Overcoming barriers in long-term, continuous monitoring of soil CO₂ flux: a low-cost sensor system

Soil CO₂ flux (F_s) is a carbon cycling metric crucial for assessing ecosystem carbon budgets and global warming. However, global F_s datasets often suffer from low temporal-spatial resolution, as well as from spatial bias. F_s observations are severely deficient in tundra and dryland ecosystems due to financial and logistical constraints of current methods for F_s quantification. In this study, we introduce a novel, low-cost sensor system (LC-SS) for long-term, continuous monitoring of soil CO₂ concentration and flux. The LC-SS, built from affordable, open-source hardware and software, offers a cost-effective solution (∼ USD 700 and ∼ 50 h for assembling and troubleshooting), accessible to low-budget users, and opens the scope for research with a large number of sensor system replications. The LC-SS was tested over ∼ 6 months in arid soil conditions, where fluxes are small, and accuracy is critical. CO₂ concentration and soil temperature were measured at 10 min intervals at depths of 5 and 10 cm. The LC-SS demonstrated high stability during the tested period. Both diurnal and seasonal soil CO₂ concentration variabilities were observed, highlighting the system's capability of continuous, long-term, in-situ monitoring of soil CO₂ concentration. In addition, F_s was calculated using the measured CO₂ concentration via the gradient method and validated with F_s measured by the flux chamber method using the well-accepted LI-COR gas analyzer system. Gradient method F_s was in good agreement with flux chamber F_s (RMSE = 0.15 µmol m⁻² s⁻¹), highlighting the potential for alternative or concurrent use of the LC-SS with current methods for F_s estimation – particularly in environments characterized by consistently low soil water content, such as drylands. Leveraging the accuracy and cost-effectiveness of the LC-SS (below 10 % of automated gas analyzer system cost), strategic implementation of LC-SSs could be a promising means to effectively increase the number of measurements, spatially and temporally, ultimately aiding in bridging the gap between global F_s uncertainties and current measurement limitations.

1 Introduction

Soil is the largest terrestrial carbon pool (Lal, 2004). Soil carbon can be subdivided into two general pools: organic and inorganic, with the global storage of each pool at approximately 1530 and 940 PgC, respectively (Curtis Monger et al., 2015). Both organic and inorganic soil carbon exchange with the atmosphere through soil CO₂ flux (F_s). F_s is one of the largest carbon fluxes in the Earth system (Bond-Lamberty, 2018; Friedlingstein et al., 2022). Compared with human-caused increases in atmospheric CO₂, annual CO₂ efflux from the soil into the atmosphere is much larger (Oertel et al., 2016). Therefore, F_s is considered a crucial carbon cycling metric, important for the determination of an ecosystem's carbon budget, calibration, validation, development of (agro)ecosystem, soil carbon models, and assessment of the current global warming scenarios (Bond-Lamberty et al., 2024; Klosterhalfen et al., 2017; Xiao et al., 2012).

For decades, there has been a lack of F_s monitoring in different parts of the globe. Various initiatives have been undertaken to integrate dispersed F_s observations worldwide into publicly accessible datasets (Bond-Lamberty et al., 2020; Bond-Lamberty and Thomson, 2010; Jian et al., 2021). However, global F_s datasets often exhibit low temporal-spatial resolution and spatial bias (Stell et al., 2021; Warner et al., 2019). These limitations constrain our understanding of the mechanisms governing soil carbon dynamics and bias regional-to-global F_s estimation. The largest uncertainties in F_s estimates are found in tundra and dryland ecosystems primarily situated at the two poles, across Africa, Central Asia, South America, and Australia (Stell et al., 2021; Warner et al., 2019; Xu and Shang, 2016). These gaps can be primarily attributed to logistical constraints in manual data collection and the high costs of commercial measuring devices (Bouma, 2017; Forbes et al., 2023; Xu and Shang, 2016). Addressing logistical and financial constraints is crucial because critical questions concerning carbon dynamics can only be answered through extensive F_s quantification (Kim et al., 2022).

Field methods commonly used worldwide to quantify F_s are the eddy covariance method (Baldocchi et al., 1988; Massman and Lee, 2002), the flux chamber method (CM) (Davidson et al., 2002; Lundegårdh, 1927), and the gradient method (GM) (De Jong and Schappert, 1972; Hirano et al., 2003; Tang et al., 2003). These methods substantially differ in principles, thus deviating in cost and F_s estimation. The eddy covariance method provides F_s from a relatively large surface area (Gu et al., 2012), whereas the CM and GM yield single-point F_s (Bekin and Agam, 2023; Maier and Schack-Kirchner, 2014). The CM allows F_s to be measured directly from the soil surface, while the GM measures subsurface soil CO₂ concentration and estimates F_s using Fick's law (Maier and Schack-Kirchner, 2014).

