Deep Relearning in the Geospatial Domain for Semantic Remote Sensing Image Segmentation

We present a classification postprocessing (CPP) technique based on fully convolutional neural networks (CNNs) for semantic remote sensing image segmentation. Conventional CPP techniques aim to enhance the classification accuracy by imposing smoothness priors in the image domain. Contrary to that, here, a relearning strategy is proposed where the initial classification outcome of a CNN model is provided to a subsequent CNN model via an extended input space to guide the learning of discriminative feature representations in an end-to-end fashion. This deep relearning CNN (DRCNN) explicitly accounts for the geospatial domain by taking the spatial alignment of preliminary class labels into account. Hereby, we evaluate to learn the DRCNN in a cumulative and noncumulative way, i.e., extending the input space based on all previous or solely preceding model outputs, respectively, during an iterative procedure. Besides, the DRCNN can also be conveniently coupled with alternative CPP techniques such as object-based voting (OBV). The experimental results obtained from two test sites of WorldView-II imagery underline the beneficial performance properties of the DRCNN models. They can increase the accuracies of the initial CNN models on average from 72.64% to 76.01% and from 92.43% to 94.52% in terms of
$\kappa $
statistic. An additional increase of 1.65 and 2.84 percentage points can be achieved when combining the DRCNN models with an OBV strategy. From an epistemological point of view, our results underline that CNNs can benefit from the consideration of preliminary model outcomes and that conventional CPP techniques can profit from an upstream relearning strategy.