Land Use Classification with Back Propagation Neural Network and The Maximum Likelihood Method: A Case Study in Ciliwung Watershed, West Java, Indonesia.

Abstract

Ciliwung Watershed is located at West Java Province; this area ranges from 106°47'29” to 107°0'25" E and from 6°24'16" to 6°46'23"N, having hard relief with slope 15% – 30%. Ciliwung watershed area has 117 km in length and 347 km2, it is one of water catchments that suffering damage relatively serious from the upstream to the downstream. Nevertheless, in less than 10 years since 1985, the conservation area has changed into settlement area or cultivation area. The objectives of the research are to apply back propagation neural network and the maximum likelihood classification method for land use classification and compare the performance of the two methods. A one periods of Landsat-7 ETM+ (December 22, 2001) images with the path/row 122/065 was used for classification in the Ciliwung Watershed. This research compared parametric method (maximum likelihood) and non-parametric method (back propagation neural network) that occupied the same Landsat-7 ETM+ and the same training area. Six bands (band 1, 2, 3, 4, 5, 7) from Landsat Image were used as input data for both classification methods. In this case, band 6 and 8 on the image not utilized for this research because both of band have different resolution with the others. Before Landsat-7 ETM+ used for classification process, it was corrected geometrically, atmospherically, and topographically. The purpose of these corrections was to decrease the error that can occur during the making of training area and classification process afterward. Landuse was classified into 8 classes: tea garden, settlement, paddy field, grass, forest, farm, bush, and water body. The target of the training area was based on Ikonos image interpretation, instead of using field data. There were 4000 pixels used for whole class categories. There were several experiments was performed by using back propagation neural network method, which were by changing number of pixels (from 100 to 4000 pixels) and number of iteration (100-2000 iterations). Error matrix was used as accuracy measurement in order to compare both methods. Error matrix showed that both methods had difficulty in classifying water class and paddy field class due to the closeness of spectral value between those particular classes. This research is also showed that back propagation neural network giving a better accuracy level (81.5%) rather than maximum likelihood method (73.4%). Kappa statistic showed that classification result by using back propagation neural network method is nearly close to the real field condition due to the value is 0.759 (75%), which is close to 1 as an ideal case.