Despite the increasing popularity of the eddy covariance and CM, the GM remains a useful, widely used method (Chamizo et al., 2022; Hirano et al., 2003; Tang et al., 2003; Vargas et al., 2010). In comparison to the other two methods, the GM offers several advantages. First, it mitigates issues associated with eddy covariance, such as turbulence insufficiency, and with CM, such as the microclimate alterations from chamber deployment (Bekin and Agam, 2023; Maier and Schack-Kirchner, 2014). Moreover, GM offers additional insights into the depth profile of gas production, consumption, and exchange in the soil (Maier and Schack-Kirchner, 2014). The most significant advantage of the GM is its lower purchase and installation costs (1–2 orders of magnitude less than the CM or eddy covariance method for continuous F_s monitoring).

The development of small, low-cost, low-power, environmental sensors, microcontrollers, and microcomputers has significantly advanced (Chan et al., 2021; Levintal et al., 2021b). This advancement has led to the extended adoption of low-cost environmental sensing systems in the scientific community (e.g., Helm et al., 2021). Attempts to monitor soil CO₂ concentration using low-cost CO₂ sensors have been made (Blackstock et al., 2019; Hassan et al., 2023; Heger et al., 2020; Osterholt et al., 2022). Others monitored CO₂ fluxes, such as stem, terrestrial, and aquatic fluxes, by implementing the CM using low-cost CO₂ sensors and data loggers (Bastviken et al., 2015; Gagnon et al., 2016; Brändle and Kunert, 2019; Carbone et al., 2019; Forbes et al., 2023; Helm et al., 2021). Implementing the GM using soil CO₂ concentrations measured by underground CO₂ sensors was also reported (Osterholt et al., 2022). However, these studies primarily focused on comparing the precision and accuracy of the low-cost systems with high-end reference systems, typically conducting short-term in-situ examinations lasting from days to weeks, which limits insights into their stability and practicality for long-term use.

To narrow the gap between the uncertainties in the regional-to-global F_s estimations and the capabilities of current measurement methods, in this study, we introduce an open-source, low-cost sensor system (LC-SS) for continuous, long-term monitoring of soil CO₂ concentrations and F_s. The LC-SS was field-tested over ∼ 6 months in arid soil conditions to examine its stability and accuracy compared to a commercial automated flux chamber. Detailed, step-by-step, do-it-yourself guides describing the design, assembly, and installation are provided to assist non-engineer end-users with easy replication and customization.

2 Materials and methods

2.1 Hardware

The LC-SS consists of two units: the control unit and the sensing unit (Figs. 1a and S1 in the Supplement). The control unit includes a microcontroller (Feather M0 Adalogger, Adafruit, USA) accompanied by Secure Digital (SD) card, a latching relay for power control (Latching mini FeatherWing, Adafruit, USA), a clock for accurate time readings (DS3231 RTC, Adafruit, USA), a screen to display real-time results (24.4 mm 128 × 64 OLED Graphic Display, Adafruit, USA), and a multiplexer allowing communication to the sensing unit (Gravity 1-to-8 I2C Multiplexer, DFRobot, China). For power, the microcontroller uses a 3.7 V lithium-ion polymer battery (3.7 V 6000 mAh, Adafruit, USA) charged by solar energy via a solar charger (bq24074, Adafruit, USA), and a 6 W 6 V solar panel (Adafruit, USA). The sensing unit includes seven sensors: six CO₂ sensors (SCD30, Sensirion, Switzerland, 0–10 000 ppm, accuracy between 400 to 10 000 ppm: ±30 ppm + 3 % of full range), and an atmospheric microclimate sensor (pressure, relative humidity, and temperature, MS8607, DFRobot, China). The SCD30 CO₂ sensor also measures temperature and relative humidity (accuracy: ±0.4 °C and ±3 %, respectively).

Figure 1 The design of the low-cost sensor system (LC-SS) (a), two waterproof designs for the SCD30 CO₂ sensors (b), field installation of the CO₂ sensor line at 5 cm (c), and the site after installation (d).

The LC-SS used in this study featured two waterproof designs of CO₂ sensors (Fig. 1b): a 50 mL Falcon tube design and a thin coating design. The 50 mL Falcon tube design is an easy-made and long-lasting option, while the thin coating design is suitable for near-surface deployment, effectively reducing errors associated with measurement depths. Both designs included a hydrophobic membrane to keep water from penetrating the sensor while allowing gas exchange with the surrounding soil. Providing two designs offers end users the flexibility to adopt the option that best fits their needs and accessibility.

