smol: Sensing Soil Moisture using LoRaDaniel Kiv∗, Garvita Allabadi∗, Berkay Kaplan, Robin Kravets

{dkiv2,garvita4,berkayk2,rhk}@illinois.eduUniversity of Illinois at Urbana-Champaign

ABSTRACTTechnologies for environmental and agricultural monitoring areon the rise, however there is a lack of small, low-power, and low-cost sensing devices in the industry. One of these monitoring toolsis a soil moisture sensor. Soil moisture has significant effects oncrop health and yield, but commercial monitors are very expensive,require manual use, or constant attention. This calls for a simpleand low cost solution based on a novel technology.

In this work we introduce smol: Sensing Soil Moisture usingLoRa, a low-cost system to measure soil moisture using received sig-nal strength indicator (RSSI) and transmission power. It is compactand can be deployed in the field to collect data automatically withlittle manual intervention. Our design is enabled by the phenom-enon that soil moisture attenuates wireless signals, so the signalstrength between a transmitter-receiver pair decreases. We exploitthis physical property to determine the variation in soil moisture.We designed and tested our measurement-based prototype in bothindoor and outdoor environments. With proper regression calibra-tion, we show soil moisture can be predicted using LoRa parameters.

CCS CONCEPTS•Applied computing→ Physical sciences and engineering; •Hard-ware→ Sensor applications and deployments.

KEYWORDSinternet of things, lora, sensing, soilACM Reference Format:Daniel Kiv∗, Garvita Allabadi∗, Berkay Kaplan, Robin Kravets. 2022. smol:Sensing Soil Moisture using LoRa. In 1st ACM Workshop on No Power andLow Power Internet-of-Things (LP-IoT’21), January 31-February 4 2022, NewOrleans, LA, USA. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3477085.3478991

1 INTRODUCTIONData-driven agriculture techniques like precision agriculture promiseto cut input costs (e.g. reduce fertilizer requirement), improve yields,and improve sustainability in agriculture [15, 23]. These techniquesrely on accurate estimates of soil and plant properties, e.g. soilmoisture, soil pH, crop density, leaf wetness, etc. using specialized∗ Equal contribution.

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than theauthor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from permissions@acm.org.LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA© 2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.ACM ISBN 978-1-4503-8702-6/22/01. . . $15.00https://doi.org/10.1145/3477085.3478991

sensors. Such sensors cost between hundreds and thousands ofdollars each and become expensive to deploy at scale.1

To avoid the cost of sensors, there have been multiple attemptsat using radio signals to infer soil properties like soil moisture.Essentially, radio signals from already deployed sensors (or justwireless transmitters) are used to infer additional soil properties,reducing the cost of sensing. As RF signals travel through soil, theirproperties, like attenuation and phase, change as a function of soilproperties. Therefore, past work [8, 27] has shown that it is possibleto measure soil moisture and electrical conductivity by measuringthe characteristics of the wireless signal itself.

Past work has used Wi-Fi [8] and RFIDs [27] for such analysis.We argue that such a design is infeasible for large scale deployments.First, Wi-Fi and RFIDs have limited range and, therefore, are nottypically used in farm networks or for connecting farm sensors.Second, these designs rely on access to phase information of thesignal which requires specialized hardware. FewWi-Fi devicesmakethis information available to application developers and RFIDs needspecialized hardware to read the tags, let alone read phase values.

In this paper, we investigate the use of LoRa, a low-power long-range communication protocol, for sensing soil moisture. LoRa,given its low-power and long-range capabilities, is emerging asthe front-runner for connecting Internet-of-things devices in out-door environments [6]. Therefore, we do not need to deploy anyadditional hardware and can use devices with LoRa connectivityalready present on the farms. Furthermore, we base our design tothe use of RSSI (Received Signal Strength Indicator) parametersreported by commercial LoRa chipsets present in both gatewaysand end devices.

