Extreme Learning Machine-based Crop Classification using ALOS/PALSAR Images

Classification maps are required for agricultural management and the estimation of agricultural disaster compensation. The extreme learning machine (ELM), a newly developed single hidden layer neural network is used as a supervised classifier for remote sensing classifications. In this study, the ELM was evaluated to examine its potential for multi-temporal ALOS/PALSAR images for the classification of crop type. In addition, the k-nearest neighbor algorithm (k-NN), one of the traditional classification methods, was also applied for comparison with the ELM. In the study area, beans, beets, grasses, maize, potato, and winter wheat were cultivated; and these crop types in each field were identified using a data set acquired in 2010. The result of ELM classification was superior to that of k-NN; and overall accuracy was 79.3%. This study highlights the advantages of ALOS/PALSAR images for agricultural field monitoring and indicates the usefulness of regular monitoring using the ALOS-2/PALSAR-2 system. Discipline: Agricultural engineering Additional key words: Hokkaido, machine learning, sigma naught *Corresponding author: [email protected] Received 12 May 2014; accepted 4 February 2015. Introduction Land-cover classification is one of the most common applications of remote sensing. Crop-type classification maps are useful for yield estimation and agricultural disaster compensation, in addition to the management of agricultural fields. Optical remote sensing remains one of the most attractive options for the accumulation of biomass information and forest monitoring (Samreth et al. 2012, Sarker and Nichol 2011). In addition, while optical satellites such as ALOS/AVNIR-2 (Sonobe et al. 2014a), Landsat (Hartfield et al. 2013), MODIS (Sakamoto et al. 2009), and NOAA (Hirano and Batbileg 2013) have been employed in the identification of species and conditions of vegetation, cloud cover significantly limits the number of available optical images, radar is unaffected by cloud cover or low solar zenith angles (Bindlish and Barros 2001). And significant information about soil and vegetation parameters can also be obtained through microwave remote sensing (Sonobe et al. 2014b). These techniques are employed increasingly to manage land and water resources for agricultural applications (Sonobe et al. 2014c). The number of studies on rice and wheat monitoring and mapping using SAR data has increased, and some studies utilizing multi-temporal SAR data have reported high correlations between backscattering coefficients, and plant height and age (Chakraborty et al. 2005, Sonobe et al. 2014d, Waisurasingha et al. 2008). These examples highlight possible uses in the area of agricultural management, specifically for the identification of paddy fields. They indicate the potential use of SAR data for the discrimination of crop types. Furthermore, a number of examples have shown the practical usefulness of supervised learning for classification (Sonobe et al. 2014b). This paper reports a comparison of crop classification using PALSAR data performed by extreme learning machine (ELM) and k-nearest neighbor algorithm (k-NN).