Aerial Spectral Super-Resolution using Conditional Adversarial Networks

TL;DR

This paper presents a conditional adversarial network that reconstructs 31 spectral bands from aerial images, enabling spectral analysis despite low resolution and noise, trained on a novel high-res dataset.

Contribution

It introduces a new aerial hyperspectral dataset and demonstrates the effectiveness of adversarial networks in spectral super-resolution from aerial imagery.

Findings

Achieved a root mean square error of 2.48 on synthesized RGB data.

Successfully generated spectral signatures from low-res aerial images.

Validated the approach on a novel high-resolution aerial hyperspectral dataset.

Abstract

Inferring spectral signatures from ground based natural images has acquired a lot of interest in applied deep learning. In contrast to the spectra of ground based images, aerial spectral images have low spatial resolution and suffer from higher noise interference. In this paper, we train a conditional adversarial network to learn an inverse mapping from a trichromatic space to 31 spectral bands within 400 to 700 nm. The network is trained on AeroCampus, a first of its kind aerial hyperspectral dataset. AeroCampus consists of high spatial resolution color images and low spatial resolution hyperspectral images (HSI). Color images synthesized from 31 spectral bands are used to train our network. With a baseline root mean square error of 2.48 on the synthesized RGB test data, we show that it is possible to generate spectral signatures in aerial imagery.