We rely on LoRa devices already deployed in the farm and em-bedded in the soil. A LoRa receiver outside the soil measures thesignal from such devices and estimates the signal strength. Intu-itively, our design relies on the physical principle that wirelesssignals become progressively weaker as they traverse moist soil.The signal strength loss is more for moist soil as opposed to dry soil.Therefore, we should be able to use RSSI to identify soil moisturelevels. The use of LoRa and RSSI offer us several advantages – LoRais low power, so sensors do not need frequent battery replacements.LoRa is ubiquitous, so our design does not need additional hard-ware. Finally, RSSI is an easy to use metric and available on multipledevices.

However, as compared to past work, restricting our measure-ments to signal strength indicators reduces the amount of avail-able information. The phase values are more informative and havebeen used in most recent papers on wireless sensing for this rea-son [4, 14, 20, 24]. To make up for the lack of phase information,we rely on computational tools and multiple measurements. Specif-ically, LoRa devices can transmit at multiple power levels. Low1Some hobbyist sensors cost tens of dollars each, but they are typically not used asagricultural sensors due to their limited accuracy.

arX

iv:2

110.

0150

1v1

[cs

.LO

] 4

Oct

202

1

LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA Daniel Kiv∗ , Garvita Allabadi∗ , Berkay Kaplan, Robin Kravets

power levels support longer battery life, while the higher onesare needed for long range connections. We leverage this ability toperform multiple measurements of the signal strength at differentpower levels. We, then, use regression techniques to identify themoisture value from the multiple signal strength measurements.

In this paper, we make the following contributions:• We quantify the relationship between soil moisture and sig-nal strength using a cheap, off-the-shelf LoRa chipset.

• We evaluate multiple setups (indoors vs outdoors, differentheights) for measuring soil moisture based on LoRa signalstrength.

• We leverage multiple transmit powers of LoRa devices toimprove soil moisture estimation.

We implemented our design using the Adafruit M0 RFM9x 915MHzchipsets, which supports LoRa, and evaluated our design in multiplesoil types and experimental settings – soil in pots (indoors), farms(outdoors), etc. Our results demonstrate:

• LoRa received signal strength correlates with volumetricwater content, as measured by our ground truth.

• Machine learning regression models can map RSSI and trans-mission power to soil moisture.

We envision that future iterations of our design will integratewith robots and drones to perform continuous soil moisture mea-surements. We also aim to leverage the motion of these drones orrobots to improve our soil moisture estimates.

2 RELATEDWORKDifferentmetrics formeasuring soil health include electrical conduc-tivity. In recent years, wireless signals have been used in varioussensing applications. In environmental and agricultural applica-tions, wireless signals can be used to examine soil health with thebasic premise that signal attenuation varies with different media[8, 26]. However, specialized devices to measure these metrics, thatare currently employed commercially, are relatively expensive (sub-1000 USD) [8].

WiFi time-of-flight (ToF) can be used to estimate electrical con-ductivity and soil moisture [8]. The bandwidth problem arisesmainly from the system using time-of-flight as it requires highbandwidth from 100s of Mhz to few Ghz [8]. Strobe maps the prop-agation time and amplitude of WiFi signals received by differentantennas to electrical conductivity, which depends on the soil mois-ture property [8]. Using this method, strobe can accurately estimatethe soil moisture up to an error margin of 15% [8]. A limitation isthat to achieve good performance, ToF systems require ultra-widebandwidth, which needs specialized equipment that will ramp upthe cost to thousands of dollars [8].

Previous research has also explored the relationship betweenRSSI and moisture level in the passive RFID domain. GreenTag [27]explores the relationship between RSSI and soil moisture by usingthe reflection of the soil embedded tags’ Differential Minimum Re-sponse Threshold (DMRT) value [27]. Researchers embed every taginto a plant pot, and the tags use a low-pass filtered DMRT to showthat the DMRT is subject to changes in radio frequency environ-ments and in the pot locations as water is added [27]. However, theshortcoming of this method is that the RFID reader is customizedand expensive as like the previous example [27].

Previous research using transmission power to aid in sensing islimited. One study leveraged transmission power, along with RSSI,in order to increase localization accuracy for wireless nodes [5].Along a similar vein of RF-based localization, another study createdPowerscan which switches through different transmission powersand records different RSSI values [17]. The authors noticed a fixedRSSI offset between transmission powers had some saturation atlow distances, which led to their decision to select a near mediantransmission power. One study models wireless sensor networksand investigates wave propagation of underground networks byexamining the effect of volumetric water content on power density.They find that soil moisture attenuates wireless signals and de-creases the power. However, their implementation doesn’t examinethe affects in an underground to above-ground environment [9].Finally, one study investigates LoRa link quality for undergroundto above-ground transmissions, but the signal strength was notmapped to soil moisture [21].

