Classification of floodplain land cover using Landsat image data and an advanced recurrent neural network

Keywords: Deep learning, classification, satellite imagery, recurrent neural network (RNN)
Abstract Type: Paper Abstract

Abstract

Given their intrinsic physical and biological properties, floodplains play various critical roles. Floodplain environments can be characterized and monitored in a spatially-explicit, synoptic manner via remote sensing. Image classification is a commonly-used method for extracting information pertaining to various features/objects/land-cover types within floodplain, and such information can enhance understanding of the physical and chemical properties of floodplain. During the past decade, deep learning (DL), an advanced machine-learning technology, has been well-investigated for the application of satellite remote-sensing image classification. In this research, based on our previous research, we use a novel recurrent neural network (RNN)-based classification framework with a fast sequential-feature extraction to extract land-cover information from Landsat 7 images collected over a portion of the Río Beni floodplain, located in the Bolivian Amazon. Differing from conventional utilizations of RNNs for multi-temporal remote-sensing image classification, for a given target pixel, we propose extracting sequential features from an individual image directly, based on the similarity measurement between the target pixel and all other pixels within the image. To accelerate such a similarity-measurement calculation, especially when processing broad-scale satellite images, we also exploit a multi-resolution segmentation algorithm, which shrinks the search range, yielding a reduced/low computational time cost. Experimental results indicate that our proposed approach performs better than other commonly-used machine-learning and deep-learning classification algorithms.