Soil moisture sensing method, system, device and medium

Abstract

The present invention provides a soil moisture sensing method, system, device, and medium. These methods utilize radar mounted on an unmanned aerial vehicle (UAV) to detect soil and collect radar signals for estimating soil moisture. Soil moisture features, including the soil's refractive index (RI) and relative amplitude ratio (RAR), are extracted from the collected radar signals. A distortion signal filtering algorithm is used to remove features contaminated by the uncontrolled motion of the UAV. A multipath interference cancellation algorithm is used to eliminate interference from multipath propagation of radar signals on soil surface cover. The features after distortion and interference cancellation are input into a trained deep neural network model, which outputs a soil moisture estimation result. A meta-model of the deep neural network model is trained, and parameters are adjusted to adapt to different soil types. This method utilizes UAVs and radar to sense soil moisture, enabling rapid adaptation to moisture sensing across diverse soil types.

Description

Technical Field

[0001] The present invention relates to the fields of wireless sensing and ubiquitous computing, and in particular to a soil moisture sensing method, system, device and medium based on an unmanned aerial vehicle and a radar. Background Art

[0002] Real-time, accurate, and wide-area soil moisture sensing plays an important role in smart agricultural irrigation. First, it helps conserve irrigation water resources. It is reported that due to excessive irrigation in agriculture, more than 15% of the earth's freshwater resources are wasted. Therefore, real-time perception of when the soil contains enough water and stopping irrigation accordingly will greatly alleviate the waste of such freshwater resources. At the same time, only when the irrigation time and water amount are appropriate can crops grow in the best way. Therefore, real-time and accurate soil moisture sensing will also help smart irrigation systems dynamically optimize irrigation plans to meet the requirements of specific types of crops, helping to increase crop yields. In addition to agricultural applications, soil moisture sensing also plays a role in other tasks, such as ecological environment monitoring and maintenance of outdoor sports fields (such as golf courses and football fields).

[0003] To date, numerous soil moisture sensing technologies have been developed, categorized as sensor-based and radio frequency (RF)-based. Sensor-based soil moisture sensing technologies rely on embedding dedicated sensor nodes in the soil. However, these technologies have high deployment and maintenance costs, making them unsuitable for smart agriculture applications. RF-based soil moisture sensing technologies use radio frequency (RF) signals to measure soil moisture, eliminating the need for deploying dedicated sensors in the soil. However, existing RF-based technologies still have various limitations. Some require embedding power devices, such as batteries, in the soil, posing the risk of soil contamination from battery damage and the labor costs of replacing batteries. Remote sensing-based soil moisture sensing methods have a coarse-grained geographic resolution and can only estimate the surface moisture of the soil. Recent RF-based soil sensing technologies have achieved finer-grained geographic resolution and greater sensing depth than remote sensing methods. However, these technologies are limited in the distance from their antennas to the soil surface, making them unsuitable for wide-area soil moisture sensing by mobile aerial platforms, such as drones.

[0004] The present invention aims to design a real-time, accurate, and wide-area soil moisture sensing system to meet the needs of application scenarios such as smart agricultural irrigation. The present invention proposes a soil moisture sensing method based on drones and radars, which realizes real-time and wide-area detection based on the high maneuverability of drones, and realizes accurate soil moisture measurement through radar. However, the design of such a method faces the following challenges. The first challenge is that the original radar signal is not sufficient to extract accurate soil moisture features for accurate soil moisture measurement, and the original radar signal needs to be enhanced. The second challenge is that drone movement and environmental multipath interference will contaminate the data, and these contaminations need to be eliminated. The third challenge is that different types of soils have different physical and chemical properties. The proposed method needs to be able to adapt to different soils and maintain a high soil moisture perception accuracy on different soils. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention aims to provide a soil moisture sensing method, system, device and medium.

[0006] A soil moisture sensing method provided by the present invention includes:

[0007] Data collection steps: Use radar mounted on a drone to detect soil and collect radar signals for estimating soil moisture;

[0008] Feature extraction step: extracting soil moisture features from the collected radar signals, wherein the soil moisture features include the refractive index RI and the relative amplitude ratio RAR of the soil;

[0009] Distortion filtering step: The features contaminated by the uncontrolled motion of the drone are filtered out through the distortion signal filtering algorithm;

[0010] Interference elimination step: using a multipath interference elimination algorithm to eliminate the interference of the radar signal's multipath propagation on the soil surface cover on the extracted features;

[0011] Moisture estimation step: The features after distortion filtering and interference elimination are input into the trained deep neural network model, and the soil moisture estimation result is output;

[0012] Meta-model training step: training the meta-model of the deep neural network model to adapt to different soil types by adjusting parameters.

