Estimating subpixel turbulent heat flux over leads from MODIS thermal infrared imagery with deep learning

The turbulent heat flux (THF) over leads is an important parameter for climate change monitoring in the Arctic region. THF over leads is often calculated from satellite-derived ice surface temperature (IST) products, in which mixed pixels containing both ice and open water along lead boundaries reduce the accuracy of calculated THF. To address this problem, this paper proposes a deep residual convolutional neural network (CNN)-based framework to estimate THF over leads at the subpixel scale (DeepSTHF) based on remotely sensed images. The proposed DeepSTHF provides an IST image and the corresponding lead map with a finer spatial resolution than the input IST image so that the subpixel-scale THF can be estimated from them. The proposed approach is verified using simulated and real Moderate Resolution Imaging Spectroradiometer images and compared with the conventional cubic interpolation and pixel-based methods. The results demonstrate that the proposed CNN-based method can effectively estimate subpixel-scale information from the coarse data and performs well in producing fine-spatial-resolution IST images and lead maps, thereby providing more accurate and reliable THF over leads.

Can Landsat thermal imagery and environmental data accurately estimate water temperatures in small streams?

The ability to monitor water temperature is important for assessing changes in riverine ecosystems resulting from climate warming. Direct in situ water temperature collection efforts provide point-samples but are cost-prohibitive for characterizing stream temperatures across large spatial scales, especially for small, remote streams. In contrast, satellite thermal infrared imagery may provide a spatially extensive means of monitoring riverine water temperatures, however, the accuracy of these remotely sensed temperatures for small streams is not well understood. Here, we investigated the utility of Landsat 8 thermal infrared imagery and both local and regional environmental variables to estimate subsurface temperatures in high latitude small streams (2 – 30 m wetted width), from a test watershed in southcentral Alaska. Our results suggested that Landsat-based surface temperatures were biased high, and the degree of bias varied with hydrological and meteorological factors. However, with limited in-stream validation work, results indicated it is possible to reconstruct average in situ water temperatures for small streams at regional-scales using a regression modelling framework coupled with publicly-available Landsat or air temperature information. Generalized additive models built from stream stage information from a single gage and air temperatures from a single weather station in the drainage fit to a limited set of in situ temperature recordings could estimate average stream temperatures at the watershed-level with reasonable accuracy (root mean square error = 2.4°C). Landsat information did track closely with regional air temperatures and could also be incorporated into a regression model as a substitute for air temperature to estimate in situ stream temperatures at watershed scales. Importantly, however, while average watershed-scale stream temperatures may be predictable, site-level estimates did not improve with the use of Landsat information or other local covariates, indicating that additional information may be necessary to generate accurate spatially explicit temperature predictions for small order streams.