluation result, the system will also output a "high confidence" or "requires review" status label to guide the user's acceptance of the evaluation result. For example, if the overall average weight of the evaluation result is higher than 0.8 and the prediction variance is less than a preset threshold (e.g., the evaluation value fluctuation range is less than 2), it is marked as "high confidence". If the average weight is lower than 0.6 or the prediction variance is too large, it is marked as "requires review" and the user is prompted that on-site survey may be necessary.

[0066] Step S5: Feedback optimization and long-term monitoring adaptation.

[0067] First, step S501 is performed: recording the historical connection stability trend of each depth layer. The data processing and analysis unit continuously stores the loop impedance fluctuation data of each depth layer sensing unit and its corresponding confidence weight value at each sampling time. This data is organized into a structured "interface health profile" and stored in a persistent database, managed in a time-series format. This profile includes a timestamp, probe ID, depth layer ID, original impedance value, impedance fluctuation standard deviation, confidence status determination, and assigned weight factors. If the interface health profile shows that a specific depth layer is in a "low confidence state" for multiple consecutive times (e.g., 5 consecutive samplings with a 1-hour interval between each sampling, i.e., 5 consecutive hours), the data processing and analysis unit will automatically activate a maintenance warning. The maintenance warning is presented through the human-machine interface (such as a web client or mobile app) and can optionally be sent to one or more designated users via email or SMS. The maintenance warning will indicate that the depth layer (e.g., probe ID: P001, depth layer: 100-300mm) may be in a high-disturbance area, such as frequent animal activity, local irrigation water scouring, or frequent agricultural machinery operations, and that on-site inspection or maintenance should be carried out, and the integrity of the stress relief interface should be checked.

[0068] The next step, S502, involves dynamically adjusting the subsequent sampling strategy. The AI ​​evaluation engine dynamically adjusts the sampling strategy in subsequent soil health assessment cycles based on the historical trend of interface connection stability recorded in the interface health profile. For depth layers that have historically exhibited a persistently "low confidence" state, their sampling frequency will be reduced. For example, if more than 50% of the samples from a certain depth layer are marked as low confidence over a continuous 72-hour period, its sampling frequency will be adjusted from once per hour to once every six hours, thereby avoiding invalid sampling, reducing energy consumption, and extending the sensor's lifespan. Simultaneously, the system will prioritize activating adjacent depth layers with historically high stability for intensive sampling (e.g., changing from once per hour to once every half hour), and will use advanced algorithms such as Gaussian process regression and Bayesian linear interpolation to perform compensatory interpolation for low-stability layers to fill in data gaps or uncertainties. For example, if the 100-300mm layer is unstable, the system will collect data from the 0-100mm and 300-600mm layers in encrypted form, and combine this data with historical data to predict the possible value of the 100-300mm layer within that time period using Gaussian process regression.

[0069] Through the aforementioned feedback mechanism, the system achieves closed-loop optimization from "passive data acquisition" to "active adaptive monitoring." The AI ​​evaluation engine adjusts data acquisition behavior based on the physical health of the sensor interfaces, thereby optimizing resource utilization efficiency and extending the overall system's effective service life while ensuring evaluation accuracy. This adaptive capability significantly enhances the system's autonomous operation and robustness in long-term, complex field environments, increasing the average maintenance cycle and the system's effective data rate.

[0070] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A comprehensive soil health assessment system based on artificial intelligence, characterized in that, include: One or more segmented sensing probes are used to deploy in the soil to acquire multi-depth vertical profile data; A main control module is used to control the operation of the segmented sensing probe, and to perform data preprocessing and connection stability sensing. A data processing and analysis unit is used to receive and process soil data with confidence labels from the main control module, and perform artificial intelligence assessment based on the confidence labels; A human-computer interaction interface is used to display evaluation results and system status.

2. The system according to claim 1, characterized in that, The segmented sensing probe has a cylindrical outer shell, inside which multiple independent sensing units are integrated along its axial direction. The number of sensing units is no less than three, and they are evenly or non-uniformly distributed along the axial direction to correspond to different soil depth layers. Each sensing unit independently contains at least one soil parameter sensing electrode, which includes at least one of a conductivity sensing electrode, a humidity sensing electrode, and a temperature sensing electrode. The signal cable of each sensing unit is led out through an independent wiring channel and connected to a multi-channel electronic switch array located inside the probe base. The multi-channel electronic switch array is controlled by the main control module to realize the on-demand conduction or disconnection of the power supply circuit and signal path of any sensing unit.

