Measuring the effectiveness of different types of imagery and image-derived products in land-cover classification – An approach based on Shapley values and game theory

PDF

Authors
  1. Nandlall, S.D.
  2. Millard, K.
Corporate Authors
Defence Research and Development Canada, Ottawa Research Centre, Ottawa ON (CAN)
Abstract
Land-cover mapping consists of determining the type and usage of particular tracts of land, and is often accomplished with remote sensing and classifiers. The maps generated with landcover analysis are used for applications that include space-based Intelligence, Surveillance and Reconnaissance (ISR); Geospatial Intelligence (GEOINT); and Intelligence Preparation of the Operational Environment (IPOE). Land-cover classifiers frequently employ different types of input data, such as imagery obtained from several types of sensors—including Synthetic Aperture Radar (SAR), LiDAR, and optical satellites—as well as ancillary datasets such as Digital Elevation Models (DEMs). However, it can be challenging to determine which inputs have the greatest impact on the accuracy of the classifier, as well as assess how important each input is relative to the others. In this work, a method of quantifying the relative importance of each input is developed and demonstrated using previously developed land-cover classifiers. The proposed method employs concepts from game theory and relies on the Shapley value, which provides a quantitative assessment of each input’s importance in terms of its average contribution to the accuracy of the classifier. The approach described herein thus provides a robust method of determining which types of images and image-derived products are most important in classifying terrain.

Il y a un résumé en français ici.

Keywords
Shapley value;game theory;remote sensing;image classification;variable selection;variable importance;random forest;land-cover mapping
Report Number
DRDC-RDDC-2018-R301 — Scientific Report
Date of publication
01 Jan 2019
Number of Pages
26
DSTKIM No
CA048557
CANDIS No
809206
Format(s):
Electronic Document(PDF)

Permanent link

Document 1 of 1

Date modified: