Machine Learning for Winter Crop Mapping Using High Spatiotemporal Time Series Satellite Imagery: Case Study – Jendouba, Tunisia

Abstract

Reliable information on crops is required to improve agriculture management and face food security challenges. The work aims at experimenting different machine learning algorithms to identify major crops using time-series Sentinel-2 data covering the region of Jendouba, Tunisia. This chapter describes the workflow for automatic extraction of “semantic information” using a supervised classification approach, applied on a region characterized by a persistent cloud cover during the winter growing season. The results indicated that SVM outperforms the other classifiers, and the best accuracy was achieved using SVM on MSI spline temporal gap-filled with an overall accuracy of 0.89 and kappa 0.86, and that most of the classifiers are robust to noise caused by clouds coverage and handle the high dimensionality of input time-series except Bayes classifier. MSI time-series provides a slightly better results than NDVI time-series, and it appears relevant to consider spline temporal interpolation instead of linear temporal interpolation because of the continuous cloud coverage.