Cultivating Insights: Unsupervised Mapping of Inter-row Management inVineyards Using Bezier Curve Properties on Sentinel-2 Time Series

Inter-row management of vineyards has various implications including on soil stability, and thus geo-
risk[2]. Two prominent classes of inter-row management are permanent grass cover (PGC), and total
tillage (TT), where the inter-row spaces are tilled to keep the soil bare. Which practice is used impacts
soil stability, and very few papers explored large-scale mapping using remotely sensed data[1]. In multi-
spectral acquisitions, reflection from vine leaves and from inter-row responses mix together, challenging
to distinguish vineyard foliage from possible inter-row vegetation. It has been indeed observed that
PGC and TT are likely more distinguishable in winter, when vines shed most of their leaves or are left
bare[1]. This increases the weight of inter-row vegetation in the spectral mix. Based on the above, here
we propose some novel discriminating features by treating the Sentinel-2 time series in winter as the
Bezier curves, which appear to increase separability.
The method has been tested on reference data collected in N-W Italy by a previous project. Data
from 130 and 141 ground truth polygons, representing PGC - and TT -managed vineyards respectively,
were collected from 10 wineries in 2015 and 2022. Sentinel-2 data from November to March with < 20%
cloud cover and ground truth for 2015 were primarily used in this work. Monthly NDVI and NDWI
data were generated using the earliest suitable S-2 acquisition each month, and their sequences of values
were used to form B`ezier curves. 5 features were considered for each index time series: arc length, area
of the bounding box, centroid of the bounding box, curvature, mean and standard deviation.
Different clustering strategies including K-Means, DBSCAN, Mean-shift, Hierarchical, and Gaussian
Mixture Model were employed. Accuracy and adjusted rand index (ARI) were used as performance
metrics. ARI ranges between [−1,1], where higher values mean better separation.
Traditional time series features such as mean, variance, maximum, average slope, ... achieve lower
accuracy levels. It can be observed from results that DBSCAN performs better with the
properties of Bezier curves in terms of accuracy and ARI. DBSCAN seems thus to be more effective at
identifying clusters of varying densities, and it is robust to noise. Hence, the proposed features generate
well-defined density-based clusters that other algorithms struggle to identify. Traditional clustering
algorithms typically assume clusters of elliptical shapes. This high disparity suggests non-spherical or
irregularly shaped clusters, where DBSCAN performs better.

This publication is part of the project NODES which has received funding from the MUR–M4C2 1.5 of PNRR funded
by the European Union-NextGenerationEU(Grant agreement no. ECS00000036).

[1] C. Garau, D. Marzi, M. Bordoni, and F. Dell’Acqua. Satellite detection of inter-row management
practices in a north-italy vineyard: Preliminary results. In IGARSS 2024-2024 IEEE International
Geoscience and Remote Sensing Symposium, pages 4325–4328. IEEE, 2024.

[2] C. Meisina, M. Bordoni, A. Vercesi, M. Maerker, C. Ganimede, M. C. Reguzzi, E. Capelli, E. Mazzoni,
S. Simoni, and E. Gagnarli. Effects of vineyard inter-row management on soils, roots and shallow
landslides probability in the apennines, lombardy, italy. In Proceedings, volume 30, page 41. MDPI,
2019.