University of Twente Student Theses


Automated Delineation of Smallholder Farm Fields using Generative Adversarial Network

Yan, Qiuyu (2020) Automated Delineation of Smallholder Farm Fields using Generative Adversarial Network.

[img] PDF
Abstract:Smallholder farms play a vital role in agricultural production in many developing countries around the world. As basic geographic information of agricultural resources, accurate boundaries of smallholder farm fields are important and indispensable geo-information for farmers, managers and policymakers to help them manage and utilize their agricultural resource. Beyond that, accurate delineation of smallholder farm fields could promote the sustainable development of agriculture. However, traditional manual methods such as image digitisation by visual inspection of satellite images or filed campaigns are inefficient and time-consuming. Therefore, this research aims to propose an automated algorithm by fully convolutional neural networks (FCN) in combination with generative adversarial networks (GAN) to improve the delineation accuracy of smallholder farms using Very High Resolution (VHR) images. This research consists of two parts. In the first part, we investigate three state-of-the-art fully-convolutional deep network architectures (U-Net, PSPNet, SegNet) to find the optimal architecture in the contour detection task of smallholder farm fields. After that, we aim to conduct the optimal FCN architecture in combination with GAN methods to improve the accuracy of contour detection. Thus, the second part explores the potentials of two GAN methods (ContourGAN and pixel2pixelGAN) for this specific task. The study area is in the Sudano-Sahelian savanna region of northern Nigeria, around the city of Kofa, Bebeji Local Government Area, Kano State. It is a 3×2 km area which composes of abundant small fields and most of them have three or more crops. The VHR image dataset consists of six 1000×1000 pixels tiles extracted from a WorldView-3 image acquired on September 25th 2015. By comparing different methods in this research based on the F1-score, we aim to propose an optimal method for this contour detection task of smallholder farm fields.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page