[0013] Preferably, the data collection step includes:

[0014] Select measurement points within the sensing area, and bury a reflector at each measurement point at a preset depth from the soil surface;

[0015] The drone is hovered above each measurement point and raised vertically within a preset altitude range, while continuously transmitting signals and receiving reflected radar signals through the radar installed on the drone.

[0016] Preferably, the feature extraction step comprises: performing interpolation on the radar signal, and extracting soil moisture features from the time of flight ToF and peak amplitude of the interpolated radar signal;

[0017] The radar baseband signal s(t) is expressed in the form of a Gaussian pulse:

[0018]

[0019] where α tx is the amplitude that determines the pulse intensity, σ 2 is the variance that determines the pulse width, t is time, e is a natural constant, and the soil surface reflection signal r1(t) is expressed as:

[0020]

[0021] in is the attenuation term of the signal propagating in the air, α air represents the air attenuation factor, n is the refractive index of the soil, d1 is the distance from the radar to the soil surface, c is the speed of light, j is a complex number, π is the circumference of the circle, and f c is the center frequency of the radar signal. Assuming the reflector is buried at a depth of d2, the reflector reflection signal r2(t) is expressed as:

[0022]

[0023] in represents the attenuation term of the signal propagating in the soil, α s represents the soil attenuation factor, is the attenuation term caused by the signal penetrating the soil-air boundary and reflecting off the reflector, where m represents the refractive index of the reflector;

[0024] The relative amplitude ratio p is defined as the ratio of the peak amplitude of r2(t) to the peak amplitude of r1(t), that is,

[0025]

[0026] Assume that the radar receives k frames of signal at the measurement point. For the i-th frame, find the sampling points r1(t) and r2(t) with the highest amplitude and get the amplitude a 1，i and a 2，i , and time of flight ToF t 1，i and t 2，i , calculate the refractive index RI and relative amplitude ratio RAR corresponding to the i-th frame, respectivelyi =0.5c(t 2，i -t 1，i ) / d2 and pi =a 2，i / a 1，i , apply this operation to all received signal frames, and obtain the set of peak amplitudes, refractive indexes RI and relative amplitude ratios RAR of r1(t) and r2(t), expressed as A1={a 1，1 , a 1，2 ,···,a 1，k}, A2={a 2，1 , a 2，2 ,···,a 2，k}, N={n1, n2, ···, n k}, P={p1, p2, ..., p k}.

[0027] Preferably, the distortion filtering step includes:

[0028] Calculate the average of the peak amplitudes of r1(t) and r2(t) per frame:

[0029] For each i-th frame i∈[1+h,kh], calculate the sliding average of the relative amplitude ratio RAR in the h previous and h subsequent frames of i and

[0030] If the i-th frame is collected when the drone is stable, the condition |p i -m i，1 |<δ and |p i -m i，2 |<δ, δ is the outlier detection threshold;

[0031] If the drone does not deviate from the measurement point, the condition is met and

[0032] The refraction index RI and relative amplitude ratio RAR of the i-th frame that simultaneously meet the above four conditions are retained to obtain a filtered refraction index RI set N′ and relative amplitude ratio RAR set P′.

[0033] Preferably, the interference elimination step includes:

[0034] Calculate the average of the refractive index RI and relative amplitude ratio RAR in N′ and P′:

[0035] Check each pair (n i , p i), if the condition and Established, it is believed that the pair (n i , p i ) is valid and retained, and outputs a valid refractive index RI set N″ and a relative amplitude ratio RAR set P″;

[0036] is the average value of the refractive index RI of all frames, is the average value of the relative amplitude ratio RAR of all frames, ∈ n and ∈ p is the outlier detection threshold.

[0037] Preferably, the humidity estimation step comprises:

[0038] Randomly select K elements from N″ and K elements from P″ to form two K-dimensional vectors n and p. n and p are input into the two branches of the encoder module of the deep neural network model respectively to obtain two feature vectors:

[0039] h n =f(W n，2 f(W n，1 n)),

[0040] h p =f(W p，2 f(W p，1 p)),

[0041] Where f is the ReLU activation function, W n，i and W p，i Represents the weight matrix of the i-th MLP layer of the two encoders respectively;

[0042] h n and h p Spliced together, as the input of the inference module of the deep neural network model, to obtain the final soil moisture estimation result

[0043] y=f(W inference [h n ||h p ]),

[0044] Among them, W inference represents the weight of the reasoning module MLP, and || represents the concatenation operation.