The total time required to build and calibrate the LC-SS is ∼ 50 h, but could vary depending on the user's familiarity with electronics and sensor integration. The detailed do-it-yourself guide of the LC-SS assembly with time estimation for each major step and sensor waterproof designs can be found on our GitHub page (https://github.com/OpenDigiEnvLab/soil-CO2-sensor-system, last access: 18 September 2025). The hardware details are summarized in Table 1.

Table 1 Summary of hardware components with examples for potential suppliers (components can be purchased from other suppliers).

^a https://www.adafruit.com/ (last access: 18 September 2025). ^b https://www.dfrobot.com/ (last access: 18 September 2025). ^c https://www.digikey.com/en/products/detail/sensirion-ag/SCD30/8445334 (last access: 18 September 2025).

The hardware is controlled using open-source Arduino code written in C$++$ (https://www.arduino.cc, last access: 18 September 2025). The complete code for the LC-SS can be downloaded from our GitHub page (https://github.com/OpenDigiEnvLab/soil-CO2-sensor-system, last access: 18 September 2025). At every measurement cycle, all sensors are activated, and measurement readings are logged onto the SD card with a corresponding timestamp and displayed on the user screen. The default measurement interval is 10 min and can be easily customized if required.

2.2 Field installation

The LC-SS was installed at the Wadi Mashash Experimental farm located in the Northern Negev desert of Israel (31°04^′14^′′ N, 34°51^′62^′′ E; 360 m above sea level). The local climate is arid, with an average annual precipitation of 116 mm, primarily occurring between October and April. The daily average maximum and minimum temperatures in January (winter) are 15.9 and 8.0 °C, and in August (summer) are 33.3 and 20.7 °C, respectively. Soil is characterized as sandy-loam loess soil (72.5 % sand, 15 % silt, and 12.5 % clay). Soil organic carbon content between 0–5 and 5–10 cm is 9.37 and 9.13 mg g⁻¹, respectively. CaCO₃ content between 0–5 and 5–10 cm is 50 % and 47 %, respectively.

The LC-SS was installed from 24 May to 14 November 2023, providing continuous measurements for 175 successive days, spanning both summer and winter. Three CO₂ sensors were installed at each depth (5 and 10 cm) to allow comparison and statistical calibration, as detailed in Sect. 2.3. At each depth, two sensors with the thin coating design (labeled as sensor#1_5cm, sensor#2_5cm and sensor#1_10cm, sensor#2_10cm) and one sensor with the 50 mL Falcon tube design (labeled as sensor#3_5cm and sensor#3_10cm) were deployed (Fig. 1c). To enable manual gas sampling for field calibration, a 60 cm Polyurethane tube (outer diameter × inner diameter = 6×4 mm) was inserted at each depth. One end of the tube was aligned with the CO₂ sensors, while the other end extended above the soil surface and was sealed with a valve (Fig. 1d). Additional measurements included soil water content (SWC) using time-domain reflectometers (TDR-315, Acclima, Inc., USA) installed at 3 and 10 cm depths. Air temperature, atmospheric pressure, and precipitation data were taken from a meteorological station located at the same field where the LC-SS was installed (https://ims.gov.il, last access: 18 September 2025; Zomet Hanegev station).

F_s measured using the CM (F_CM) was measured at 1 h intervals using a non-dispersive infrared (NDIR) gas analyzer (LI-8100A, LI-COR, USA) connected to four automated non-steady-state chambers (104C, LI-COR, USA). F_CM was determined as the average readings obtained from the four chambers. The F_CM measurements were conducted for the periods 24 May–18 June, 17–23 August, and 5 September–17 October 2023.

2.3 Two-step calibration of the CO₂ sensors

Calculating F_s based on the GM (F_GM) (Sect. 2.4) requires accurate soil CO₂ concentrations. Therefore, we developed a two-step calibration process for the underground CO₂ sensors: a field calibration and a statistical calibration.

For the field calibration, CO₂ concentrations from the low-cost SCD30 CO₂ sensors (C_SCD30) were calibrated against reference CO₂ concentrations (C_ref). C_ref were obtained by measuring the CO₂ concentrations sampled from the sampling tube either by a high-end CO₂ sensor (GMP252, Vaisala Inc., Finland) or by LI-COR gas analyzer (LI-8100A, LI-COR, USA) with three replicates from each depth (the choice of calibration devices can be adjusted depending on local availability). C_ref by the Vaisala CO₂ sensor was measured every 5 h between 06:00 and 16:00 UTC+2 on two days, 12 June and 17 July 2023. C_ref by the NDIR gas analyzer was measured every 3 h from 12:00 to 21:00 on 10 September 2023 and from 00:00 to 12:00 on 11 September 2023. In total, the calibration was determined with 21 and 17 measurement points for each sensor at 5 and 10 cm, respectively, over the range of concentrations from ∼ 300 to ∼ 650 ppm.

