Monitoring grapevine water status using Landsat 8 images: a two-year case study in a Merlot vineyard

Abstract

Context and purpose of the study – Viticulture needs for spatial and temporal information are increasing to improve vineyard management, especially concerning water efficiency. Remote sensing, particularly from satellites, can be a powerful tool to assess vineyard characteristics such as vigor or water status in space-time. In this study, we use Landsat 8, an American Earth observation satellite with six bands from the visible (VIS) to the Short-Wave Infrared (SWIR) domains with 30m spatial resolution and two thermal bands with 100m spatial resolution.

Material and methods – In 2020 and 2021, we conducted an experiment in a Merlot vineyard in Bakersfield, CA with 24 experimental units spatially distributed according to the pixel location of the Landsat 8 image. Contemporary to the satellite overpass on the vineyard, we measured water status every two weeks from June to August with midday stem water potentials (Ψ_stem), stomatal conductance (g_s), and net carbon assimilation rate (A_N). Berry weight, pH, total soluble solids (°Brix), and titratable acidity (TA) were also measured every two weeks from July to August. We converted the satellite images to reflectance values, extracted data for each experimental unit, and computed several vegetation indices (VI). Minimum relative humidity and maximum temperature were obtained from a neighboring weather station within the California Weather Irrigation System. We used linear regression and a gradient-boosting machine learning model with block-out and date-out cross-validation to predict water status from different combinations with band reflectance values, vegetation indices (VI), and minimum relative humidity and maximum temperature. We evaluated the results using the correlation coefficient (r), the determination coefficient (R²), and the root mean square error (RMSE). We interpreted the model by calculating feature importance for each predictor to understand the role of each Landsat 8 band in the model.

Results – The Ψ_stem data correlated with two VI: the Moisture Index, MSI, and the Normalized Difference Moisture Index, NDMI, (r = -0.64 and r = 0.63). The machine learning model predicted water status (Ψ_stem, A_N, and g_s) with an R² of ~0.8 using only the Landsat 8 bands. The R² was higher than 0.8 when remote sensing data was combined with ground weather data as evaluated by block-out cross-validation, thus showing the ability of the model to predict in space. Results were less encouraging when evaluated through date-out validation, therefore assessing the ability of the model to predict in time. The feature importance extraction identified the near-infrared (NIR) and SWIR domains as the most important bands to predict vine water status. These results confirm the interest in using satellite imagery to monitor vineyards at a large scale and with a good temporal resolution.