[0045] Preferably, the metamodel training step includes:

[0046] Pre-training a meta-model of the deep neural network model and initializing weight parameters of the meta-model;

[0047] Each time, a set of data is sampled from the training data of a type of soil in the deep neural network model, and the meta-model is gradient calculated and weight updated on the support set of this set of data. Then, the gradient is calculated on the replica using the query set of this set of data to update the weight parameters of the meta-model. This process is repeated multiple times to obtain a meta-model that can adapt to new types of soil.

[0048] After the meta-model training is completed, the meta-model is adapted to the target soil type through the following process:

[0049] First, a dataset of the target soil type is collected at one or two soil moisture levels. Then, the meta-model is adjusted using stochastic gradient descent on this dataset to minimize the MAE loss between the estimated soil moisture and the true value. During the adjustment process, the de-weighted parameters of the encoder module are frozen, and only the weight parameters of the inference module are updated. Through the meta-model training process, the encoder module of the meta-model learns the encoding rules applicable to different types of soil, and only adjusting the inference module can adapt to new types of soil.

[0050] A soil moisture sensing system provided by the present invention includes:

[0051] Data collection module: uses the radar mounted on the drone to detect the soil and collect radar signals for estimating soil moisture;

[0052] Feature extraction module: extracts soil moisture features from the collected radar signals, wherein the soil moisture features include the refractive index RI and relative amplitude ratio RAR of the soil;

[0053] Distortion filtering module: Filters out features contaminated by uncontrolled drone motion through a distortion signal filtering algorithm;

[0054] Interference elimination module: eliminates the interference of multipath propagation of radar signals on soil surface cover on the extracted features through multipath interference elimination algorithm;

[0055] Moisture Estimation Module: This module inputs the features after distortion filtering and interference elimination into the trained deep neural network model and outputs the soil moisture estimation result.

[0056] Metamodel training module: trains the metamodel of the deep neural network model to adapt to different soil types by adjusting parameters.

[0057] According to the present invention, a computer-readable storage medium storing a computer program is provided. When the computer program is executed by a processor, the steps of the soil moisture sensing method are implemented.

[0058] According to the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the soil moisture sensing method are implemented.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. Compared with the traditional sensor-based soil moisture estimation method, the present invention reduces the labor cost of burying peripherals such as gateways, and also reduces the expenses caused by purchasing a large number of expensive soil moisture sensors.

[0061] 2. Compared with the existing soil moisture sensing method based on radio frequency signals, the present invention designs the first mobile radio frequency soil moisture sensing system based on drones, and has higher measurement accuracy.

[0062] 3. The present invention has a reasonable structure, is easy to use, and can overcome the defects of the prior art.

[0063] 4. This invention adopts a framework of data collection-feature extraction-data selection-humidity estimation, which can accurately extract humidity-related features from radar signals, eliminate the influence of undesirable UAV motion and multipath reflection on radar signals, and accurately estimate soil moisture at multiple measurement points over a large area of land. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0065] Figure 1 This is an example diagram of the system flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0067] like Figure 1 As shown, a soil moisture sensing method includes:

[0068] Step S1: Data collection: Use the IR-UWB radar carried by the UAV to detect the soil and collect radar signals that can be used to estimate soil moisture.

[0069] Step S2: Radar signal processing and feature extraction, extracting features related to soil moisture from the radar received signal, namely soil refractive index (RI) and relative amplitude ratio (RAR).

[0070] Step S3: Propose a distortion signal filtering algorithm to filter out data contaminated by the uncontrolled motion of the drone;

[0071] Step S4: Propose a multipath interference elimination algorithm to eliminate the interference of radar signal multipath propagation on soil surface cover on the extracted features;

[0072] Step S5: Design a deep neural network model that takes RI and RAR as input and outputs soil moisture estimation;

[0073] Step S6: Design a meta-learning-based neural network fast migration framework to ensure the versatility of the model on different types of soils.

