Developing and testing a PM 2.5 low-cost sensor in Ethiopia under ambient and indoor air pollution conditions

PM 2.5 low-cost sensors are a promising trend for low-income countries, where the PM 2.5 associated burden of disease is high and few measurement instruments are available. Commercially available Sensor Systems (SSys) are relatively affordable and easy to use. They are, however, not designed for or evaluated in contexts characterized by much biomass burning, regular power interruptions and/or low internet coverage, typical for low-income countries. Alternatively, local teams can build a sensor system with PM 2.5 sensors from Original Equipment Manufacturers (OEM). Existing African OEM projects depend on international partners and funding. This puts the affordability for local teams without funding into question. Furthermore, field comparisons of such sensors for ambient concentrations and indoor settings are rarely conducted in low-income contexts. In Arba Minch, Ethiopia, we developed a sensor system (SPSA) with the OEM Sensirion SPS30 and other components, together with an Arduino microprocessor, with LoRaWAN data transmission. We used the hardware and software in multiple configurations. The SPSA was used in 14 contexts typical for Ethiopia. During these tests we encountered problems that were easily solved by maintenance on location. On seven locations we collocated the SPSA with itself, gravimetric instruments and/or SSys. Amongst SPSA we found coefficients of determination (R 2 ) of at least 0.98 for three ambient and one indoor location. The accuracy in comparison with the gravimetric method was 16% under ambient and 13% under indoor circumstances. This is lower than the internationally required 25%. The R 2 in comparison to two SSys was 0.91-0.98 under ambient and 0.88-1.00 under indoor circumstances. The SPSA is a versatile sensor system that can be used in both ambient and indoor air pollution circumstances. Local development without international partners and funding resulted in local experience gaining, low costs, local ownership, and the possibility of tailoring the system to local needs regarding power and connectivity.


Introduction
Air pollution is amongst the top risk factors for the global disease burden (Babatola, 2018;Shaddick et al., 2018), placed second for data of 2019 (IHME, 2021).This burden is relatively higher in lowincome countries, due to sources like open waste burning and cooking with biomass fuel (World Health Organization, 2021, 2022).Paradoxically, resources for measurements are lowest in those countries.A primary indicator for air pollution in indoor and ambient situations is particulate matter with a diameter smaller than 2.5 µm (PM 2.5 ) (World Health Organization, 2021).The reference method for monitoring PM 2.5 is filter-based gravimetry.This method typically assesses concentrations on a 24-hour average level (EPA, 2006;European Commission, 2010), and is associated with high operating costs (Sousan, Regmi and Park, 2021).Various continuous monitors (monitoring concentrations at hour-or even second level) are recognized as equivalent to the reference method, such as the Beta Attenuation Monitor (BAM), the Tapered Element Oscillating Microbalance (TEOM), and Palas Fidas.These are expensive as well, with costs of $11 500 -$30 000 per monitor (Mooney, Willis and Stevenson, 2006), and technically still need calibration to the gravimetric reference method.In high-income countries the application of such monitors is widespread, and measured data is usually openly available.For African countries, however, only few have some coverage (Subramanian and Garland, 2021).
A promising trend is the development of low-cost sensors (LCS).Karagulian et al. (2019) distinguish two types of LCSs: only a sensor as developed by original equipment manufacturers (OEM) and sensor systems (SSys), which combine an OEM with encasing, sampling system, power system, hardware, software, and data acquisition methods.SSys cost between $200 and $500, whereas self-built systems with OEMs for measuring PM 2.5 can be created for $50.Using SSys does not require skills in programming and electronics, but their internet and power requirements are challenging (Sewor, Obeng and Amegah, 2021).Moreover, SSys are still expensive for African researchers (Subramanian and Garland, 2021), and data ownership of data collected through the network functionality lies with the producer.These issues plea for a focus on OEM.Creating an own system based on an OEM is cheaper, and the users can decide to combine the OEM with alternative energy sources (power bank, solar panel) and network connectivity (LoRaWAN, GSM), according to their context.This makes OEMs more compatible to local climates (Subramanian and Garland, 2021).Working with OEMs requires programming and electronics proficiency.When a local team develops a setup with an OEM, there is local experience gaining, local ownership of instruments and data, and local decision making.Apart from making a system with an OEM locally, it is important to evaluate it under the conditions where it will be used.Parameters such as particle density, particle hygroscopicity, refraction index, and particle composition strongly affect the operational principle of PM 2.5 LCS (light scattering).All these factors vary from site to site and with seasonality (Karagulian et al., 2019).Open biomass burning, cooking with biomass fuel and older vehicles are air pollution sources typical for a low-income country.The data quality of an OEM should be evaluated under such circumstances.
Three examples of OEM projects in the African continent are https://sensors.africa/air(OEM SDS011), Ngom et al. (2018) (OEM HK-A5) and Airqo.net(Bainomugisha, Ssematimba and Okure, 2023;OEM Plantower PMS 5003).Sensors.Africa and AirQo have international partners and funding.This raises the question whether local development without funding is possible.Furthermore, we only found field tests of data quality in a low-income country context for AirQo.Adong et al. (2022) compared the AirQo to a BAM at two ambient locations in Kampala, Uganda.We did not find any field comparisons under indoor air pollution circumstances for these OEM projects.Dingemanse (2022) presented the evaluation of three LCS under field conditions in Arba Minch, Ethiopia.One of these LCS was a locally developed system with the OEM Sensirion SPS30 and an Arduino microprocessor.In this article, we present this system and show the field evaluation in more detail.We refer to this system as SPSA, named after the sensor (SPS) and the microprocessor, being Arduino (A).The SPSA has a power backup and can transmit its data with LoRaWAN technology.This makes the setup compatible with a context that often lacks stable power supply and WiFi.We conducted extensive field comparisons in Ethiopia, at three ambient and four indoor high concentration locations.We collocated the SPSA with itself, with a gravimetric measurement instrument, and with two SSys: IQAir Airvisual Pro (the SSys of Airvisual, which is the largest real-time air quality databank (Wernecke et al., 2021)) and PATS (designed for indoor air pollution measurements).We developed the system and conducted the field comparisons without funding.
The next part of the article includes a description of the system we used and evaluated, followed by an evaluation of its operational reliability and an evaluation of the data quality.This article ends with a discussion and conclusion on the implications of the operational reliability, data quality and field evaluations in general.While we included all parts of a working sensor system with network aspect, our aim is not to present a finished product.Rather, it is a showcase of the potential of Do-It-Yourself development with OEM within Ethiopia.Our unique contributions are the presentation of a sensor system developed locally without international partners or funding, and field tests of this system across both ambient and indoor circumstances in Ethiopia.These are the first field tests of the OEM SPS30 in a low-income country.We hope it inspires the reader to conduct similar projects, including field comparisons, in low-income country contexts.All software and data analysis code and raw data, used in this study, is shared on an OSF repository: https:// doi.org/10.17605/OSF.IO/DXEZ8.

