To enhance the processing of multispectral images under adverse conditions, such as the presence of clouds, this study compared the neural networks U-Net, DeepLabV3+, and SegFormer using images from the Sentinel-2 satellite, evaluating their performance through metrics such as Intersection over Union (IoU) and Dice coefficient. The results indicate that SegFormer achieved the highest accuracy in cloud detection but with a longer inference time, while U-Net demonstrated a balance between accuracy and efficiency, and DeepLabV3+ stood out for its shorter processing time but lower performance. The study highlights the relevance of neural networks in remote sensing and indicates the need for model optimization.
Accelerated urbanization has expanded urban heat islands, necessitating the precise monitoring of Land Surface Temperature (LST). This study spatially predicts LST in Butantã (SP) for 2018 and 2021 using a high-resolution multiscale approach. To capture the relationships between urban morphology and LST, XGBoost and LightGBM models were compared, with XGBoost demonstrating superior predictive performance (R2 of 0.890 and 0.848). Applying the SHapley Additive exPlanations (SHAP) method to the winning model ensured explainability, revealing the dominance of impervious surfaces in warming, the strong mitigating potential of water bodies and vegetation, and the heat retention induced by topography in valleys.
This study evaluates fusion strategies between ERA5-Land orbital data and INMET surface observations to estimate daily air temperature in Belém, using data from 1994 to 2024. Different Random Forest models were compared, along with SARIMAX, LSTM, and stacking (SARIMAX and RF). The RF model with multivariate fusion achieved the best performance (RMSE = 0.5157 °C; R2 = 0.708; Skill = 6.19%), reducing bias relative to the orbital baseline, while chronological validation confirmed the temporal stability of RF-based models. The results indicate that, under low equatorial thermal variability, tree-based models are the most robust for satellite–surface fusion and thermal downscaling.
This study investigates the imputation of missing values in mean air temperature time series in the state of Pará, Brazil, by integrating observational data from INMET and the ERA5-Land reanalysis. A multivariate non-linear approach based on the MissForest algorithm was developed and compared with models using only station data and with linear interpolation. The evaluation employed synthetic gaps together with the RMSE, MAE, and MAE in quantile 99 metrics. The results indicate that the inclusion of ERA5-Land reduced RMSE to 0.664–0.744 °C, MAE to 0.509–0.569 °C, and ∆P99 to 0.909–1.013 °C, while increasing model stability and outperforming the other strategies, thus supporting the methodological consistency of the approach for tropical climate time series.