Remote Sensing Signature Classification of Agriculture Detection Using Deep Convolution Network Models

Abstract

Categorical signature classification plays a vital role in environmental monitoring, disaster response, etc. In view of Geographical expansions, manual identification of the object is time consuming and task of classification is difficult owing to limited trained images. Conventional categorical signature classification using Machine learning techniques requires higher level of abstract features. Deep learning is a successful technique that is widely used in extracting minute level multiple features of the object representation for automatic learning of the data representation. In this paper, performance analysis of the state of the art eight pre trained Convolutional Neural Networks (CNN) networks (Alexnet, Resnet34, Resnet 50, Resnet-101 Resnet-152, VGG-16, VGG-19 and Densenet-121) are tested for the benchmarked datasets (UC Merced and EUROSAT Datasets). Common signature between the datasets like Agriculture, Residential, River and salt lake is considered with the objective of search of the agriculture in an area composed of residential, river and lake. It is concluded from the results that the use of more number of images to train for the search of particular dataset and the use of shallow CNN network increase the accuracy, precision, recall and F-Score, closer to unity (one). Densenet-121 performs better when compared to other CNN networks for both the datasets with an accuracy of 99.67% (EUROSAT) and 97.05% (UC Merced) respectively. Hence, Densenet-121 is recommended for the search of the particular object in a remote sensed scene and classification into respective labels.