Machine learning classifiers are being increasingly used nowadays for Land Use and Land Cover (LULC) mapping from remote sensing images. However, arriving at the right choice of classifier requires understanding the main factors influencing their performance. The present study investigated firstly the effect of training sampling design on the classification results obtained by Random Forest (RF) classifier and, secondly, it compared its performance with other machine learning classifiers for LULC mapping using multi-temporal satellite remote sensing data and the Google Earth Engine (GEE) platform. We evaluated the impact of three sampling methods, namely Stratified Equal Random Sampling (SRS(Eq)), Stratified Proportional Random Sampling (SRS(Prop)), and Stratified Systematic Sampling (SSS) upon the classification results obtained by the RF trained LULC model. Our results showed that the SRS(Prop) method favors major classes while achieving good overall accuracy. The SRS(Eq) method provides good class-level accuracies, even for minority classes, whereas the SSS method performs well for areas with large intra-class variability. Toward evaluating the performance of machine learning classifiers, RF outperformed Classification and Regression Trees (CART), Support Vector Machine (SVM), and Relevance Vector Machine (RVM) with a >95% confidence level. The performance of CART and SVM classifiers were found to be similar. RVM achieved good classification results with a limited number of training samples.

Due to its comparatively high spatial resolution and its daily repeat frequency, the tropospheric nitrogen dioxide product provided by the TROPOspheric Monitoring Instrument (TROPOMI) onboard the Sentinel-5 Precursor platform has attracted significant attention for its potential for urban-scale monitoring of air quality. However, the exploitation of such data in, for example, operational assimilation of local-scale dispersion models is often complicated by substantial data gaps due to cloud cover or other retrieval limitations. These challenges are particularly prominent in high-latitude regions where significant cloud cover and high solar zenith angles are often prevalent. Using the example of Norway as a representative case for a high-latitude region, we here evaluate the spatiotemporal patterns in the availability of valid data from the operational TROPOMI tropospheric nitrogen dioxide (${\mathrm{NO}}_2$) product over five urban areas (Oslo, Bergen, Trondheim, Stavanger, and Kristiansand) and a 2.5 year period from July 2018 through November 2020. Our results indicate that even for relatively clean environments such as small Norwegian cities, distinct spatial patterns of tropospheric ${\mathrm{NO}}_2$ are visible in long-term average datasets from TROPOMI. However, the availability of valid data on a daily level is limited by both cloud cover and solar zenith angle (during the winter months), causing the fraction of valid retrievals in each study site to vary from 20% to 50% on average. A temporal analysis shows that for our study sites and the selected period, the fraction of valid pixels in each domain shows a clear seasonal cycle reaching a maximum of 50% to 75% in the summer months and 0% to 20% in winter. The seasonal cycle in data availability shows the inverse behavior of ${\mathrm{NO}}_2$ pollution in Norway, which typically has its peak in the winter months. However, outside of the mid-winter period we find the TROPOMI ${\mathrm{NO}}_2$ product to provide sufficient data availability for detailed mapping and monitoring of ${\mathrm{NO}}_2$ pollution in the major urban areas in Norway and see potential for the use of the data in local-scale data assimilation and emission inversions applications.

Fine particulate matter (PM_2.5) is a key air quality indicator due to its adverse health impacts. Accurate PM_2.5 assessment requires high-resolution (e.g., atleast 1 km) daily data, yet current methods face challenges in balancing accuracy, coverage, and resolution. Chemical transport models such as those from the Copernicus Atmosphere Monitoring Service (CAMS) offer continuous data but their relatively coarse resolution can introduce uncertainties. Here we present a synergistic Machine Learning (ML)-based approach called S-MESH (Satellite and ML-based Estimation of Surface air quality at High resolution) for estimating daily surface PM_2.5 over Europe at 1 km spatial resolution and demonstrate its performance for the years 2021 and 2022. The approach enhances and downscales the CAMS regional ensemble 24 h PM_2.5 forecast by training a stacked XGBoost model against station observations, effectively integrating satellite-derived data and modeled meteorological variables. Overall, against station observations, S-MESH (mean absolute error (MAE) of 3.54 μg/m³) shows higher accuracy than the CAMS forecast (MAE of 4.18 μg/m³) and is approaching the accuracy of the CAMS regional interim reanalysis (MAE of 3.21 μg/m³), while exhibiting a significantly reduced mean bias (MB of -0.3 μg/m³ vs. -1.5 μg/m³ for the reanalysis). At the same time, S-MESH requires substantially less computational resources and processing time. At concentrations >20 μg/m³, S-MESH outperforms the reanalysis (MB of -7.3 μg/m³ and -10.3 μg/m³ respectively), and reliably captures high pollution events in both space and time. In the eastern study area, where the reanalysis often underestimates, S-MESH better captures high levels of PM_2.5 mostly from residential heating. S-MESH effectively tracks day-to-day variability, with a temporal relative absolute error of 5% (reanalysis 10%). Exhibiting good performance at high pollution events coupled with its high spatial resolution and rapid estimation speed, S-MESH can be highly relevant for air quality assessments where both resolution and timeliness are critical.

