Comparison of Convolutional Autoencoder Architectures for Representation Learning of MODIS Aqua and Terra Observations

Self-supervised representation learning of Earth observation data is important to perform efficient deep neural network training for classification and regression tasks with limited labels. For example, to predict the onset date of Indian summer monsoon from remotely sensed MODIS data, we have only 21 labels (2002 to 2023), making the training of deep neural networks infeasible. Traditionally, linear methods for dimensionality reduction, such as principal component analysis and dynamic mode decomposition, have been used extensively to represent Earth observation data compactly for use in supervisd learning. Convolutional autoencoders are a class of neural networks that can be used for nonlinear dimensionality reduction and representation learning. In the present work, we compare the performance of multiple autoencoder architectures to learn an efficient representation of observations from the MODIS instrument onboard Aqua and Terra satellites. Specifically, we used and compared multiple ResNet-based autoencoders with different latent dimensions and number of layers, trained from scratch. We show that ResNet50 with 7x7x512 latent dimension serves as a good autoencoder for representation learning purposes.