Mapping perennial crops from Sentinel-2 image time series : a use case in Belgium

In recent years, climate change and population growth have resulted in pressures on agricultural systems. In this context, crop monitoring and mapping systems have become increasingly needed as decision support tools for smallholders and farmers and as global land-use evolution assessment products. So far, research aiming to develop such systems has mainly been focused on annual crops. Therefore, the objective of this master’s thesis is to design a classification approach using Sentinel-2 image time series, aiming to discriminate perennial crops from annual crops, then to classify crop types among perennial crops. In order to meet these objectives, metrics specific to crops’ temporal systems were designed, and the performance of two classification models was assessed using multiple feature composition scenarios. The first classifier considered as the baseline is a random forest, and the second one is an extreme gradient boosting classifier. Concerning the binary crop group classification, the designed metrics lead to the worst results when used alone as features but to the best results when coupled with the NDVI time series using the extreme gradient boosting classifier. The best crop type classification results were obtained using existing crop type-specific metrics. Overall, the results were mixed but promising, and improvement opportunities, as well as ideas, were proposed as perspectives for the continuation of the development of the approach.