Near-real-time crop mapping in Ningxia (China) using Sen2-Agri system : contributions of Sentinel-2 and Gaofen-1 satellite data

In the framework of the Dragon 4 cooperation (between ESA and the Chinese MOST) and the recent development of the Sentinel-2 for Agriculture system, the aim of this master thesis is to evaluate the current capacities for near-real-time crop mapping with Sentinel-2 and Gaofen-1 satellite data. The first objective was to analyse and compare the data quality of both satellites and assess their compatibility for crop mapping. The results of this joint imagery analysis showed a potential compatibility of both satellites but revealed an issue in the coregistration of Gaofen-1 and a significant geometrical shift between both satellites’ images. Concluding further research is needed before considering a combination of Sentinel-2 and Gaofen-1 for crop mapping, this study further focussed on crop mapping with Sentinel-2 only. The focus was then set on the evaluation of the Sentinel-2 for Agriculture system performances, by using it to map an agricultural area in the Ningxia Hui Autonomous region in China. As a first step, several experiments were carried out on the cropland mask and the crop type map products generated for the 2017 growing season. The general accuracy of the cropland mask was assessed as well as its sensitivity to the training data and to the classification algorithm. The experiments led to the conclusion that the training dataset has a great impact on the resulting cropland mask and that its spatial distribution and land cover diversity coverage are crucial. The accuracy of the crop type classification was assessed for the different crops of the region. The system delivered a very accurate classification for the main crops but showed difficulties for identifying the more marginal crop types. The capacity of crop mapping along the season was also evaluated. Results revealed an accurate crop map could be generated from the end of June on. The second stage was to perform near-real-time crop mapping for the 2018 growing season, considering the previous observations. Thanks to a better distributed training dataset, better crop mapping performances were achieved, confirming the importance of training data in the Sen2-Agri system. To sum up, this study showed Gaofen-1 could contribute to crop mapping in combination with Sentinel-2 in the future, if a solution is found for the geometrical correction. Secondly, it confirmed Sentinel-2 for Agriculture is a very efficient system for crop mapping. Finally, experiments showed the most crucial factor for obtaining accurate products, if enough imagery is available, is the quality of the training and validation data.