Components and connections
Core components of the system are a PM 2.5 sensor, SD card module and Real Time Clock (RTC), connected to a microprocessor.Optional components include a Relative Humidity sensor, LoRaWAN communication shield and LED control light.Table 1 gives an overview of the components.
The Dragino LoRa Shield is a premade component that fits directly on the pin configuration of the microprocessor.Other components of the system are connected to the main circuit board via shielded connecting wires.The wires are kept short, less than 10 centimeters, to avoid electromagnetic interference and crosstalk.Power supply connections are shared to avoid wiring complexity.Also, the SDA and SCL connection points are shared between the SPS30 and DS3231.Figure 1 shows a sketch of the circuit.

Configurations
The SPSA hardware was combined with a power bank in plastic boxes (lunch boxes or plastic cups).The boxes were chosen based on availability on the local market and the intended use.Three configurations have been used:     Figure 2 shows photos of the three configurations.
The air sampling and ventilation openings of the SPS30 are at the frontside of the sensor.In the plastic boxes a hole the size of the SPS30 was cut, and the SPS30 was placed through the hole for approximately 5 millimeters.In this way, the sensor was exposed to the ambient air while keeping most of the sensor safe inside the box.Also, because the hole was exactly sized to the SPS30, movement of the sensor within the SPSA was not possible.All other components were kept at their place with glue.For configuration C1, the BME280 air and humidity sensor was placed outside the box.A hole was cut inside the box for exactly the size of the BME280 wires.Configuration C3's casing was a cup, so that it could be fixed to a belt.

