Multitemporal Relearning With Convolutional LSTM Models for Land Use Classification

In this article, we present a novel hybrid framework, which integrates spatial-temporal semantic segmentation with postclassification relearning, for multitemporal land use and cover (LULC) classification based on very high resolution (VHR) satellite imagery. To efficiently obtain optimal LULC maps, the framework utilizes model to harness temporal dependency extracting high-level features. addition, principle of relearning is adopted optimize output. Thereby, initial outcome provided subsequent via an extended input space guide learning discriminative feature representations in end-to-end fashion. Last, object-based voting coupled coping intraclass low interclass variances. The was tested two different strategies (i.e., pixel-based relearning) three convolutional neural network models, i.e., UNet, simple Convolutional LSTM, UNet Convolutional-LSTM. experiments were conducted datasets labels that contain rich information variant building morphologic features (e.g., informal settlements). Each dataset contains four time steps from WorldView-2 Quickbird experimental results unambiguously underline proposed efficient terms classifying complex maps VHR images.