This paper demonstrates the usage of neural network with considerable reduction in training size, large spatial extensions and fine tuning the classification. The illustration has been made through two case studies viz. Land use/Land cover classification of Delhi Ridge, and Species classification of floral resources in Shimla and Chopal Regions. The results have been compared with the statistical methods viz. Maximum Likelihood Classifier, Mahalonobis Distance Classifier, Minimum Distance Classifier
The investigations with different neural network architectures viz. number of hidden neurons, constrained neuronal connections (Hierachical Networks) and fuzzy aggregation based synaptic neuronal functions (Fuzzy Neural Network) are also reported in the present study.
The Ridge Forest Area in the National Capital Territory of Delhi
The Delhi Ridge is the northern extension of Aravalli hill range, the oldest mountain chain in the country, entering the National Capital Territory of Delhi in the Southeast at the Tughlakabad - Bhatti mines - Dera Mandi axis and terminating in the north near Wazirabad on the right bank of Yamuna. Delhi Ridge is divided into four administrative regions viz. Northern, Central, Southern Central and Southern ridges.
Shimla and Chopal Regions
The Shimla and Chopal regions consist of three administrative districts in the state of Himachal Pradesh, namely, Solan, Shimla, and Shirmour.
The study area (Shimla region) has four forest types viz. upland hardwoods mixed with coniferous, mixed coniferous, blue pine, fir/spruce besides exposed rocks, barren lands; and agricultural land and habitation.
The data from Indian Remote Sensing Satellites (IRS IA & IB with LISSII and LISSI) sensors cover the 0.45-0.52, 0.52-0.59, 0.62-0.68, 0.77-0.86 mm spectral ranges. The data set was split, by random selection, into independent training and testing sets comprising of different pixels of each class. The distribution characteristics of spectral response are chosen to be different for different spectral bands and classes. The distribution characteristics are also chosen to be different for training and test data sets. The choice of training and testing data sets is chosen to test the hypotheses:
Neural networks based classification methods are not dependent on statistical distributions of spectral response of different classes in the test and training data sets
Neural networks based classification methods are not dependent upon choice of training data set sizes and spatial characteristics, as also the upscaling ratio
Neural Networks based classification methods are accurate enough with small size of training data set and large upscaling ratio
Neural Network Models
In this investigation, the issues of choice of different representations including data transformation are not addressed as the conclusive negative evidence reported in the literature. However, the investigations regarding effectiveness of different models of neural networks, dependence on architecture (number of hidden neurons), training set size and scale up factor (ratio of the size of testing and training data set) are the focus of this investigation.
In addition, the following neural network models for classification of satellite imageries of Delhi Ridge and Shimla regions are investigated:
- Crisp Neural networks
- Neural network with fuzzy synaptic operation
- Neural networks with constrained connections based on spectral relevance (hierarchical neural networks)
The classification results obtained using the models are compared to those obtained using parametric and non-parametric statistical classifiers, viz. Maximum likelihood classifier, Mahalonobis distance classifier and Minimum distance classifier.
The confusion matrices for land use classification in Delhi Ridge and species classification of floral resources in Shimla region for the case of one choice of training / verification set upscaling ratio 12 are delineate in Tables 1 through 8 respectively..
The study conclusively established the effectivity of crisp neural network based classifiers for satellite imagery analysis. The specific advantage with this set of classifiers is higher accuracy and reduced ground truth data size in comparison to statistical classifiers. The above conclusion is not dependent on application, geographical locations of training and verification data sets and upscaling requirements. The reduction in the requirement of ground truth data is a significant advantage in practical applications. In addition, the crisp neural network is also effective in fine tuning the classification as demonstrated by the forest classification based on dominant species in Shimla and Chopal regions.