Power consumption
Based on their respective data sheets, the power consumption of the SPSA components is as follows: 1. Arduino Mega 2560: 5 V x 50 mA = 250 mW.   2. SPS30: 5 V x 80 mA = 400 mW.
3. DS3231: 5V with a maximum of 0.65 mA = 3.26 mW.A rechargeable battery is included as back-up.4. Micro SD card with adapter: maximum standby current of 1 mA x 5 V = 5 mW. 5. BME280: 0.0036 mA at 3.6 V = 0.013 mW. 6. LoRaWAN: 5 V x 32 mA = 160mW during transmission, and 5 V x 1 mA = 5 mW during sleep mode.This transmission is used once every five minutes.We assume a transmission time of maximum 20 seconds.This results in per hour 20 * 12 = 240 seconds of transmission mode, and remaining time sleep mode.7. LED: 40 mA x 5 V = 200 mW when in use.A blink time of 0.2 seconds is used per minute, leading to a usage time of 0.2 * 60 = 12 seconds per hour.
Table 2 gives an overview of the SPSA hourly power consumption.
Most commercial power banks have a capacity of 6-10 Ah at 5 V.
In case of connection to an irregular power grid, a 6 Ah power bank can provide back-up power during power interruptions for 6 Ah / (0.674 W / 5 V) = 44 hours.In combination with a 1.5 W solar panel, the power bank can be charged during daytime and used during nighttime.The battery capacity for 12 hours, including a maximum depth of discharge of 20%, should then be (0.674 W * 12 h) * 1.2 / 5 V = 1.94 Ah.
Introduction of a sleep mode for the SPS30 can decrease power consumption.Under certain conditions, it is possible to achieve the same monitoring quality through point measurements taken every 15 minutes.This would reduce the power consumption of the SPS30 sensor with at least 90%.As a result, the total system consumption would drop to 314 mW from its current level of 674 mW.

Software
Figure 3 shows a flowchart of the SPSA operation process.

LoRa network software
With LoRa functionality, data is sent to a gateway.The gateway is registered at The Things Network (TTN; https://www.thethingsnetwork.org/).If the receiving gateway is connected to the internet, the data is stored on the TTN server for seven days.Within those seven days, data can be retrieved from the server.For proof of concept, we created Python code for the following tasks: • Retrieving data from TTN • Creating a graph of the data • Uploading a graph to a website.
Example code is uploaded to the OSF repository (https:// doi.org/10.17605/OSF.IO/DXEZ8).If this code runs on an online server, data sent over LoRa can be turned into online visualizations.This creates a real-time measurement network.This data transmission method can also be used to check the SPSA, such as the general operation, the real-time clock, and the measured values.

Operational reliability evaluation
This section discusses to what extent the SPSA, our setup, is operationally reliable in the context of Arba Minch and Addis Ababa.The evaluation methodology section describes how the SPSA is used under different circumstances.The operational reliability section describes the findings on operational reliability based on this use.

Evaluation methodology
The SPSA is used under various circumstances: indoor and outdoor, stationary, and mobile and in high concentration and low concentration environments.For stationary measurements, the instruments were placed between 1.5-2.5 meters above the ground.Mobile measurements have been taken by placing the instrument in the frontside of a public transport tricycle, or by letting students walk around with the instrument on a belt (personal sampling).Most measurements have taken place in Arba Minch, but two set-ups have been installed in Addis Ababa.
A full overview of measurement circumstances is given in Table 3.
Measurements at locations A1-A3 and K1-K4 were conducted for an evaluation of the SPSA data quality.These locations are selected for their distinct ambient and indoor concentration levels.Regarding these locations, more information is given under the measurement locations section of the data quality evaluation.Other measurements were part of student measurement projects.Locations and durations were chosen by the students.
In situations 4, 8, 9 and 14, the instruments were used with a power bank that was charged before each use.In all other situations a continuous power supply was maintained by connecting the instruments to the grid via the power bank.
In situations 1, 2, 10 and 11 the SPSA was deployed with LoRa functionality.Two gateways were installed in Arba Minch.A full network structure including database was not yet built.Instead, the focus was on 'proof of concept': for a certain time having the gateways up and testing the reception at different distances.

Operational reliability
This section describes the experiences related to the operational reliability of the SPSA which were gained while conducting the measurements as described in the evaluation methodology section.

