You are here:
Comparison of Support Vector Machine, Neural Network, and CART Algorithms for the Land-Cover Classification Using Limited Training Data Points
Citation:
Shao, Y. AND R. S. LUNETTA. Comparison of Support Vector Machine, Neural Network, and CART Algorithms for the Land-Cover Classification Using Limited Training Data Points. ISPRS Journal of Photogrammetry and Remote Sensing. Elsevier BV, AMSTERDAM, Netherlands, 70:78-87, (2012).
Impact/Purpose:
MODIS (Moderate Resolution Imaging Spectroradiometer) data have been increasingly used to characterize land-cover and monitor vegetation phenology at regional and global scales, since being launched in 2000. One of the most appealing aspects of MODIS data is its unique combination of spectral, spatial, radiometric, and temporal resolutions, which are considered to be substantially improved over other similar observation systems [1]. It has become common practice to utilize MODIS time-series data to monitor vegetation characteristics and condition using phenology information and derived metrics that can provide additional information to differentiate spectrally confusing cover types [2]–[4]. At global scales, Friedl et al. [4] developed an operational annual land-cover mapping approach using MODIS time-series data. For regional applications, a large number of researchers have been exploring MODIS-based classification protocols, algorithms, and validation approaches [5]–[8].
Description:
Support vector machine (SVM) was applied for land-cover characterization using MODIS time-series data. Classification performance was examined with respect to training sample size, sample variability, and landscape homogeneity (purity). The results were compared to two conventional nonparametric image classification algorithms: neural networks (NN) and classification and regression trees (CART). For training sample sizes ranging from 20 - 800 pixels per class, SVM generated overall accuracies ranging from 79% - 83%, compared to 70% - 80% and 58% - 75% for NN and CART, respectively. These results indicated that SVM’s had superior generalization capability, particularly with respect to small training sample sizes. There was also less variability of SVM performance when classification trials were repeated using different training sets. Additionally, classification accuracies were directly related to sample homogeneity/heterogeneity. The overall accuracies for the SVM algorithm were 91% (Kappa = 0.75), 83% (Kappa = 0.57), and 63% (Kappa = 0.33) for homogeneous, intermediate (real world) and heterogeneous pixels, respectively. The inclusion of heterogeneous pixels in the training sample did not increase overall accuracies. Also, the SVM performance was examined for the classification of multiple year MODIS time-series data at annual intervals. Finally, using only the SVM output values, a method was developed to directly classify pixel purity. Approximately 67% of pixels within the Albemarle-Pamlico Basin study area were labeled “functionally homogeneous” with an overall classification accuracy of 91% (Kappa = 0.79). The results indicated a high potential for the operational regional level land-cover characterization.
URLs/Downloads:
viewer.htmLUNETTA 10-108 FINAL JOURNAL ARTICLE..PDF (PDF, NA pp, 3074 KB, about PDF)