Soil salinity and moisture sensing (SSAMS) technology

Abstract

A sensing device, associated method, and associated system can determine soil salinity and moisture via multiple different sensor types. A dielectric sensor measures a dielectric permittivity of a soil sample. An electrical conductivity sensor is configured to measure an electrical conductivity of the soil sample. A piezoelectric acoustic sensor is configured to measure an acoustic impedance of the soil sample. The data collected by the sensors is applied as an input data set to a neural network, and salinity and moisture content of the soil sample is determined based on an output of the neural network.

Description

5362.1009001Soil Salinity and Moisture Sensing (SSAMS) TechnologyRELATED APPLICATION

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 727,483, filed on December 3, 2024, the entire teachings of which are incorporated herein by reference.BACKGROUND

[0002] Groundwater is a vital source of freshwater in numerous coastal regions, rendering these areas particularly vulnerable to seawater intrusion (SWI). SWI occurs when saltwater from oceans infiltrates freshwater aquifers, primarily induced by excessive groundwater extraction. The challenge of SWI is being exacerbated by rising sea levels attributed to climate change.

[0003] Sea water intrusion further presents substantial risks to coastal agriculture, infrastructure, and communities. Coastal agriculture, crucial for food production and the socioeconomic stability of millions residing in these regions, faces significant threats from saltwater intrusion, which diminishes agricultural productivity. Infrastructure, predominantly constructed from concrete and steel, is also highly susceptible to damage from saltwater intrusion. Concrete deterioration, often resulting from salt crystallization, affects structures such as service tunnels, walls, bridge abutments, and piers, a phenomenon referred to as a “physical salt attack.” Furthermore, SWI impacts soil freeze-thaw cycles, exacerbating stress and pothole development in pavements. Saltwater accelerates the corrosion of steel and steel-reinforced underground structures. Coastal regions with high groundwater usage for industry and domestic consumption are particularly at risk from SWI. Examples include Chennai in India and the Central Valley of California, USA, and coastal cities with low-lying topography and high population densities.SUMMARY

[0004] Example embodiments include a device for determining soil salinity and moisture. The device can include multiple sensors. A dielectric sensor may be configured to measure the dielectric permittivity of a soil sample. An electrical conductivity sensor may be configured to measure the electrical conductivity (resistivity) of the soil sample. A piezoelectric acoustic sensor may be configured to measure an acoustic impedance of the soil sample. A frame may be configured to house the dielectric sensor, the electrical conductivity sensor, and the piezoelectric acoustic sensor, and may be adapted to be inserted into the soil sample.- 1 -4251900. vl5362.1009001

[0005] The device may further include a temperature sensor configured to measure the temperature of the soil sample. A transmitter may be configured to transmit measurement data from the dielectric sensor, the electrical conductivity sensor, and the piezoelectric acoustic sensor to a remote server. The dielectric sensor and the electrical conductivity sensor may share at least one common electrode. The piezoelectric acoustic sensor may include at least two acoustic transducers having different resonant frequencies.

[0006] Further embodiments include a system for determining soil salinity and moisture. The system may include a sensor device as described above, as well as a computer processor configured to 1) apply an input data set to a neural network, the input data set including values from the sensor device representing the dielectric permittivity, electrical conductivity, and acoustic impedance of the soil sample; and 2) determine salinity and moisture content of the soil sample based on an output of the neural network.

[0007] The neural network may be trained via a reference data set including corresponding values of dielectric permittivity, electrical conductivity, acoustic impedance, salinity, and moisture of a plurality of reference soil samples. The processor may be further configured to generate a confidence score based on the output of the neural network, the confidence score indicating an estimated accuracy of the determined salinity and moisture content. The sensor device may further include a temperature sensor, and the input data set may further include a value representing a temperature of the soil sample.

[0008] Further embodiments include a method of determining soil salinity and moisture. Dielectric permittivity, electrical conductivity, and acoustic impedance of a soil sample may be measured. An input data set may be applied to a neural network, the input data set including values representing the measured dielectric permittivity, measured electrical conductivity, and measured acoustic impedance of the soil sample. Salinity and moisture content of the soil sample may then be determined based on an output of the neural network.

