Science Inventory

DeepCoast: Quantifying Seagrass Distribution in Coastal Water through Deep Capsule Networks

Citation:

Perez, D., K. Islam, V. Hill, B. Schaeffer, R. Zimmerman, AND J. Li. DeepCoast: Quantifying Seagrass Distribution in Coastal Water through Deep Capsule Networks. In Proceedings, Chinese Conference on Pattern Recognition and Computer Vision (PRCV), Guangzhou, CHINA, November 23 - 26, 2018. Springer, Heidelberg, Germany, 404-416, (2018). https://doi.org/10.1007/978-3-030-03335-4_35

Impact/Purpose:

The goal of this project is to develop a model that is able to quantify LAI from high resolution satellite imagery with limited field observations that may be ubiquitously applied to other localities. To achieve this goal, the following two questions need to be answered: (1) Do the satellite multispectral images contain enough information for a machine learning model to learn and quantify LAI in the same region? (2) Can a machine learning model trained at one location be generalized to other locations for LAI mapping?

Description:

Seagrass is a highly valuable component of coastal ecosystems ecologically and economically, yet reliable mapping of seagrass density is not available due to the high cost of data processing and spatial mapping. This paper presents a deep learning approach for quantification of leaf area index (LAI) levels of seagrass in coastal water using high resolution multispectral satellite images. Specifically, a deep capsule network (DCN) is developed for simultaneous classification and quantification of seagrass based on the multispectral images. The DCN is jointly optimized for classification and regression, and is capable of performing end-to-end seagrass quantification. We separately validated the proposed method on three images taken in Florida coastal area and achieved much better results with DCN when compared against a deep convolutional neural network (CNN) model and a linear regression model. In addition, transfer learning strategies are developed to transfer knowledge in a DCN trained at one location for seagrass quantification to different locations with minimum field observations, which saves a significant amount of time and resources in the mapping of seagrass LAI. Our experimental results show that the developed capsule network achieved superb performances in few-shot transfer learning as compared to direct linear regression and traditional CNN models.

Record Details:

Record Type:DOCUMENT( PAPER IN NON-EPA PROCEEDINGS)
Product Published Date:11/02/2018
Record Last Revised:09/06/2019
OMB Category:Other
Record ID: 346422