Air Quality Monitoring & Prediction

Air Quality Monitoring & Prediction

V. R. Dhanavarshini

Department of AI&DS Panimalar Engineering, college Chennai, India

R. Kavya

Department of AI&DS Panimalar Engineering, college Chennai, India

S. Akshayprasan

Department of AI&DS Panimalar Engineering, college Chennai, India

S. Laxmipriya

Department of AI&DS, Panimalar Engineering college Chennai, India

Mrs. M Megala

Assistant Professor Department of AI& DS, Panimalar Engineering college Chennai, India

AbstractAir pollution is still one of the biggest problems for the environment around the world. It causes millions of deaths every year because it makes breathing, heart disease, and neurological problems worse. The issue is worse in cities with a lot of people because industrial emissions, car exhaust, and seasonal biomass burning cause pollutant concentrations to change quickly and unpredictably. Standard air quality monitoring stations are accurate, but they are expensive and xed and only give you a general idea of pollution levels in a certain area. To x these problems, this work suggests a new Air Quality Monitoring and Prediction System that uses cheap Internet of Things (IoT) sensor networks and a hybrid Machine Learning (ML) framework. Long Short-Term Memory (LSTM) networks are used by the system to nd temporal dependencies in pollution patterns. Gradient boosting machines improve short-term forecasts, striking a good balance between trend modeling and high-resolution accuracy. Also, Explainable AI (XAI) methods like SHAP supply clear information about how pollutants affect Air Quality Index (AQI) predictions, which builds trust in the results. A new crowdsourced calibration system lets community devices check and improve sensor readings, which lowers drift errors over time. Also, real- time geospatial AQI heatmaps show pollution hotspots in great detail, making it easier to make targeted policy changes. Exper- imental evaluation shows that the proposed hybrid-XAI method makes AQI forecasting more accurate by up to 18% compared to baseline models. It also greatly improves situational awareness for both citizens and policymakers. This integrated methodology connects affordability, interpretability, and predictive reliability, providing a scalable way to manage urban air quality based on data.

Index TermsAir Quality Monitoring, IoT, Machine Learning, LSTM, Gradient Boosting, Explainable AI, AQI Prediction, Environmental Monitoring.

1. Introduction

   One of the most important environmental and public health concerns of the twenty-rst century is air pollution. According to the World Health Organization (WHO), more than 99 percent of people worldwide breathe air that is polluted above acceptable limits, which causes 7 million preventable deaths annually. Pollutants such as carbon monoxide (CO), sulfur dioxide (SO2), nitrogen dioxide (NO2), ne particulate matter PM2.5, and ozone (O3) not only reduce air quality but also

   increase the risk of infections, cardiovascular disease, respi- ratory conditions, and cognitive decline. The health burden is particularly high in industrial hubs, densely populated urban areas, and areas affected by dust storms or seasonal biomass burning. Poor air quality not only endangers human health but also destroys ecosystems and reduces agricultural produc- tivity. Despite the urgency, the current air quality monitoring system has signicant aws. These shortcomings can lead to delayed responses in addressing hazardous conditions, poten- tially putting public health at risk. It is crucial to invest in more advanced technologies and data analytics to enhance the ac- curacy and reliability of air quality assessments. Despite their high accuracy, traditional monitoring stations are prohibitively expensive, require regular calibration by trained personnel, and are often deployed in small quantities due to high operating costs. These stations, which typically provide readings at xed sites, do not capture the ne-grained geographical variability of pollution, especially in diverse metropolitan contexts. This limited coverage leads to data blind spots, which prevent citizens and decision-makers from accessing current, localized information on air quality conditions. Furthermore, most of the systems in use today are passive monitors with little to no predictive ability to anticipate pollution spikes before they occur.Low-cost, scalable, and predictive air quality monitoring systems are increasingly being researched in light of these difculties. The rapid development of the Internet of Things (IoT) has enabled the deployment of affordable sensor net- works capable of collecting data continuously and widely. When paired with machine learning (ML), these systems are able to measure and forecast trends in the Air Quality Index (AQI). This makes it possible to take swift action, such as setting industrial emission limits, changing trafc patterns, or sending out public health alerts.The development of affordable, scalable, and predictive air quality monitoring systems is becoming more and more popular in light of these difculties. The Internet of Things (IoT) explosive growth has made it possible to deploy reasonably priced sensor networks that can gather data widely and continuously. These systems

   can measure and predict trends in the Air Quality Index (AQI) when combined with machine learning (ML). This enables prompt action, like imposing limits on industrial pollution, altering trafc patterns, or issuing public health alerts.

2. Related Work

   Air quality monitoring and prediction has gained signicant attention in the last decade due to its public health and environmental implications. Various approaches have been ex- plored, including government-grade monitoring stations, low- cost IoT networks, machine learning-based forecasting, and community-driven sensing. While each approach offers unique advantages, they also have notable limitations, which we aim to address.

