A Bayesian-inspired, deep learning-based, semi-supervised domain adaptation technique for land cover mapping

Land cover maps are a vital input variable to many types of environmental research and management. While they can be produced automatically by machine learning techniques, these techniques require substantial training data achieve high levels accuracy, which not always available. One technique researchers use when labelled scarce is domain adaptation (DA)—where from an alternate region, known as the source domain, used train classifier this model adapted map study or target domain. The scenario we address in paper semi-supervised DA, where some samples available In present Sourcerer, Bayesian-inspired, deep learning-based, DA for producing land satellite image time series (SITS) data. takes convolutional neural network trained on then trains further with novel regularizer applied weights. adjusts degree modified fit data, limiting change few number increasing it quantity increases. Our experiments Sentinel-2 images compare Sourcerer two state-of-the-art four baseline models. We show that different source-target pairings outperforms all other methods any fact, results more difficult starting accuracy (when no available), 74.2%, greater than next-best method 20,000 instances.