A Comparison of Neural Networks and Maximum Likelihood Remotely Sensed Data Classifiers to Detect Logged-Over Tropical Rain Forest in Indonesia

要旨

Selective logging has been applied in the Indonesian tropical rain forest since the 1960. This has resulted in thousands of hectares of logged-over forest. In Labanan, Berau, East Kalimantan, selective logging will enter the second rotation in 2010. A comprehensive analysis of the forest condition should be made before harvesting the logged over forest. One aspect that should be considered is the forest structure. The objective of this study is to compare two classification techniques (Maximum Likelihood and Neural Network Classifiers) in characterizing the condition of the logged over and unlogged tropical rain forest using satellite remotely sensed data, namely: Landsat-7 ETM, JERS-1 SAR, ERS-2 SAR and Radarsat-1 SAR images. The results indicated a significant difference in structure condition between logged over and unlogged forest. The canopy closure, stem density, and basal area of logged over forest in the study area are 84%, 511 trees/hectare (ha), and 26 m2/ha, respectively. The corresponding results for the unlogged forest are 90%, 583 trees/ha, and 32 m2/ha, respectively.The use of a neural networks classifier is found to improve the accuracy of classification result, compared to the maximum likelihood classifier. Moreover, using neural networks, it is possible to classify two classes of logged over forest with significant difference in stem density and basal area per hectare.