Gradual drift was assessed by evaluating whether the pairwise differences in CO₂ concentration among three sensors placed at the same depth (sensor#2-sensor#1, sensor#3-sensor#1, sensor#2-sensor#3) changed over time. To quantify this, the pairwise concentration differences were plotted against time, and linear regression was applied to determine the relative drift rate (ppm d⁻¹). The cumulative deviation was then estimated as the product of the drift rate and the number of days. If this cumulative deviation exceeded a predefined threshold – set at 10 % of the mean concentration in our study – separate field calibration curves were applied to account for the drift.

The statistical calibration consisted of two sequential algorithms. The first algorithm (Fig. 2a) addressed abrupt anomalies or jumps of each sensor reading by flagging data points where the difference between measured and smoothed data exceeded 10 % of the measured data point. The smoothed data was executed using the LOESS smoothing algorithm (Jacoby, 2000), which fits multiple locally weighted least squares regressions to estimate a smooth curve through a scatterplot of data points. The second algorithm (Fig. 2b) focused on correcting deviation of between three sensors at the same depth, utilizing user-defined thresholds to determine when the difference between one sensor and the other two sensors becomes significant enough to require correction. Thresholds of 5 % and 10 % relative to the average for sensors at 5 and 10 cm, respectively, were defined. All calibration algorithms were applied post-data acquisition, ensuring accurate CO₂ concentrations essential for calculating F_GM.

Figure 2 Flowchart of the two statistical calibration algorithms. The algorithm to correct jumps (a) and the algorithm to correct deviation between three sensors at the same depth (b).

2.4 Calculating the F_GM using the LC-SS data

To calculate F_GM, CO₂ concentrations were first corrected for temperature and pressure (Eq. S1 in the Supplement) and then converted to mole density (Eq. S2). The GM is based on Fick's first law, where F_GM from depth z to the soil surface is calculated as (De Jong and Schappert, 1972):

$$\begin{array}{}\text{(1)}& F_{\mathrm{GM}}=-D_{\mathrm{s}}{\displaystyle \frac{C_z-C_{\mathrm{0}}}{z}}\end{array}$$

where F_GM [µmol m⁻² s⁻¹] is assumed to be equal to F_s from the soil surface (a positive F_GM indicates CO₂ efflux and a negative F_GM indicates CO₂ influx), D_s [m² s⁻¹] is the CO₂ diffusion coefficient between depth z [m] (negative) and the soil surface (0 m), C_z [µmol m⁻³] is the CO₂ mole density at depth z, and C₀ [µmol m⁻³] is the atmospheric CO₂ mole density (C₀ = 18 741.63 µmol m⁻³ or 420 ppm). The reference value of 420 ppm was based on the average atmospheric CO₂ concentrations measured by a LI-COR gas analyzer between 16 May–18 June and 2 July–13 August 2023. F_GM in this study was calculated using CO₂ concentration gradients between 0 and 5 cm depth, as recommended by Chamizo et al. (2022).

The relative CO₂ diffusion coefficient in the soil ($D_{\mathrm{s}}/D_{\mathrm{a}}$ where D_a [m² s⁻¹] is the CO₂ diffusion coefficient in free air) is estimated based on soil air content-dependent models M(ε), with ε being the volumetric air-filled porosity:

$$\begin{array}{}\text{(2)}& {\displaystyle \frac{D_{\mathrm{s}}}{D_{\mathrm{a}}}}=M\left(\mathit{\epsilon }\right)\end{array}$$

D_a needs to be corrected to in-situ environmental conditions (Jones, 2013) using Eq. (S3). Models used in this study to calculate M(ε), including the most common models, are listed in Table 2.

Table 2 Classical soil diffusion coefficient models used for the GM. Porosity (φ) values were calculated as described in Eq. (S4), and equal to 45 %.

From the ten listed diffusion models, ten F_GM time series were calculated. The total net flux over the observed period for each F_GM time series was calculated by determining the total area under the curve of CO₂ efflux minus the total area above the curve of CO₂ influx. The average daily cumulative flux [g C m⁻² d⁻¹] was calculated by dividing the total net flux by the total number of days (n=175).