While there is some work mapping RSSI to another feature usingmodels, such as machine learning[3, 7], none blend the elementsof LoRa, soil moisture, and transmission power. In this paper, weleverage LoRa and simple easy to access metrics using a variety ofmodels to attain an estimate of soil moisture.

3 DESIGNsmol is a LoRa based soil moisture sensing system. It uses twoLoRa transceivers. The transmitter device is placed inside the soil,while the receiver can be at varying heights from the soil bed, inair. A Raspberry Pi hosts the communication and data collectionfor the devices. Figure 5 shows a depiction for one of the usescase of smol. The transmitter sweeps through the different powerlevels on the device and broadcasts the power information to thereceiver. The receiver records the Received Signal Strength and thepower information. Once the data is collected, smol leverages thisdata for soil moisture estimation using regression models. Figure 1describes the two modes of operation for smol. In the training mode,the receiver combines the measured values with the ground truthsupplied by an industry scale moisture sensor to train regressionmodels. While in the test mode, it uses this information to predictthe Volumetric Water Content (VWC) of the soil using the trainedregression model.

We discuss the details of the key features and design choices forsmol in the subsequent sections.

3.1 Long Range (LoRa)LoRa uses Chirp Spread Spectrum (CSS) modulation, a spread spec-trum technique where the wireless signal is modulated by chirppulses (frequency varying sinusoidal pulses) hence improving re-silience and robustness against the Doppler and multi-path effect[2]. Previous work identifies LoRa as more stable compared to WiFior BLE in terms of higher adaptability to the changes in the environ-ment [18]. However, LoRa is not resistant to the variables that couldaffect the signal from an underground node to an above-groundnode. Some of these include attenuation, reflection, and multi-pathfading [25].

The radio can be modulated by a variety of parameters suchas transmission power, spread factor, and bandwidth [21]. Spread

smol: Sensing Soil Moisture using LoRa LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA

(a) Training Mode (b) Inference Mode

Figure 1: smol Software Design: In the training phase, the re-ceiver measures the RSSI of the signal and the ground truth.smol uses this data to train a regression model. In the testphase, the trainedmodel predicts VolumetricWater Content(VWC) from RSSI measurements.

factor determines the number of chirps per second. A higher spreadfactor increases the sensitivity and energy cost. Bandwidth repre-sents the frequency range of the signals. A higher bandwidth yieldsa higher data rate but also at the cost of distance. Transmissionpower affects the communication performance and range of the sig-nal at the cost of energy. That is, a higher transmission power yieldsa lower signal-to-noise ratio and higher range but also requiresmore energy. At the opposite end, we can save energy but havea lower range. By sweeping across the full range of transmissionpowers, multiple measurements can be taken for some soil moisturelevel. Instead of selecting a single transmission power, such as thehighest one, a balance between the two extremes can be reached.

LoRa typically operates in the sub-GHz frequency band makingit ideal for long range transmissions. A higher frequency bandalso ensures that LoRa has deeper soil penetration. All these factorsmake LoRa a good choice for using in a soil moisture sensing system.

3.2 Ground Truth MeasurementsTypically soil is composed of soil, water and, air. The Volumet-ric Water Content (VWC) of soil is the ratio of volume of waterin a given volume of soil to the total soil volume. This value isrepresented as a decimal or percentage.

For our ground truth values, we use an industry scale soil mois-ture sensor - FieldScout TDR 300 Soil Moisture Meter [11]. Thesensor uses Time Domain Reflectometry (TDR) to measure theVWC of soil. The underlying principal of TDR involves measuringthe travel time of an electromagnetic wave along a wave-guide[12, 19]. The speed of the wave in soil is dependent on the relative

(a) RSSI decreases with volumetric water content. Air (bar on the left) andwater (bar on the right) are provided as baselines. The ground truth was cali-brated against tap water.

(b) All possible transmission power supported by the transceiver.

