Classification of Small Region of Interest from Remote Images Using Neural Networks

The classification and segmentation of small region of interest from ground images is important in many domains: flood detection and evaluation, crop monitoring, environment monitoring, defense, etc. The paper uses a fusion scheme to classify small regions from remote images in four classes (different for satellite and unmanned aerial vehicle UAVs). As novelty, two neural networks, considered as primary classifiers, AlexNet and Perceptron (the last based on ten textural image features) are combined in a convolutional scheme to make up the global classifier. The weights from the convolutional layer are calculated according to the performances of the primary classifiers in a validation phase. Two datasets with different images are used for classifier learning, validation, and testing, one consisting in images from satellite and another consisting in images from UAVs. The global system performances were better than those of the individual neural network (system components) accuracies.