A Comparitive Analysis of Remote Sensing Land Use Land Cover Image Classification using Deep Convolutional Neural Networks based on Super Pixels

Abstract

The categorization of human activities and natural elements on the landscape over time using scientific and statistical methodologies is referred to as land use or land cover (LULC). Satellite imagery is a great tool for studying terrestrial resources, and land cover classification is critical in this analysis. A benchmark dataset “EuroSAT dataset”, based on Sentinel-2 satellite is used for study purpose. Transfer learning was utilized rather than building training models from scratch. Deep learning approaches, particularly those trained on the ImageNet dataset, have grown in popularity for identifying land use and land cover. In this paper, the pretrained Deep Convolutional Neural Network models, like VGG16, ResNet50, and InceptionV3 have been studied. For comparison, parameters like accuracy, precision, recall and F1-Score have been used. The InceptionV3 model achieved an overall classification accuracy(88.18%),Precision(0.9711), Recall(0.9800) and F1-Score(0.9887) outperforming both VGG16 and ResNet50 models. In this paper, a Modified Mean Shift Algorithm(MMSA) is used to generate valuable features known as Super pixels from the input image to enhance accuracy and reduce execution time. Those super pixels were subsequently fed to the pretrained models as input. A combination of pretrained models and the MQSA is trained and evaluated on the same dataset. The Experimental results shows the combination slightly improves the performance.