Machine learning methods for remote sensing applications: An overview
Machine learning algorithms have shown a surprisingly successful development within the last years. Several data intensive technical and scientific fields - like search engines, speech recognition, and robotics - have an enormous benefit of these developments. Remote sensing tasks belong to data intensive applications as well. Today, remote sensing provides data over a wide range of the electromagnetic spectrum (UV, VIS, NIR, IR, and Radar). The capabilities of the sensors include single band images as well as multi- and even hyperspectral data. Due to the fact that remote sensing applications are often monitoring tasks, long time series data are in the focus of image exploitation. Several machine learning algorithms have been used in the remote sensing community since decades, ranging from basic algorithms such as PCA and K-Means to more sophisticated classification and regression frameworks like SVMs, decision trees, Random Forests, and artificial neural networks. Through a combination of data availability, algorithmic progress, and specialized hardware, deep learning methods and convolutional networks (ConvNets) came in the focus of the image exploitation community during the last years and are now on the verge between revolutionary success and illusionary hype. This overview aims to explore in which situations these new approaches are useful in remote sensing applications, which problems are actually solved, and which are still open.