The measurement of UV radiation is usually based on the inner photoelectric effect occurring in photodiodes. These sensors are built by p-n junctions sensible to the light. The p-n junction is operated in reverse bias preventing any significant current flow unless electron-hole pairs are created by the incidence of photons (Cox, 1998). Thus, the radiation produces an electric current which is, to first order, linearly proportional to the incident irradiance. The material used to construct the photodiode determines the wavelength range to which the sensor is sensitive. For the ultraviolet range (wavelength below 400 nm) silicon is commonly used since it is sensitive between 190 and 1,100 nm (Werner et al., 2000). This sensitivity is not equal for all wavelengths within the interval and, therefore, a specific spectral response function for each sensor must be considered.

The main distribution channels of electronic devices offer a large number of sensors that could be implemented in the design. However, these distribution channels are specialized in high-end sensors and, therefore, not affordable for every user. Since the main idea is to promote the widespread use of low-cost UV sensors, the search focused on more general vendors, looking for sensors designed for mobile recreational electronics, with low-power consumption and being affordable for an average user. Among the variety of sensors available on the market, three different commercial UV-sensitive detectors were chosen because of their high popularity and low cost (less than 5 euros): the ML8511, the UVM30A and the VEML6075 (Figure 1). The great success of these sensors is due in part to their easy control via an Arduino or similar board. For each sensor model, two prototypes were built, labeled 1 and 2.

The ML8511 sensor is manufactured by LAPIS Semiconductor Co. (LAPIS Semiconductor Co., Ltd., 2013). It features a photodiode and an inner amplifier, providing a voltage analog output. This sensor has become notably popular, being mainly used for UV Index calculation (Zhang et al., 2013; Haryanto and Pratomo, 2017), learning electronics (Pramono et al., 2020) and studying the effect of UV radiation over materials (Wijatna et al., 2019).

Figure 2A shows the spectral response function of the different sensors as provided by the manufacturers. The figure also depicts a typical solar spectrum for day March 19, 2019 at 12:00 UTC weighted by the spectral response function of each sensor, in order to show the actual contribution of each wavelength to the final signal given by the sensors (Figure 2B).

The ML8511 shows significant response mainly in the 280–400 nm wavelength range, and some sensitivity above 400 nm (Figure 2A). This sensitivity to the visible wavelengths results in a substantial contribution of the 33% of the total signal output due to the solar visible radiation. The peak sensitivity of the ML8511 is found within the interval 308–376 nm (values above 0.9) (Figure 2A).

The second type of UV module analyzed is the UVM30A, featuring a GUVA-S12SD sensor. This sensor has a simple photodiode GUVA-S12SD and provides analog output. It has been used for vitamin-D analysis (Gao et al., 2014) and to monitor the UV Index (Paredes Ccama, 2017; Quintero Cardenas, 2018; Htwe, 2020). However, most of these studies are limited to academic projects. The GUVA-S12SD sensor is sensitive in the whole UV range (Figure 2A) expanding from 280 to 380 nm, including both UVB and UVA intervals. Its peak sensitivity (values above 0.9) lies between 326 and 361 nm. Only UV wavelengths contributes to its output signal (Figure 2B), better representing the UV radiation than the previously analyzed ML8511.

The third sensor analyzed was the VEML6075. This is an advanced UV sensor manufactured by Vishay Semiconductors (Vishay Intertechnology Inc., 2016). It includes several channels and digital I²C output. It embodies various photodiodes with spectral responses sensitive to different parts of the ultraviolet wavelength range. The so-called channels UVA and UVB are sensitive to 350–370 nm and 320–340 nm, respectively. It also includes two compensation channels designed to counteract the contribution to the output signal of the visible and infrared wavelengths, respectively.

These sensors were used to build a device to measure UV radiation. Several prototypes were manufactured depending on the characteristics to be analyzed and the development stage of the device.