3. The system according to claim 2, characterized in that, The probe housing integrates an annular groove at the signal cable outlet corresponding to each sensing unit to form a local stress relief structure. The design of the groove allows the signal cable to have greater bending flexibility in the local area. When the soil generates micro-displacement stress on the probe body due to creep, wet-dry cycles, or thermal expansion and contraction, the signal cable preferentially undergoes elastic bending deformation at the groove, thereby absorbing and dispersing mechanical stress and ensuring the stability of the signal connection.

4. The system according to claim 1, characterized in that, When the target sensing unit is activated, the main control module injects a weak test current into the signal circuit of the unit and simultaneously monitors the equivalent impedance of the signal circuit and its fluctuation amplitude and trend. The main control module compares the monitored impedance fluctuation with a preset threshold. If the fluctuation exceeds the threshold, the signal connection is determined to be unstable and a "low confidence state" mark is generated. Otherwise, it is marked as a "high confidence state". The main control module binds this confidence state information with the original data of the corresponding sensing unit.

5. The system according to claim 4, characterized in that, The data processing and analysis unit assigns a dynamic credibility weight factor to the raw sensing data of each depth layer based on the connection stability perception result; the artificial intelligence evaluation engine receives the weighted profile data and calculates the vertical gradient of specific parameters between adjacent depth layers, and then multiplies the gradient value with the combined value of the credibility weights of the corresponding two depth layers to obtain the credibility gradient feature; if the credibility weight of a certain depth layer is too low, the contribution of the data of that depth layer in participating in the vertical gradient calculation is suppressed or compensated by interpolation.

6. A comprehensive soil health assessment method based on artificial intelligence, applicable to the system described in any one of claims 1-5, characterized in that, Includes the following steps: Step S1: Deploy a soil probe with segmented sensing and stress relief structure. The probe has multiple independent sensing units integrated along the axial direction, and each sensing unit has a stress relief interface formed by an annular groove integrated at the signal cable outlet. Step S2: Perform segmented selective activation and synchronous sensing of connection status. Select the target depth layer according to the evaluation requirements and activate the corresponding sensing unit, and synchronously collect the connection stability status information of the activated segment. Step S3: Construct a hierarchical confidence-weighted soil profile dataset, and assign dynamic confidence weight factors to the original sensor data of each activated depth layer according to the connection stability state information; Step S4: Perform a hierarchical fusion-based artificial intelligence comprehensive assessment of soil health. The artificial intelligence model analyzes the data based on weighted profile data and constructed reliable gradient features, and outputs a hierarchical health index and comprehensive assessment results. Step S5: Feedback optimization and long-term monitoring adaptation. The data processing and analysis unit records the historical connectivity stability trend of each depth layer and dynamically adjusts the subsequent sampling strategy.

7. The method according to claim 6, characterized in that, The connection status synchronous sensing in step S2 includes: while activating the target sensing unit, injecting a weak test current into the signal circuit of the unit, and monitoring the equivalent impedance of the signal circuit and its fluctuation amplitude and trend; if the impedance fluctuation amplitude continues to exceed a preset threshold, the connection status of the signal segment is determined to be unstable, and a "low confidence status" mark is generated.

8. The method according to claim 6, characterized in that, Step S3 includes: for sensing units determined to be in a "high confidence state", directly reading their output parameter values; for sensing units determined to be in a "low confidence state", starting a data verification process, repeatedly activating and collecting data, and if the difference between multiple readings is still large, marking them as "data to be corrected"; and assigning dynamic confidence weight factors to each depth layer according to the determination result, and binding the original parameter values ​​with the weight factors.

9. The method according to claim 6, characterized in that, Step S4 includes: calculating the parameter change rate between adjacent depth layers, and multiplying the gradient value by the combined value of the confidence weights of the corresponding two layers to obtain the confidence gradient feature; if the weight of a certain layer is too low, its contribution to the gradient calculation is suppressed; the artificial intelligence model performs deep learning analysis based on weighted profile data and confidence gradient features.

10. The method according to claim 9, characterized in that, Step S4 further includes: outputting soil health sub-indices for each depth layer, wherein the calculation of the sub-indices has embedded the confidence weight of the data for that layer; fusing all sub-indices to generate an overall comprehensive soil health assessment value and outputting a confidence status label; Step S5 includes: continuously storing the connection stability data of each depth layer to form an interface health profile; if a certain depth layer repeatedly shows a low confidence status, a maintenance warning is triggered, and the subsequent sampling strategy is dynamically adjusted, including reducing the sampling frequency of that layer or prioritizing the activation of adjacent high-stability layers for data compensation.

Images