[0009] The neural network may be trained via a reference data set including corresponding values of dielectric permittivity, electrical conductivity, acoustic impedance, salinity, and moisture of a plurality of reference soil samples. An accuracy score may be generated based on the output of the neural network, the accuracy score indicating an estimated accuracy of the determined salinity and moisture content (e.g., via a confidence score). A temperature of the soil sample may also be measured, and the input data set may further include a value representing the temperature of the soil sample. Measuring the acoustic impedance of the soil sample may include measuring the acoustic impedance at a minimum of two distinct acoustic frequencies.- 2 -4251900. vl5362.1009001BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0011] The foregoing will be apparent from the following more particular description of example embodiments, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0012] Fig. l is a diagram of a device for determining soil salinity and moisture in one embodiment.

[0013] Fig. 2 is an illustration of the device of Fig. 1 deployed in an environment.

[0014] Fig. 3 is a diagram of a system comprising a plurality of sensors in one embodiment.

[0015] Fig. 4 illustrates a model architecture of a deep neural network (DNN) incorporating learned probability distribution (LPD) in one embodiment.

[0016] Fig. 5 is a flow diagram of a method of determining soil salinity and moisture in one embodiment.

[0017] Fig. 6 is a graph illustrating model predictions plotted against true salinity values in one embodiment.

[0018] Figs. 7A-B are graphs illustrating uncertainties in the predicted salinity of a model based on true values of volumetric water content (VWC) of soil in one embodiment.DETAILED DESCRIPTION

[0019] A description of example embodiments follows.

[0020] Monitoring seawater intrusion (SWI) is commonly performed through the logging of observation wells, which involves using specialized instruments to measure the properties of surrounding soil and pore fluids. Recent non-intrusive techniques for soil surface salinity measurements include satellite-based systems, electromagnetic sensors, and electrical resistivity sensors. They are subject to limitations, including shallow depth, coarse resolution, and the necessity for site-specific calibrations.

[0021] Example embodiments provide a sensor fusion methodology that integrates dielectric permittivity, electrical conductivity (EC), acoustic impedance, and optionally temperature measurements within a deep neural network (DNN) framework to achieve reliable and accurate predictions of soil salinity. A deep learning model is particularly suitable due to the complex- 3 -4251900. vl5362.1009001 relationship and partial dependence between any singular measurable input feature and the response (soil salinity). Additionally, quantifying uncertainty within the network is useful to minimize risky decisions based on model predictions, particularly in the face of complex and uncertain environmental conditions.

[0022] Uncertainty quantification (UQ) in deep learning is pivotal for applications where safety and reliability are critical, such as environmental monitoring and healthcare. Several studies have applied UQ in DNNs. Example embodiments can employ learned probability distribution (LPD) modeling to quantify prediction uncertainties arising from sensor data and the DNN model. Our approach categorizes uncertainty into model (epistemic) uncertainty, related to network weights, and data (aleatoric) uncertainty, resulting from sensor noise. One advantage of this solution is the ability to deliver more precise and reliable soil salinity measurements irrespective of soil composition, unlike traditional single-principled sensors, benefiting industries and research fields reliant on accurate soil data in dynamic environments.

[0023] Fig. 1 is a diagram of a device 100 for determining soil salinity and moisture in one embodiment. The device 100 may include a frame 105 housing a plurality of sensor groups1 lOa-e, and each of the sensor groups 1 lOa-e may be placed at a different location (e.g., different height) along the frame 105 to measure soil samples at different depths of the soil environment. An auger 160 may be located at a terminal (or other) portion of the frame 105 to assist in inserting the frame 105 into a soil sample. Although the frame 105 as shown has a cylindrical shape, the frame 105 may occupy other shapes suitable for placement in a soil sample. A top end of the frame 105 may support a solar panel 130 to power the device, as well as a data acquisition (DAQ) and modem module 140 configured to acquire measurement data from a sensor interface circuit 150 of each of the sensor groups 1 lOa-e and transmit the measurement data to a remote server, described below with reference to Fig. 3.