The two sensors with analog output (ML8511 and UVM30A) were connected to a PCF8591 analog-to-digital converter, manufactured by NXP Semiconductors (NXP Semiconductors, 2013). This A/D converter facilitates a 0–255 value and admits up to four analog sensors to be connected, which are managed by a multiplexer. It is a 8-bit widely used converter especially designed for low-consumption systems. It communicates through an I²C bus.

The VEML6075 sensor has already a 16-bit digital output since it integrates an A/D converter. It is also designed for low-consumption systems and communicates through an I²C bus. This sensor includes a Teflon diffuser to improve the angular response and prevent the existence of privileged directions that could result in higher response by the sensor. The software library developed by Adafruit Industries (Fried, 2019) was used for reading the measurements registered by the VEML6075 sensor.

In this work, different microcontrollers programmed with Arduino were used. They were chosen according to the needs of each step. The first prototype used a low-cost ESP8266EX microcontroller manufactured by Espressif Systems (Espressif Systems, 2020). It is easy to manage and has WiFi connectivity. Later on, an Arduino Mini Pro microcontroller was used. Finally, the final prototype included an ATmega328P microcontroller manufactured by Atmel Corporation (Atmel Corporation, 2015), with the same functionalities as the Arduino Mini Pro, but smaller size. Its cost is barely 2 euros. Different I²C buses were used to avoid address conflict when connecting the sensors to the microcontroller.

LoRaWAN was used as communication infrastructure for transmitting the measurements to Internet of Things (IoT). LoRaWAN is the protocol composed by gateways and nodes which is based on the wireless technology LoRa. Its deployment is notably increasing, especially on university campuses. This network standard addresses requirements characteristic of IoT. Since LoRaWAN does not allow the storage of information, the web ThingSpeak was used for that purpose. The first prototype also has WiFi connection, which was the communication used for the first tests and the calibration of the sensors.

The microcontrollers that have been used can not connect directly to the LoRa network and, therefore, a LoRa transceiver governed by the microcontroller is needed to communicate according to the LoRa protocol. For that aim, the low-cost transceiver RFM95W (about 5 euros), manufactured by Hope Microelectronics (Hope Microelectronics Co., 2006), was used. This transceiver is suitable as it operates in the 868.1 MHz band, which is the band specifically reserved for LoRaWAN in Europe.

In the final prototype, the sensor, the LoRa transceiver and the microcontroller were implemented on a small printed-circuit-board (PCB). For this purpose, the open source software KiCad (KiCad, 1992–2019) for electronic design automation was used. This suite allows schematic capture and PCB layout.

Adding the price of the electronic parts, the total cost of the device is barely 15 euros. This low cost makes it suitable for a large number of potential applications.

To test the performance of the low-cost sensors in measuring solar ultraviolet radiation and calibrate them, they were installed outdoors at the BJZ radiometric station of the University of Extremadura (UEx) in the Campus of Badajoz, Spain. Its geographical coordinates are 38.88°N, 7.01°E, 199 m a.s.l. Badajoz is located in Southwestern Spain, near the Portuguese border. It corresponds to a Köppen-Thornwaite Csa climate, with mild winters, and very dry and hot summers. It features one of the highest solar radiation records in Europe, with UVI (ultraviolet index) of 9–10 in summer, and occasional extreme values of 11. The ultraviolet index is an international standard measure of the solar irradiance that produces sunburn and reaches a given location at a given time. It is calculated by integrating the spectral solar irradiance weighted by the erythemal action spectrum and multiplying the resulting value by the scale factor 40 m²/W to obtain a dimensionless number. The erythemal action spectrum was determined by McKinlay (McKinlay, 1987) and measures the effectiveness of each ultraviolet wavelength in producing sunburn on human skin. Thus, UVI measures the risk of sunburn, providing very important information for preventing harmful effects due to overexposure to solar radiation. Many weather stations provide measurements not of the bare ultraviolet irradiance, but of the erythemally weighted irradiance, i.e., the UVI.