   Low-Cost IoT Sensing [1] developed a cost-effective indoor air quality monitoring system using MQ-series gas sensors and particulate matter sensors with microcontroller integration. While effective for continuous indoor monitoring, the system lacked outdoor scalability, predictive capability, and automated calibration.

   LSTM-Based Air Quality Forecasting[2] applied Long Short-Term Memory (LSTM) networks for predicting PM

   1. Comparative Analysis [13]

      Table I summarizes the reviewed studies, highlighting the differences in hardware, modeling approach, explainability, crowd calibration, and visualization features.

      TABLE I

      Paper	Hardware	ML Model	XAI	Crowd Calib.	Geo-Vis	Limit
      Othman et al. (2024)	MQ, PM sensors (indoor)		No	No	Minimal	Indoor- only; no predic- tion
      Drewil et al. (2022)	Public datasets	LSTM	No	No	No	Black- box; poor spike handling
      Ramadan et al. (2024)	Industrial sensors + edge	Domain- specic ML	No	No	No	Industrial focus only
      Chakraborty et al. (204)	Public datasets	RF, XG- Boost	LIME	No	No	Post-hoc only
      Huang et al. (2019)	Mobile citizen sensors	Statistical	No	Yes	Yes	Data quality issues
      Maah et al. (2021)	Low- cost + ref sensors	Regression	No	No	No	Calibration chal- lenges
      Rajesh et al. (2025)	Mixed sensors	Ensemble	Yes	No	Some	No crowd calibra- tion
      Carotenuto et al. (2023)	Dense city nodes	Statistical	No	No	Yes	drift

       

      Comparison of Existing Air Quality Monitoring Approaches

      2.5

      and PM10 levels in urban environments. Their work demon- strated strong temporal trend learning but lacked interpretabil- ity and struggled with abrupt pollution spikes[17].

      Industry-Oriented AI for Emission Monitoring Ramadan

      [3] presented an AI-powered industrial emission monitoring system using edge devices connected to cloud servers. Al- though suitable for industrial scenarios, the approach was domain-specic and not optimized for large-scale city-wide deployment[15].

      Explainable AI in AQI Prediction Chakraborty [4] inte- grated SHAP and LIME explainability into AQI prediction models, improving stakeholder trust. However, their approach applied explainability post-hoc rather than embedding it into the real-time prediction pipeline.

      Crowdsourced Urban Air Quality Mapping Huang [5] em- ployed portable citizen-operated sensors for urban air quality mapping. This enhanced spatial coverage but suffered from inconsistent data quality, sensor drift, and the absence of a robust calibration process.

      Calibration and Sensor Drift Studies Maah [6] reviewed calibration methods for low-cost environmental sensors, con- cluding that without regular recalibration, prediction accuracy deteriorates signicantly over time[14].

      Operational XAI-Integrated Forecasting Rajesh [7] imple- mented a hybrid ensemble forecasting model with embedded SHAP explanations for AQI predictions. While a step toward real-time interpretability, it did not incorporate crowdsourced calibration or high-resolution geo-visualization. Dense Net- work Deployments Carotenuto [8] deployed dense networks of low-cost AQI sensors in urban settings, revealing micro-scale pollution variability. However, sensor maintenance, communi- cation failures, and drift remained challenges.

   2. Identied Gaps[11]

      From the literature, four critical gaps are identied:

      1. Lack of real-time embedded explainability in AQI pre- diction.
      2. Limited robustness to sudden pollution spikes in ML models.
      3. Absence of automated crowdsourced calibration mech- anisms.
      4. Weak integration of hybrid forecasting, interpretability, and geo-spatial visualization[16].
   3. Positioning Our Work[19]

   Our proposed system integrates low-cost IoT sensors (MQ135, SDS011, DHT11, GPS on ESP32), a hy-

   brid LSTMGradient Boosting prediction model, embedded SHAP-based explainability, crowdsourced calibration, and real-time geo-spatial AQI heatmapsaddressing all identied gaps and providing a scalable, transparent, and community- driven solution[20].

3. Proposed Methodology

   To provide precise, comprehensible, and community-driven AQI forecasting, the proposed Air Quality Monitoring and

   Prediction System combines explainable articial intelligence (XAI), a hybrid deep learningensemble prediction model, and inexpensive IoT sensing hardware. Preprocessing, predictive modeling, interpretability analysis, visualization, and action- able alerting mechanisms are all part of the methodologys end-to-end framework, which starts with data acquisition at the edge. Every step is thoughtfully planned to guarantee afford- ability for broad implementation while preserving inclusivity, scalability, and dependability.

   1. System Architecture Overview

      The systems overall architecture is divided into ve main layers that cooperate to provide intelligence throughout the pipeline. First, using dispersed, inexpensive IoT sensors placed throughout urban, peri-urban, and rural regions, the Sensing and Data Acquisition Layer is in charge of gathering environ- mental data. Raw pollutant and weather-related readings are produced by these devices. Before secure cloud-based trans- mission using Wi-Fi or MQTT protocols, the Edge Processing and Transmission Layer makes sure that initial preprocessing like timestamping, noise ltering, and packet packaging takes place.

