A quantitative and qualitative assessment from official statistics to spatial statistics: Agricultural greenhouses detection over time integrating of remote sensing and transfer learning-based machine learning approach

Abstract

The availability of medium-resolution satellite data such as the open-access Sentinel-2 as well as high-resolution commercial satellite imagery from PlanetScope presents significant opportunities for the agricultural sector and allows us to gain insight into the land surface, land use, and their management. Agricultural Greenhouses (AGs) are the fastest-growing food or commercial ornamental production approach around cities, driven by different factors. The current study is aimed to conduct temporal greenhouse maps in the covered upland greenhouse region (Isparta-Deregümü region, Southwestern Türkiye) from 2016 to 2021 using open-access Sentinel 2 and PlanetScope imagery and machine learning algorithm in a high-performance computing environment (R Core Environment). As a result of the qualitative evaluation of satellite imagery with two different spatial and spectral resolutions, PlanetScope, which has a higher spatial resolution, was determined to be useful in the detection of AGs. Temporal greenhouse maps were generated using random forest algorithm at two-time periods, and the overall accuracy of the predictions was around 90%. While the total greenhouse area in the current area increased by 47% from 2016 to 2021 in official statistics, the methodology allowing to obtain spatial statistics detected this increase by 76%. The current study significantly improves the link between spatial statistics and official statistics.