Instrument
Single SPSAs functioned well in all measurement locations, except for the following three issues: • In the first period of field measurements (No. 1, 2), one type of SD cards (brand FFFAS, 16 GB) malfunctioned.After it happened in multiple SPSAs, other brands of SD cards were used.
• In some cases (No. 1,2,3 7,8,14), the battery of the Real Time Clock died, resulting in a reset of the time.In most cases, data could be restored by comparing the (wrong) time on the instrument with the actual time, and changing the time based on that difference.Validation of this was done based on comparison with collocated sensors and occurrence of daily concentration peaks.
• In two cases (No. 1 and 2), an SPS30 sensor was giving extremely high values without an air pollution source nearby.The sensors in question were replaced.Despite the above-mentioned problems, many hours of data have been collected.This shows that the SPSA is operationally reliable.The setup was able to collect data under ambient (1-7), on-road mobile (8,9), indoor (10-13) and personal ( 14) circumstances.It could be used by staff (No. 1-3, 10-13) and students (4)(5)(6)(7)(8)(9)14).Regular check-up is advisable, considering the above-mentioned problems.However, we found that also without check-up some systems could operate without errors for a long period.In Addis Ababa (No. 3) the instruments were installed in July 2022.For logistical reasons there was no checkup from the end of July 2022 until May 2023.

Network
The network potential of a group of SPSAs works well with LoRa.The transmission over LoRa is largely dependent on the proper placement of the gateway.Messages have been received over an 8 km distance.This reception is greatly influenced by obstruction of line-of-sight.One gateway received significantly less messages from the direction where a large building was obstructing line of sight to the SPSA.For about two weeks a full 'network' was running: SPSAs with LoRa functionality were deployed, gateways were up and running, and the Python scripts, as listed in the LoRa network software section, ran on an hourly interval at a Virtual Private Server.During those weeks, an hourly updated real-time graph showed live PM 2.5 concentrations on various locations in Arba Minch.

Data quality evaluation
This section discusses the field testing we conducted with the SPSA for the purpose of data quality evaluation.The evaluation methodology section discusses the methodology we use for evaluating data quality.The concentration patterns section displays what concentrations were measured (to validate that measurements were taken under various conditions).The remaining sections show the data quality metrics for the three comparison types we used: within, gravimetric and SSys.

Measurement locations
Table 3 listed 14 locations where the SPSA has been used.For data quality evaluation, three ambient locations (A1-A3) and four indoor locations (K1-K4) were selected.Only at these locations, the SPSA was collocated with itself and/or other instruments.
For the three ambient locations, sources of PM 2.5 are traffic and neighborhood biomass burning.The three ambient locations represented three distinct concentration levels.The four indoor locations were selected for their use of biomass fuel, to represent (extremely) high and variable concentrations.All four locations were in Arba Minch.Location K1 was in a small local restaurant, in a room connected to the kitchen.
In that room, charcoal fire was used for coffee preparation.In the kitchen, food was prepared with biomass fuel.Location K2 was in the kitchen of a student restaurant.In this kitchen, food was prepared with biomass fuel.Locations K3 and K4 were in a large kitchen with more than six biomass fuel cooking locations.These two locations were on opposite sides of the kitchen, and cooking fires closest to the respective locations were used at different times.For this reason, these locations are considered separate.

Measurement comparison types
We evaluate data quality by comparing measurement results of an SPSA with instruments that are collocated.Three types of measurements have been conducted for data quality evaluation.
The first type compares measurement results of SPSAs amongst each other (within comparison).We placed multiple SPSAs in the same location.
The second type compares SPSA with a gravimetry instrument (gravimetric comparison  , where σi is the standard deviation and µi is the mean of measurements of identical LCS during time period i, and n is the number of time periods., where x i is the concentration of the LCS and y i the concentration of the reference instrument for time period i, and n is the number of time periods.Corrected data xnew based on the old data x old was calculated with Accuracy (Ac) NIOSH <25% Gravimetric The upper value of the confidence interval at 90% of all , where x i is the concentration of the LCS and y i the concentration of the reference instrument for time period i.We used the names Ac _BC and Ac _AC to distinguish between accuracy before and after bias correction, respectively. 1Not all instruments were available during the whole period.Therefore, the relationship between the number of data pairs in evaluations and the period duration is not linear.recognized as equivalent to the gravimetric reference method.Data of this BAM is available online (AirNow Department of State, 2023).We used measurements of this BAM for indicative comparison: indicative because the BAM and SPSA are not at the same location.Location A3 is within the city ring roads, while the location with the BAM is closer to the outskirts of Addis Ababa.A second source of indicative comparison is a gravimetry measurement project at the same location as our measurements (A3).During November 2015 -November 2016, Tefera et al. (2020) collected 69 PM 2.5 samples.Their measurement method is the official PM 2.5 reference method (gravimetry), and their measurements are at the same location, but their measurements are not during the same time period.
We used these measurement results as indicative comparison with respect to the long-term concentration level.
The third type compares measurements of the SPSA with two SSys (SSys comparison): Airvisual Pro (IQAV) and UCB-PATS+ (PATS).Like the SPS30 in our system, both IQAV and PATS estimate the PM 2.5 concentration based on scattering of infrared light (Pillarisetti et al., 2017;Zamora, Rice and Koehler, 2020;Sousan, Regmi and Park, 2021).The PATS is designed for personal sampling and (high) indoor concentrations, but not for low ambient concentrations (lower detection limit is 10 µg/ m 3 ).Hence, we only used the PATS at indoor locations.We used the IQAV both in ambient and indoor situations.The IQAV is not meant for very high concentrations (>5000 µg/m 3 ), because the highest reported value of the IQAV is set to 4488 µg/m 3 .
Table 4 shows an overview of all measurements used for data quality evaluation.