      Fig. 1. Architecture Diagram

      The third stage, the Data Preprocessing and Storage Layer, is designed to ensure reliability and consistency of the gathered data by removing outliers, lling missing values, and normal- izing readings against environmental variations. This cleaned and geo-tagged data is then fed into the Hybrid Machine Learning Layer, which integrates Long Short-Term Memory (LSTM) networks for capturing temporal dependencies with Gradient Boosting methods for rening short-term predictions. Finally, the Visualization and Alert Layer provides intuitive dashboards, spatial heatmaps, predictive trends, and com- munity leaderboards. The inclusion of alerting mechanisms ensures that both citizens and policymakers receive proactive notications.

      This multi-layered design not only guarantees robustness and scalability but also promotes transparency and inclu- siveness. Each functional block supports modular upgrades,

      allowing the system to evolve with improvements in sen- sor technology, machine learning algorithms, or visualization frameworks.

   2. Data Acquisition

      The use of distributed and inexpensive IoT sensing hard- ware that records a variety of weather and pollution-related parameters is a crucial component of the suggested system. By carefully placing sensors throughout urban areas, residen- tial neighborhoods, industrial clusters, and rural belts, high spatial resolution is attained. The sensors include the SDS011 laser dust sensor for accurate particulate matter measurement (PM2.5 and PM10), the MQ-135 gas sensor, which detects harmful gases like NO2, NH3, benzene, and CO2 equivalents, and the DHT11 sensor, which measures ambient temperature and humidity, variables known to affect pollutant dispersion dynamics. Furthermore, by integrating latitude and longitude coordinates into the data stream, the GPS module (NEO-6M) guarantees spatial traceability and enables geospatial analysis of pollution hotspots.

      Every sensor node is controlled by an ESP32 microcon- troller, which serves as a data aggregator and a lightweight edge preprocessor before sending the data to the cloud us- ing MQTT or Wi-Fi protocols. By using a crowdsourced deployment approach, the system becomes community-driven, allowing local households, schools, and institutions to install nodes voluntarily. By increasing data granularity, this method makes it possible to gain micro-level insights into local pollu- tion and supports an ecosystem of environmental monitoring that is more democratic. In order to address the issues of air pollution, the design places a strong emphasis on accessibility, affordability, and participatory governance.

   3. Data Preprocessing

      To guarantee quality and preparedness for predictive mod- elng, the gathered data is put through a rigorous prepro- cessing pipeline. Due to hardware constraints, connectivity problems, or environmental interference, raw IoT sensor read- ings are frequently prone to a variety of anomalies, including drifts, spikes, and missing values. A multi-stage preprocessing method is used to lessen these effects. First, inconsistent read- ings brought on by eeting spikes are eliminated using outlier detection and removal techniques like z-score-based ltering and Interquartile Range (IQR). After that, temporal sequence modeling and linear interpolation are used in missing value handling to restore missing segments and preserve continuity in time-series data.

      In order to fairly account for variations brought on by atmospheric conditions, environmental normalization is used to adjust pollutant levels against contextual weather vari- ables like humidity, temperature, and wind. Sensor readings are aligned within a single spatiotemporal framework by applying geotagging and time synchronization to each data point. Using crowdsourced calibration to incorporate drift correction mechanisms is a signicant innovation. To address systematic biases, data from several inexpensive sensors is

      cross-referenced with neighboring nodes and reference-grade monitoring stations. Community involvement becomes a data quality enhancer as a result of this crowdsourced validation, which guarantees that reliability increases steadily over time.

   4. Hybrid Machine Learning Model

      A hybrid predictive model, which combines the advantages of ensemble learning and deep learning, is at the core of the system. Because of their demonstrated ability to identify temporal dependencies in sequential data, Long Short-Term Memory (LSTM) networks are used. They are very good at spotting recurrent cycles of pollution, like daily peaks during rush hour or seasonal spikes in particulate matter brought on by the winter inversion. However, when short-term non-linear inuences (like unexpected rainfall or industrial emissions) are present, LSTM predictions are vulnerable to residual errors.

      A secondary layer employing gradient boosting algorithms (such as XGBoost or LightGBM) is introduced to overcome this constraint. Gradient Boosting is an excellent method for modeling non-linear feature interactions and handling struc- tured tabular data. By learning from the residuals left by LSTM outputs, it serves as a correction layer that improves the accuracy of short-term predictions. A weighted ensemble of the nal AQI prediction is produced:

      AQIpred = · AQILST M + (1 ) · AQIGB

      where is optimized through cross-validation. This hybrid de- sign ensures that both long-term dependencies and short-term variations are captured, striking a balance between accuracy, robustness, and generalizability across diverse environmental contexts.