2.5 Validation of F_GM using F_CM

F_GM from ten gas diffusion models were validated using measured F_CM. First, we conducted a cross-correlation analysis (Horvatic et al., 2011) between F_CM and F_GM to systematically assess the lag time between measured F_CM and calculated F_GM, which reflects the time delay associated with gas transport from the 5 cm depth to the soil surface as previously reported (Sánchez-Cañete et al., 2017). Then, we shifted the F_GM using the identified lag time to align with the temporal dynamics of F_CM.

To evaluate the best-fitted diffusion model, ten shifted F_GM calculated based on ten diffusion models were compared with measured F_CM. The selection of the best-fitted diffusion model is based on a comparison of interquartile range, average daily cumulative flux, R², root mean square errors, and three components of mean squared deviations, namely squared bias, non-unity slope, and lack of correlation (Gauch et al., 2003).

3 Results and discussion

This study focuses on the development and field performance of the LC-SS for measuring soil CO₂ concentrations and calculating F_GM. Therefore, our results and discussion will focus mainly on the LC-SS capabilities, such as long-term stability and accuracy.

3.1 CO₂ sensors calibration

Over the tested period, we observed a low rate of gradual drift in all six sensors (0.06–0.72 ppm d⁻¹) (Fig. S4). The cumulative deviations for six sensors were below the predefined threshold (10 % of the mean concentration). Therefore, for the entire period of 175 d, we used one calibration curve for each sensor. The field calibration curves for the six low-cost CO₂ sensors are presented in Fig. 3a. All sensors show good linearity with high R²>0.8. The statistical calibration algorithms (Fig. 2) improved both the sudden and permanent drifts (Fig. 3b). At 5 cm, only 6.6 %, 2.1 %, and 4.4 % out of 25 200 readings of sensors #1 (thin coating), #2 (thin coating), and #3 (falcon tube), respectively, required correction. At 10 cm, 34.5 %, 1.9 %, and 1.39 % readings were corrected for sensor #1 (thin coating), #2 (thin coating), and #3 (falcon tube), respectively. Except for sensor#1_10cm, corrections required for other sensors were due to sudden jumps. 34.5 % data correction for sensor#1_10 was due to a systematic, permanent drift shifting baseline from ∼ 300 to ∼ 200 ppm from 20 September 2023 until the end of the observed period 14 November 2023. The results demonstrate the high stability of the CO₂ sensors after 6 months. However, sensor drifting is often system-specific and varies with environmental conditions. Therefore, it is important to detect the gradual drifting of raw data over time (e.g., Fig. S4) and conduct field calibration accordingly.

Figure 3 Calibration curves of the SCD30 CO₂ sensors using reference CO₂ concentration measured by Vaisala CO₂ sensor between 12 June–17 July 2023 and LI-COR gas analyzer 10–11 September 2023 (a), and distribution of CO₂ concentrations collected by six SCD30 CO₂ sensors after field and statistical calibration step (b).

3.2 Soil CO₂ concentrations

The 10 min interval time series of CO₂ concentrations at 5 and 10 cm, and precipitation for one month (24 May–24 June 2023) as an example are shown in Fig. 4a–b. The CO₂ concentrations for the entire studied period is presented in the Supplement (Figs. S2 and S3). The magnitude of CO₂ concentrations at 10 cm was greater than at 5 cm (∼ 340–∼ 730 ppm compared to ∼ 320–∼ 1000 ppm, respectively). CO₂ concentrations at both depths during daytime (∼ 07:00–∼ 21:00 in summer and ∼ 08:00–∼ 19:00 in winter) were higher than in the atmosphere, with average daytime concentrations of 545 and 621 ppm at 5 and 10 cm, respectively. However, during nighttime (all hours excluding daytime hours), soil concentrations were lower than in the atmosphere, with average nighttime concentrations of ∼ 380 ppm at both depths. This indicates an efflux of CO₂ from the soil to the atmosphere during daytime in contrast to an influx of CO₂ from the atmosphere into the soil during nighttime. Daytime efflux and nighttime influx were previously observed in arid soils (Cueva et al., 2019; Hamerlynck et al., 2013; Sagi et al., 2021). The study conducted by Sagi et al. (2021) in the Negev Desert revealed a connection between soil CO₂ influx, cooling soil temperatures, and high soil-to-air temperature gradients, specifically occurring when SWC was below the threshold of ∼ 8 %. We observed similar conditions during our study (Fig. 4c–e).

Figure 4 One month example of continuous CO₂ concentration measurements between 24 May–24 June 2023 at 5 cm (a) and 10 cm (b) depths, average daily values at 5 cm of CO₂ concentration, temperature, and volumetric soil water content (SWC) during four days with precipitation from May to September (Summer) (c), 130 d without precipitation between May and September (Summer) (d), and 44 d without precipitation between October and November (Winter) (e).