[0024] Each of the sensor groups 1 lOa-e may include multiple sensors configured to measure different properties of an adjacent soil sample. For example, a piezoelectric acoustic sensor 112 may measure an acoustic impedance of the soil sample. The piezoelectric acoustic sensor 112 may include at least two acoustic transducers having different resonant frequencies (e.g., an acoustic sensor pair as illustrated in Fig. 1). A dielectric sensor 113 may measure a dielectric permittivity of the soil sample, and an electrical conductivity sensor 114 may measure an electrical conductivity of the soil sample. As shown, the dielectric sensor 113 and electrical conductivity sensor 114 may share a common electrode. The common electrode may one of- 4 -4251900. vl5362.1009001 three electrodes arranged along a length of the frame 105 of the device 100. The device 100 may further include a temperature sensor 115 configured to measure a temperature of the soil sample.

[0025] Conventional sensors developed for moisture and salinity measurements rely on no more than one principle. Moisture measurement and salinity measurement are interdependent, and they are co-dependent on soil type. Thus, conventional sensors cannot accurately measure either salinity or moisture content without site-specific calibration.

[0026] In contrast, by utilizing multiple different sensors, referred to as a sensor fusion technique, the device 100 can effectively measure both moisture content and salinity level, irrespective of soil composition. Thus, the device 100 can accurately measure seawater intrusion in soil and monitor its spatial-temporal variation.

[0027] The soil properties measured by the device 100 respond differently to the influencing factors of moisture content, salinity, and soil composition. The dielectric sensor 113, which measures dielectric permittivity at different frequencies, captures how well a material can store electrical energy in an electric field. Such sensors are less sensitive to the impact of salinity, especially at high measurement frequencies. Conversely, resistance-based sensors measuring electric resistance, such as the electrical conductivity sensor 114, capture how easily electricity can flow through a material. Such sensors are substantially sensitive to the salinity level in the soil matrix. The acoustic sensor 112, which is affected by sound speed and attenuation in a medium at different frequencies, can provide better insights on the soil structure and degree of saturation.

[0028] While individual measurement principles cannot yield accurate readings of moisture content or salinity, the integration of multiple measurements using a neural network (NN) based sensor fusion technique offers additional dimensions of information. By training with diverse data covering various combinations of salinity, moisture content, and soil composition, the deep fusion technique captures the interplay and trends arising from these three pivotal factors. An example of such training and configuration is described in further detail below with reference to Fig. 4. Thus, by leveraging multiple different sensor measurements and a trained NN, the device 100 can provide accurate measurements of soil salinity and moisture content, regardless of the soil composition, eliminating the requirement for calibration that is specific to each site.

[0029] Fig. 2 is an illustration of the device 100 deployed in a soil sample 205. As shown, the device may be inserted into the soil sample 205 at a depth such that each of the sensor groups 1 lOa-e may each measure a respective layer of the soil sample 205 at distinct depths. As a result,- 5 -4251900. vl5362.1009001 the device 100 may measure salinity and moisture content of the soil sample 205 at multiple different depths. The insert in Fig. 2 illustrates partial insertion of the device into soil.

[0030] Fig. 3 is a diagram of a system 300 comprising a plurality of sensor devices 300a-d. Each of the sensor devices 300a-d may incorporate some or all features of the sensor device 100 described above, and may be deployed remotely from one another to measure respective soil samples at distinct locations. A computer server 310 may include a wireless transceiver 315 configured to communicate with each of the sensor devices 300a-d to collect measurement data from the sensor devices 300a-d and store it to a measurement data store 325. A model data store 330 may store a NN model as described in further detail below. From the measurement data collected from one or more of the sensor devices 300a-d, a computer processor 320 may generate an input data set, which may include values from the sensor device(s) representing the dielectric permittivity, electrical conductivity, and acoustic impedance of the soil sample(s). The input data set may also include values from the sensor device(s) representing temperature. The processor 320 may then apply the input data set to the NN represented by the NN model and determine salinity and moisture content of the soil sample based on an output of the NN.

[0031] Fig. 4 illustrates a model architecture of a deep neural network (DNN) 400 incorporating learned probability distribution (LPD) in one embodiment. In operation, the DNN 400 can take input from multiple sensors with distinct measurement principles to determine soil salinity and moisture content. By combining three distinct measurement principles (i.e., electrical conductivity (EC), permittivity, and acoustic impedance) with varying sensitivities to salt content, water content, and soil stiffness or composition, the DNN 400 can extract information on the unique combination of these variables. Sample temperature may also be used as input to account for general fluctuations in readings due to ambient heat.