   5. Explainable AI (XAI) Integration

      The black-box nature of traditional machine learning models is a major aw that restricts interpretability and erodes public condence. In order to address this, the suggested system incorporates SHapley Additive exPlanations (SHAP) to offer insight into the models decision-making procedure. The con- tribution of each feature (such as PM2.5, NO2, humidity, and temperature) to the nal AQI prediction is quantied by SHAP. This makes it possible for policymakers and end users to access predictive results and comprehend the logic underlying them.

      The system illustrates which pollutants are driving AQI degradation in particular regions or time windows by visu- alizing feature importance trends at both temporal and spatial scales. These observations can guide policy measures, like determining NO2 concentrations in industrial belts or trafc- induced PM2.5 peaks close to highways. Transparent expla- nations also increase public trust because they help citizens understand the reasons behind warnings, which increases the likelihood that they will take the suggested preventive action. Thus, XAI strengthens the legitimacy of AI-driven environ- mental governance by bridging the gap between technical prediction and social adoption.

   6. Visualization and Real-Time Alerting

      To guarantee that forecasts result in useful insights, the system offers robust visualization and alerting features. A web-based dashboard provides an easy-to-use interface with graphical representations of pollutant breakdowns, live AQI readings, and predictive trends. Trend graphs facilitate histor- ical analysis and allow citizens and policymakers to monitor long-term changes. Furthermore, the system is geographically intuitive thanks to geo-spatial heatmaps that highlight real- time pollution hotspots and are created using Leaet.js or the Google Maps API. In order to warn communities before the AQI deteriorates to dangerous levels, predictive intelligence is used to create proactive alerts via email, WhatsApp, or SMS. Vulnerable populations, including children, the elderly, and people with respiratory disorders, are empowered to take preventative action thanks to this early warning system. The addition of community leaderboards, which rank neighbor- hoods according to average AQI, further gamies clean-air initiatives. In the end, this turns environmental monitoring into a shared duty by promoting civic engagement and inspiring communities to embrace greener practices.

   7. Uniqueness of the Proposed Approach

   The novelty of this methodology lies in its holistic inte- gration of IoT sensing, hybrid learning, and explainability. Unlike conventional systems that focus solely on monitoring, this approach emphasizes prediction, explanation, and commu- nity engagement. The hybrid learning model ensures superior accuracy by balancing long-term and short-term variations, while the crowdsourced calibration mechanism continuously improves sensor reliability through collective intelligence.

   The adoption of explainable AI introduces transparency, which is critical for policy-level adoption and public trust. Unlike opaque black-box predictions, SHAP-based interpre- tations clearly highlight which factors are driving pollution at any given time and location. The system also emphasize geo-spatial and temporal intelligence, enabling simultaneous analysis of where and when pollution peaks occur. Most importantly, the focus on action-oriented alerts transforms monitoring from a reactive process into a preventive one, thereby minimizing health risks and enabling timely interven- tions.

   Thus, the uniqueness of this project lies not only in its technical innovation but also in its societal relevance, bridging the gap between smart technology and sustainable community- driven solutions.

4. Work and Implementation

   The suggested Air Quality Monitoring and Prediction Sys- tems actual implementation is covered in detail in this section. To guarantee modularity, scalability, and robustness, the work was broken up into several phases. Every step was thoughtfully planned to accommodate software and hardware limitations while preserving novelty at the research level.

   1. Hardware Implementation

      The systems foundation is the hardware implementation. Every IoT node incorporates a number of inexpensive, depend- able sensors made for various environmental circumstances. Because of its broad sensitivity range for gases like NH3, NOx, CO2, and benzene, the MQ135 sensor was selected. Based on laser scattering technology, the SDS011 sensor gives precise readings for PM2.5 and PM10, two factors that signicantly inuence AQI uctuations. Temperature and humidity readings from the DHT11 sensor have a direct impact on how pollutants spread and change chemically in the atmosphere. Lastly, all sensor readings are geotagged using a GPS module (NEO- 6M), allowing for localized air quality analysis.

      Because of its dual-core CPU, low power consumption, and buit-in Wi-Fi, the ESP32 microcontroller serves as the central processing unit. It ensures effective cloud communication by carrying out edge-level preprocessing tasks like timestamping and noise ltering. Rechargeable batteries that support solar charging power each node, making them environmentally friendly for outdoor deployments. Depending on connectivity, the nodes use MQTT over Wi-Fi or LoRaWAN to send sensor data every 60 seconds. Rapid deployment in residential neighborhoods, industrial zones, and schools is made possible by the modular design, which guarantees dense and varied coverage. With sensor redundancy and calibration techniques, this decentralized hardware conguration ensures reliability while providing exibility, scalability, and cost-effectiveness in contrast to conventional xed stations.

   2. Software Stack and Cloud Integration

      A tiered architecture is used in the software implementation to facilitate effective analytics and communication. At the edge, real-time sensor acquisition, packet formation, and sim- ple noise ltering are guaranteed by Arduino-based rmware running on the ESP32. For lightweight communication, data is serialized in JSON format. MQTT, a publish-subscribe protocol designed for Internet of Things devices, is used at the transmission layer to provide low-bandwidth and secure communication.