Data corrections
In their guidelines for Demonstrating Equivalence of Ambient Air Monitoring Methods (DEM) the European Commission (2010) allows for removal of up to 2.5% percent of outliers based on Grubb's outlier test at 99% level.We removed outliers with this method for Sp3 and Sp5 at location A2, and for Sp11, Sp12, Iq7 and Iq10 at location A3.
The Sp11 and Sp12 comparison is only done from October 2022, because up to that time there was a real time clock (RTC) issue for Sp11.In contrast, the Sp12 / IQAV comparison could only be done with measurements until October 2022, because after October 2022 both Iq7 and Iq10 malfunctioned.
At locations K3 and K4, there was data loss during the gravimetric measurements.We did not use LCS results with more than 15% data loss during measurements with the gravimetric method in the gravimetric comparison.2010).Evaluation metrics are taken from these sources and summarized in Table 5.

Data analysis
The DEM also prescribes expanded uncertainty (WCM) for ambient reference comparison.That metric applies to reference comparison at 24-hour level.With the gravimetric method available in our study (the UPAS) it was not possible to conduct 24-hour comparison measurements under ambient circumstances.Although Dingemanse (2022) reported this metric, we did not include it in our study.
For data-analysis, we used five time-averaging periods: 10-minute, hour, day, filter-duration, and month.Hour-averages are most used in monitoring networks, while daily averages represent the short-term WHO PM 2.5 guideline value.Filterduration is used in the gravimetric comparison, where the duration varied according to the time needed to get sufficient filter load.We only used monthly averages for the indicative comparison.For the SPSA inter-comparison, we also present 10-minute averages.Table 6 shows a summary of the data quality evaluation.

Concentration patterns
Figure 4 shows the average and 95-percentile concentrations per 10-minute period on a day, for locations A1, A2 and A3.This is not meant to compare locations, since they are not from the same time.It is merely to show the order of magnitude of measured concentrations.
Both slope and bias show that under ambient conditions a calibration of a factor 2 is required.For indoor / high exposure conditions this is not required (|B| < 10%).This observation implies that the SPS30 sensor responds differently to aerosols primarily originating from biomass burning compared to those originating from a variety of ambient sources.

Addis Ababa indicative comparison
Figure 8 shows daily and monthly averaged concentrations for Sp12 in comparison with a BAM at another location in Addis Ababa.Despite these not being on the same location, the correlation is good (R 2 >0.9).This indicates that the SPSA follows the city-wide concentration trend.
On the same location, between November 2015 and November 2016 the average concentration for measurements with the gravimetric reference methods was 53.8 µg/m 3 (Tefera et al., 2020).The average for 10 months of measurements with the Sp12 was 28.7 µg/m 3 .This indicates a factor two difference with the gravimetric method.This is similar to the calibration factor found between SPSA and gravimetry under ambient circumstances in Arba Minch (see the SPSA vs UPAS section).

Comparison with SSys
Table 9 shows the slope and R 2 for SPSA versus the two SSys, IQAV and PATS.It is shown for all hourly and daily averaged data pairs.For some locations there are multiple SPSA.At those locations, the SPSA which has the most data pairs with the collocated SSys is selected.
Correlations are high to both IQAV in ambient (R 2 0.91-0.98)and PATS in indoor (R 2 0.88-1.00)circumstances.For the IQAV we see similar slopes under similar circumstances.For the PATS there is a rather large variability in witnessed slopes.The IQAV under indoor circumstances gives estimations that are too low because all concentrations higher than 4488 µg/m 3 are reported as 4,488 µg/m 3 .Table 10 shows correlation results between SPSA and IQAV for only <4489 µg/m 3 data.For that data, slopes are substantially closer to 1, and R 2 s are on average higher.
Average concentrations at locations A1, A2 and A3 are 12.6, 20.6 and 27.8 µg/m 3 , respectively.10-minute average concentrations ranged up to 120 µg/m 3 , in Addis Ababa (A3) this happened both in the morning and in the evening.
Figure 5 shows box plots of 10-minute, hourly and daily averaged concentrations in the four indoor locations K1-K4.
Averages in the four indoor locations ranged from 300 to 2 000 µg/m 3 , with the highest 10-minute concentrations well over 10 000 µg/m 3 .