[0074] In step S1, in order to select a suitable radar for soil moisture sensing on a drone, we investigated the common small and light commercial radars on the market and conducted experiments on the soil penetration capabilities of these radars. Specifically, three common commercial radars were tested, with central operating frequencies of 77 GHz, 24 GHz, and 7.29 GHz, respectively. The experiments were conducted on four types of soil, including sand, loam, silt, and clay. The radar signal needs to penetrate tens of centimeters of soil at a drone hovering altitude of several meters, and be able to receive radar signals reflected by reflectors buried in the soil. Based on the experiments, a 7.29 GHz IR-UWB radar with good penetration capability was selected;

[0075] First, several representative measurement points are selected within the sensing area. At each measurement point, a reflector is pre-buried at a predetermined depth below the soil surface. This reflector is highly reflective of IR-UWB radar signals. During data collection, a drone hovers above each measurement point and vertically ascends within a predefined altitude range. Simultaneously, the drone's IR-UWB radar continuously transmits and receives reflected signals. This vertical elevation design helps mitigate multipath interference.

[0076] In step S2, the signal collected from the IR-UWB radar is essentially a discrete sampling of the continuously received signal. Since features need to be extracted from the ToF and peak amplitude of the received radar signal, the ToF accuracy of the discrete sampling points determines the accuracy of the extracted features. To obtain more accurate ToF and peak amplitude of the received signal, we propose upsampling the received signal through interpolation. Specifically, 16x spline interpolation is performed on the received sampled signal, improving ToF accuracy from 0.343ns to 21.3ps.

[0077] Establish the IR-UWB radar signal transmission model. Specifically, the baseband signal s(t) of the IR-UWB radar is expressed in the form of Gaussian pulses, that is,

[0078]

[0079] where α tx is the amplitude that determines the pulse intensity, σ 2 is the variance that determines the pulse width. The soil surface reflection signal can be expressed as

[0080]

[0081] in is the attenuation term of the signal propagating in the air, α air represents the air attenuation factor, and n is the refractive index (RI) of the soil. Similarly, assuming the reflector is buried at a depth of d2, the reflector reflection signal can be expressed as

[0082]

[0083] in represents the attenuation term of the signal propagating in the soil, α s represents the soil attenuation factor, is the attenuation term caused by the signal penetrating the soil-air boundary and reflecting off the reflector, where m represents the refractive index of the reflector.

[0084] The first feature extracted from the signal is n, the refractive index of the soil, whose value is affected by the relative dielectric constant and conductivity of the soil, both of which are closely related to soil moisture. The second feature used is the attenuation factor α in the soil s Since it is difficult to obtain the air attenuation α1 term, the relative amplitude ratio (RAR) is used to represent α s .

[0085] Definition 1 (Relative Amplitude Ratio RAR): The relative amplitude ratio p is defined as the ratio of the peak amplitude of r2(t) (i.e., α1α2α3) to the peak amplitude of r1(t) (i.e., ), that is,

[0086]

[0087] According to Definition 1, if the reflector material and its burial depth are fixed, that is, m and d2 are fixed, the value p will be determined only by α s and n. And α s Both nitrogen and nitrogen are closely related to soil moisture.

[0088] Extract soil moisture features from the interpolated radar received signal. Assume that the IR-UWB radar receives k frames of signal at the measurement point. For the i-th frame, first find the r1(t) and r2(t) sampling points with the highest amplitude, and obtain the amplitudes a of these two samples. 1，i and a 2，i , and ToF t 1，i and t 2，i Then calculate the soil RI and RAR values corresponding to the i-th frame, respectively, i =0.5c(t 2，i -t 1，i ) / d2 and p i =a 2，i / a 1，i Applying this operation to all received signal frames, we obtain the set of peak amplitudes of r1(t) and r2(t), soil RI and RAR, which is expressed as A1 = {a 1，1 , a 1，2 ,…,a 1，k}, A2={a 2，1 , a 2，2 ,…,a 2，k}, N={n1, n2, ..., n k}, P={p1, p2, ..., p k}.

[0089] In step S3, the uncontrollable motion of the drone causes some of the received signals to be contaminated, making them unsuitable for direct soil moisture estimation. Therefore, a distorted signal filtering algorithm is proposed. This algorithm takes as input the four sets A1, A2, N, and P extracted in step S2, a sliding window size h, and an outlier detection threshold δ. It outputs the filtered soil RI set N′ and RAR set P′.

[0090] Step S3.1: Initialize N′ and P′ to be empty.

