Improving remote sensing of extreme events with machine learning: land surface temperature retrievals from IASI observations

Retrieving weather extremes from observations is critical for weather forecasting and climate impact studies. Statistical and machine learning methods are increasingly popular in the remote sensing community. However, these models act as regression tools when dealing with regression problems and as such, they are not always well-suited for the estimation of the extreme weather states. This study firstly introduces two error types that arise from such statistical methods: (a) ‘dampening’ refers to the reduction of the range of variability in the retrieved values, a natural behavior for regression models; (b) ‘inflating’ is the opposite effect (i.e. larger ranges) due to data pooling. We then introduce the concept of localization that intends to better take into account local conditions in the statistical model. Localization largely improves the retrievals of extreme states, and can be used both for retrieval at the pixel level or in image processing techniques. This approach is tested on the retrieval of land surface temperature using infrared atmospheric sounding interferometer observations: the dampening is reduced from 1.9 K to 1.6 K, and the inflating from 1.1 K to 0.5 K, respectively.