[0032] Electrical conductivity (pS / cm) refers to the ability of a material to conduct electricity, which is correlated to the amount of salt in soils. Furthermore, soil electrical conductivity changes with varying soil moisture content even at constant absolute salt content. Permittivity (F / m) is a material's ability to store electrical energy when an electric field is applied and is proportional to the amount of moisture in a soil mass, and it is also affected by soil salinity. Acoustic impedance (□) is a fundamental characteristic of any material medium and can be expressed as its bulk density (p) multiplied by the speed of a mechanical wave within it (c), i.e., Z = pc. More fundamentally, acoustic impedance is a measure of a material medium's resistance to acoustic particle velocity (u) for a given acoustic pressure (p), Z=p / u. Acoustic impedance is well correlated to soil moisture content and porosity, and invariably soil stiffness.- 6 -4251900. vl5362.1009001The performance of electrical conductivity, permittivity, and acoustic sensors can be affected by the host's ambient temperature, and, thus, temperature information can improve the accuracy of the determined soil salinity and moisture content.

[0033] To train the DNN 400, electrical conductivity, permittivity, and acoustic sensors may be used to collect data on the various soil combinations with varying salinity and volumetric water content (VWC). The data collected is used to train a model to predict soil salinity.

[0034] By incorporating LPD, the DNN 400 exhibits a model architecture that can process high-dimensional data and capture complex nonlinear relationships within the dataset. Based on the five input features as shown in Fig. 4, the architecture can predict a mean salinity (p) and the variance (c2) in the prediction. In the example illustrated in Fig. 4, the input layer includes data for electrical conductivity, permittivity, temperature, acoustic impedance at 140 KHz (Impl40KHz), and acoustic impedance at 340 KHz (Imp340KHz). These input features are represented as nodes in the input layer, each corresponding to a different measurable property of the soil.

[0035] As illustrated in Fig. 4, the DNN network 400 may include multiple hidden layers, each comprising a different number of neurons. Each neuron in one layer may be connected to every neuron in the subsequent layer. The first hidden layer includes 30 neurons (denoted as ai(1), a2(1), ..., a3o(1)); the second, 60 neurons (denoted as ai(2), a2(2), ..., aeo(3)); the third, 120 neurons (denoted as ai(3), a2(3), ..., ai2o(3)); the fourth, 240 neurons (denoted as ai(4), a2(4), ..., a24o(4)); and the fifth hidden layer, 100 neurons (denoted as ai(5), a2(5), ..., aioo(5)). The network's output layer is designed to provide two crucial pieces of information: the predicted mean value of the output and the predicted variance. The predicted variance may be connected to a softplus function, ensuring that the variance is a non-zero value. This configuration can prevent the network from predicting a zero variance, which could lead to inaccurate results. The hidden layers, in contrast, may use ReLU (Rectified Linear Unit) activation functions to address the issue of vanishing gradients and facilitate convergence. The dense connection in the network allows it to learn complex, non-linear relationships between the input features and the target outputs, enhancing its predictive capabilities.

[0036] There are two primary types of uncertainty: aleatoric and epistemic. Aleatoric uncertainty is the inherent variability in the data, while epistemic uncertainty captures the uncertainty of the model itself, which can be reduced with more training data. Aleatoric uncertainty can be modeled by providing that the output y for a given input x follows a normal distribution with a mean and variance <72, both dependent on the input x as shown in Eq.1 :- 7 -4251900. vl5362.1009001Eq. lwhere 6 are the parameters of the neural network. For N data points, the loss function, defined as negative log-likelihood (NLL), incorporates this variance, and mean, as shown in Eq.2:Eq.2

[0037] This formulation, referred to as heteroscedastic loss, allows the model to express different levels of uncertainty for various inputs. The variance in predictions across the ensemble can represent uncertainty by training multiple models from different initializations, architecture, and treating the ensemble as an approximation to the Bayesian posterior. The ensemble prediction and its variance are given by Eqs.3-5:E 3Where pmand a2mdenote the predicted mean and variance of the m-th model in the ensemble with size M. Both aleatoric and epistemic uncertainties are beneficial for understanding a model's predictions.