The BJZ station is located on the roof of the Physics building, guaranteeing suitable conditions for radiometric measurements, with an open horizon free from obstacles and daily maintenance of the instrumentation. It is the main site of a radiometric network that covers Extremadura and Western Andalusia with 11 stations equipped with UV broadband radiometers measuring erythemally weighted irradiance (Schmalwieser et al., 2017). This network is operated in cooperation with the National Institute for Aerospace Technology (INTA), and daily reports measured and simulated UVI values. The BJZ station is maintained by the Department of Physics of UEx since it started up in 2002. It is fully instrumented with several Kipp & Zonen CMP11, UVS-E-T and CGR1 radiometers to measure solar and terrestrial radiation and its different components (global, direct, diffuse) in different wavelength ranges. Its is also well equipped with a Jenoptik 15 k-Nimbus ceilometer and a SONA 202U all-sky for cloud monitoring. As part of the AERONET network for aerosol measurement (Holben et al., 1998), the station is also instrumented with a CIMEL-318 and follows AERONET’s measurement and calibration protocols.

Regarding UV radiation, the station is equipped with three Kipp & Zonen UVS-E-T broadband radiometers to measure global erythemal solar radiation and its diffuse component independently. For diffuse measurements, the radiometers are shaded by different devices, such as a shadow band and a sun tracker with shading balls.

The UVS-E-T radiometers were calibrated during the 14th Intercomparison Campaign of the Regional Brewer Calibration Center-Europe held in “El Arenosillo” Atmospheric Sounding Station in Mazagon (Spain) from 17 to June 28, 2019. Their spectral and angular responses were characterized in laboratory following procedures that have been contrasted with those used at the PMOD/WRC (Gröbner et al., 2006). The European standard for spectral UV measurements, the QASUME unit (Bais et al., 2003), was used as reference for the comparison outdoors. The calibration followed the procedure recommended by the COST Action 726 for broadband UV radiometers (Webb et al., 2006; Vilaplana et al., 2007).

Once the sensors have been leveled and their absolute calibration and angular response have been studied, the next step is to investigate whether a portable device can be built. The main requirements for this aim are: to be compact, small, wireless, and with long distance connection and low power consumption. With these requirements, a second version of the device with an ATMega328P microcontroller was built (Figure 4 (left)). In this step, the VEML6075 sensor was used because it is more demanding than the other sensors. Thus, the fulfillment of the requirements by this sensor guarantees the compliance of the other sensors.

The WiFi connection had to be replaced by LoRa communication to make it suitable for distance and energy requirements. Thus, LoRa communication was preferred in this study since, due to its shorter bandwidth, it requires lower energy consumption than other technologies such as WiFi or GSM. Additionally, LoRa modulation provides more link budget and, therefore, longer reach. Thus, the LoRa 868.1 MHz propagates better than the 2.4 GHz used by WiFi. The existing LoRaWAN gateway on the University Campus was used to upload the measurements to the Internet. This gateway operates in the 868.1 MHz band. A LoRa RFM95W module connected to the Arduino Mini Pro was used to communicate over the LoRaWAN network. This module serves as the interface between the microcontroller and the LoRaWAN gateway. The module was registered with credentials on the TTN (The Things Network) website and programmed according to the instructions given by (Telkamp and Kooijman, 2019). The payload transmitted through LoRa included 2-bytes data of UVA, UVB, compensation channel 1, compensation channel 2 and battery voltage, totaling 10 bytes. The LoRa setup recommended by LoRaWAN was used, consisting of a spreading factor of 7 and 125 KHz bandwidth. The resulting transmission rate is 5.5 Kbits/s, low but enough for this type of applications, and the time on air 61 ms. Since data were sent every minute, the duty cycle was 0.1%, being much lower than the limit of 1% allowed for the 868.1 MHz, as established by ETSI EN300.220.