Figure 2: Initial indoor bucket tests with the receiver on topof the soil and the transmitter buried 15 cm.

permittivity of the soil. Recall that relative permittivity or the di-electric permittivity constant (Yr), is the ratio of the permittivity ofthe substance to the permittivity of free space, Y0.

Yr =Y

Y0(1)

The fact that water (Yr = 80) has a much greater dielectric con-stant than air (Yr = 1) or soil solids (Yr = 3-7) is exploited to determinethe VWC of the soil. The VWC measured by TDR is an averageover the length of the waveguide.

Wemake two observations here. First, the ground truth sensor re-lies on the same principle which guides wireless signal attenuation.Wireless signal attenuation depends on the electric permittivityof the medium – in this case, soil. As soil moisture increases, theelectric permittivity increases and the signal strength decreasesfor typical wireless transmissions. Therefore, we are relying on aproxy measure for our system similar to traditional sensor. Second,the ground truth sensor costs approximately 1500 US dollars persensor for errors of ±3%.

LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA Daniel Kiv∗ , Garvita Allabadi∗ , Berkay Kaplan, Robin Kravets

3.3 Calibration using regression modelsIn order to map the features of the transceiver to soil moisture, thedevices need to be in the communication range to exchange packets.LoRa works fairly long range compared to other communicationdevices, even in an underground setting [26]. However, most ofthe experiments we perform will have a distance of less than 300cm between the two transceivers. After setting up the distance, wecalibrate the ground truth instrument. The 0% value is calibratedto air while 100%, full moisture, is calibrated to water.

Once the ground truth is calibrated, we begin the calibrationprocess for the LoRa transceivers. In the experiments, a transmit-ter is buried 15 cm in the soil and the receiver is vertically placedat a height at or above the soil. At each level of moisture, we in-crementally increase the VWC by moderately adding water. Theground truth reading is recorded by taking an average of 10 mea-surements in various spots around the area where the transmitterwas buried. Subsequently, a Python script runs to collect data ateach step. During the process, the transmitter sweeps through the18 available power levels, ranging from 5 to 23 dBm, broadcasting asingle packet at each level. Each packet contains information aboutthe transmission power. One can also embed a unique identifier intothe packet for multiple devices. The receiver receives the packetand records the transmission power level and RSSI.

After collecting the data, a plot will look similar to Figure 4where there is a visible trend between RSSI and VWC. It is evidentthat the soil moisture is within the baseline calibration we set forthe ground truth.

smol then fits a model with the RSSI and transmission power asinputs in order to predict the VWC. The models are trained using80% of the dataset and evaluated using the remaining 20% of thedataset. A variety of regression models are used including linearregression, ridge regression, and random forest. Performances arecompared in the results section. To make things more accessibleand easy for the end user, smol implementation uses Python andthe scikit-learn library [22] for calibration. Machine learning andregression is used in a variety of calibration situations [27, 28]. Oncethe best fit curve is constructed and the best algorithm is chosenbased on the metrics, RSSI and transmission can be continuouslymapped to VWC.

4 IMPLEMENTATIONWe implement smol using two Adafruit M0 RFM9x 915MHz LoRatransceivers that use RSSI as the main metric [1]. One device actsas a transmitter while the other acts as a receiver. We also addeda 7.5cm antenna to each of the devices to increase the range. Thechoice of hardware for our experiments was based on the avail-able radio frequency bands, range of transmission powers, ease ofprogramming and finally the cost of the device. Apart from thesedevices, smol also uses a Raspberry Pi 4 for communicating withthe LoRa devices and collecting the data.

smol is an easy to use, low-cost system to measure the soil mois-ture content. It uses a variety of machine learning models like linearregression, ridge regression, and random forests to train and predictthe moisture content of soil. For this purpose, it measures threevariables:

(a) Vertical setup in thelab.

(b) Vertical setup at re-search field.

(c) Transmitter placed15 cm in a hole.

Figure 3: Experimental setups

• Received signal strength indicator (RSSI). The primaryquantity of interest for correlating with soil moisture.