Collocated SPSA (within comparison)
Figure 6 shows data pairs for all collocated SPSA measurements, on different time averaging periods.Table 7 shows the data quality metrics.
Both Figure 6 and Table 7 show that multiple SPSA under the same circumstances show an extremely similar signal across all concentration ranges.At location K4, for the hourly and 10-minute averages the R 2 is 0.78-0.87.In all other cases the R 2 is at least 0.98.The CV is lower than the required 10%, and the ubs is lower than the required 2.5 µg/m 3 .

SPSA vs UPAS
Figure 7 shows the UPAS gravimetry and SPSA measurement results.Table 8 shows the data quality metrics.
Accuracy is sufficient in all cases (<25%).For ambient circumstances, the Pearson correlation is too low (<0.9).This most probably has to do with the fact that there have only been

Discussion
The implications of our showcase involve two areas: the operational reliability and the data quality in contexts as encountered in Ethiopia.In the operational reliability and data quality sections, we discuss these two areas subsequently.We discuss field testing in general in the field testing section.

Operational reliability
Regarding operational reliability it was found that not all systems were without errors.However, the problems were solved by replacing materials such as a sensor or a real time clock battery.
A malfunctioning SSys requires shipment to the producer for repair.Our system simply needs replacement of the broken part, which can be done on location.However, to get reliable repair services, knowledge of both air quality and electronics, or a close collaboration with electronics staff is necessary.

Data quality
Regarding the data quality, this case shows high concentration correlations between collocated instruments.This is especially the case for Arba Minch, where abundant measurements were conducted.Less evidence is gathered for Addis Ababa.The strong correlations between SPSAs show high instrument reliability.This implies calibration can happen in retrospect or with other studies.For ambient conditions, the results suggest a strong equivalence with the gravimetric method if calibration is conducted, but more evidence is required.For indoor circumstances, the strong correlation with the gravimetric reference method, in combination with the low bias, shows that the SPSA can be used without calibration.The strong correlations between measured concentrations of collocated     SPSA and SSys imply that SSys can be used as unofficial calibration instruments in local and remote contexts that have no resources for calibration with a gravimetric method.The large difference in required calibration between ambient and indoor circumstances is most likely due to particle differences (Karagulian et al., 2019).For the SPS30, Sousan et al. (2021) found under laboratory conditions highly linear results (r = 0.99) with a reference instrument for three aerosol types.The slopes for each of the aerosol types, however, were different, ranging from 0.75 to 2.0.This underlines the importance of testing an LCS at the location of intended use.We further discuss this under the field testing section.
Of all OEMs created, we have only used the SPS30 in our setup.
Other studies find that of the different OEMs, SPS30 performs well.In their laboratory study, Sousan et al. (2021) compared the SPS30 along with other LCS to a reference instrument.They conclude that the 'SPS30 and OPC-N3 (…) demonstrated the best performance, with high correlation and lowest bias values, for all aerosol types and PM metrics in environmental and occupational settings.(…) In contrast, AirBeam2 and PMS A003 exhibited low accuracy for all aerosol types and PM metrics in both settings.'(Sousan, Regmi and Park, 2021, pp. 23-24).The AirBeam2 uses instruments and calibration like the PurpleAir.The PMS A003 is a Plantower instrument, used both in the PurpleAir and AirQo.A common problem for a PM 2.5 LCS is influence by relative humidity (RH).However, compared to OEMs Plantower and Honeywell, SPS30 has the lowest reaction to RH (Hassani et al., 2023).According to Budde et al. (2018), the SDS011 (used in sensors.Africa) has a significant variance amongst similar sensors and is strongly influenced by relative humidity.Like Sousan et al. (2021), we observed a robust performance of the SPS30 across different concentration ranges.The sensor showed a small variance amongst collocated identical sensors.The sensor is versatile for both indoor and ambient situations, across all concentration ranges.This was less the case for the two SSys used as comparison in our study (IQAV and PATS).A low influence of RH also follows from the fact that over a period of 10 months the SPSA followed the city-wide concentration trend in Addis Ababa, across both dry and wet seasons.

