Nationwide operational mapping of grassland first mowing dates combining machine learning and Sentinel-2 time series

Abstract

Grassland dynamics are modulated by management intensity and impact overall ecosystem functioning. In mowed grasslands, the first mowing date is a key indicator of management intensification. The aim of this work was to assess several supervised regression models for mapping grassland first mowing date at national-level using Sentinel-2 time series. Three deep-learning architectures, two conventional machine learning models and two threshold-based methods (fixed and relative) were compared. Algorithms were trained/calibrated and tested from field observations, using a spatial cross-validation approach. Our findings showed that time aware deep-learning models – Lightweight Temporal Attention Encoder (LTAE) and 1D Convolutional Neural Network (1D-CNN) – yielded higher performances compared to Multilayer Perceptron, Random Forest and Ridge Regression models. Threshold-based methods under-performed compared to all other models. Best model (LTAE) mean absolute error was within six days with a coefficient of determination of 0.52. Additionally, errors were accentuated at extreme (late/early) mowing dates, which were underrepresented in the data set. Oversampling techniques did not improve predicting extreme mowing dates. Finally, the best prediction accuracy was obtained when the number of clear dates surrounding the mowing event was greater than 2. Our outputs evidenced time aware deep-learning models’ potential for large-scale grassland first mowing event monitoring. A national-level map was produced to support bird-life monitoring or public policies for biodiversity and agro-ecological transition in France.