Probabilistic neural network based on multinomial model for remote sensing image classification

This paper proposes a Probabilistic Neural Network (PNN) based on a multinomial model for multisensor remote sensing image classification. PNN based on Gaussian model has been widely used for optical-sensor remote sensing image classification. This classifier performs well because the optical-sensor data has a Gaussian distribution model. The problem is that there are applications in tropical countries where the use of optical-sensor data hampered by the presence of clouds. In these cases, either radar images or combination of optical and radar-sensor images is required. The proposed PNN classifier has used the more adaptable multinomial model that could be applicable for both optical and radar-sensor image classification. The important innovative contribution of this work is that the multinomial-based PNN is applicable to work as a uniform classifier for both the optical and the radar-sensor data. The experimental results using optical and SAR-sensor images show that the multinomial-based PNN has slightly better classification accuracies than the Gaussian-based PNN. Using the multinomial-based PNN, the average producer's and user's classification accuracies are 95.360% to 99.88% for Landsat TM images and 95.96% to 98.77% for SAR images. Using the Gaussian-based PNN, the average producer's and user's classification accuracies are 89.10% to 100.00% for Landsat TM images and 85.95% to 98.73% for SAR images. The overall accuracies using multinomial-based PNN ranges 96.39% to 96.40% for Landsat TM images and 97.05% to 97.07% for SAR images, The overall accuracies using Gaussian-based PNN ranges from 95.68% to 95.84% for Landsat TM images and 92.28% to 92.54% for SAR images. The multinomial-based PNN has also a slightly better generalization capability (0.02% to 0.04%) compared to Gaussian-based PNN (0.15% to 0.28%). There are no significant different in their computation times.