University of Twente Student Theses


Polygonal delineation of greenhouses using a deep learning strategy

Pote, Ranju (2021) Polygonal delineation of greenhouses using a deep learning strategy.

[img] PDF
Link to full-text:
(only accessible for UT students and staff)
Abstract:Geoinformation update and maintenance are crucial for planning, decision-making processes and geospatial analysis. In the Netherlands, the Dutch cadaster (Kadaster) handles the geodata maintenance, and updates those datasets. As per the Dutch Kadaster, “The digital map is still being built”. ‘Basisregistratie Topografie’ (BRT) registry of the Kadaster contains the geospatial information of objects such as buildings, the agricultural field, roads, and tracks, which are freely available as open data. One of the objects of interest is the greenhouses used for horticulture purposes. The greenhouses are being manually delineated for updating the geodata set. Kadaster has been using deep learning approaches for object recognition. However, state of the art image segmentation models applied in Kadaster typically output segmentation in raster format. The applications of geographic information systems often require vector output. Additionally, there is a considerable research gap for the delineation of greenhouses through the deep learning (DL) method in vector format. Thus, this study aims at developing a DL technique to extract the greenhouses in a vector format. There are two state of the art methods for vectorization using deep building segmentation. First is an end-to-end method that learns the vector representation directly, and secondly, vectorizing the classification map by a network. In this study, the second state of the art method was utilized. Girard et al. (2020) introduced a building delineation method based on frame field learning to extract the regular building footprints in polygonal vector format using aerial RGB imagery. The method was utilized in the greenhouse, where a fully convolution network (FCN) was trained to simultaneously learn the mask of the greenhouse, contours and the frame field, followed by polygonization. The contours information in the frame field produces regular outlines which accurately detects the edges and the corners of the greenhouse. The study was conducted within the three provinces of the Netherlands. Two orthoimage datasets of summer and winter images with the resolution of 0.25 m and 0.1 m, respectively, were used. The normalized digital surface model (nDSM) was added to the winter RGB images to extract the accurate and regular greenhouse polygons. The addition of nDSM improved the prediction and outlines of the greenhouses compared to using only 0.1 m winter RGB images. The mean intersection over union (IoU) of (RGB + nDSM) for 0.1m images was 0.751, while for the same resolution dataset, the IoU was 0.673, indicating the improvement of greenhouse delineation accuracy with the addition of height information. The IoU for 0.25m RGB image was 0.745 and could predict the greenhouses, which 0.1m RGB image could not. The qualitative analysis of the result shows the regular and precise predicted polygons.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Spatial Engineering MSc (60962)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page