Images acquired by TIRS sensor, on Landsat 8, have an original spatial resolution of 100 m. These images are distributed at a spatial resolution of 30 m after cubic convolution over-sampling. In my opinion, this pre processing step lowers the quality of TIR values and a nearest neighbor or natural neighbor resampling algorithm should be used.
For this reason I’m looking for TIRS images distributed with the original spatial resolution of 100 m. Do you know where these images can be found?
Dear Sergio,
yes, cubic convolution is not invertible, I was just suggesting a very simple option to have your data at 100m. This approach should provide a good result (more or less) if your study area is homogeneous.
Regarding the TIRS bands of Landsat-8 (may be you already now this), there is a calibration problem because undesired (and unexpected) noise (apart from the inherent thermal noise) was introduced into the sensor measurements. The USGS performed a new recalibration of these bands, and they were processed again. However we found that even with this USGS calibration results may differ from "true" values, which suggests that recalibration with in-situ measurements are still needed. At least this is what we found in our test sites.
Sergio very true. you are right I too will be interested to see the image if you get one.
you are right, but not sure if original images are currently available. In the meanwhile, as a simple option, you can resample the 30m image to 100m using a "pixel average", and then resample again to 30m using nearest neighbor.
Dear Juan-Carlos,
Tank you forr your answer.
The method that you suggest would be only a very approximate estimate of the original TIRS measurements:
- Cubic convolution is not invertible (i.e. deconvolution is not possible);
- Pixel average works using a moving average box: what is its dimension? 3 or 4 pixels (100/30=3.333) ?
With this method temperature patterns, temperature gradients and noise would be very smoothed. I suspect that the original TIRS data are very noisy, so the cubic convolution over-sampling is a way to mask this problem.
Sergio
Dear Sergio,
yes, cubic convolution is not invertible, I was just suggesting a very simple option to have your data at 100m. This approach should provide a good result (more or less) if your study area is homogeneous.
Regarding the TIRS bands of Landsat-8 (may be you already now this), there is a calibration problem because undesired (and unexpected) noise (apart from the inherent thermal noise) was introduced into the sensor measurements. The USGS performed a new recalibration of these bands, and they were processed again. However we found that even with this USGS calibration results may differ from "true" values, which suggests that recalibration with in-situ measurements are still needed. At least this is what we found in our test sites.
Dear Juan-Carlos,
what you say is very interseting. Do you have some reference paper showing this TIRS problem?
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing