GIS and Remote Sensing

A single action of changing land use (LU) in peri-urban areas looks trivial, but cumulative effect of numerous individual acts would pose significant threat to the people and its surroundings. In Yogyakarta Special Region, Indonesia, the LU change trend was analyzed by observing Landsat 5 Thematic Mapper (TM) images in 2000, 2005 and 2011 with the aid of remote sensing technology and machine learning technique. First, a LU classifier was built by the stochastic gradient boosting method with nested 10-fold cross validation on R 3.4.0 using the 10416 normalized pixels of four spectral bands in Landsat 5 TM 2005 and 16 LU classes supervised by a provincial LU map in 2004. Kappa coefficient value was 0.34, showing overestimation of paddy field as the largest and widely variated class that attracted other LUs pixels. Kappa was improved to 0.50 by decomposing the paddyfield into four equal sized sub-classes by normalized difference vegetation index (NDVI). Aggregating the classified LUs into a simple urban/non-urban classification increased accuracy into 77.1% with positive best fit relation. The classifier was then employed for urban/non-urban distinction in all observation years by using separately normalized digital numbers in the same manner; thus, the urban/non-urban maps were mutually comparable. Result of classification showed urban areas were increasing from 24.4% in 2000 through 29.8% in 2005 to 32.04% in 2011, and this change was dominated by the housing class increase and paddy field decrease. This empirical evidence shall alert the local government to control LU change permitfor the individuals to curb the sprawl in peri-urban areas.