[0091] Step S3.2: Calculate the average of the peak amplitudes of r1(t) and r2(t) for each frame:

[0092] Step S3.3: For each i-th frame i∈[1+h,kh], check whether it is available. First calculate the sliding average of the RAR in the h previous and h next frames of i and If the i-th frame is collected when the drone is stable, the RAR of this frame will have a value similar to the calculated RAR sliding average, which means that the condition |p i -m i，1 |<δ and |p i -m i，2|<δ holds. In addition, if the drone does not deviate from the measurement point, the peak amplitudes of r1(t) and r2(t) should not be less than the average peak amplitude of all frames calculated, that is, the condition is met. and Therefore, if all four conditions above are met, the algorithm will retain the soil RI and RAR of the i-th frame.

[0093] In step S4, in addition to the undesirable movement of the drone, multipath interference caused by reflections from objects other than the reflector (such as bushes and stones that may appear near the measurement point) may also contaminate the collected data. Multipath signals may be mixed with r1(t) and r2(t), affecting the obtained peak amplitude and ToF accuracy, making the extracted soil RI and RAR unusable for accurate soil moisture perception. Therefore, a multipath interference elimination algorithm is proposed, which uses the filtered soil RI set N′ and RAR set P′ output from step S3, the outlier detection threshold ∈ n and ∈ p As input, the valid soil RI set N″ and RAR set P″ are output.

[0094] Step S4.1: Initialize the sets N″ and P″ to be empty.

[0095] Step S4.2: Calculate the average values of soil RI and RAR in input sets N′ and P′:

[0096] Step S4.3: Check each pair (n i , p i ), evaluate whether their values are affected by multipath interference. Since in practice, multipath interference only occurs in a few frames at a specific height, the average values of soil RI and RAR of all frames are and It should be close to the average value without multipath interference. and Established, it is believed that the pair (n i , p i ) is valid and retain it.

[0097] Step S5 includes:

[0098] Step S5.1: After the above processing, randomly select K RIs from N″ and K RARs from P″ to form two K-dimensional vectors n and p. n and p are respectively input into the two branches of the encoder module of the neural network to obtain two feature vectors, namely

[0099] h n =f(W n，2 f(W n，1 n)),

[0100] h p =f(W p，2 f(W p，1 p)),

[0101] Where f is the ReLU activation function, W n，i and W p，i Represents the weight matrix of the i-th MLP layer of the two encoders respectively;

[0102] Step S5.2: h n and h p They are concatenated and used as the input of the inference module to obtain the final soil moisture estimation as the output, i.e.

[0103] y=f(W inference [h n ||h p ]),

[0104] Among them, W inference represents the weight of the reasoning module MLP, and || represents the concatenation operation.

[0105] The step S6 comprises:

[0106] Step S6.1: Initialize the mSoilIdNet weight parameters.

[0107] Step S6.2: Each time a set of data is sampled from the training data for a particular soil type, an instance of mSoilIdNet is created. Gradients are calculated and weights are updated for this instance on the support set of the sampled data. Gradients are then calculated on the replica using the query set of the sampled data to update the weight parameters of mSoilIdNet.

[0108] Step S6.3: Repeat step S6.2 multiple times to obtain an mSoilIdNet that can quickly adapt to new types of soil.

[0109] Step S6.4: After the meta-model mSoilIdNet is trained, it is adapted to the target soil type through the following process. First, a dataset of the target soil type is collected at very few (as few as one or two) soil moisture levels. Afterwards, mSoilIdNet is fine-tuned using stochastic gradient descent on this dataset to minimize the MAE loss between the estimated soil moisture and the true value. In order to improve the efficiency of the above adaptation process, it is proposed to freeze the de-weighted parameters of the encoder module and only update the weight parameters of the inference module. The rationale of this partial update method lies in that through the meta-model training process, the encoder module of mSoilIdNet learns encoding rules applicable to different types of soil. Therefore, fine-tuning the inference module alone is sufficient to adapt to new types of soil.

[0110] The present invention also provides a soil moisture sensing system based on a drone and an IR-UWB radar. The soil moisture sensing system based on a drone and an IR-UWB radar can be implemented by executing the process steps of the soil moisture sensing method based on a drone and an IR-UWB radar. That is, those skilled in the art can understand the soil moisture sensing method based on a drone and an IR-UWB radar as the operating mode of the soil moisture sensing system based on a drone and an IR-UWB radar.