CO₂ diurnal cycles at 5 cm showed differences between days with and without precipitation (Fig. 4c–d) and between summer months (May–September) and winter months (October–November) (Fig. 4d–e). On days with precipitation, the average CO₂ concentration increased from 400 ± 20 ppm around 08:00–9:00 to a daily peak of 530 ± 70 ppm at 16:00. On days without precipitation, the morning increase occurred earlier around 11:00–13:00, reaching 662 ± 16 ppm. Inter-season patterns were also observed, with a winter daily peak lower than the summer daily peak by 106 ± 22 ppm. The occurrence of diurnal cycles during all seasons is a typical phenomenon previously reported (Spohn and Holzheu, 2021; Chamizo et al., 2022).

Our results showcase the ability of the underground CO₂ sensors to capture typical diurnal and seasonal changes of soil CO₂ concentration. The results also highlight the capability of the sensor system to capture “hot moments”, such as the effect of precipitation events on CO₂ concentration in arid soils, significantly contributing to the understanding of the driving mechanisms underlying these moments.

3.3 F_GM calculations

The calculated F_s using the GM (F_GM, Eq. 1) and the measured F_s using the CM (F_CM) are presented in Fig. S5; for simplicity, continuous results from only three representative days without precipitation are shown. Calculated F_GM using different soil gas diffusion models (Table 2) were compared to the F_CM. We observed a time lag in all calculated F_GM compared to the F_CM. Since the F_GM was calculated using the CO₂ concentration gradient between 5 cm and the soil surface, F_GM can only represent subsurface F_s. Cross-correlation analysis was used to evaluate the lag time between the surface F_CM and the sub-surface F_GM (dashed lines) resulting in a lag time of three hours. To establish temporal alignment between F_GM and F_CM, F_GM was shifted three hours to the past (Fig. S5, solid lines).

A delay was also observed in the nocturnal influx F_GM compared to the nocturnal influx F_CM. Given the direction of nocturnal CO₂ exchange – moving from the atmosphere into the soil – at any given moment, the volume of CO₂ traversing a unit surface area at a given time (CO₂ influx in units of µmol m⁻² s⁻¹) must exceed that passing through the subsurface region at 5 cm depth. This leads to a more negative nocturnal influx F_CM than nocturnal influx F_GM. Therefore, we used the average daily minimum of nocturnal influx F_CM as a reference to shift the magnitude of F_GM. The time lag between F_GM and F_CM associated with measurement depth was also reported in previous studies (Sánchez-Cañete et al., 2017); the delay generally increases with sensor depth.

The magnitude and distribution of F_CM and F_GM (box plots), and average daily cumulative flux (blue scatters) are presented in Fig. 5a. The diffusion model evaluation using components of mean squared deviation is presented in Fig. S6. In comparison to F_CM, Buckingham F_GM was the most comparable, for both magnitude and distribution, average daily net flux, as well as based on components of mean squared deviation. A representative nine-day time series of Buckingham gradient flux (original and shifted) and chamber flux are presented in Fig. 5b. Seven models, including Penman (1940), Marshall (1959), Millington (1959), Millington and Quirk (1961), Currie (1970), Lai et al. (1976), Moldrup et al. (2000) overestimated and two models including, Campbell (1985) and Sadeghi et al. (1989), underestimated F_CM. In generalization, ten models can be classified into two categories based on their assumptions: (1) soil-type/SWC-independent models including Buckingham (1904), Penman (1940), Millington (1959), Campbell (1985), Marshall (1959) and Lai et al. (1976) which depends solely on air porosity, and (2) SWC-dependent models including Millington and Quirk (1961), Currie (1970), Sadeghi et al. (1989) which also includes a water-induced linear reduction term, equal to the ratio of air-filled porosity to total porosity ($\mathit{\epsilon }/\mathit{\phi }$). The first category can be generalized in the form bε^m (with ε being air-filled porosity, b and m being fitting constants). Currie (1965) has shown that an equation of the form bε^m represents well diffusion in dry porous materials, with m typically falling between 1 and 2, and b from 0.5 to 1, depending on the shape of the soil particles. The second category can be generalized in the form $b{\mathit{\epsilon }}^m(\mathit{\epsilon }/\mathit{\phi }{)}^n$ (with b, m, n being fitting constants). The addition of the term $\mathit{\epsilon }/\mathit{\phi }$, according to Moldrup et al. (2000), helps to better predict diffusion in wet soils. The reasons for the difference of fitting constants ($b,m,n$), for example, Penman (1940) found b=0.66 and m=1, Marshall (1959) b=1 and m=1.5, are that different tortuosity models were used to develop the diffusion model, and the developed diffusion models were validated under varying soils and soil conditions where soil properties such as the pore geometry and the length of gas passage were different. The majority of models were validated against a wide spectrum of soil texture (e.g., Moldrup et al., 2000, tested on 21 differently textured and undisturbed soils, or Sadeghi tested on 7 soils with clay content 7 %–51 %), fitting constants ($b,m,n$) were therefore concluded as soil-type independent. However, biases were frequently observed, and there is no unique solution holding true for any given specific soil type (Pingintha et al., 2010; Sánchez-Cañete et al., 2017; Yan et al., 2021). For example, in our case, dry, undisturbed soil with 12.5 % clay content, matching soil type examined by Sadeghi et al. (1989), Lai et al. (1976), and Moldrup et al. (2000); however, Sadeghi et al. (1989) underestimated F_CM, while Lai et al. (1976) and Moldrup et al. (2000) overestimated F_CM. The Buckingham model (b=1, m=2), one of the models of the first category for dry porous materials, showed the best prediction. However, under higher SWC, increased tortuosity and reduced flow cross-section suggest that higher m in bε^m models – or $b{\mathit{\epsilon }}^m(\mathit{\epsilon }/\mathit{\phi }{)}^n$ models – may yield better performance. When selecting the most suitable empirical diffusion model for estimating soil gas transport, it is recommended to prioritize bε^m models for dry soils and $b{\mathit{\epsilon }}^m(\mathit{\epsilon }/\mathit{\phi }{)}^n$ models for wet soils. Testing multiple models in the same category but differing in formulation ($b,m,n$ values) can help assess their sensitivity and applicability to a specific site.

