Identifying Winter Crop Harvest Dates with Machine Learning and Sentinel Time-Series Imagery

Abstract

The date of harvest is information critical to food availability, harvesting logistics and crop modelling. However, acquiring farm management information at national scopes through grower surveys is time-intensive and costly, and harvest dates are estimated by generalising the farming systems in a region and forgoing local variability. This study outlines an approach to predict harvest dates of winter grains at a pixel level from satellite imagery, trained and validated with legacy yield monitor data. A total of 297 yield maps were collected from 2019 to 2021 across 22 farms (56,000 ha) in eastern Australia. Each day in the harvest period (October-December) was defined as pre- or post-harvest and attributed with a time-series of 8 Sentinel-2 bands or indices and 2 Sentinel-1 bands. Data was split into 70-15-15% into calibration, test and validation and modelled with a random forest binary classification model. Recursive backwards elimination reduced the number of variables from an initial 135 to just 12 based on classification accuracy, retaining the day of year, and pre-harvest time series of Normalised Difference Red Edge Index (NDRE) and the red edge 1 band. The model was further optimised by lowering the classification threshold for harvest from 0.50 to 0.42. At a daily timestep, the validation split achieved a RMSE of 10.9 days. As this approach solely utilises publicly available imagery, with an adequate training dataset it is readily scalable to the regional or national scopes required by end users.

Keywords: yield monitor data, random forest model, broadacre cropping, winter crops