Comparing ARTMAP Neural Network with the Maximum-Likelihood Classifier for Detecting Urban Change

Urbanization has profound effects on the environment at local, regional, and global scales. Effective detection of urban change using remote sensing data will be an essential component of global environmental change research, regional planning, and natural resource management. This paper presents results from an ARTMAP neural network to detect urban change with Landsat TM images from two periods. Classification of urban change, and, in particular, conversion of agriculture to urban, was statistically more accurate with ARTMAP than with a more conventional technique, the Bayesian maximum-likelihood classifier (MLC). The effect of different levels of class aggregation on the performance of change detection was also explored with ARTMAP and MLC. Because ARTMAP explicitly allows “many-to-one” mapping, classification using coarse class resolution and fine class resolution training data generated similar results. Together, these results suggest that ARTMAP can reduce labor and computational costs associated with assembling training data while concurrently generating more accurate urban change-detection results. Introduction The world is undergoing an urban transformation unprecedented in human history. Human settlements, which for tens of thousands of years were mainly rural, are becoming increasingly urban. Globally, urban agglomerations have expanded into the countryside, transforming natural ecosystems, converting agricultural land, and enveloping agrarian communities. Although urban areas cover less than 2 percent of the Earth’s total land surface (Grübler, 1994), half of the world’s population reside in urban regions, and, according to recent United Nations estimates, 60 percent of global population will reside in urban areas by 2030 (United Nations, 2001). This urban revolution has profound environmental impacts at multiple scales, including local and regional climate change, loss of wildlife habitat and biodiversity, and increases in pressure on water, energy, and agricultural resources. From the provision of clean drinking water to the construction of transportation infrastructure, every aspect of the urbanization process presents huge environmental challenges. Successful monitoring of temporal and spatial patterns of urban change will be imperative to anticipate—and hopefully mitigate— negative environmental, social, and economic impacts of Comparing ARTMAP Neural Network with the Maximum-Likelihood Classifier for Detecting Urban Change Karen C. Seto and Weiguo Liu urban development. There is a need to monitor not only urban expansion, but also changes within built-up areas, such as intensification of land use within urban regions. Therefore, the use of remote sensing to detect both expansion and intensification of urban areas will be an essential component of global environmental change research, regional planning, and natural resource management. Among the myriad studies that have assessed urban landuse change with remote sensing (Barnsley and Barr, 1996; Ridd and Liu, 1998; Zhang and Foody, 1998; Jensen and Cowen, 1999; Masek et al., 2000; Ji et al., 2001; Lopez et al., 2001; Stefanov et al., 2001; Yeh and Li, 2001; Seto et al., 2002), three common issues emerge. First, the detection of urban change often is confounded with variations in vegetation and ground reflectance associated with the agricultural crop cycle of planting, growth, and harvesting. This confusion is pervasive throughout tropical and subtropical regions, in areas with multi-crop fields, and in places where agricultural plot size is small. Fallow or barren fields confound the classification of urban areas or urban change. Accurate estimates of agricultural land loss to urban expansion are vital, yet difficult to acquire, even through annual compendium (Seto et al., 2000). Because agriculture by definition is diverse and includes a variety of crops and cropping patterns, it can be easily misclassified as urban change. In Asia, where a majority of the urban growth in the 21 century will occur, multi-crop fields, agricultural terracing, and small field sizes produce textures and tones that can be difficult to differentiate from patches of urban expansion. Separation of urban change from agricultural phenology has been achieved with the use of synthetic aperture radar (SAR) and fusion techniques (Henderson and Xia, 1997; Kuplich et al., 2000; Dell’Acqua and Gamba, 2001), but there has been limited success using only optical data. Second, the spatial and temporal patterns of urban features are difficult to characterize. Urban land use occurs along a continuum and is manifested in different shapes, sizes, styles, and trajectories. As such, there is no sole archetype for urban change. Interand intra-class spectral and land-cover variability of urban and agriculture types limit the success of a single approach to urban change mapping. For example, urban development in the western United States often involves the complete removal of existing land cover and replacement with concrete (Jensen and Cowen, 1999). This suburban model of urban change usually occurs along roads and other infrastructure development. In contrast, urbanization in PHOTOGRAMMETR IC ENGINEER ING & REMOTE SENS ING September 2003 981 K.C. Seto is with the Department of Geological and Environmental Sciences and the Institute for International Studies, Stanford University, Encina Hall, E413, Stanford, CA 943056055 ([email protected]). W. Liu is with ACI Worldwide Inc., Riverside, RI 02915 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 69, No. 9, September 2003, pp. 981–990. 0099-1112/03/6909–981$3.00/0 © 2003 American Society for Photogrammetry and Remote Sensing 03-914.qxd 8/7/03 5:26 PM Page 981