Crop-type mapping using a machine-learning strategic approach and a combination of vegetation indices

Mapping and identifying crop types are essential for predicting crop potential in agricultural fields. The crop-type map can be obtained using remote sensing data. The study area consists of two districts, namely Pasrujambe and Candipuro Districts, located in Lumajang Regency. Sentinel-2 images with 10% cloud cover recorded 12 days were used as input data. The images were processed using the Google Earth Engine (GEE) platform with eight scenarios of vegetation indices combination. Moreover, the most common classification algorithm widely used to produce crop-type maps, Random Forest (RF), was applied to generate crop-type classification maps from Sentinel-2 data. The training and testing data have been collected from field data collection using Global Positioning System (GPS), digital cameras and Google Earth images. In general, the accuracy of overall and kappa of thematic maps achieved more than 80%. The highest accuracy is produced on images with Band-2, Band-3, Band-4, Band-8, Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), and Modified Bare Soil Index (MBI), with a 90.48% overall and 87.79% kappa accuracy. Adding vegetation indices NDVI, SAVI, and MBI (Scenario 2) can significantly improve map accuracy compared to satellite imagery that only uses bands 2, 3, 4, and 8 (Scenario 8).