An approach for the automatic detection of agricultural field sub-boundaries from high resolution satellite images

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Hacettepe Üniversitesi, Fen Bilimleri Enstitüsü, Geomatik Mühendisliği A.B.D., Türkiye

Tezin Onay Tarihi: 2014

Tezin Dili: İngilizce

Öğrenci: SAMAN GHAFFARİAN

Danışman: Mustafa Türker

Özet:

Agricultural crop mapping is quite important for crop yield estimation in regional and national scale. Remote sensing images are popular data to identify and classify land cover types in the agricultural areas. The recent image classification techniques for agricultural areas use approaches which work on field-by-field basis by means of assigning a crop label for each agricultural field individually. In field-based classification approaches, the classification is performed within the permanent agricultural field boundaries that are stored in a geographical information system (GIS) as vector polygons. However, crop variation within the fields is an important problem to be solved. To solve this problem, image segmentation is needed to be executed to extract the sub-boundaries within the permanent boundaries of the fields and subsequently to achieve higher accuracy in field-based classification operations.

In this study, a field-based segmentation approach is proposed to extract within-field sub-boundaries from high resolution remotely sensed images. The within-field sub-boundary extraction operation is carried out one field at a time by means of processing each field separately. First, the within-field edges are detected using the Canny edge detection algorithm and the image is clustered using an automatic Fuzzy C-means (FCM) clustering algorithm. To automate the FCM clustering algorithm, the algorithm is executed iteratively starting with the assumption that the fields contain maximum six sub-fields that correspond to six clusters. After the first iteration, the Euclidean distances between the cluster centers are computed. If at least one distance stays below a defined threshold value, the number of sub-fields is decreased by one and the clustering operation is repeated with the reduced number of clusters. This iterative execution of the FCM clustering algorithm is carried out until all the distances between the cluster centers stay above the threshold value. Next, the external forces for the Gradient Vector Flow (GVF) Snake are calculated based on the detected edges and the clusters. To calculate and construct the gradient vectors, the distances between the pixels are computed within each cluster so as to steer the contours toward the correct boundaries. After computing the external forces, a novel cluster-based method is used to seed the GVF Snake by means of constructing an ellipse for each cluster that fall within the field. After that, the GVF Snake is executed for detecting the within-field sub-boundaries. As the final step, the detected sub-boundaries are simplified through a line simplification algorithm and thus the final appropriate sub-boundaries are extracted.

The developed approach was implemented in an agricultural area in Karacabey, Bursa plain located in north-west of Turkey. The high resolution satellite images used include the Ikonos multispectral (XS) image acquired in 15 July 2004 and the Quickbird multispectral (XS) and pansharpened (PS) images acquired in 13 August 2004.The results achieved are quite promising. The overall sub-boundary extraction accuracies through the proposed automatic approach were computed to be 93.61%, 84.96% and 88.78% for the Ikonos (XS), Quickbird (XS) and Quickbird (PS) images, respectively.