Dimensionality Reduction of Deep Learning for Earth Observation: Smaller, Faster, Simpler

As deep learning attracts earth observation (EO) community's interest, the challenge to derive explainable, actionable information creates a bottleneck in EO models' development. Computer vision proved that effective results of the DL algorithms imply a considerable amount of training datasets. This is not the case in EO, where images are characterized by a broad variety of sensor data, ranging from multispectral to Synthetic Aperture Radar (SAR) with variable number of spectral bands, polarization, or spatial resolution. In this article, we present effective methodologies for fast training with reduced datasets of simple deep neural networks, while preserving the similar performance of state-of-the-art methods. The hybrid solutions we provide imply the reduction of input data dimension in a convolutional neural network. We replaced dataset's patches with histograms of pixel intensity, Bag-of-Words, or downsampled images. Following the proposed approaches, the training time and the dataset size are significantly reduced, while the performance of classification is preserved. These optimized implementations enable the deployment of lightweight deep learning models for real-time processing tasks able to exhibit accurate results, for instance, a disaster management scenario. We demonstrated the computational efficiency of these approaches on various, complex data, both multispectral and SAR, with different resolutions.