Remote Sensing Example for Deep Learning

Remote sensing has become heavily dependent on machine learning algorithms such decision trees, random forests, support vector machines and artificial neural networks. However, there is an increasing recognition that deep learning, which has been applied successfully in other areas such as computer vision and language processing, is a viable alternative to traditional machine learning. This chapter will work through a specific example of the application of deep learning algorithms to one important area of remote sensing data analysis, namely land cover classification. Land cover and land use change analysis is of importance in many practical areas such urban planning, environmental degradation monitoring and disaster management. The main goal of this chapter is to provide a detailed understanding of the performance of various deep learning models applied to the problem of land cover classification starting from a known dataset. Although we use remote sensing as an example, the key point is to show the level of hyper-parameter tuning required to get desired results from any multiclass problem to which deep learning is applied. To emphasize the hands-on approach the full script is provided and can be flexibly targeted to other datasets. We divide the presentation into 5 main parts that cover preliminary information on the models including input data restrictions, followed by exploration of the EuroSAT data contents, pre-processing steps and performance evaluation results for several selected models. Finally, we test the performance of the models with a new data set to get a clear picture of the limitations of the presented approach in the face of unseen data. The Keras library is mainly used throughout this chapter due to its strong user-orientation and intuitive syntax. Finally, some self-evaluation exercises are given to emphasize the key takeaways from the chapter. We also provide a list of references for further reading.