Field testing
It is important to evaluate LCSs in the same context as where they will be used to monitor air quality.For both SSys and OEMs, field validation in low-income country contexts is limited.Karagulian et al. (2019) conducted a review of 112 different LCS (64 independent studies, 31 OEMs and 81 SSys).From all 64 studies, only one concerned a PM 2.5 sensor which is evaluated in a low-income country: the SSys PATS inside kitchens in Guatamala (Pillarisetti et al., 2017).Most projects using SSys and OEMs in Africa do not mention field tests at the location of the study.Awokola et al. (2020) present the data of measurements with a PurpleAir 'as is' and refer for field tests to studies conducted at regulatory sites in the USA (Malings et al., 2020).AfricaAir, using the SSys PurpleAir and Met-One, likewise refer to these studies.We were not able to find any published results of this yet.For indoor contexts, we did not find any field validation of OEMs in a low-income country.Also, none of the sparse ambient field validations make mention of comparison with the gravimetric method.In other words, thus far the field validation of LCS in African contexts has been poor, both for indoor and ambient studies circumstances.Our study is the first to compare the SPS30 under indoor field circumstances to a gravimetric method.Our ambient field testing is also a first of its kind in this size (multiple ambient locations across multiple cities).Still also in our study the ambient testing is limited.We have only used the gravimetric method three times at one ambient location.Currently, data collection for the SPSA is ongoing at ambient locations in Addis Ababa and Adama, with 20 collocated gravimetric samples planned.We used the UPAS for both ambient and indoor measurements for comparison, but the UPAS is designed for indoor concentrations.Using a gravimetric instrument made for ambient monitoring (such as a Low Volume Sampler) would make the reliability of the data quality testing stronger.At this moment, such an instrument is not available in Ethiopia.
In any case, a large advantage of using OEMs instead of SSys is that, when a better sensor enters the market, one can decide to self-assemble the newer sensor in the system.For example, sensors.Africa at this moment uses the SDS011.They are likely to have the infrastructure to switch to an OEM with lower variance and less reaction to RH, like the SPS30 seems to have.

Conclusion
Low-cost sensors are a promising direction for PM 2.5 monitoring in Ethiopia, and self-developing the sensor system with an OEM even more so.While low-cost sensor projects have started in the African continent, most are externally funded and field validation at the location of deployment is still very poor.Our results from Arba Minch and Addis Ababa show the potential for a cheap, flexible, and reliable PM 2.5 measurement instrument, across both ambient and indoor pollution circumstances, based on the OEM Sensirion SPS30.To use it in larger quantities (such as a network), or for others to use it, either a more official approach is required (with the danger of losing ownership), or good cooperation between experts of different professions.
Challenges of working with an OEM are the required expertise and lower ease of use compared to a SSys.These challenges should be seen as opportunities for increasing local skill, local ownership, instruments tailormade for local use, and for achieving the highest data quality for the lowest costs.

Figure 1 :
Figure 1: Sketch of the SPSA circuit.LoRaWAN is not included in the sketch; the Dragino LoRa shield is premade to fit directly on the normal Arduino Uno or Mega pin layout.

Figure 2 :
Figure 2: Configurations of the SPSA.(A) Stationary sampling box with LoRa and BME280.(B) Stationary sampling box without LoRa and BME280.(C) Personal sampling box without LoRa and BME280.

Figure 3 :
Figure 3: Flowchart of software on the SPSA.

Figure 4 :
Figure 4: Average and 95-percentile per 10-minute period of the day for three ambient measurement locations.

Figure 5 :
Figure 5: Box plots of time averaged data for four kitchen locations (K1-K4).The boxes extend from the first quartile (Q1) to the third quartile (Q3) of the data.The whiskers extend from the boxes by 1.5 times the inter-quartile range (Q3-Q1).
For OEMs, Adong et al. (2022) compared the AirQo (OEM Plantower) to a BAM at two ambient locations in Kampala city in Uganda, with average concentrations of 37.5 and 45.1 µg/m 3 .Wernecke et al. (2021) announced tests of LCS and reference-grade instruments collocated at one ambient location, for half a year in 2021.

Figure 8 :
Figure 8: Daily and monthly averaged concentration for one SPSA and a BAM, at distinct locations in Addis Ababa.

