Inhaltszusammenfassung

Accurate assessment of the urban heat island (UHI) phenomenon relies on land surface temperature (LST) maps with high spatialtemporal resolution. Although many studies have focused on enhancing the level of spatial detail in LST from MODIS by using Landsat 8 imagery along with advanced deep learning models such as CNNs and GANs, these methods typically require significant computational resources. In addition, the potential of drone-based thermal imagery to enhance the spatial resolution of La...Accurate assessment of the urban heat island (UHI) phenomenon relies on land surface temperature (LST) maps with high spatialtemporal resolution. Although many studies have focused on enhancing the level of spatial detail in LST from MODIS by using Landsat 8 imagery along with advanced deep learning models such as CNNs and GANs, these methods typically require significant computational resources. In addition, the potential of drone-based thermal imagery to enhance the spatial resolution of Landsat 8 insufficiently investigated. In response to these challenges, we present an innovative approach to generate high-resolution LST (LSTSR) maps. This approach combines low-resolution Landsat 8 LST data (LSTLR) with high-resolution Planet-Scope imagery (IHR) and drone-collected thermal data (THR) using Kolmogorov-Arnold networks (KAN) augmented by spline-based feature extraction techniques. In this approach, 3-metre Planet-Scope imagery is used to transfer spatial detail to the Landsat-8 LST, while THR uses as a ground truth data. The LSTLR and IHR datasets serve as inputs to the KAN models. We designed and tested KAN architectures: (1) shallow-linear, (2) deep-linear, and (3) deep-nonlinear using training sites in Germany and evaluated their performance at an independent test site. The results showed that the deep-linear model with two hidden layers and five neurons per layer achieved the best performance. This model had an RMSE of 3.65°C, a MAE of 2.70°C, an SSIM of 0.82, a PSNR of 23.13, and a MAPE of 8.33%. It demonstrated a superior ability to capture thermal patterns and reconstruct fine-scale features, such as edges. In contrast, the shallow-linear model performed the worst, with an RMSE of 6.22°C, a MAE of 5.12°C, an SSIM of 0.75, a PSNR of 18.51, and a MAPE of 15.51%.» weiterlesen» einklappen

Autoren

Klassifikation

DFG Fachgebiet:
3.43-02 - Geodäsie, Photogrammetrie, Fernerkundung, Geoinformatik, Kartographie

DDC Sachgruppe:
Geowissenschaften