Proceedings Article | 21 October 2019
KEYWORDS: Agriculture, Satellites, Data acquisition, Satellite imaging, Cartography, Image classification, Accuracy assessment, Associative arrays, Remote sensing, Data modeling, Data acquisition, Satellite imaging, Cartography, Image classification, Accuracy assessment, Associative arrays, Remote sensing, Data modeling
With the recent launch of Sentinel-2, the possibilities of image analysis increased especially in the field of land cover and
crops mapping. The identification of different crop types is important particularly for agricultural field management and
water resource monitoring. This paper aims to select a limited number of acquisition dates, from Sentinel-2 time series,
to improve the accuracy of crop mapping method in the Haouz plain Marrakech, Morocco. First, the collected reference
data was exploited to build the phenological shifts of the different crops present in the area of interest, namely: Apricot,
Citrus, Olive, Cereals, watermelon and bare soil/fallow, using the derived NDVI layers. As a second step, the NDVI
profiles were employed to distinguish different periods based on the spectral separability between the thematic classes.
Then, crop classification independently for the 20acquired images, per period, was performed and evaluated in order to
extract Overall Accuracy (OA) and kappa coefficient temporal evolution. The optimal acquisition dates per period were
determined, relying on the calculated OA values, and classified based on the corresponding NDVI time series. By
minimizing the number of optimal acquisition dates, different combinations were tested and analysed (3 dates per period,
2 dates per period, 1 date per period). The classification results, proved that the proposed approach found out that three
acquisition dates, well distributed through the agricultural season achieve the best differentiation between five crops and
bare soil with an OA of 85%.