To be portable and autonomous, the device must be powered by batteries instead of being plugged into the electrical grid. Therefore, its energy consumption and autonomy must be assessed. To achieve low power consumption and consequently long autonomy, all components were selected to be energy efficient. The second prototype was powered by a 750 mA⋅h Li-Ion battery. The microcontroller is put into a sleeping mode to reduce power consumption when it is neither receiving nor sending data. A first study of each individual component identified the VEML6075 sensor as a major source of power consumption. Then, it was forced to shut down when not measuring, resulting in a reduction in consumption from 500 to 6.8 μA during the sleeping mode. Following this individual study, a comprehensive study of the consumption of the complete device was conducted for a full cycle (including active and sleeping stages). Three active modes were identified: measurement, communication check and data sending. The frequency of measurements determines the number of cycles and, subsequently, the consumption, which limits the autonomy time. For this second version device, it was decided to measure every 30 s. However, the third and final version included a 120 mA⋅h battery. Also it used a lower frequency of one measurement per minute in order to enlarge its autonomy.

With all the knowledge acquired, a third and final device was developed. It was designed from scratch for better optimization. Its main requirements concerned battery supply, minimum power and physical resources, compact and small size. A minimum set of electronic components was selected and they were individually required to demand low power consumption.

KiCad software was used to design the two (front and rear) faces of the PCB, including the microcontroller, sensor, LoRa module and charge module for the battery. Besides the ATMega328P, the microcontroller included also a resonator, in charge of generating the periodic clock signals used by the microcontroller. In addition, a button was included to reset the device when necessary. The charge module integrates several components: an USB OTG port, TP4056 charge regulator, DW01-P protector, FS8205A double transistor, HT7333 linear regulator of voltage that guarantees 3.3 V supply, and LIR2450 Ion-Li battery with a maximum charge of 120 mA-h. The PCB board was designed to be as small as possible. It was manufactured using a LPKF ProtoMat H100 milling machine. The front face contains a larger number of components and was soldered using an INSA M962A reflow oven. Subsequently, the rear face was soldered by hand. This final device is shown in Figure 4 (right).

The leveling results provided the first insight into the angular response issues shown by the low-cost UV sensors. Figure 5 depicts the output signal obtained when the two ML8511 sensors and the two UVM30A sensors were horizontally leveled.

It should be noted that, although the leveling was performed in two perpendicular directions that are supposed to be independent, adjusting the leveling along the azimuth direction 0°–180° slightly affected the already leveled 90°–270° values. Therefore, the best leveling could only be achieved by successive adjustments, and the final output signals are shown in Figure 5. It can be seen that no perfect leveling was possible, with small differences remaining between specific opposite directions. In addition to the impossibility of leveling each pair of directions independently, once leveled, a notable difference was observed between the output signal measured at the 90°–270° and 0°–180° directions. This difference is approximately 12% for the ML8511.1, 8% for the ML8511.2, 8% for the UVM30A.1, and 13% for the UVM30A.2. These facts are the first signal of an irregular angular response.

As mentioned before, the VEML6075 module contains a sensor for UVB (labeled channel B) and another one for UVA (labeled channel A). Figure 6 show the output signal for the two VEML6075 sensors once leveled. The leveling procedure showed that no simultaneous leveling of the channels A and B could be achieved. Thus, leveling the output signal of channel A caused unleveling the output signal of channel B, and viceversa. This is due to the fact that each channel uses its own photodiode. Similar behavior was found for the compensation channels, which could not be leveled simultaneously to channels A or B. Therefore, a compromise solution had to be taken. Since two VEML6075 modules were available, it was decided to level the channel A of one sensor, and the channel B of the other. An additional problem was found for the already leveled channels; its output signal in directions 90°–270° and 0°–180° differed. Thus, once leveled, channel A showed a difference of 5.1% between 90°-270° and 0°–180°, and channel B showed a difference of 11.1% between 90°-270° and 0°–180°.