• Transmission power from the transmitter. The lowerthe transmission power, the lower the signal strength andpossible range. The higher the transmission power, the higherthe signal strength and possible range. This value can varyfrom 5 to 23 dBm. In the results, we mainly look at 5 to 22,since 23 dBm appears to wrap around to 5. This parametercan be easily modified with the provided Adafruit libraries.

• Volumetric water content. This is the ground truth valueas measured by an industry scale soil moisture sensor.

5 EXPERIMENTAL SETUPThe goal of our experiments is to examine how RSSI is affected bydifferent levels of soil moisture and the distance between the sensors.We evaluate our system in two setups - inside the laboratory andin a real world scenario outside on a farm.

In the lab experiment, the transmitter was placed at the bottomof the bucket, 15cm in the soil. The receiver was placed at heights0, 195 and 265cm away from the soil. It was attached to the arm ofthe PVC pipe so that it would be directly above the transmitter. Thesetup is depicted in Figure 3a. In this way, it could emulate the waythat a drone or vehicles passes over the area to take a measurement,similar to Figure 5. The devices were faced towards each other. Thesoil used in the lab was Miracle Gro Potting Mix [16].

At the farms, we dug a hole approximately 15cm in an area at theplot. We secured the PVC pipes into the ground as shown in Figure3b. The transmitter was placed inside the hole and then buried. Thiswas done at the Illinois Autonomous Farm [10].

The transmitter was wrapped inside a plastic sandwich bag. Thebag was twisted and taped in order to prevent water from damagingthe electronics. Using a plastic bag is a low-cost way to protectelectronics.2 Special care was taken so that the front facing part ofthe device had a single layer of the plastic bag in the line of sightbetween the transceivers. A visualization of the experimental setupcan be seen in Figure 3c.

The soil was artificially moistened with water. This was doneby sprinkling about 16 ounces of water in the soil. A few minutes

2No devices were harmed in the making of this project.

smol: Sensing Soil Moisture using LoRa LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA

(a) Laboratory measurements.

(b) In-situ measurements at the research farm.

Figure 4: Laboratory and in-situ results of our experiments.Each point on the graph represents the mean RSSI of themedian transmission power. Each line is the height abovethe top of the soil that the recording occurred at.

were given to let the water settle, so that all the moisture was notat the top.

6 RESULTS6.1 Laboratory and in-situFigure 4 shows both the laboratory and in-situ results of our exper-iments. Each point on the graph represents the mean RSSI of themedian transmission power. Each line is the height above the topof the soil that the recording occurred at. Since the offsets of thetransmission powers, as shown in Figure 2, are generally equal, theoverall curve would look similar. Selecting a different transmissionpower other than the median would just result in an offset.

Two baselines were obtained in air and only water as shown inFigure 2. As expected, with no obstructions in the line of sight, thesignal strength was at its highest. In all water, the signal strengthwas at its lowest. Surprisingly, there was a decent signal strength

even when completely submerged in water, which wireless signalstend to travel poorly through.

During the laboratory experiments, particularly for higher heightsand VWC, the multi-path effect may affect RSSI readings. This isbecause the radius of the bottom of the pot, where the transmitterwas placed, is smaller than the distance to the top of the soil fromthe bottom. The signal may take the path along the bottom radiusand through the air, which could be faster than travelling straightthrough the soil.

For comparison, we also performed the experiment in-situ at theresearch farms of the University of Illinois at Urbana-Champaign.The multi-path effect would be reduced since the signals wouldhave to take a route through the soil, since all around is dirt. TheRSSI for the outdoor experiments for the same height are slightlylower as a result. For this experiment, we watered a larger areaaround where the transmitter was buried in order to better simulaterain in an open field.

6.2 Regression modelsWe evaluated and compared a variety of regression models. Weuse a random forest, polynomial with degree more than and equalto 2 (>= 2), and a linear regression. In the models appended with"all TX powers," the inputs are transmission power and RSSI. Inmodels appended with "median TX power," the input is simply theRSSI from the median TX power. The output is the VWC. Basically,we are seeing how much of a performance gain we can get fromadding an additional feature, the transmission power. Table 1 showsa comparison of the calibration models we trained and evaluated.Each model in the table provides the coefficient of determination(𝑅2) and mean absolute error (MAE). In general, higher 𝑅2 valueand a lower MAE value is better.