[0038] Fig. 5 is a flow diagram of a method 500 of determining soil salinity and moisture in one embodiment. With reference to Fig. 1, the sensor device 100, via one or more of the sensor groups HOa-e, may measure dielectric permittivity (505), electrical conductivity (510), and acoustic impedance (515) of a soil sample. This measurement date can optionally include temperature measured by the sensor device. The sensor device 100 may transmit this measurement data, via the modem 140 or other means, to a computer processor such as the server 310 of Fig. 3. The computer processor may then apply the measurement data as an input data set to a neural network (520). Salinity and moisture content of the soil sample may then be determined based on an output of the neural network (525).4251900. vl5362.1009001

[0039] Results in Example Embodiment

[0040] Figs. 6 and 7A-B illustrate soil salinity prediction results of a real-world embodiment, comparable to the embodiments described above, in comparison with true salinity values of the measured soil. Fig. 6 is a graph illustrating model predictions plotted against true salinity values. Using a random initialization approach, ten models with identical architecture and model parameters are trained with random weight initializations, and the best-performing models are selected to form the ensemble. In this example, the top four models are used to form the ensemble. The selection criterion is a DeepHyper ensemble function that prioritizes models with the lowest validation loss across all outputs for inclusion, ensuring that only the top-performing models contribute to the final ensemble's predictions. The ensemble is used to make predictions and determine the uncertainties of the data and model. The test set contained 479 data Points. The model predictions are plotted against the ground truth values, as shown in Fig. 6. The high R2value (0.997) suggests that the model explains most of the variability in the data. The low root mean square error (RMSE) of 0.817 indicates that the model's predictions are very close to the actual values. The mean absolute percentage error (MAPE) of 1.584% further reflects the model's good performance.

[0041] The ensemble modeling approach demonstrates excellent predictive performance. It implies that the model can capture the underlying patterns and variability in the data and generalize to unseen data within the sample space. Furthermore, regarding the predicted versus ground truth salinity, the data's total uncertainty (c) is more significant at mid-high range salinity values (20 - 40 g / 1).

[0042] Figs. 7A-B are graphs illustrating uncertainties in the predicted salinity of a model based on true values in one embodiment. Each figure includes two subplots: a heatmap on the left, and a curve plot on the right. The heatmap displays the average uncertainty in the predicted response as a function of the soil's ground truth volumetric water content (VWC). The plot illustrates how VWC affects the model's salinity response across different soil types. It suggests that within a certain range of saturation, the predicted salinity response of soils is relatively uncertain (widely spread), given the model inputs. This uncertainty may arise because the inputs, although originating from different soils and corresponding to different salinity levels, are relatively similar in this region. Each cell in the heatmap is color-coded to represent the magnitude of the uncertainty (c) associated with the predictions, with darker shades of blue indicating higher uncertainty. The average uncertainty values are also numerically indicated within each cell. The curve-plot on the right in each figure illustrates the average uncertainty (c)- 9 -4251900. vl5362.1009001 as a function of predicted response, i.e., across all ground truth values magnitudes. It provides a continuous representation of how the average uncertainty varies with the predicted response.

[0043] Fig. 7A visualizes the aleatoric uncertainty in the ensemble model (Eq.4) predicting salinity across various ground truth VWCs. The aleatoric uncertainty (c) values range from 0.19 to 1.40. The higher data uncertainty is observed at the dryer region (0 - 0.28 m3 / m3), while relatively lower at the wetter region (0.28 - 0.7 m3 / m3) across all salinity values. This suggests that data quality improvement is required within this region, as sensors cannot adequately function in dry soils. Alternatively, other modeling technique may be used to consider the data uncertainties. The curve plot provides a continuous representation of how data uncertainty varies with predicted VWC.

[0044] The epistemic uncertainty of the ensemble, as described in Eq.5 for the predicted salinities, ranges from 0.05 to 1.17, and it is visualized in Fig. 7B. As a result of the data uncertainties, higher uncertainty is observed at lower VWCs (0 - 0.28 m3 / m3) within the salinity range of 8 to 40 g / 1). This is because the model cannot adequately capture the dispersion within the data set in this region. Lower uncertainty is observed as VWC increases beyond 0.28 m3 / m3, indicating that the model predictions are most reliable in this range. As shown, there is higher uncertainty at the higher end of the predicted salinity range. This suggests some variability in model confidence for very high salinity predictions, although the overall uncertainty remains within an acceptable range.

