Preliminary Evaluation of Bespoke Spatiotemporal Embeddings for Sub-Pixel Landscape Feature Detection

High availability of openly accessible Earth observation (EO) data has unlocked vast potential for low-cost land monitoring. Some monitoring tasks, however, remain challenging to perform using data of 10 metre resolution (currently the best available for open satellite imagery) or lower. This study was performed under the assumption that spatiotemporal raster embeddings hold potential to alleviate the aforementioned difficulties for one specific task: in-pixel detection of landscape features (as defined by the Common Agriculture Policy of the EU) using Sentinel-1 and Sentinel-2-based data, as a low-cost alternative to processing high-resolution orthophotographic imagery and on-site surveys. Evaluation of this principal assumption was performed by using spatiotemporal embeddings of Sentinel-1 and Sentinel-2-derived datasets (produced by a naive CNN autoencoder) as covariates for pixel-based classification. Preliminary evaluation demonstrated potential applicability of the method, with F1 scores above 0.8 for over half of the evaluated landscape feature classes on the validation dataset.