Stacking Approach for CNN Transfer Learning Ensemble for Remote Sensing Imagery

In this paper we propose a stacking approach for Convolutional Neural Network (CNN) transfer learning ensemble for remote sensing imagery, in particular for the task of scene classification. We propose to use a combination of features produced by an ensemble of CNNs as one feature vector for classification. At the same time the original data set can be processed with different up-sampling and image enhancement methods and then used to obtain more features from pretrained networks. We investigate both fine-tuning and non fine-tuning approaches for transfer learning. We have selected Brazilian Coffee Scenes data set as a benchmark to measure the classification accuracy. Proposed method in case of a non fine-tuned model shows 89.18% classification accuracy. For a fine-tuned model the best classification rate is 96.11%. We analyzed how networks that have appeared recently (VGG-19 and SqueezeNet), can be applied to the task of transfer learning for remote sensing. Also we describe a method of decreasing processing time and memory consumption while preserving classification accuracy by using feature selection based on feature importance.