Genetic Programming for Classification of Remote Sensing Data

Abstract

本研究之目的是在探討機器學習方法應用於遙測影像分類之可行性，並利用SPOT衛星資料以遺傳程式設計法進行分類，以區分植生、裸土及火災跡地。分類結果顯示遺傳程式設計法可以有效分類遙測影像，以訓練樣本進行分類之精確度可達99%，遺傳程式設計法所自動產生之電腦程式並可用於選取分類所需之重要變數。機器學習方法分類結果並與傳統之統計方法分類結果相互比較，結果顯示二者之分類效果相似。

The overall objective of this research was to develop an adaptive machine learning technique for the classification of remote sensing data. The genetic programming paradigm was implemented to classi1 vegetation, bare soil, and burnt-over areas using SPOT multispectral data. Two SPOT imageries obtained on 31 Dec. 1986 and 15 Jan. 1988 were used in this study. The results show that the genetic programming paradigm was very effective in classi1’ing the data set (e.g., the best classification accuracy obtained was 99% for the training samples). Moreover, the computer programs derived from genetic programming allowed important variables for classification to be identified. Classification results for the machine learning approach were then compared to the results obtained using a conventional statistical approach (i.e., the Gaussian maximum likelihood classifier). The comparison shows that the classification results for both approaches are similar.