29 ISEIS Journal of Environmental Informatics Letters 3(1) 29-39 www.iseis.org/jeil Multi-Layer Perceptron Neural Network and Markov Chain Based Geospatial Analysis of Land Use and Land Cover Change L. Shen 1 , J. B. Li 1 * , R. Wheate 2 , J. Yin 3 , and S. S. Paul 4 1 Environmental Engineering Program, University of Northern British Columbia, Prince George, BC V2N 4Z9, Canada. 2 Geography Program, University of Northern British Columbia, Prince George, BC V2L 1R5, Canada. 3 Ministry of Forests, Lands, Natural Resources Operations & Rural Development, Prince George, BC V2N 4W5, Canada. 4 Faculty of Land and Food Systems, University of British Columbia, Vancouver, BC V6T 1Z4, Canada. Received 2 March 2020; revised 16 March 2020; accepted 24 March 2020; published online 31 March 2020 ABSTRACT. We combined multi-layer perceptron (MLP) neural network and Markov Chain (MC) modeling with object-based image analysis (OBIA) to map and predict land use and land cover (LULC) changes in Stoney Creek Watershed (SCW), British Columbia, Canada. Unsupervised classification was performed using Landsat Thematic Mapper (TM) and Operational Land Imager (OLI) images to produce LULC maps of years 1986, 1999 and 2016. The classification resulted in an overall accuracy of 91.50%. The results show that coniferous forest in SCW experienced a sharp loss while agriculture area increased (4.77% land gain) from 1986 to 2016. LULC scenarios were predicted through MLP neural network and MC modeling based on LULC change analysis data and transition potential. The results indicated that ‘Coniferous Forest’ LULC type had the highest (3.38% land loss) transition potential and ‘Water’ and ‘Urban Area’ LULC types had the lowest transition potential. Application of the proposed method provided valuable information of LULC pat- terns and dynamics for planners and researchers. The method also has the potential for improved management in other watersheds with similar LULC types. Keywords: geospatial analysis, land use and land cover (LULC) change, landsat imagery, markov Chain (MC) model, multi-layer perceptron (MLP) neural network, object-based image analysis (OBIA). 1. Introduction Land use and land cover are two different terms but are often used interchangeably by many researchers. Land use re- fers to the human use of the natural landscape for habitat and livelihood, while land cover represents the biophysical charac- teristics such as vegetation, soil and water distributed on earth’s surface. The conversion of land use in the human subsystem driven by social activities will change land cover, and land co- ver changes could impact natural environment and biosphere (Rawat and Kumar, 2015; Islam, 2018a; Chen et al., 2018). As a result, land use and land cover (LULC) change resulting from anthropogenic activities has led to various concerns for envi- ronmental degradation around the globe (Islam et al., 2018b; Paul et al., 2018). The assessment of LULC change is thus of critical importance for effective environmental management and sustainable development of land resources. Remote Sensing (RS) techniques can be used for LULC chan- ge detection and understanding the dynamics of the change. Due to the spatial data management, creation, and analysis func- * Corresponding author. Tel.: +1-250-9606397; fax: +1-250-9605845. E-mail address: Jianbing.Li@unbc.ca (J.B. Li). ISSN: 2663-6859 print/2663-6867 online © 2020 ISEIS All rights reserved. doi:10.3808/jeil.202000023. tions of geographic information systems (GIS), the combina- tion of RS and GIS has been successfully applied as an effecti- ve technique in LULC change detection (Paul, 2013; Srivastava et al., 2013; Nguyen et al., 2016). Satellite remote sensing tech- nology is especially popular as it is supported by satellite sen- sors which could provide time-series image data with high spa- tial resolution and geometric precision, and can capture tempo- ral variation (Stabile, 2012; Pervaiz et al., 2016). Landsat satel- lite images which provide a continuous inventory of imagery since 1972 have been widely applied for LULC analysis (USGS, 2016). The Landsat sensors have proved sensitive eno- ugh to categorize different spectral patterns related to the LU- LC classes in many complex landscape conditions (Zhao et al., 2012; Butt et al., 2015). RS analysis for change detection is usually relying on digital satellite image classification by assi- gning image pixels to real-world LULC feature types (Paul et al., 2018). Pixel-based classification (PBC) is a conventional method and has been broadly applied as supervised and unsu- pervised classification based on characteristics of single pixel (MacLean et al., 2013; Rwanga and Ndambuki, 2017). Howe- ver, when a pixel-by-pixel classification algorithm is applied to all available image signals, the pixels with similar spectral re- flectance are grouped together, while some spatial and conte- xtual information of image pixels are neglected. Thus, the pi- xels may not represent true geographical objects when using PBC method, and its accuracy would be affected (MacLean et