Feature Extraction for Remote Sensing Image Classification using Variants of Deep Learning Pre-trained Models Densenet-169, Densenet-121 and Densenet-201

Abstract

Accurately recognizing the objects in an image is known as image classification. The effective classification of high-resolution spatial images for extensive remote sensing archives is known as remote sensing image classification. Efficient extraction of features from images is directly related to high classification performance. Prior to deep learning being widely used in the field of remote sensing picture classification, the majority of feature extraction stages relied on manually created low level features, mostly concentrating on fundamental aspects like colour, form, and texture. However, CNNs (Convolution Neural Networks), which effectively extracted abstract information, soon supplanted these conventional handcrafted methods due to their poorer performance. Significant training restrictions affect deep convolution neural networks, including pre-trained models created on the enormous ImageNet dataset for the main objective of image classification and picture retrieval. These limitations must be appropriately considered during the training phase. This paper attempts to simplify all of the crucial elements that need to be considered while training an extremely deep neural network so that the model can produce the best classification results. In order for the model to generate optimal classification results, This paper examines the experiments conducted on four remote sensing datasets—UC-Merced, AID, NWPU-RESISC-45, and Patternet using the pre-trained model DenseNet and its variants, DenseNet169, DenseNet-121, and DenseNet-201. On the UC-Merced, AID, NWPU-RESISC45, and PatterNet datasets, the DenseNet pre-trained model DenseNet-201 achieved highest test accuracy of 97.60% and least Test Loss of 0.6402% respectively. This indicates that the PatterNet dataset is the most effective at classifying remote sensing images.There are several potential applications for the content-based remote sensing information retrieval system, including forestry and agricultural. With merely an aerial view, CBIR might be a huge help in agricultural regions to identify sick crops. By remotely monitoring the impacted region deforestation may be tracked.