Greenhouse Mapping using Object Based Classification and Sentinel-2 Satellite Imagery

Bektaş Balçık F., Senel G., Göksel Ç.

8th International Conference on Agro-Geoinformatics (Agro-Geoinformatics), İstanbul, Turkey, 16 - 19 July 2019 identifier identifier


Efficient methodologies to map greenhouses are very important for the implementation of sustainable agricultural practices, natural resource management, and sustainable urban and rural development. Remote sensing imagery provides a great potential with different spatial and spectral resolutions for greenhouse monitoring and mapping. The conventional techniques for greenhouse mapping are time consuming, and expensive. Because of this reason, many different image processing methods such as classification methods including pixel-based or object based classification and remote sensing indices have been applied for greenhouse mapping. In this study, greenhouses in Anamur, Mersin, Turkey were determined by using object based classification and selected remote sensing indices. Freely available new generation 2018 dated Sentinel-2 MSI data which has 10-meters spatial resolution was used to detect the greenhouse in the selected region. Multi-resolution segmentation (MRS) method was conducted to Sentinel-2 NISI data for object-based image analysis (OBIA). In the first stage, the image segmentation process was performed. Spectral features (mean values of the layers) and remote sensing indices such as Normalized difference vegetation index (NDVI), Normalized difference water index (NDWI) and Retrogressive plastic greenhouse index (RPGI) were extracted from the segmented image. Then, four different datasets were created and the OBIA classification process was performed by applying the nearest neighbor classifier to the created data sets. Reference dataset for training and validation has been created by field survey, apart from this some of the sample are taken with the help of high resolution Google earth images. On the final stage, the accuracy assessment analysis was performed to test the agreement between classification results and ground truth data using error matrix. Dataset-4 (mean values of the layers, NDVI, NDWI and RPGI) has the highest producer and overall accuracies with 82% and 74%, respectively.