Constructing a pollen proxy from low-cost Optical Particle Counter (OPC) data processed with Neural Networks and Random Forests

Local spatiotemporal dynamics of particulate matter and oak pollen measured by machine learning aided optical particle counters

Conventional techniques for monitoring pollen currently have significant limitations in terms of labour, cost and the spatiotemporal resolution that can be achieved. Pollen monitoring networks across the world are generally sparse and are not able to fully represent the detailed characteristics of airborne pollen. There are few studies that observe concentrations on a local scale, and even fewer that do so in ecologically rich rural areas and close to emitting sources. Better understanding of these would be relevant to occupational risk assessments for public health, as well as ecology, biodiversity, and climate. We present a study using low-cost optical particle counters (OPCs) and the application of machine learning models to monitor particulate matter and pollen within a mature oak forest in the UK. We characterise the observed oak pollen concentrations, first during an OPC colocation period (6 days) for calibration purposes, then for a period (36 days) when the OPCs were distributed on an observational tower at different heights through the canopy. We assess the efficacy and usefulness of this method and discuss directions for future development, including the requirements for training data. The results show promise, with the derived pollen concentrations following the expected diurnal trends and interactions with meteorological variables. Quercus pollen concentrations appeared greatest when measured at the canopy height of the forest (20-30 m). Quercus pollen concentrations were lowest at the greatest measurement height that is above the canopy (40 m), which is congruent with previous studies of background pollen in urban environments. The attenuation of pollen concentrations as sources are depleted is also observed across the season and at different heights, with some evidence that the pollen concentrations persist later at the lowest level beneath the canopy (10 m) where catkins mature latest in the season compared to higher catkins.

Machine learning methods for low-cost pollen monitoring - Model optimisation and interpretability

Pollen is a major issue globally, causing as much as 40 % of the population to suffer from hay fever and other allergic conditions. Current techniques for monitoring pollen are either laborious and slow, or expensive, thus alternative methods are needed to provide timely and more localised information on airborne pollen concentrations. We have demonstrated previously that low-cost Optical Particle Counter (OPC) sensors can be used to estimate pollen concentrations when machine learning methods are used to process the data and learn the relationships between OPC output data and conventionally measured pollen concentrations. This study demonstrates how methodical hyperparameter tuning can be employed to significantly improve model performance. We present the results of a range of models based on tuned hyperparameter configurations trained to predict Poaceae (Barnhart), Quercus (L.), Betula (L.), Pinus (L.) and total pollen concentrations. The results achieved here are a significant improvement on results we previously reported: the average R2 scores for the total pollen models have at least doubled compared to using previous parameter settings. Furthermore, we employ the explainable Artificial Intelligence (XAI) technique, SHAP, to interpret the models and understand how each of the input features (i.e. particle sizes) affect the estimated output concentration for each pollen type. In particular, we found that Quercus pollen has a strong positive correlation with particles of optical diameter 1.7-2.3 μm, which distinguishes it from other pollen types such as Poaceae and may suggest that type-specific subpollen particles are present in this size range. There is much further work to be done, especially in training and testing models on data obtained across different environments to evaluate the extent of generalisability. Nevertheless, this work demonstrates the potential this method can offer for low-cost monitoring of pollen and the valuable insight we can gain from what the model has learned.

The underpinning factors affecting the classroom air quality, thermal comfort and ventilation in 30 classrooms of primary schools in London

