Air pollution in urban areas remains a threat, with particulate matter being one of the key indicators of pollution. Numerous studies show that particulate matter with a diameter of 2.5 μm or smaller (PM2.5) affects public health, hence the need for more accurate models. Traditional numerical methods fail to capture pollutants’ spatiotemporal patterns; however, neural networks can learn from environmental data inputs to obtain correlations. In this study, a Temporal Convolutional Network–Bidirectional Long Short-Term Memory (TCN–Bi-LSTM) hybrid model is proposed. TCN- Bi-LSTM is for PM2.5 monthly estimation in 2023 using Taiwan’s Ministry of Environment air quality historical data from July 2015 to December 2022 – including weather parameters and land use data – to produce 1km PM2.5 maps via k nearest neighbors – inverse distance weighting (KNN – IDW). The results compare the proposed model against classical models like Convolution Neural Network and Long Short-Term Memory. The proposed model performs better compared to the others with a metric evaluation of R2 = 0.7960 and root mean square error (RMSE) = 2.1606 μg/m3 in the GroupKFold cross-validation. This study aims to contribute to air quality monitoring research by reducing reliance on aerosol optical depth (AOD) derived PM2.5 estimates, which are usually affected by cloud cover gaps during satellite AOD retrieval. This study, therefore, addresses the missing data issue by producing continuous PM2.5 concentration exposure maps. Future work aims to incorporate independent surface PM2.5 observations to check whether new findings and results will be obtained.

Total Hydrocarbon (THC) is a critical air pollutant composed of various hydrocarbon compounds, commonly emitted from vehicle exhaust, industrial activities, and fuel combustion. Accurate estimation of THC concentrations is essential for understanding pollution patterns and supporting environmental management. This study applies a Light Gradient Boosting Machine (LightGBM) model to estimate the spatiotemporal variation of THC levels in a part of the Northern Air Quality Zone (AQZ) from June 25, 2015, to December 31, 2020, with 2021 used for testing with unseen data. LightGBM, a machine learning algorithm known for its high accuracy and computational efficiency, achieved strong predictive performance with R2 values of 0.84 for training, 0.82 for validation, and 0.77 for testing. The average THC concentration in the study area was 2.07 ppm. SHAP (SHapley Additive exPlanations) analysis revealed that nitrogen dioxide (NO?) and nitrogen oxides (NOx) were the most influential predictors, highlighting their close association with combustion-related emissions. Other important variables included population density, land use types such as plantation forests and paddy fields, and wind speed. Spatially, THC concentrations were higher in the western part of the study area, which is more urbanized and industrialized, compared to the more rural eastern region. These results demonstrate the potential of LightGBM to model complex air pollution patterns and support data-driven air quality strategies.

This study investigates the application of deep learning models for gain–loss detection in land surface change using bi-temporal satellite imagery. An end-to-end framework is adopted to directly capture semantic and spectral differences between pre- and post-change images. The proposed approach is based on the Densely Attentive Refinement Network (DARNet) and extends it to a multi-class change detection architecture. To improve consistency between multi-source imagery, including Landsat and Sentinel-2, statistical matching and super-resolution techniques are incorporated into the preprocessing workflow. Coastal change detection is selected as the case study to evaluate and compare the performance of several state-of-the-art deep learning models on a multi-class land-cover change dataset. The analysis focuses on class discrimination capability, boundary delineation accuracy, and overall classification performance to provide guidance for future model selection and development. Experimental results demonstrate that the multi-class DARNet integrated with statistical matching and super-resolution achieves superior performance in detecting fine-scale land changes. Furthermore, the proposed data integration strategy enhances both model accuracy and generalization, indicating its strong potential for practical applications in land change detection.

With technological advancements, LiDAR sensors have been integrated into some Smart handheld devices. This study assesses the feasibility of using the iPhone 15 Pro paired with the viDoc RTK Rover for under-canopy positioning and scanning within a Swietenia macrophylla plantation. The results indicate that this setup can reduce smartphone positioning error to centimeter-level accuracy, demonstrating its potential to replace traditional RTK GNSS receivers for under-canopy point collection. To obtain valid point cloud data, connection to the viDoc for scanning is necessary, and marking ground control points (GCPs) can improve accuracy. Regarding specific measurements, a 90% overlap rate yielded optimal results for Diameter at Breast Height (DBH) estimation, while a 75% overlap was optimal for determining tree positions. Based on a marginal effect analysis of processing time and various measurement outcomes, the recommended configuration is to connect the viDoc with a 70% overlap rate and mark ground control points. This equipment combination effectively reduces costs while acquiring usable point cloud data, though it remains constrained by scanning range and shows limited detection capability for trees with DBH less than 5 cm.