[0111] Module M1: Use a drone equipped with an ultra-wideband pulse radar to collect relevant information. Sampling points are set up in areas where soil moisture needs to be sensed. Aluminum reflectors are pre-buried at the sampling points. The drone hovers above each sampling point and vertically increases its altitude within a determined altitude range to collect reflected signals to estimate soil moisture.

[0112] Module M2: Extract soil moisture related features, design a distorted data filtering algorithm to filter out unusable data distorted by the uncontrollable motion of the drone, and design a multipath interference elimination algorithm to detect and discard data affected by multipath interference.

[0113] Module M3: A deep neural network is used to fit the mapping relationship between soil moisture features and soil moisture. A meta-learning framework is used to enable the training model based on known soils to quickly learn based on a small number of training samples and be applicable to moisture estimation of unknown soils.

[0114] Preferably, the module M1 includes:

[0115] Module M1.1: Implements a UAV system equipped with an ultra-wideband pulse radar with a center frequency of 7.29 GHz. The system also has a mini computer that can send commands through the serial port to control the radar's data transmission and reception;

[0116] Module M1.2: A 30cm*30cm aluminum metal plate is used as a reflector and buried 30cm from the soil surface at each sampling point. This reflector has the advantages of high reflectivity, corrosion resistance, and low cost.

[0117] Module M1.3: Collect data by having the drone hover over the sampling point and then ascend vertically, laying the foundation for subsequently eliminating the impact of multipath interference.

[0118] Preferably, the module M2 includes:

[0119] Module M2.1: Upsamples the collected radar data through interpolation, improving the time-of-flight accuracy of the collected data to 21.3ps;

[0120] Module M2.3: Extract features related to soil moisture, soil refractive index n and relative amplitude ratio p, where soil refractive index n is related to soil moisture and can be calculated using the following formula

[0121]

[0122] Where d2 is the depth of the reflector, which is a constant, t1 is the flight time of the radar signal reflected by the soil surface, and t2 is the flight time of the radar signal reflected by the reflector.

[0123] The relative amplitude ratio is the ratio of the amplitude of the signal reflected from the reflector surface to the amplitude of the signal reflected from the soil surface, which can be calculated by the following formula

[0124]

[0125] Where α1 is the amplitude of the radar signal reflected from the soil surface, α2 is the amplitude of the radar signal reflected from the reflector surface, m is the refractive index of the reflector, is a constant, and α s is the soil attenuation factor, which is related to soil moisture.

[0126] Module M2.4: A distortion data filtering algorithm is used to filter out the distorted data generated by environmental factors during the flight of the drone.

[0127] The algorithm collects k frames of data and calculates the set of α1 A1={α 1，1 , α 1，2 …α 1，k}, the set of α2 A2 = {α 2，1 , α 2，2 …α 2，k}, the set N of n, the set P of p and the sliding window size h, the outlier detection threshold As input, N' and P' are given as output after filtering the data.

[0128] The designed algorithm first calculates the mean of all α1 in set A1 and the mean of all α2 in set A2

[0129]

[0130] Then, the following same operation is performed on each frame data i from the 1+hth frame to the khth frame to determine whether the frame data is available. The operation is: calculate and

[0131] Finally, make a judgment, if |m 1，i -p1|<δ and|m2，i -p1|<δ and and The frame data is considered usable data, otherwise it is distorted data and is filtered out.

[0132] Module M2.5: A multipath interference cancellation algorithm is used to eliminate the impact of multipath interference caused by other reflective objects encountered by the drone during flight.

[0133] The algorithm combines the outputs N′ and P′ of the distorted data filtering algorithm and the outlier detection threshold ∈ n and ∈ p As input, N″ and P″ are taken as output after the data is filtered again.

[0134] The algorithm first calculates the mean of all elements in the set N′ and the mean of all elements in the set P′

[0135]

[0136] Then for each element n in the sets N′ and P′ i and p i Make a judgment, if and Then the element n is considered i and p i The final available data is the data; otherwise, it is regarded as the data affected by multipath and filtered out.

[0137] Preferably, the module M3 includes:

[0138] Module M3.1: Fit the mapping relationship between soil moisture characteristics and soil moisture through a deep neural network.

[0139] First, K elements are randomly selected from each set N″ and P″, and then they are combined into two K-dimensional feature vectors n and p as the input of the neural network;

[0140] The two feature vectors are then fed into two encoders, both of which are multi-layer perceptrons consisting of alternating ReLU activation layers and fully connected layers.