Figure 7A shows the output signal from the ML8511.1 and ML8511.2 sensors when measuring solar radiation outdoors during the cloud-free day March 19, 2019, along with the irradiance estimates provided by the model as weighted by the spectral response of the ML8511. The stepped aspect of the figure is due to the number of bits in the converter which, in the case of the PCF8591, is 8, resulting in values between 0 and 255. It is worth remembering that the sensors show similar behavior but different amplitude, indicating a different value of the calibration factor. In addition, a slight asymmetry is observed suggesting an azimuth-dependent angular response.

The output signal of both ML8511 sensors increase approximately linearly with the solar irradiance (Figure 7B), but with different slopes, indicating the need for the manufacturer to provide an individual calibration for each sensor. The relative difference in the calibration slope between the two ML8511 sensors is 13%. Table 1 shows the main statistics of the linear regressions of the different sensors used in this study. The ML8511 sensors present the best values of the coefficient of determination (R²) and a root-mean-square error (RMSE) around 2 (W/m²).

Figure 8A shows the diurnal evolution of the solar UV radiation as measured by the two sensors UVM30A along the March 19, 2019. The modeled irradiance weighted by the UVM30A spectral response has been also plotted, showing the expected behavior. Although the output signal does not follow exactly that expected evolution, a generally symmetric pattern is found. This symmetry is clearer in UVM30A.1, while the UVM30A.2 pattern is slightly skewed to the left. A symmetric pattern with respect to the noon indicates no azimuth dependence, as it is required for a reliable monitoring.

The UVM30A sensors present the best RMSE values, being only 1.08 W/m² for UVM30A.1 (Table 1). Although both sensors show a similar daily pattern, their amplitudes differ substantially (Figure 8A). This is confirmed by the regressions (Table 1), with a slope for UVM30A.1 being 0.6 times the value of UVM30A.2.

Figures 9A,10A show the diurnal evolution of the solar UV irradiance as measured by sensors VEML6075.1-A and VEML6075.1-B, along with the irradiance estimated by the SBDART model and weighed by the specific spectral response function of each sensor. Both sensors roughly follow the diurnal pattern of the radiation. However, their diurnal curves do not show the expected symmetry with respect to the solar noon, indicating some azimuth dependence. Additionally, some fluctuations can be observed that are more evident in sensor VEML6075.1-A than in VEML6075.2-B.

The regression of the output signal of the sensors versus the irradiance is shown in Figures 9B, 10B. Although the sensitivity of the sensors to the UV radiation is confirmed, the agreement is worse than for the other sensors analyzed, with R² values of only 0.944 and 0.863 for sensors VEML6075.1-A and VEML6075.2-B, respectively. It is worth to remember that channel A is studied for sensor VEML6075.1, and channel B for sensors VEML6075.2 and, therefore, no similar slope values can be expected.

The diurnal evolution of the output signal of the sensors during a cloud-free day resulted in asymmetric curves for some of the tested sensors. Since the day was rather stable in terms of ozone and aerosols, and no asymmetries were found in the measurements performed by a high-quality secondary standard Kipp & Zonen UV-S-E-T radiometer 3, it can be concluded that these asymmetries correspond to the angular response of the low-cost sensors. To investigate this topic in more detail, a specific experiment was conducted following the procedure described in the Materials and methods section.

Figures 11–13 illustrate the azimuth response of the different sensors as measured outdoors while rotating the turntable in order to test the response to different azimuth angles with respect to the sun. The figure on the left illustrates the output signal measured at different times in the day (13:00, 15:00 and 17:00 UTC). Normalized values of the output signal, instead of absolute values, have been plotted to allow comparison for different times of the day. Thus, the values have been divided by the mean value and the curve has been interpolated by splines in order to better show the angular response. The polar figure on the right shows the time-averaged values for each available azimuth angle.

The angular responses of ML8511.1 and 0.2 are shown in Figure 11A,B, respectively. Both sensors show nearly identical response with a marked maximum around 100° and a minimum around 0°, with a variation amplitude of 77%. This high difference between minimum and maximum highlights the importance of paying attention to the angular response in the design and manufacturing processes of a radiation sensor. The dependence on the azimuth angle can be clearly observed in Figure 11D, where the mean azimuth response is shown.