Figure 5 Comparison of measured chamber flux (green) and calculated gradient flux (red) using ten published gas diffusion models, and average daily cumulative flux (blue scatter) (a), and diurnal cycles of measured chamber flux (blue scatters) and calculated gradient flux using Buckingham diffusion model (dashed orange) and Buckingham gradient flux shifted by 3 h lag time (solid orange) during nine representative days without precipitation (b).

The linear regression between Buckingham F_GM and measured F_CM is presented in Fig. 6a (R²=0.70, RMSE = 0.15 µmol m⁻² s⁻¹). F_s obtained by these two methods correlated most strongly on days without precipitation (Fig. 6b). In contrast, on days with precipitation, large variations between the two methods were observed (outliers in Figs. 5a, 6a and c – A precipitation event on 13 June 2023 with 2 mm d⁻¹). The instantaneous increase of F_CM due to precipitation was a well-recognized phenomenon when rewetting occurs in water-limited arid soils (Andrews et al., 2023; Barnard et al., 2020; Fierer and Schimel, 2003). The observed CO₂ pulse, as measured by the CM, agrees with the observed pattern of very high rates right after rewetting and slowly declines over time (Kim et al., 2012). These precipitation-induced CO₂ pulses were underestimated by the GM. Previous studies also reported that the GM did not capture the abrupt CO₂ pulse increases after water application (Jiang et al., 2022; Yang et al., 2018). Rewetting of arid soils after a dry period triggers the sudden increase of microbial activity, leading to a burst in carbon mineralization (Barnard et al., 2020). In arid soil, the top ∼ 1 cm is often the most microbially active due to the presence of biocrust (Belnap et al., 2016). The increased CO₂ efflux from the topsoil was captured by the CM, yet underestimated by the GM (Jiang et al., 2022; Yang et al., 2018). Under rewetting events, the assumptions of the GM, such as one-directional gas movement and linear concentration gradient with soil depth, are invalid. Greater soil CO₂ on the topsoil than in the deeper soil leads to bidirectional concentration gradients and fluxes (Tang and Baldocchi, 2005). The application of the GM, therefore, is not recommended for F_s estimation of dry soils upon rewetting. It is important to note that this is a well-known methodological limitation, extensively reported in the literature, and it persists regardless of the type of NDIR CO₂ sensor used (Fan and Jones, 2014; Tang and Baldocchi, 2005). Even though F_GM under rewetting events is unreliable, it does not limit the application of the GM under relatively steady moisture conditions (i.e., SWC can be moderate to high but no abrupt changes due to rainfall or irrigation) (Fan and Jones, 2014; Turcu et al., 2005).

Figure 6 Comparison between the gradient flux (F_GM) calculated by the best-fitted Buckingham diffusion model and the LI-COR chamber flux (F_CM) for the whole tested period of 175 d (a), the Buckingham gradient flux (orange) and the LI-COR chamber flux (blue) during two representative days without precipitation (23–24 July 2023) (b), and during two representative days with precipitation (13–14 June 2023) (c).