The health and academic performance of children are significantly impacted by air quality in classrooms. However, there is a lack of understanding of the relationship between classroom air pollutants and contextual factors such as physical characteristics of the classroom, ventilation and occupancy. We monitored concentrations of particulate matter (PM), CO2 and thermal comfort (relative humidity and temperature) across five schools in London. Results were compared between occupied and unoccupied hours to assess the impact of occupants and their activities, different floor coverings and the locations of the classrooms. In-classroom CO2 concentrations varied between 500 and 1500 ppm during occupancy; average CO2 (955 ± 365 ppm) during occupancy was ∼150% higher than non-occupancy. Average PM10 (23 ± 15 μgm-3), PM2.5 (10 ± 4 μgm-3) and PM1 (6 ± 3 μg m-3) during the occupancy were 230, 125 and 120% higher than non-occupancy. Average RH (29 ± 6%) was below the 40-60% comfort range in all classrooms. Average temperature (24 ± 2 °C) was >23 °C in 60% of classrooms. Reduction in PM10 concentration (50%) by dual ventilation (mechanical + natural) was higher than for PM2.5 (40%) and PM1 (33%) compared with natural ventilation (door + window). PM10 was higher in classrooms with wooden (33 ± 19 μg m-3) and vinyl (25 ± 20 μgm-3) floors compared with carpet (17 ± 12 μgm-3). Air change rate (ACH) and CO2 did not vary appreciably between the different floor levels and types. PM2.5/PM10 was influenced by different occupancy periods; highest value (∼0.87) was during non-occupancy compared with occupancy (∼0.56). Classrooms located on the ground floor had PM2.5/PM10 > 0.5, indicating an outdoor PM2.5 ingress compared with those located on the first and third floors (<0.5). The large-volume (>300 m3) classroom showed ∼33% lower ACH compared with small-volume (100-200 m3). These findings provide guidance for taking appropriate measures to improve classroom air quality.

Isolating the species element in grass pollen allergy: A review

Grass pollen is a leading cause of allergy in many countries, particularly Europe. Although many elements of grass pollen production and dispersal are quite well researched, gaps still remain around the grass species that are predominant in the air and which of those are most likely to trigger allergy. In this comprehensive review we isolate the species aspect in grass pollen allergy by exploring the interdisciplinary interdependencies between plant ecology, public health, aerobiology, reproductive phenology and molecular ecology. We further identify current research gaps and provide open ended questions and recommendations for future research in an effort to focus the research community to develop novel strategies to combat grass pollen allergy. We emphasise the role of separating temperate and subtropical grasses, identified through divergence in evolutionary history, climate adaptations and flowering times. However, allergen cross-reactivity and the degree of IgE connectivity in sufferers between the two groups remains an area of active research. The importance of future research to identify allergen homology through biomolecular similarity and the connection to species taxonomy and practical implications of this to allergenicity is further emphasised. We also discuss the relevance of eDNA and molecular ecological techniques (DNA metabarcoding, qPCR and ELISA) as important tools in quantifying the connection between the biosphere with the atmosphere. By gaining more understanding of the connection between species-specific atmospheric eDNA and flowering phenology we will further elucidate the importance of species in releasing grass pollen and allergens to the atmosphere and their individual role in grass pollen allergy.

Monitoring and apportioning sources of indoor air quality using low-cost particulate matter sensors

Air quality is one of the most important factors in public health. While outdoor air quality is widely studied, the indoor environment has been less scrutinised, even though time spent indoors is typically much greater than outdoors. The emergence of low-cost sensors can help assess indoor air quality. This study provides a new methodology, utilizing low-cost sensors and source apportionment techniques, to understand the relative importance of indoor and outdoor air pollution sources upon indoor air quality. The methodology is tested with three sensors placed in different rooms inside an exemplar house (bedroom, kitchen and office) and one outdoors. When the family was present, the bedroom had the highest average concentrations for PM_2.5 and PM₁₀ (3.9 ± 6.8 ug/m³ and 9.6 ± 12.7 μg/m³ respectively), due to the activities undertaken there and the presence of softer furniture and carpeting. The kitchen, while presenting the lowest PM concentrations for both size ranges (2.8 ± 5.9 ug/m³ and 4.2 ± 6.9 μg/m³ respectively), presented the highest PM spikes, especially during cooking times. Increased ventilation in the office resulted in the highest PM₁ concentration (1.6 ± 1.9 μg/m³), highlighting the strong effect of infiltration of outdoor air for the smallest particles. Source apportionment, via positive matrix factorisation (PMF), showed that up to 95 % of the PM₁ was found to be of outdoor sources in all the rooms. This effect was reduced as particle size increased, with outdoor sources contributing >65 % of the PM_2.5, and up to 50 % of the PM₁₀, depending on the room studied. The new approach to elucidate the contributions of different sources to total indoor air pollution exposure, described in this paper, is easily scalable and translatable to different indoor locations.