We evaluate the added value of integrating validated Low-Cost Sensor (LCS) data into a Machine Learning (ML) framework for providing surface PM_2.5 estimates over Central Europe at 1 km spatial resolution. The synergistic ML-based S-MESH (Satellite and ML-based Estimation of Surface air quality at High resolution) approach is extended, to incorporate LCS data through two strategies: using validated LCS data as a target variable (LCST) and as an input feature via an inverse distance weighted spatial convolution layer (LCSI). Both strategies are implemented within a stacked XGBoost model that ingests satellite-derived aerosol optical depth, meteorological variables, and CAMS (Copernicus Atmospheric Monitoring Service) regional forecasts. Model performance for 2021-2022 is evaluated against a baseline trained on air quality monitoring stations without any form of LCS integration. Our results indicate that the LCSI approach consistently outperforms both the baseline and LCST models, particularly in urban areas, with RMSE reductions of up to 15-20 %. It also exhibits higher accuracy than the CAMS regional interim reanalysis with a lower annual mean absolute error (MAE) of 2.68 μg/m³ compared to 3.32 μg/m³. SHapley Additive exPlanations based analysis indicates that LCSI information improves both spatial and temporal representativeness, with the LCSI strategy better capturing localized pollution dynamics. However, the LCSI's dependency on the spatial LCS layer limits its ability to capture inter-urban pollution transport in regions with sparse or no LCS data. These findings highlight the value of large-scale sensor networks in addressing spatial coverage gaps in official air quality monitoring stations and advancing high-resolution air quality modeling.

Nitrogen dioxide (NO2) is among the major air pollutants in Europe posing severe hazard to environmental and human health. The concentrations of surface NO2 are measured by ground monitoring stations which are fairly limited in representation and distribution. While NO2 estimates from chemical transport models are realistic, their complexity makes them computationally intensive. Satellite observations from instruments such as TROPOMI provide high spatiotemporal distribution of NO2. However, these instruments capture NO2 density only along the tropospheric column and not on the surface. Exploiting the availability of ground station measurements and spatially continuous information from TROPOMI, this study estimates surface NO2 concentrations over Europe at 1km spatial resolution for 2019-2021 using XGBoost machine learning model. While ground measurements are used as target reference features, satellite observations such as tropospheric column density of NO2 (from TROPOMI), night light radiance (from VIIRS), NDVI (from MODIS) and modelled meteorological parameters such as planetary boundary layer height, wind velocity, temperature are used as input features to the model. We find an overall mean absolute error of 7.87µg/m3, mean bias of -3.13µg/m3 and spearman correlation of 0.61 during model validation. We found that the performance of the model is influenced by NO2 concentration levels and is most reliable for predictions at concentration levels <40µg/m3 with a relative bias of <40%. The spatial error analysis also indicates the spatial robustness of the model across the study area. The importance of input features is evaluated using SHapley Additive exPlanations (SHAP), which shows TROPOMI NO2 being the most important source for the modelled NO2 predictions. Furthermore, SHAP values also highlight the role of VIIRS night light radiance in deriving finer detailed spatial patterns of surface NO2 estimates. Despite the complex non-linear relationship of the input features, the trained XGBoost model requires an average of 570 seconds to predict single day surface NO2 concentrations for the large study area of continental scale. Thus, this work evaluates the importance of TROPOMI data and reliability of machine learning models for estimating surface NO2 concentrations on a larger spatial scale.