      Incoming sensor packets are ingested by a Flask-based REST API running on AWS EC2 on the cloud side. A NoSQL database (MongoDB), selected for its adaptability in managing time-stamped, location-based sensor data, is linked to the backend. For traceability, every data packet is indexed using a distinct sensor ID and GPS coordinate. Additionally, to handle high-frequency sensor uploads across numerous nodes, a stream processing engine based on Kafka is integrated. When the system is implemented in big cities with hundreds of active sensors, this guarantees real-time scalability. Preprocessing, machine learning prediction, and explainability modules are coordinated by the analytics layer, which is implemented in Python. Platform independence and quick scaling are ensured through the use of Docker containers for deployment. Dash and Plotly are used to deploy visualization components for smooth dashboard integration, while cloud-native services like AWS S3 are used for data archiving. Efciency, scalability, and

      maintainability are all balanced by this modular software stack. The suggested stack is extensible, meaning that new sensors or models can be added with little overhead, in contrast to traditional monolithic designs.

   3. Data Preprocessing Pipeline

      Sensor drift, environmental interference, and connectivity problems make raw sensor data extremely prone to incon- sistencies. A multi-step preprocessing pipeline is used to guarantee reliable forecasting. First, statistical techniques like interquartile range (IQR) and machine learning techniques like isolation forests are used to remove outliers. These methods identify abrupt spikes or irrational values, which are frequently caused by calibration errors or sensor noise.

      Second, a two-level approach is used to handle missing data imputation. Forward or backward ll techniques are used for brief intervals (one to three minutes). Temporal K- nearest neighbor (KNN) imputation is used for longer gaps; this method uses similarities from nearby time intervals and locations to estimate missing values. This hybrid strategy prevents bias while guaranteeing continuity.

      Third, normalizationwhich is essential for deep learning convergenceis implemented using min-max scaling to bring all pollutant concentrations into a consistent range between 0 and 1. Crowdsourced calibration is another method used to correct sensor drift, in which low-cost sensors operated by citizens are routinely compared to government-grade reference stations. The deployed IoT nodes long-term dependability is guaranteed by this ongoing recalibration.

      Ultimately, the multi-sensor streams are aligned into a logical spatiotemporal dataset through geotagging and time synchronization. When compared to using raw data, this preprocessing pipeline greatly improves forecasting accuracy by ensuring that the prediction model receives high-quality, consistent, and contextualized inputs.

      Fig. 2. Work And Implementation

   4. Machine Learning Model Implementation

      The prediction module uses a hybrid architecture that com- bines Gradient Boosting Regressors (GBR) and Long Short- Term Memory (LSTM) networks. The selection of LSTM networks is based on their capacity to identify long-term trends and sequential dependencies in pollutant uctuations. For instance, LSTM cells efciently learn the weekly and daily cycles in emissions related to trafc. LSTMs, however, frequently have trouble with sudden, transient uctuations brought on by weather or industrial emissions. GBR is in- corporated as a renement layer to overcome this restriction. The hybrid architecture operates in two phases. First, using historical time series data, the LSTM network forecasts a base- line AQI value. Second, GBR learns corrective patterns from the residuals (errors) of the LSTM predictions. A weighted

      ensemble is used to determine the nal AQI prediction:

      AQIpred = · AQILST M + (1 ) · AQIGBR

      where is optimized during model training. This approach reduces both long-term bias and short-term error.

      Real-time sensor data is used to rene the model after it has been trained on historical AQI datasets from public gov- ernment repositories. Grid search and Bayesian optimization are used to tune the hyperparameters of both models. The LSTM is set up with three hidden layers, an Adam optimizer for quicker convergence, and dropout for regularization. With a maximum depth of six and a learning rate of 0.05, GBR employs 300 estimators. In comparison to standalone LSTM or GBR models, experimental results demonstrate that this hybrid model reduces Root Mean Square Error (RMSE) by almost 18%. This illustrates how well sequential learning and ensemble renement work together to predict air quality.

   5. Explainable AI Integration

      The lack of interpretability of advanced machine learn- ing models is a signicant problem that undermines public and policymaker condence. SHapley Additive exPlanations (SHAP) is used to integrate Explainable AI (XAI) in order to address this. Each input feature is given a contribution value by SHAP, which measures how much of an inuence it has on the nal prediction. For instance, the system indicates that particulate pollution is the primary cause of poor AQI during winter evenings if PM2.5 has the highest SHAP score during that period.

      There are three benets to this interpretability. In the rst place, it gives stakeholders transparency so they can com- prehend the rationale behind a particular alert. Secondly, it provides policy insights by exposing regionally specic trends in pollutants. For example, PM10 dust pollution predominates in semi-urban areas, whereas trafc-dominated NOx emissions may emerge as the key factor in urban centers. Third, by mak- ing the AI decision-making process transparent, it increases public trust, which is essential for adoption in applications that interact with citizens.