[0141] The last two feature vectors are encoded by two encoders to output two new vectors. These two new vectors are connected and further passed through a ReLU activation layer and a fully connected layer to finally output the estimated soil moisture.

[0142] Module M3.2: Using a meta-learning framework and designing a new training strategy, we enable a model trained on known soils to quickly learn and apply moisture estimation to unknown soils using a small number of new training samples. By improving the existing meta-learning framework MAML, each training round is trained using only training data collected from the same soil type. After training the meta-learning model, we collect a small dataset of the target soil type and fine-tune the model using stochastic gradient descent to minimize the loss between the model's estimated moisture and the true value.

[0143] The present invention can effectively and quickly sense soil moisture in a large area with the help of drones and radars, has a certain degree of robustness against environmental interference, and can quickly adapt to moisture perception of different soil types.

[0144] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0145] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A soil moisture sensing method, characterized in that: include: Data collection steps: Use radar mounted on a drone to detect soil and collect radar signals for estimating soil moisture; Feature extraction step: extracting soil moisture features from the collected radar signals, wherein the soil moisture features include the refractive index RI and the relative amplitude ratio RAR of the soil; Distortion filtering step: The features contaminated by the uncontrolled motion of the drone are filtered out through the distortion signal filtering algorithm; Interference elimination step: using a multipath interference elimination algorithm to eliminate the interference of the radar signal's multipath propagation on the soil surface cover on the extracted features; Moisture estimation step: The features after distortion filtering and interference elimination are input into the trained deep neural network model, and the soil moisture estimation result is output; Meta-model training step: training the meta-model of the deep neural network model to adapt to different soil types by adjusting parameters; The meta-model training step includes: Pre-training a meta-model of the deep neural network model and initializing weight parameters of the meta-model; Each time, a set of data is sampled from the training data of a type of soil in the deep neural network model, and the meta-model is gradient calculated and weight updated on the support set of this set of data. Then, the gradient is calculated on the replica using the query set of this set of data to update the weight parameters of the meta-model. This process is repeated multiple times to obtain a meta-model that can adapt to new types of soil. After the meta-model training is completed, the meta-model is adapted to the target soil type through the following process: First, a dataset of the target soil type is collected at one or two soil moisture levels. Then, the meta-model is adjusted using stochastic gradient descent on this dataset to minimize the MAE loss between the estimated soil moisture and the true value. During the adjustment process, the de-weighted parameters of the encoder module are frozen, and only the weight parameters of the inference module are updated. Through the meta-model training process, the encoder module of the meta-model learns the encoding rules applicable to different types of soil, and only adjusting the inference module can adapt to new types of soil.

2. The soil moisture sensing method according to claim 1, characterized in that: The data collection steps include: Select measurement points within the sensing area, and bury a reflector at each measurement point at a preset depth from the soil surface; The drone is hovered above each measurement point and raised vertically within a preset altitude range, while continuously transmitting signals and receiving reflected radar signals through the radar installed on the drone.

3. The soil moisture sensing method according to claim 1, characterized in that: The feature extraction step includes: performing interpolation on the radar signal, and extracting soil moisture features from the time of flight ToF and peak amplitude of the interpolated radar signal; The radar baseband signal s(t) is expressed in the form of a Gaussian pulse: where α tx is the amplitude that determines the pulse intensity, σ 2 is the variance that determines the pulse width, t is time, e is a natural constant, and the soil surface reflection signal r1(t) is expressed as: in is the attenuation term of the signal propagating in the air, α air represents the air attenuation factor, n is the refractive index of the soil, d1 is the distance from the radar to the soil surface, c is the speed of light, j is a complex number, π is the circumference of the circle, and f c is the center frequency of the radar signal. Assuming the reflector is buried at a depth of d2, the reflector reflection signal r2(t) is expressed as: in represents the attenuation term of the signal propagating in the soil, α s represents the soil attenuation factor, is the attenuation term caused by the signal penetrating the soil-air boundary and reflecting off the reflector, where m represents the refractive index of the reflector; The relative amplitude ratio p is defined as the ratio of the peak amplitude of r2(t) to the peak amplitude of r1(t), that is, Assume that the radar receives k frames of signals at the measurement point. For the i-th frame, find the sampling points r1(t) and r2(t) with the highest amplitude and get the amplitude a 1,i and a 2,i , and time of flight ToFt 1,i and t 2,i , calculate the refractive index RI and relative amplitude ratio RAR corresponding to the i-th frame, respectively, i =0.5c(t 2,i -t 1,i ) / d2 and p i =a 2,i / a 1,i , apply this operation to all received signal frames, and obtain the set of peak amplitudes, refractive indexes RI and relative amplitude ratios RAR of r1(t) and r2(t), expressed as A1={a 1,1 ,a 1,2 ,···,a 1,k }, A2={a 2,1 ,a 2,2 ,···,a 2,k }, N={n1,n2,···,n k }, P={p1,p2,···,p k }.

