Segmentation of Cropland Field Parcels from SAR-based Phenological Stage Probability Maps
Segmentierung von Feldschlägen aus SAR-basierten Schätzungen der Wachstumsstadien von Nutzpflanzen
Boundary data on cropland field parcels plays an increasing role for applications including precision farming, distribution of subsidies and yield estimation statistics. The demand on this data is so high and its acquisition area so large that automating the process is worth investigating. This thesis shows a novel method to derive arbitrarily shaped field parcel extents from synthetic probability maps originating from Sentinel-1 Synthetic Aperture Radar imagery. This instance segmentation process is completely automatic and generates a map of universally applicable GIS-Features from the given input rasters. In a two-tier process, the outlines of parcel candidates are first extracted from a multi-temporal stack of Phenological Sequence Pattern probability maps (PSP) using the Maximally Stable Extremal Regions blob detection algorithm for grayscale images. A Constrained Triangulation is applied to yield a 2D intermediate representation of the blob features. Secondly, classification results are used to merge adjacent triangle segments back into contiguous parcels. In this step, an adapted form of Statistical Region Merging (SRM) is applied, merging segments based on their cultivation record over multiple growing periods. An area-based accuracy assessment revealed an overall IoU of 0.58, under-segmentation of 0.18 and over-segmentation of 0.33 which is comparable to performance data presented for an alternative approach by Graesser and Ramankutty 2017. While F1-Scores range from 0.41 for parcels >7.5ha to 0.67 for those smaller than 4ha, the overall minimum precision lies at 43% (recall: 96%). This performance is comparable to another study on the topic (Nyandwi et al. 2019), whereas we exceed its recall of 45%. Our approach offers a new technique of field parcel delineation by using exclusively active remote sensing data. The findings of this work could guide the way towards accurate and reproducible acquisition of field's shapes and sizes to ultimately replace cadastral and manually acquired data.
Darmstadt, TU, Master Thesis, 2020