      The dashboard, which generates force plots, bar charts, and pollutant ranking tables in real time, incorporates the SHAP

      analysis directly. Instead of providing decision-makers with ambiguous forecasts, these visualizations give them action- able intelligence. In contrast to conventional black-box AQI forecasting systems, this works integration of XAI offers a unique contribution by bridging the gap between accuracy and interpretability.

   6. Visualization and Alert System

   A thorough visualization and alert platform that guarantees insights are accessible to both citizens and policymakers is the last phase of implementatio. Real-time AQI readings, pollutant breakdowns, and anticipated trends are shown on a dashboard hosted in the cloud. By integrating GPS coordinates with geospatial heatmaps created with Leaet.js, users can interactively explore pollution hotspots. Modules for historical trend analysis offer weekly, monthly, and seasonal reports for in-depth research.

   An automated alert system is created for proactive engage- ment. The system noties users in real time via email, What- sApp, and SMS when anticipated AQI levels are expected to surpass WHO thresholds. While government agencies receive location-specic hotspot alerts for prompt action, citizens receive health advisories like Avoid outdoor activity or Use N95 masks.

   The suggested platform incorporates explainability, proac- tive alerts, and predictive analytics, in contrast to traditional dashboards that solely show static values. As a result, the system serves as a public awareness and decision-support tool in addition to a monitoring tool. The visualization and alert platform completes the suggested solutions implementation cycle by making real-time, interpretable, and actionable air quality data more widely available.

5. Experimental Results and Discussion

   The experimental assessment of the suggested Air Quality Monitoring and Prediction System is shown in this section. Both benchmark datasets from public government air quality repositories and real-time sensor data gathered from IoT nodes placed in urban and semi-urban areas were used in the exper- iments. To give an organized assessment, the conversation is broken up into six subsections.

   1. Dataset Description and Deployment

      Two main data sources were used in the experimental evaluation. First, the developed IoT nodeswhich included MQ135, SDS011, and DHT11 sensorswere used to gather real-time data. Nodes were placed in semi-industrial zones, close to busy roads, and in residential areas. Over the course of 60 days of continuous monitoring, each node sent data at one-minute intervals. In addition to meteorological factors like temperature and humidity, the dataset contained pollutant concentrations (PM2.5, PM10, NOx, CO, CO2).

      Second, to guarantee extensive validation, publicly accessi- ble datasets from the UCI repository and the Central Pollution Control Board (CPCB) of India were combined. These datasets provided a wide range of pollution patterns for model training

      because they included hourly AQI data spanning several years. More than 1.2 million data records were included in the nal dataset used for the experiments after preprocessing. This variety made it possible to test the suggested hybrid model in a range of geographic locations, seasonal uctuations, and pollution levels. The evaluation was thorough and realistic be- cause it included both benchmark and eld-collected datasets.

   2. Performance Evaluation Metrics

      Several evaluation metrics were used to gauge the sug- gested hybrid LSTM-GBR models predictive accuracy. Since it penalizes large deviations more severelya crucial factor in health-sensitive AQI predictionsRoot Mean Square Error (RMSE) was selected as the primary indicator. In order to measure average prediction accuracy without undue weighting, Mean Absolute Error (MAE) was also computed. Additionally, the models ability to capture the variance in observed AQI values was evaluated using the Coefcient of Determination (R2 score). Predictionobservation plots, which graphically illustrate the correspondence between anticipated and actual AQI trends, were used for qualitative evaluation in addition to numerical metrics. Additionally, classication-based eval- uation was made possible by the generation of confusion matrices for the AQI categories (Good, Moderate, Poor, and Severe). These matrices shed light on the models ability to accurately classify risk levels, which is essential for sending out health alerts. The evaluation made sure that the system was both practically useful for real-world deployment and numerically accurate by combining statistical and categorical metrics.

   3. Comparison with Baseline Models

      Traditional machine learning models like Random Forest, Support Vector Regression (SVR), and standalone LSTM were compared to the suggested hybrid LSTM-GBR model. According to experimental results, classical models such as SVR did not capture complex temporal dependencies, but they did perform fairly well for linear patterns. Although Random Forest performed better than SVR, it was unstable during periods of high pollution and frequently underestimated severe AQI events.

      The effectiveness of standalone LSTM in capturing tem- poral patterns was hindered by overtting during short-term uctuations. The hybrid model, on the other hand, performed noticeably better than any baseline. For example, the suggested models RMSE was 18% lower than that of standalone LSTM and 26% lower than that of Random Forest. The hybrid design made good use of GBR for residual error correction and LSTM for long-term sequence learning. Because of this, it contin- uously produced accurate forecasts, even during unexpected spikes in pollution, proving its resilience and exibility under a variety of circumstances.

   4. Explainable AI Analysis

      The integration of Explainable AI (XAI) using SHAP-based feature importance analysis is one of the systems distinctive

      features. Insightful pollutant contributions across time periods and regions were revealed by the experiments. According to SHAP analysis, for example, stagnant atmospheric conditions during the winter months were the main cause of PM2.5s contribution to AQI deterioration. On the other hand, NOx emissions from automobile trafc became the main source during the summer.

