A performance comparison of Landsat TM land cover classification based on neural network techniques and traditional maximum likelihood and minimum distance algorithms
Traditionally it has been difficult to obtain image classification accuracies above 60% when performing landcover mapping with established statistical image classification algorithms such as those currently available on modern image processing systems. The advent of neural network techniques in other scientific fields has improved the processing of multi-dimensional data. Preliminary trials on small data sets have yielded classification accuracies of the order of 80%, indicating that these techniques have significant potential in the field of multi-dimensional image data classification. The application of neural network software to landcover classification of satellite imagery has been investigated. This paper presents a summary of the techniques involved and illustrates the results to date by means of a case study. This compares land cover classification of a Landsat TM scene of the Kennet catchment performed using Erdas maximum likelihood and minimum distance algorithms with the result of classifications performed using neural network techniques.
Bibliographic Reference: Paper presented: 1992 Annual Conference of the Remote Sensing Society, Dundee (GB), September 15-17, 1992
Availability: Available from (1) as Paper EN 37096 ORA
Record Number: 199211491 / Last updated on: 1994-11-29
Original language: en
Available languages: en