Determining True Sensor Spatial Resolution of Very High Resolution Optical Imagery

Some satellite data is delivered in images with gridded pixels. This gridded pixel size is often assumed to be the spatial resolution of the satellite sensor; however, this is not always the case. An image can be grided to any arbitrary pixel size, but the sensor resolution will remain constant. For example, an image with a pixel grid size much smaller than the sensor resolution will appear blurry along what should be sharp transitions. This discrepancy between an image’s pixel size and true sensor spatial resolution can be the source of much confusion and even misinformation among data users, which may lead them to waste time and resources on using images that do not suit their spatial resolution needs.

This presentation will highlight our evaluation of the true spatial resolution of various government and commercial images in the pixel size range of 0.3 m to 60 m. Images evaluated include ESA’s Sentinel-2 (60 m, 20 m, 10 m pixels), USGS’s Landsat 8/9 (30 m & 15 m pixels), Planet’s SuperDoves (3 m pixels), BlackSky’s Globals (~1 m pixels), and the optical bands of Maxar’s WorldView-2 (2.4 m – 0.41 m pixels) and WorldView-3 (1.38 m – 0.31 m pixels). Our evaluation of true sensor spatial resolution, or ‘footprint size’ is based on the sensor’s line spread function (LSF). We calculate the width at half the height of the LSF to find the full width at half maximum (FWHM). The FWHM is how we report sensor spatial resolution.

Different objects are examined for constructing the LSF depending on the sensor spatial resolution. Coarser resolution sensors in this evaluation such as Sentinel-2 and Landsat 8/9 are examined at bridges over a dark water background. The bright bridge acts as a line impulse, giving a sensor’s line spread function (LSF) in one direction. Additionally, we simulate the impacts of bridge width on the apparent LSF to obtain a true LSF without the effects of bridge width for these sensors.

Finer resolution sensors will image the irregularities in bridges such as trusses, sidewalks, and in some cases painted lines, interfering with the LSF construction. Instead, these sensors are evaluated at large (60 m – 140 m) black and white checkerboards known as Cal/Val sites. At these locations, the image’s transition from black to white is extracted as an edge spread function (ESF). We calculate the derivative of this ESF to obtain the sensor’s LSF. From there, we find the FWHM as we do for the coarser resolution images.

With the FWHM and pixel size, we determine how over- or under-sampled the images are. When the ratio of a sensor’s spatial resolution and the gridded image’s pixel size is less than 1, the image is considered under-sampled. In this case, each pixel’s information is unique but only a portion of that pixel’s ground area has been measured. On the other side, if the ratio is greater than 1, the image is considered over-sampled. That is, each pixel’s information is sourced from within the ground extent of the pixel and some extent outside additionally.

We will show the true spatial resolution and the extent of over-/under-sampling in the imagery from ESA’s Sentinel-2 (60 m, 20 m, 10 m pixels), USGS’s Landsat 8/9 (30 m & 15 m pixels), Planet’s SuperDoves (3 m pixels), BlackSky’s Globals (~1 m pixels), and the optical bands of Maxar’s WorldView-2 (2.4 m – 0.41 m pixels) and WorldView-3 (1.38 m – 0.31 m pixels).