4. The soil moisture sensing method according to claim 3, characterized in that: The distortion filtering step comprises: Calculate the average of the peak amplitudes of r1(t) and r2(t) per frame: For each i-th frame i∈[1+h,kh], calculate the sliding average of the relative amplitude ratio RAR in the h previous and h subsequent frames of i and If the i-th frame is collected when the drone is stable, the condition |p i -m i,1 |<δ and |p i -m i,2 |<δ, δ is the outlier detection threshold; If the drone does not deviate from the measurement point, the condition is met and The refraction index RI and relative amplitude ratio RAR of the i-th frame that simultaneously meet the above four conditions are retained to obtain a filtered refraction index RI set N′ and relative amplitude ratio RAR set P′.

5. The soil moisture sensing method according to claim 4, characterized in that: The interference elimination step includes: Calculate the average of the refractive index RI and relative amplitude ratio RAR in N′ and P′: Check each pair (n i ,p i ), if the condition and Established, it is believed that the pair (n i ,p i ) is valid and retained, and outputs a valid refractive index RI set N″ and a relative amplitude ratio RAR set P″; is the average value of the refractive index RI of all frames, is the average value of the relative amplitude ratio RAR of all frames, ∈ n and ∈ p is the outlier detection threshold.

6. The soil moisture sensing method according to claim 5, characterized in that: The humidity estimation step comprises: Randomly select K elements from N″ and K elements from P″ to form two K-dimensional vectors n and p. n and p are input into the two branches of the encoder module of the deep neural network model respectively to obtain two feature vectors: h n =f(W n,2 f(W n,1 n)), h p =f(W p,2 f(W p,1 p)), Where f is the ReLU activation function, W n,i and W p,i Represents the weight matrix of the i-th MLP layer of the two encoders respectively; h n and h p Spliced together, as the input of the inference module of the deep neural network model, to obtain the final soil moisture estimation result y=f(W inference [h n ||h p ]), Among them, W inference represents the weight of the reasoning module MLP, and || represents the concatenation operation.

7. A soil moisture sensing system, characterized in that: include: Data collection module: uses the radar mounted on the drone to detect the soil and collect radar signals for estimating soil moisture; Feature extraction module: extracts soil moisture features from the collected radar signals, wherein the soil moisture features include the refractive index RI and relative amplitude ratio RAR of the soil; Distortion filtering module: Filters out features contaminated by uncontrolled drone motion through a distortion signal filtering algorithm; Interference elimination module: eliminates the interference of multipath propagation of radar signals on soil surface cover on the extracted features through multipath interference elimination algorithm; Moisture Estimation Module: This module inputs the features after distortion filtering and interference elimination into the trained deep neural network model and outputs the soil moisture estimation result. Metamodel training module: training the metamodel of the deep neural network model to adapt to different soil types by adjusting parameters; The meta-model training module includes: Pre-training a meta-model of the deep neural network model and initializing weight parameters of the meta-model; Each time, a set of data is sampled from the training data of a type of soil in the deep neural network model, and the meta-model is gradient calculated and weight updated on the support set of this set of data. Then, the gradient is calculated on the replica using the query set of this set of data to update the weight parameters of the meta-model. This process is repeated multiple times to obtain a meta-model that can adapt to new types of soil. After the meta-model training is completed, the meta-model is adapted to the target soil type through the following process: First, a dataset of the target soil type is collected at one or two soil moisture levels. Then, the meta-model is adjusted using stochastic gradient descent on this dataset to minimize the MAE loss between the estimated soil moisture and the true value. During the adjustment process, the de-weighted parameters of the encoder module are frozen, and only the weight parameters of the inference module are updated. Through the meta-model training process, the encoder module of the meta-model learns the encoding rules applicable to different types of soil, and only adjusting the inference module can adapt to new types of soil.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the soil moisture sensing method according to any one of claims 1 to 6 are implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the soil moisture sensing method according to any one of claims 1 to 6 are implemented.

Images