Non-linear unmixing as an innovative tool to detect vine diseases in UAVs, airborned and satellite images: preliminary results

Abstract

Context and purpose of the study – Vine diseases have a strong impact on vineyards sustainability, which in turns leads to strong economic consequences. Among those diseases, Flavescence dorée spreads quickly and is incurable, which led in France to the setup of a mandatory pest control implying the systematic use of pesticides and the prospection and uprooting of every infected plants. Remote sensing could be a very powerful tool to optimize prospection as it allows to produce quickly accurate maps over large areas. Recent studies have shown that high spatial resolution (10cm/pixel) multispectral images acquired from UAVs allow to map Flavescence dorée in vineyards using leaves discolorations [e.g. Albetis et al., Remote Sensing, 2017]. Nevertheless, confusion and misdetections still exist, especially with other diseases showing similar leaves discolorations and with mixtures of different materials occurring within one pixel. Mixture effects are also crucial when dealing with satellite images where spatial resolution is much lower (≥10m/pixel). This study aims at improving the detection of vine diseases in UAVs, airborned and satellite images using an innovative tool that identifies the spectral signatures of every elementary materials (e.g. healthy and sick leaves) and their relative contribution at a subpixel level.

Material and methods – We use three distinct datasets acquired in 2016 over the same vineyard located in the Southwest of France (AOC Gaillac): a multispectral image acquired with MicaSense sensor onboard an UAV (5 bands, 10cm/pixel), a Sentinel-2 multispectral image (12 bands, 10m/pixel) and an airborned hyperspectral image (256 bands, 1m/pixel). Ground truth for validation is available through exhaustive centimetric locations of every sick vines for several plots in the studied area. On the methodological perspective, we use an innovative method that performs an unsupervised unmixing jointly with anomalydetection capacities and has a global linear complexity [Nakhostin et al., TGRS, 2016]. Nonlinearities are handled by decomposing the data on an overcomplete set of spectra, combined with a specific sparse projection, which guarantees the interpretability of the analysis.

Results – This paper reports preliminary results obtained with the unmixing algorithm ran over one selected plot available in the dataset. Initial results show the algorithm can detect and separate multiple sources within the plot. Analysis of retrieved endmembers shows a good correlation with the components that can be found in the field, especially with the evidence of healthy and sick leaves’ signatures. Nevertheless, initial mapping still shows some discrepancies with ground truth and further work needs to be done to fine tune the model parameters.