Thermal images of water bodies often show a radiance gradient perpendicular to the banks. This effect is frequently due to mixed land and water thermal pixels. In the case of the Landsat images, radiance mixing can also affect pure water pixels due the cubic convolution resampling of the native thermal measurements. Some authors recommended a general-purpose margin of two thermal pixels to the banks or a minimum river width of three pixels, to avoid near bank effects in water temperature retrievals. Given the relatively course spatial resolution of satellite thermal sensors, the three pixel margin severely restricts their application to temperature mapping in many rivers. This study proposes a new algorithm to enhance the retrieval of stream surface temperature using Landsat 8 thermal data, although it is also applicable to Landsat 7 and Landsat 5. The aim is not to perform a subpixel radiance unmixing but to refine the selection of unmixed, reliable pixels for temperature mapping. For this purpose, the spatial arrangement of native Landsat thermal pixels is approximated, and pure water pixels in the downscaled thermal band are selected accordingly. The least-favourable cubic convolution near-bank radiance mixing is simulated on image basis. Only pure thermal water pixels unaffected by the simulated worst-case resampling are selected. The algorithm allowed retrieving water surface temperature in reaches down to 120 m wide, clearly improving the existing three pixel, i.e. 300 m for Landsat 8, recommendation. The enhancing algorithm was applied to a reach in the Ebro River reach, Spain. It provided spatially distributed temperatures in narrow parts, upstream and downstream of a wide reservoir, offering new insight of the overall impact of the reservoir over the river thermal regime.