      Additionally, location-specic drivers of air pollution were identied by the interpretability framework. Whereas CO and PM2.5 were more signicant in residential areas, SO2 and PM10 had higher SHAP scores in industrial zones. By offering pollutant-specic attribution, the system enabled policymakers to create focused interventions, like limiting the movement of heavy vehicles in high NOx areas or implementing dust suppression techniques in semi-urban areas. Thus, the explain- able framework closed the gap between public trust and AI accuracy by converting black-box predictions into actionable intelligence.

   5. Visualization and Alert System Outcomes

      The visualization dashboards usability, responsiveness, and information-dissemination efcacy were assessed. High gran- ularity pollution hotspots were effectively highlighted by real-time geospatial heatmaps. Users could view pollutant breakdowns in almost real-time and zoom into particular neighborhoods. Time-series plots helped users comprehend changing trends by offering comparative views of historical and projected AQI.

      Setting threshold triggers in accordance with WHO guide- lines allowed the alert system to be tested during experi- mental deployments. With an average latency of less than ve seconds, notications were sent by WhatsApp and email. Actionable health advisories, like limit outdoor exercise or wear protective masks, had a greater impact than generic alerts, according to test participants feedback. The auto- matic hotspot detection, which facilitated quicker decision- making during periods of high pollution, was also valued by government stakeholders. The assessment veried that by guaranteeing that insights were accessible to non-technical users, the visualization and alert modules added usefulness.

   6. Limitations and Future Scope

   AEven though the suggested system showed notable ad- vancements, some drawbacks were noted. Even though they worked well for dense deployment, low-cost sensors contin- ued to show calibration drift after extended use. Even with crowdsourced calibration, extree weather conditions showed discrepancies of up to 57%. Additionally, in remote de- ployments, real-time predictions were constrained by network connectivity, which occasionally resulted in packet loss.

   Although accurate from a modeling standpoint, the hybrid approach uses more computing power than standalone models. This was lessened by deployment on cloud infrastructure, but on-device prediction for fully ofine use is still difcult. The integration of lightweight transformer-based models tailored for edge devices will be the main focus of future research.

   In order to guarantee privacy-preserving model training with citizen-owned sensors, federated learning techniques will also be investigated. Promising avenues for system scaling in- clude integrating satellite remote sensing data and extending coverage to include indoor air quality monitoring. These improvements will increase the suggested solutions scalability and practical impact even more.

6. Conclusion

One of the biggest environmental problems endangering ecosystems, human health, and general urban livability is air pollution. In this work, we proposed and implemented an Air Quality Monitoring and Prediction System that combines real-time visualization with community-driven calibration, a hybrid machine learning framework, and IoT-based sensing hardware in a way that is economical, intelligent, and expli- cable. This projects efcacy rests in its capacity to blend affordability with precision and comprehensibility. Because traditional monitoring stations are expensive and have limited coverage, the systems ESP32 microcontroller and inexpensive sensors like MQ135, SDS011, and DHT11 allow for scalable deployment in a variety of environments.

The gathered pollutant and meteorological readings were prepared for reliable modeling through cloud-based prepro- cessing, data cleaning, and normalization. Predictive accuracy and resilience against abrupt pollution spikes were improved by the hybrid prediction architecture, which outperformed traditional standalone models by utilizing Gradient Boosting for residual renement and LSTM for temporal sequence learning.

The systems emphasis on explainability and transparency is another important asset. By incorporating SHAP-based Explainable AI, the project transcends the conventional black- box nature of machine learning models, giving researchers, policymakers, and users a clear understanding of the contri- butions of individual pollutants to variations in the AQI. This interpretability is essential for directing focused interventions like industrial emission control, trafc regulation, or public health advisories, as well as for building condence in AI- driven decisions. Additionally, regular citizens can benet from the systems visualization dashboard and alert mecha- nisms. The system is very user-centric and accessible thanks to its real-time geospatial heatmaps, user-friendly AQI dash- boards, and instant alerts via email, SMS, or WhatsApp. These characteristics guarantee that useful information reaches the public and inuences their behavior during periods of high pollution, not just researchers.

Notwithstanding its noteworthy contributions, the project has certain drawbacks. Although inexpensive sensors enable dense and cost-effective deployment, they are intrinsically vul- nerable to calibration drift and have lower long-term stability. Extreme weather and high humidity levels still introduced de- viations that require additional compensation techniques, even though the suggested crowdsourced calibration mechanism decreased these errors. Reliance on network connectivity for cloud processing and real-time data transfer presents another

drawback. Occasional packet losses and delays were noted in areas with shaky internet infrastructure. Real-time edge deployment in resource-constrained environments may be lim- ited by the hybrid models computational requirements, which are comparatively higher than those of lightweight algorithms despite its high accuracy. Furthermore, the current system prioritizes outdoor air quality over indoor air pollution.