•
C1: A larger box for stationary sampling, including LoRa and

Table 2 :
Power consumption calculation for the SPSA.

Table 3 :
Details on measurement locations of the SPSA, including per location the SPSA type, number of SPSAs (#) and the amount of data collected at that location (data quantity) A1 was in front of a house in a quiet neighborhood in Arba Minch (latitude 6.01589 N, longitude 37.55480 E).At the compound of this house, cooking only is done with electric stoves.The nearest road to this location is approximately 120 meters, and this road is only used by local traffic.Location A2 was at the entrance of a hotel compound in the center of Arba Minch (latitude 6.03311 N, longitude 37.55783 E).The instrument was approximately 15 meters from the nearest road.This road is the main road through Arba Minch.It is used by traffic driving through Arba Minch, and by stopping and starting traffic visiting businesses.Location A3 was in Addis Ababa, sub-city Lideta, at the city monitoring station of the Ethiopian Meteorology Institute (latitude 9.01888 N, longitude 38.74728 E).This location is approximately 10 meters from the nearest road.
(Abiye, 2020) Addis Ababa is higher than in Arba Minch, and the number of vehicles per person is also higher.As per 2020, Addis Ababa had registered 630 440 vehicles, while the region SNNP (of which Arba Minch is a part), with a population estimation of 13 044 044, had registered 118 424 vehicles(Abiye, 2020).Location Volckens et al. (2017)nt methods for PM 2.5 are based on gravimetry.As gravimetric instrument, we used the Ultrasonic Personal Aerosol Sampler (UPAS), as this instrument was the only available gravimetric instrument in Arba Minch, Ethiopia.The UPAS is designed for measuring medium to high concentrations.Over ranges of 20-1000 µg/ m 3 ,Volckens et al. (2017)found strong correlations with the EPA federal reference method.Afshar-Mohajer et al.Developing and testing a PM 2.5 low-cost sensor in Ethiopia Page 7 of 15 For Addis Ababa (3), there were no gravimetric measurements taken by us.At another location (Jacros area, Addis Ababa, distance to location A3 8 km, latitude 9.01163 N, longitude 38.82151 E), the US Embassy operates a BAM.A BAM isTechnical article:1 Not all metrics apply in all situations.Where applicable, the table lists the scope of application with respect to type of comparison (within, gravimetric, SSys), time-averaging period (10-minute, hourly, daily, filter-duration or monthly) and/or context (ambient, indoor).

Table 6 :
Data quality evaluation summary.

Table 4 :
Overview of measurements used for the data evaluation of the SPSA.Individual SPSAs are labeled Spx, where x refers to the setup ID number.IQAV and PATS are coded Iqx and Pax, respectively.

Table 7 :
Number of data pairs (N) and data quality metrics for two collocated SPSAs.The results are ordered according to location and time averaging period.

Table 8 :
Number of data pairs (N) and data quality metrics for SPSA compared to UPAS gravimetric measurements.
(Giordano and Jaramillo, 2021) worked well if the sensors were placed outside.A network of LoRa will more likely become operational if additional setups, measuring other environmental data like soil and water quality, are used in the study area.As an alternative to LoRa, in bigger cities with stable internet connection, WiFi is an option.Most SSys use this to turn their systems into a network.Alternatively, GSM communication can be used.More generally, with an OEM-based self-built system the data transmission method can be chosen according to the local context.A SSys with WiFi instead forces the user to provide good internet connections or face data loss.For upscaling the SPSA to a whole network of SPSAs, a higher level of organization is needed.Production, installation, maintenance, data collection and data distribution of many more systems are not likely run by a single initiative taker, such as was the case in this showcase.There are some examples of organized LCS application in the African continent.Awokola et al. (2020)describe the use of the SSys PurpleAir at 13 different sites across 7 African countries.Afriqair.org(GiordanoandJaramillo, 2021)is another African initiative that uses existing SSys (PurpleAir and MetOne).Working with a SSys, however, results in dependency on foreign funding, foreign expertise, specifications that are not necessarily designed for the local context, and it raises questions on data ownership.Local development with OEM sensors, instead keeps the costs lower, gives local staff the opportunity to gain expertise, gives the possibility for systems tailor-made for the local context and gives clear local ownership.We found three African initiatives that create their own sensor system with OEMs: https://sensors.

Table 9 :
Data pairs comparison between the SPSA and SSys.
Technical article: Developing and testing a PM 2.5 low-cost sensor in Ethiopia Page 11 of 15