Evaluation of Resampling Techniques to Provide Better Synthesized Input Data to Super-Resolution Deep Learning Model Training

Hyperspectral images often have low spatial resolution due to the sensor sizes required to capture the required spectral response. Super-resolution (SR) techniques try to mitigate this by injecting more detail in the upscaled image, either with numerical methods or deep learning and Convolution Neural Networks. In the deep learning methods, the models learn image details by inferring a high-resolution (HR) image from a synthetic low-resolution (LR) image that simulates the natural degradation of sensors by applying resampling (to reduce the image detail) and noising (to add small errors and interference). Often disregarded in the literature, the resampling method applied to generate the synthetic image can impact greatly the deep learning model training. This work, evaluate several resampling techniques to measure this impact using the Harvard hyperspectral dataset. Results showed that the Lanczos filter was the best among eight other resampling methods. The Nemenyi and Friedman ranking statistical tests also indicated that the Cubic-Spline, Bicubic, and RMS achieved good results.