[0045] Using sensor fusion enabled by ensemble modeling to predict salinity, example embodiments demonstrate high reliability for moderate-high salinity soils. Such performance benefits SWI applications, as predictions can be made with greater confidence. Example embodiments further provide a multimodal sensing approach that includes specific calibrations for accurate soil salinity characterization, which is essential for informed environmental management and agricultural decision-making.

[0046] While example embodiments have been particularly shown and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments encompassed by the appended claims.- 10 -4251900. vl

Claims

5362.1009001CLAIMSWhat is claimed is:

1. A device for determining soil salinity and moisture, the device comprising: a dielectric sensor configured to measure a dielectric permittivity of a soil sample; an electrical conductivity sensor configured to measure an electrical conductivity of the soil sample; a piezoelectric acoustic sensor configured to measure an acoustic impedance of the soil sample; and a frame configured to house the dielectric sensor, the electrical conductivity sensor, and the piezoelectric acoustic sensor, the frame adapted to be inserted into the soil sample.

2. The device of claim 1, further comprising a temperature sensor configured to measure a temperature of the soil sample.

3. The device of claim 1 or 2, further comprising a transmitter configured to transmit measurement data from the dielectric sensor, the electrical conductivity sensor, and the piezoelectric acoustic sensor to a remote server.

4. The device of any one of claims 1-3, wherein the dielectric sensor and the electrical conductivity sensor share at least one common electrode.

5. The device of any one of claims 1-4, wherein the piezoelectric acoustic sensor includes at least two acoustic transducers having different resonant frequencies.

6. A system for determining soil salinity and moisture, the system comprising: a sensor device comprising: a dielectric sensor configured to measure a dielectric permittivity of a soil sample; a electrical conductivity sensor configured to measure an electrical conductivity of the soil sample; and a piezoelectric acoustic sensor configured to measure an acoustic impedance of the soil sample; and a computer processor configured to:- 11 -4251900. vl5362.1009001 apply an input data set to a neural network, the input data set including values from the sensor device representing the dielectric permittivity, electrical conductivity, and acoustic impedance of the soil sample; and determine salinity and moisture content of the soil sample based on an output of the neural network.

7. The system of claim 6, wherein the neural network is trained via a reference data set including corresponding values of dielectric permittivity, electrical conductivity, acoustic impedance, salinity, and moisture of a plurality of reference soil samples.

8. The system of claim 6 or 7, wherein the processor is further configured to generate a confidence score based on the output of the neural network, the confidence score indicating an estimated accuracy of the determined salinity and moisture content.

9. The system of any one of claims 6-8, wherein the sensor device further includes a temperature sensor, and wherein the input data set further includes a value representing a temperature of the soil sample.

10. A method of determining soil salinity and moisture, comprising: measuring dielectric permittivity of a soil sample; measuring an electrical conductivity of the soil sample; measuring an acoustic impedance of the soil sample; applying an input data set to a neural network, the input data set including values representing the measured dielectric permittivity, measured electrical conductivity, and measured acoustic impedance of the soil sample; and determining salinity and moisture content of the soil sample based on an output of the neural network.

11. The method of claim 10, further comprising training the neural network via a reference data set including corresponding values of dielectric permittivity, electrical conductivity, acoustic impedance, salinity, and moisture of a plurality of reference soil samples.

12. The method of claim 10 or 11, further comprising generating an accuracy score based on the output of the neural network, the accuracy score indicating an estimated accuracy of the determined salinity and moisture content.- 12 -4251900. vl5362.100900113. The method of any one of claims 10-12, further comprising measuring a temperature of the soil sample, wherein the input data set further includes a value representing a temperature of the soil sample.

14. The method of any one of claims 10-13, wherein measuring the acoustic impedance of the soil sample includes measuring the acoustic impedance at a minimum of two distinct acoustic frequencies.4251900. vl

Images