3.4 Limitations and modifications

The LC-SS system can be built for approximately USD700, taking ∼ 50 h depending on the user's familiarity with electronics and sensor integration. This relatively low cost and manageable time commitment make the LC-SS a practical and scalable option for long-term, continuous CO₂ monitoring, especially in remote or underfunded research settings. However, we acknowledge that this work has certain limitations. The first limitation involves using high-end LI-COR chambers and gas analyzers for the validation of calculated F_GM. This practice may pose a cost constraint for resource-limited research. Even though using F_CM measured by high-end gas analyzers to validate F_GM is a recommended practice (Chamizo et al., 2022; Sánchez-Cañete et al., 2017) and applied in this study, it is not inherently obligatory. Several alternatives can be considered. First, the site-specific diffusion coefficient can be measured directly for the calculation of F_GM without using published gas diffusion models. For example, Osterholt et al. (2022) suggested an approach to inject CO₂ to estimate the diffusion coefficient. Furthermore, high-end, expensive chambers and gas analyzers can also be replaced with a low-cost, open-source chamber system (e.g., Forbes et al., 2023). The same CO₂ sensor SCD30, as used in this study, can also be used to manually build a low-cost chamber. When used with the LC-SS, only one chamber-gas analyzer system per several LC-SSs is needed since only a short duration of F_CM measurements is required for validation. Additionally, conventional CO₂ quantification techniques – such as gas chromatography or the alkali absorption method – can be used to monitor CO₂ concentration changes inside a static chamber to quantify F_s (Yan et al., 2021; Pumpanen et al., 2004; Yim et al., 2002; Christiansen et al., 2015). Integrating the LC-SS with the alkali absorption method could be a promising approach that balances affordability, automation, and long-term monitoring of CO₂ concentration and F_s, while enhancing accuracy; particularly in remote or resource-limited locations where access to high-end instruments like gas analyzers or gas chromatography is not accessible.

The second limitation is that the system was tested only in dry, arid soils. Although a few precipitation events were captured and analyzed, the system's performance under persistently high SWC conditions was not evaluated over the long term. In general, the use of the GM may not be suitable under conditions of sustained soil saturation, frequent rainfall typical of humid climates, or frequent irrigation.

Last, the LC-SS presented here relies exclusively on an SD card for data logging and storage, which requires manual data retrieval and lacks real-time accessibility for monitoring and troubleshooting. Alternatively, we introduce an updated version of LS-SS equipped with a modem for real-time data updates and immediate troubleshooting whenever necessary (e.g., Levintal et al., 2021a). A detailed, step-by-step, do-it-yourself guide for the updated version is also available on our GitHub page.

4 Conclusions

This study introduces an innovative LC-SS developed for continuous, long-term monitoring of soil CO₂ concentration and F_s, facilitating in-situ soil-gas-related research. The LC-SS was built from low-cost, readily available hardware and open-source software components. The LC-SS design emphasizes modularity, with publicly available, comprehensive, technical documentation for each module, allowing straightforward replication and customization for non-engineering, low-budget end-users worldwide.

The LC-SS was field-tested for ∼ 6 months, showcasing high stability and capabilities to capture the temporal dynamics of soil CO₂ concentrations, including diurnal and seasonal variabilities. Furthermore, the agreement observed between the calculated F_GM and measured F_CM, both in the short term (i.e., sub-daily fluctuation) and in the long term (i.e., net CO₂ exchange over ∼ 6 months), demonstrate the potential of the LC-SS as a new approach for F_s quantification. The use of LC-SSs and GM is recommended in soils with consistently dry to moderate SWC conditions. For reliable F_s results, the diffusion coefficient can be measured directly, or several methods of F_s quantification (high-end/low-cost chambers, gas chromatography, or alkali absorption method) were suggested for the validation of the calculated gradient flux.

In conclusion, the LC-SS, priced at ∼ USD 700, not only provides high accuracy of F_s but also offers higher temporal resolution and the potential for improved spatial resolution if widely adopted. This, in turn, could contribute to a more comprehensive dataset for regional-to-global estimation of F_s and advancing our understanding of the global soil carbon cycle.

Appendix A: Abbreviations

CM	chamber method
Cref	reference CO2 concentration measured by Vaisala CO2 sensor and LI-COR gas analyzer
CSCD30	CO2 concentration measured by low-cost SCD30 CO2 sensors
FCM	soil CO2 flux measured by chamber method
FGM	soil CO2 flux calculated by gradient method
Fs	soil CO2 flux
GM	gradient method
LC-SS	low-cost sensor system
NDIR	non-dispersive infrared
SD	secure digital
SWC	soil water content