We can see that the random forest algorithm outperforms theother regression models, since it is better suited for multiple fea-tures. When comparing the two random forest models, we can seethat having transmission power along with RSSI as an input canimprove performance. Since transmission powers also scale withsignal strength, devices don’t need to be hard-coded with a singletransmission power. Taking the highest transmission power is as-sociated with higher energy cost. Taking the lowest transmissionpower causes the signal strength to be weak, which also means themaximum distance is shorter. By taking a spectrum, we can coverthe entire range while also having comparable performance to justtaking a single transmission power.

7 FUTUREWORK AND LIMITATIONSsmol is a first step to leverage the strengths of LoRa for agriculturalmonitoring. In this paper, we have shown experiments in a con-trolled setting, therefore there are some caveats that would requirefurther investigation to make smol more robust against factors thatcan affect the measurement.

An additional experiment can discuss the impact of differentsoil types such as sandy soil or clay soil etc. on the RSSI valuesfor LoRa. It would be interesting to see the results of smol for longscale continuous deployments in outdoor settings. One possibleoutcome for the system is to detect a rainy or dry day based on thesignal strength. This can help in large scale deployment of smol,

LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA Daniel Kiv∗ , Garvita Allabadi∗ , Berkay Kaplan, Robin Kravets

Model 𝑅2 Mean Absolute Error (MAE)Random Forest w/ all TX powers 0.92 1.63Random Forest w/ median TX power 0.90 0.94Polynomial w/ all TX powers 0.80 4.43Polynomial w/ median TX power 0.88 1.45Linear Regression w/ all TX powers 0.18 7.66Linear Regression w/ median TX power 0.88 3.23

Table 1: Comparison of model performances.

making the system more easily accessible to those who have LoRadevices but are on a budget and want to deploy additional systemsin their fields.

LoRa is known for long ranges. A study investigating how thesignal strength varies at different soil depths, distances, and soilmoisture would be interesting. Comparing the performance of mea-suring soil moisture using different wireless signals would be ofgreat benefit. Each wireless technology has their own pros and cons.Depending on the situation, an optimal technology can be used tosuit the user’s intent.

Since LoRa signals were able to be received, by some extent,through water, it would be interesting to investigate the feasibilityof using LoRa underwater. A work detailing the signal strength atdifferent distances underwater could pave the way for a networkthat integrates with existing infrastructure, such as sound. Alongthe same vein, localization using LoRa signals underwater wouldbe quite interesting.

Our setup calibrates each LoRa device that is used in the field.This may be a time consuming process, therefore work on improv-ing the calibration process and generalizing to multiple deviceswould be of great benefit. It is important to note that the the re-gression model for the device was trained using data collectingin a controlled environment. There was no rain, foliage, or anysignificant obstruction between the devices. It is possible that raincould cause the air to become moist and shuffle the soil, whichcould affect the estimation of volumetric water content since the aircould moist and occupy the space between the soil and the device.We leave this discussion for future work.

8 CONCLUSIONThis paper presents smol, a low-power and low-cost LoRa basedsystem to estimate the volumetric water content of soil. It leveragesthe received signal strength and transmission power informationto estimate soil moisture. Our indoor and outdoor experimentsdemonstrate the correlation between RSSI and actual VWC. Resultsshowed that smol predicts the soil moisture levels accurately usingtrained regression models.

9 ACKNOWLEDGEMENTSThanks to Deepak Vasisht, Robin Kravets, and Vikram Adve forhelpful discussions and providing hardware to support this work.We alsowant to thank the leadership at the Illinois Center for DigitalAgriculture [13] for granting us access to the Illinois AutonomousFarm [10] to conduct our experiments.

Figure 5: An illustration of a possible extension for smol,where a drone could take measurements based on the wire-less signal strength from LoRa. This requires consideringthe movement of the drone, multi-path, obstructions, andother factors.

REFERENCES[1] Adafruit. 2016. Adafruit Feather M0 with RFM95 LoRa Radio - 900MHz - Ra-

dioFruit. https://www.adafruit.com/product/3178. Accessed on 04.30.2021.[2] Ferran Adelantado, Xavier Vilajosana, Pere Tuset-Peiro, Borja Martinez, Joan

Melia-Segui, and ThomasWatteyne. 2017. Understanding the limits of LoRaWAN.IEEE Communications magazine 55, 9 (2017), 34–40.

