Mapping rice and wheat crops in Egypt’s Nile Delta using Sentinel-2 from 2018 to 2022

Mapping crop types is vital for quantifying cultivated areas and supports agricultural statistics, land use analysis, and crop yield predictions. This study focuses on creating accurate rice and wheat crop type masks in Egypt’s Nile Delta, specifically in the Gharbia governorate, from 2018 to 2022, in the context of EO Africa initiative. Crop type maps for both, summer and winter seasons were developed using a Random Forest method. A set of ground truth points for each year, majorly of rice and wheat but also a few of onion, maize, clover, citrus and grape crops were filtered and completed analytically to be used as training and test data. To face imbalance, Synthetic Minority Oversampling Technique (SMOTE) method was also applied, generating synthetic data of the non-interest crops. The masking method consisted in extracting the spectral information of these points for all the selected cloud-masked Sentinel-2 images of the season. This information was stacked, obtaining several features equal to the number of spectral bands per number of images of the season considered. This set was optimized through a PCA analysis applied over the spectral bands, reducing the number of features to the 25-30 first components. For each image, also different vegetation and water indices (NDVI, DVI, SAVI, NDWI, AWEI, EVI) were computed and stacked to the previous set to conform the final features space, reducing dimensionality and showing the best results compared with the application of PCA to the complet set (bands and indices). This process was applied to each season, obtaining an accuracy between 0.85 and 0.95 and consistent commission and omission errors, meaning balanced estimations. From the final classification map, rice and wheat classes were extracted, obtaining preliminary masks. Finally, isolated pixels were removed, and possible detection holes were covered with the use of morphology techniques (opening and closing), obtaining final operational masks.