Fusing deep convolutional neural network features with Thepade’s SBTC for land usage identification

Abstract

Rapid progress in deep learning (DL) is benefitting almost all fields including ‘Remote Sensing’. Until now several deep convolutional neural networks (DCNN) are proposed, every one of them gives distinct results according to its depth as well as hyperparameters. It’s very arduous to develop these DCNNs from scratch because it requires huge efforts and computing resources. Transfer learning (TL) is a technique of customizing such pre-trained DCNNs and molding them for newer classification tasks like ‘Land Usage Identification’ (LUI). The current paper presents an empirical performance assessment of six individual finetuned DCNNs namely VGG16, VGG19, MobileNet, Xception, InceptionV3, and DenseNet121 for LUI. Further, the best performing DCNN’s features are fused with Thepade’s Sorted Block Truncation Coding (SBTC) 10-ary and passed to different machine learning (ML) classifiers like RandomForest, Support Vector Machine (SVM), DecisionTree, NaiveBayes, K-Nearest Neighbours (KNN), ExtraTree and GradientBoosting classifier (GBC) for improving LUI capability of the method. The proposed methodology is experimented on UC Merced Land Use (UCM) dataset by considering accuracy, F1-score, recall, and precision as performance metrics. From computed results, it’s evident that finetuned InceptionV3 + Thepade’s SBTC 10-ary + ExtraTree classifier gives overall remarkable results for LUI across both 70–30 and 80–20 dataset split.

Keywords

• Land usage identification
• Transfer learning
• Pretrained deep CNN
• Remote sensing
• Thepade’s SBTC