Revealing long-Term physiological trajectories of grasslands from legacy b&w aerial photographs

Abstract

Landscape reconstruction is crucial to measure the effects of climate change or past land use on current biodiversity. In particular, retracing past phenological changes can serve as a basis for explaining current patterns of plant communities and predict the future extinction of species. Old spatial data are currently used to reconstruct vegetation changes, both morphologically (with landscape metrics) and semantically (grasslands to crops for instance). However, poor radiometric properties (single panchromatic channel, illumination variation, etc.) do not offer the possibility to compute environmental variables (e.g. NDVI and color indices), which strongly limits long-Term phenological reconstruction. In this study, we propose a workflow for reconstructing phenological trajectories of grasslands from 1958 to 2011, in the French central Vosges, from old aerial black and white (B&W) photographs. Noise and vignetting corruptions were first corrected in B&W photographs with non-local filtering algorithms. Panchromatic scans were then colorized with a Generative Adversarial Network (GAN). Based on the predicted channels, we finally computed digital greenness metrics (Green Chromatic Coordinate, Excess Greenness) to measure vegetation activity in grasslands. Our results demonstrated the feasibility of reconstructing long-Term phenological trajectories from legacy photographs with insights at different levels: (1) the proposed correction methods provided radiometric improvements in old aerial missions; (2) the colorization process led to promising and plausible colorized historical products; (3) digital greenness metrics were useful for describing past vegetation activity.