To sum up, the suggested Air Quality Monitoring and Prediction System is a signicant step in closing the gap between affordable sensing, precise forecasting, and decision support that can be explained. It proves that accurate and comprehensible AQI forecasting is possible without incurring the exorbitant expenses linked to conventional monitoring systems. Even though issues with connectivity, computational overhead, and sensor calibration still exist, these constraints open the door for further developments. Stronger scalability and resilience can be achieved by coupling ground-sensor data with satellite remote sensing, extending the framework to include indoor air quality, and integrating federated learning for privacy-preserving distributed model updates. All things considered, the project demonstrates a distinctive and compre- hensive strategy that combines explainable AI, hybrid machine learning, and IoT to produce a useful.

References

1. A. Alareqi, R. Jiang, and W. Zhou, Advancements in air quality monitoring: a systematic review of IoT-based air quality monitoring and AI technologies, *Articial Intelligence Review*, vol. 58, no. 2, pp. 123156, 2025.
2. M. T. Ghayvat, P. Mukhopadhyay, and S. A. Rajasegarar, Application of articial intelligence in air pollution monitoring and forecasting: A systematic review, *Environmental Modelling & Software*, vol. 174,

   pp. 106124, 2024.

3. H. Abdel-Raouf and A. Youssef, A systematic survey of air quality prediction based on deep learning, *Alexandria Engineering Journal*, vol. 67, no. 3, pp. 451472, 2024.
4. C. C. Zhang and L. Guo, Machine learning approaches for outdoor air quality modelling: A systematic review, *Applied Sciences*, vol. 8, no. 12, pp. 25702588, Dec. 2018.
5. K. V. Reddy, S. Mishra, and R. S. Rao, A review on emerging articial intelligence techniques for air pollution forecasting: Fundamentals, ap- plication and performance, *Journal of Cleaner Production*, vol. 320,

   pp. 128145, 2021.

6. A. Sharma and N. Patel, Applications of machine learning and IoT for outdoor air pollution monitoring and prediction: A systematic literature review, *arXiv preprint arXiv:2401.01788*, 2024.
7. Y. Wang, M. Li, and P. Kumar, Systematic review of machine learning and deep learning techniques for spatiotemporal air quality prediction,

   *Atmosphere*, vol. 15, no. 11, pp. 13521369, 2024.

8. J. Lee, K. H. Park, and T. Nguyen, A review of machine learning for modeling air quality: Overlooked but important issues, *Atmospheric Research*, vol. 292, pp. 106120, 2024.
9. S. Li, A. Chen, and H. Sun, Air quality forecasting with articial intelligence techniques: A scientometric and content analysis, *Envi- ronmental Modelling & Software*, vol. 149, pp. 105117, 2022.
10. P. Das, A. Dutta, and V. Sharma, Explainable AI insight: An orderly survey, in *Recent Trends in Computing*, LNNS vol. 748. Springer, 2023, pp. 123140.
11. N. Gupta and R. K. Singh, A survey of explainable articial intelligence

    for smart cities, *Electronics*, vol. 12, no. 4, pp. 10201035, 2023.

12. M. I. Khan and J. P. Li, Enhancing accuracy in urban air quality prediction: A comparative study of predictive algorithms for air pol- lutant concentrations, *International Journal of Intelligent Systems and Applications in Engineering*, vol. 12, no. 2, pp. 7889, 2024.
13. K. Sundar and P. R. Kumar, An intelligent IoT-cloud-based air pollution forecasting model using univariate ime-series analysis, *IEEE Access*, vol. 11, pp. 4567845689, 2023.
14. X. Liu, Z. Chen, and R. Wu, High-resolution air quality prediction using low-cost sensors, *arXiv preprint arXiv:2006.12092*, 2020.
15. H. Luo, Y. Zhang, and C. Wu, Federated learning in the sky: Aerial- ground air quality sensing framework with UAV swarms, *arXiv preprint arXiv:2007.12004*, 2020.
16. M. F. Hossain, T. Zaman, and M. A. Rahman, A comprehensive survey on Internet of Things (IoT) based air pollution monitoring systems, *Sustainable Computing: Informatics and Systems*, vol. 37,

    pp. 100121, 2023.

17. R. Dutta, K. K. Singh, and S. S. Shukla, Articial intelligence for sustainable air quality management: A review of techniques, challenges, and future directions, *Environmental Science and Pollution Research*, vol. 31, no. 2, pp. 14501472, 2024.
18. C. Li, X. Wang, and P. Zhang, Review of deep learning applications in air quality prediction: Current trends and future perspectives, *Eco- logical Informatics*, vol. 77, pp. 102122, 2023.
19. S. Banerjee, N. Mehta, and V. Kumar, Air quality prediction using hybrid models: A systematic review of machine learning and statistical approaches, *Journal of Environmental Management*, vol. 335, pp. 117139, 2023.
20. Y. K. Gaur, R. K. Mishra, and T. Gupta, Explainable articial in- telligence for environmental monitoring: A systematic review, *IEEE Access*, vol. 12, pp. 4569045715, 2024.