Deep learning for very high-resolution imagery classification

Document Type

Article

Publication Date

1-1-2017

Abstract

Very high-resolution (VHR) land cover classification maps are needed to increase the accuracy of current land ecosystem and climate model outputs. Limited studies are in place that demonstrate the state-of-the-art in deriving VHR land cover products [1-4]. Additionally, most methods heavily rely on commercial softwares that are difficult to scale given the area of study (e.g., continents to globe). Complexities in present methods relate to (1) scalability of the algorithm, (2) large image data processing (compute and memory intensive), (3) computational cost, (4) massively parallel architecture, and (5) machine learning automation. VHR satellite data sets are of the order of terabytes and features extracted from these data sets are of the order of petabytes. This chapter demonstrates the use of a scalable machine learning algorithm using airborne imagery data acquired by the National Agriculture Imagery Program (NAIP) for the Continental United States (CONUS) at an optimal spatial resolution of 1 m [5]. These data come as image tiles (a total of quarter million image scenes with ~ 60 million pixels) that are multispectral in nature (red, green, blue, and near-infrared [NIR] spectral channels) and have a total size of ~ 60 terabytes for an individual acquisition over CONUS. Features extracted from the entire data set would amount to ~ 8 -10 petabytes. In the proposed approach, a novel semiautomated machine learning algorithm rooted on the principles of “deep learning” is implemented to delineate the percentage of canopy tree cover. In order to perform image analytics in such a granular system, it is mandatory to devise an intelligent archiving and query system for image retrieval, file structuring, metadata processing, and filtering of all available image scenes. This chapter showcases an end-to-end architecture for designing the learning algorithm, namely deep 114belief network (DBN) (stacked restricted Boltzmann machines [RBMs] as an unsupervised classifier) followed by a backpropagation neural network (BPNN) for image classification, a statistical region merging (SRM)-based segmentation algorithm to perform unsupervised segmentation, and a structured prediction framework using conditional random field (CRF) that integrates the results of the classification module and the segmentation module to create the final classification labels. In order to scale this process across quarter million NAIP tiles that cover the entire CONUS, we provide two architectures, one using the National Aeronautics and Space Administration (NASA) high-performance computing (HPC) infrastructure [6,7] and the other using the Amazon Web Services (AWS) cloud compute platform [8]. The HPC framework describes the granular parallelism architecture that can be designed to implement the process across multiple cores with low-to-medium memory requirements in a distributed manner. The AWS framework showcases use-case scenarios of deploying multiple AWS services like the Simple Storage Service (S3) for data storage [9], Simple Queuing Service (SQS) [10] for coordinating the worker nodes and the compute-optimized Elastic Cloud Compute (EC2) [11] along with spot instances for implementing the machine learning algorithm

Publication Source (Journal or Book title)

Large Scale Machine Learning in the Earth Sciences

First Page

113

Last Page

130

This document is currently not available here.

Share

COinS