[3] Mahnoor Anjum, Muhammad Abdullah Khan, Syed Ali Hassan, Aamir Mahmood,Hassaan Khaliq Qureshi, and Mikael Gidlund. 2020. RSSI Fingerprinting-BasedLocalization Using Machine Learning in LoRa Networks. IEEE Internet of ThingsMagazine 3, 4 (2020), 53–59. https://doi.org/10.1109/IOTM.0001.2000019

[4] Atul Bansal, Akshay Gadre, Vaibhav Singh, Anthony Rowe, Bob Iannucci, andSwarun Kumar. 2021. OwLL: Accurate LoRa Localization Using the TV Whites-paces. In Proceedings of the 20th International Conference on Information Processingin Sensor Networks (Co-Located with CPS-IoT Week 2021) (Nashville, TN, USA)(IPSN ’21). Association for Computing Machinery, New York, NY, USA, 148–162.https://doi.org/10.1145/3412382.3458263

[5] Martin Bor and Utz Roedig. 2017. LoRa Transmission Parameter Selection. In2017 13th International Conference on Distributed Computing in Sensor Systems(DCOSS). 27–34. https://doi.org/10.1109/DCOSS.2017.10

[6] Martin Bor, John Edward Vidler, and Utz Roedig. 2016. LoRa for the Internet ofThings. (2016).

[7] Ioannis Daramouskas, Vaggelis Kapoulas, and Michael Paraskevas. 2019. UsingNeural Networks for RSSI Location Estimation in LoRa Networks. In 2019 10thInternational Conference on Information, Intelligence, Systems and Applications(IISA). 1–7. https://doi.org/10.1109/IISA.2019.8900742

[8] Jian Ding and Ranveer Chandra. 2019. Towards Low Cost Soil Sensing UsingWi-Fi. In The 25th Annual International Conference on Mobile Computing and

smol: Sensing Soil Moisture using LoRa LP-IoT’21, January 31-February 4 2022, New Orleans, LA, USA

Networking (Los Cabos, Mexico) (MobiCom ’19). Association for ComputingMachinery, New York, NY, USA, Article 39, 16 pages. https://doi.org/10.1145/3300061.3345440

[9] Xin Dong and Mehmet C. Vuran. 2011. A Channel Model for Wireless Under-ground Sensor Networks Using Lateral Waves. In 2011 IEEE Global Telecommuni-cations Conference - GLOBECOM 2011. 1–6. https://doi.org/10.1109/GLOCOM.2011.6134437

[10] Illinois Autonomous Farm. 2021. Illinois Autonomous Farm. https://digitalag.illinois.edu/autonomous-farm/.

[11] Fieldscout. [n.d.]. FieldScout TDR 350 Soil Moisture Meter. https://www.specmeters.com/soil-and-water/soil-moisture/fieldscout-tdr-meters/fieldscout-tdr-350-soil-moisture-meter-with-case/.

[12] Fieldscout. 2021. FieldScout TDR 350 Soil Moisture Meter Product Manual.https://www.specmeters.com/assets/1/22/6430TDR300_(web).pdf.

[13] Center for Digital Agriculture. 2021. Center for Digital Agriculture. https://digitalag.illinois.edu/.

[14] Jon Gjengset, Jie Xiong, Graeme McPhillips, and Kyle Jamieson. 2014. Phaser:Enabling Phased Array Signal Processing on Commodity WiFi Access Points. InProceedings of the 20th Annual International Conference on Mobile Computing andNetworking (Maui, Hawaii, USA) (MobiCom ’14). Association for Computing Ma-chinery, New York, NY, USA, 153–164. https://doi.org/10.1145/2639108.2639139

[15] H. Charles J. Godfray, John R. Beddington, Ian R. Crute, Lawrence Haddad, DavidLawrence, James F. Muir, Jules Pretty, Sherman Robinson, Sandy M. Thomas,and Camilla Toulmin. 2010. Food Security: The Challenge of Feeding 9 BillionPeople. Science 327, 5967 (2010), 812–818. https://doi.org/10.1126/science.1185383arXiv:https://science.sciencemag.org/content/327/5967/812.full.pdf

[16] Miracle Gro. 2021. Miracle-Gro® Potting Mix. https://www.miraclegro.com/en-us/products/soils/miracle-gro-potting-mix.

