A fast instance selection method for support vector machines in building extraction

• Training support vector machines (SVMs) for pixel-based feature extraction purposes from aerial images requires selecting representative pixels (instances) as a training dataset. In this research, locality-sensitive hashing (LSH) is adopted for developing a new instance selection method which is referred to as DR.LSH. The intuition of DR.LSH rests on rapidly finding similar and redundant training samples and excluding them from the original dataset. The simple idea of this method alongside its linear computational complexity make it expeditious in coping with massive training data (millions of pixels). DR.LSH is benchmarked against two recently proposed methods on a dataset for building extraction with 23,750,000 samples obtained from the fusion of aerial images and point clouds. The results reveal that DR.LSH outperforms them in terms of both preservation rate and maintaining the generalization ability (classification loss). The source code of DR.LSH can be found in https://github.com/mohaslani/DR.LSH.

Subject headings

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

NATURVETENSKAP  -- Geovetenskap och miljövetenskap -- Annan geovetenskap och miljövetenskap (hsv//swe)

NATURAL SCIENCES  -- Earth and Related Environmental Sciences -- Other Earth and Related Environmental Sciences (hsv//eng)

NATURVETENSKAP  -- Data- och informationsvetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences (hsv//eng)

Keyword

Support vector machines

Data reduction

Instance selection

Big data

Building extraction

Computerized Image Processing

Datoriserad bildbehandling

Publication and Content Type

ref (subject category)

art (subject category)