Résumé
Tree decline affects many perennial orchards with potentially heavy impacts on production. In five years 3 % of national production of kiwifruit orchards were lost in France, leading to economic issues and jeopardizing the value chain. Some specific management practices could help to mitigate kiwifruit decline, but rapid and simple tools are needed to assess the development of the decline in response to these practices. As kiwifruit decline is characterized by a low-vigor canopy along with changes in canopy color and density, Sentinel-2 images were used to detect vine decline over large areas. We first selected 28 orchards, with varying characteristics (e.g. row grass cover, hail protection nets, fertilization practices) and characterized each vine (14000 in total) according to its agronomic status (i.e. vigor and presence of decline symptoms). Sentinel-2 images are made up of 10 x 10 m pixels. To classify them we fitted two models using a random forest procedure. The spectral model (SM) used only spectral inputs, while the agronomic and spectral model (ASM) used both spectral inputs from satellite images and agronomic inputs obtained from orchard characteristics. Spectral inputs included raw spectral bands (e.g., red, near-infrared, green) and vegetation indices (e.g., NDVI, GNDVI). Results show that vigorous and dead areas were well detected (more than 80 % of correct predictions). Declining areas were correctly detected when patches of decline were larger than 500 m 2 . Mixed pixels containing vines with different agronomic status were poorly predicted. Accuracy improves when agronomic information, such as soil texture, is incorporated into the model. Both models (SM and ASM) could enable growers to adjust their practices in real time according to the health status of their vines.