It is to note the displacement in the angle corresponding to the maximum from 90° at 15:00–120° at 17:00. This combined dependence on the zenith and azimuth angles indicates a likely issue related to tilting or misalignment of the photodiodes with respect to the light entrance.

The azimuth responses of the sensors UVM30A.1 and 0.2 are shown in Figure 12. Measurements are only available at one time (15:00) for UVM30A.1 and at two different times (15:00 and 17:00) for UVM30A.2 due to a connection failure.

The UVM30A.1 sensor shows very little dependence with the azimuth angle, while the sensor UVM30A.2 shows a marked azimuth response with an amplitude range about 45%. This amplitude increases with the zenith angle, going from around 40% at 15:00 to around 50% at 17:00. The two sensors agree in the location of maximum response at 120° and minumum at 270°.

Figure 13 shows the normalized signal output of the channels A and B of the VEML6075.1 and 0.2 sensors, respectively, at three different times (13:00, 15:00 and 17:00) of the day. It can be clearly observed that the response highly depends on the azimuth, with four local minima around 0°, 90°, 180° and 270° and four local maxima around 30°, 150°, 220° and 330°. The amplitude of these minima and maxima increases with the zenith angle, indicating the angular response depends simultaneously on both solar angles, zenith and azimuth. The behavior shown by channels A and B is very similar. The mean azimuth response is plotted in Figure 13 (right), clearly indicating the existence of privileged orientations.

The third and final version (Figure 4 (right)) features LoRa connectivity for long distance. This LoRa connectivity was tested sending messages from different locations within the University Campus, covering a 1.4 km × 0.37 km area (Figure 14). The LoRaWAN gateway was installed on the roof of the Physics building, with an open horizon free from obstacles. The strength of the connection to the LoRa network is measured by the RSSI (Receiver Signal Strength Indication). This parameter is directly related to the remaining link budget, i.e., with the intensity of the signal received at the receptor. Given the LoRa link budget, if the payload, the power transmitted to the signal, the antenna gain and the distance and obstacles between the transmitter and the receiver are known, the RSSI can be obtained. The RSSI value is a negative number scaled in decibels, whose absolute value increases with decreasing signal, i.e. with increasing distance to the gateway or interfering obstacles.

Thus, the LoRa connection of the device was assessed at different locations within the University Campus. First, the device was placed next to the gateway, obtaining an RSSI of −27 dB. Then the device was moved to other locations and RSSI values were recorded (Figure 14). It was observed that RSSI decreases rapidly as moving away from the transmitter. An RSSI value of -84 dB is obtained at tens of meters from the transmitter even without obstacles. Thus, the strength ranged from −27 dB at 1 m from the gateway, to −122 dB at 698 m at the farthest location to the East in the University Campus. This reduction was due to distance and also to the attenuation by physical obstacles such as buildings and tall trees. Thus, while RSSI is −108 dB at 749 m to the West, it decreases to −122 dB at only 698 m to the East, being this difference mainly due to the attenuation by the buildings and tall trees. Therefore, the LoRa communication proved to be suitable for the purpose of the study.

This third and final version device benefited from the knowledge acquired during the development of the previous versions, as described in the Materials and Methods section. Power consumption was another focus area specifically improved during the development process. The consumption per cycle of this final version in the active mode was accurately measured by a Picotech Picoscope 5203 digital oscilloscope and an Aim-TTi I-prober 520 clamp ammeter, obtaining a mean value of (40.8 ± 3.1) mA⋅s. In the sleeping mode, it was reduced to (556 ± 6) μA⋅s, as measured by an AMPROBE 37XR-A microammeter. In total, the consumption over a full cycle is (41.4 ± 3.3) mA⋅s. Thus, if a LIR2450 battery with a charge of 120 mA⋅h is used, the device can perform (10400 ± 800) 1-min cycles. This means an autonomy of seven complete days, or 14 days if working only during daylight hours.