[17] Wolfgang Hönig and Nitin Kamra. 2016. RF-Based Relative Localization for RobotSwarms.

[18] Bashima Islam, Md Tamzeed Islam, Jasleen Kaur, and Shahriar Nirjon. 2019.Lorain: Making a case for lora in indoor localization. In 2019 IEEE InternationalConference on Pervasive Computing and Communications Workshops (PerComWorkshops). IEEE, 423–426.

[19] Colleen Josephson, Bradley Barnhart, Sachin Katti, Keith Winstein, and Ran-veer Chandra. 2019. RF Soil Moisture Sensing via Radar Backscatter Tags.arXiv:1912.12382 [eess.SP]

[20] Manikanta Kotaru, Kiran Joshi, Dinesh Bharadia, and Sachin Katti. 2015. SpotFi:Decimeter Level Localization Using WiFi. In Proceedings of the 2015 ACM Confer-ence on Special Interest Group on Data Communication (London, United Kingdom)(SIGCOMM ’15). Association for Computing Machinery, New York, NY, USA,269–282. https://doi.org/10.1145/2785956.2787487

[21] Kaiqiang Lin, Tong Hao, Zhouwei Yu, Wuan Zheng, and Wenchao He. 2019. APreliminary Study of UG2AG Link Quality in LoRa-based Wireless UndergroundSensor Networks. In 2019 IEEE 44th Conference on Local Computer Networks (LCN).51–59. https://doi.org/10.1109/LCN44214.2019.8990756

[22] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour-napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: MachineLearning in Python. Journal of Machine Learning Research 12 (2011), 2825–2830.

[23] Deepak Vasisht, Zerina Kapetanovic, Jongho Won, Xinxin Jin, Ranveer Chandra,Sudipta Sinha, Ashish Kapoor, Madhusudhan Sudarshan, and Sean Stratman.2017. FarmBeats: An IoT Platform for Data-Driven Agriculture. In 14th USENIXSymposium on Networked Systems Design and Implementation (NSDI 17). USENIXAssociation, Boston, MA, 515–529. https://www.usenix.org/conference/nsdi17/technical-sessions/presentation/vasisht

[24] Deepak Vasisht, Swarun Kumar, and Dina Katabi. 2016. Decimeter-Level Local-ization with a Single WiFi Access Point. In 13th USENIX Symposium on Net-worked Systems Design and Implementation (NSDI 16). USENIX Association,Santa Clara, CA, 165–178. https://www.usenix.org/conference/nsdi16/technical-sessions/presentation/vasisht

[25] M Vuran and Agnelo Silva. 2009. Communication Through Soil in Wireless Under-ground Sensor Networks – Theory and Practice. 309–347. https://doi.org/10.1007/978-3-642-01341-6_12

[26] Xue-fen Wan, Yi Yang, Jian Cui, and Muhammad Sohail Sardar. 2017. Lorapropagation testing in soil for wireless underground sensor networks. In 2017Sixth Asia-Pacific Conference on Antennas and Propagation (APCAP). 1–3. https://doi.org/10.1109/APCAP.2017.8420657

[27] Ju Wang, Liqiong Chang, Shourya Aggarwal, Omid Abari, and Srinivasan Keshav.2020. Soil Moisture Sensing with Commodity RFID Systems. In Proceedings ofthe 18th International Conference on Mobile Systems, Applications, and Services(Toronto, Ontario, Canada) (MobiSys ’20). Association for Computing Machinery,New York, NY, USA, 273–285. https://doi.org/10.1145/3386901.3388940

[28] Lakitha O.H.Wijeratne, Daniel R. Kiv, Adam R. Aker, Shawhin Talebi, and David J.Lary. 2020. Using Machine Learning for the Calibration of Airborne ParticulateSensors. Sensors 20, 1 (2020). https://doi.org/10.3390/s20010099