Kevin Ying

Project: CNN-LSTM Methods for Satellite Imagery: River Water Masking

Lab Group: Ganti Lab

Graduate Mentor: Feifei Zhao

Abstract: Recent developments in deep learning methods have displayed the potential for certain segmentation architectures to be effectively applied to binary river water masking. The majority of proposed solutions to water masking have been based in convolutional neural methods, which extract features from an image to develop a binary mask. We propose an extension of this spatial feature extraction method by leveraging the capacity of geospatial imagery as a time series to introduce a temporal dimension to the model architecture. Our goal in this is to see whether binary river water masking, or broadly any spatiotemporal segmentation task, may significantly improve performance when considering temporality.  Specifically, we employ Convolutional Long Short-Term Memory (ConvLSTM) nodes to simultaneously analyze spatial and temporal features of images, and measure its performance against a U-Net, a common convolutional segmentation architecture. We looked at each model’s F1 score and Jaccard Index (IoU) to determine comparative performance. We found that the ConvLSTM model performed marginally better than U-Net on our data (F1 improving from 0.958 to 0.963; IoU from 0.982 to 0.987), suggesting that the temporal element may have some positive impact on river water masking.