Convolutional Neural Network Considering Physical Processes and its Application to Diatom Detection

Kohei Arai

doi:10.14569/ijacsa.2019.0101214

Abstract

Convolutional Neural Network (CNN) considering physical processes with time series of stages for diatom detection with remote sensing satellite derived physical data (Chlorophyll-a, Photosynthesis Available Radiance (PAR), Turbidity, Sea Surface Temperature (SST)) and meteorological data is proposed. Diatom is bloomed under the condition of suitable sea water temperature, nutrition rich water (Chlorophyll-a derived from river water flow), photosynthesis available radiance derived from solar irradiance, transparency of the sea water for photosynthesis (turbidity), and sea water convection between bottom sea water and sea surface water. Almost all the conditions can be monitored by remote sensing satellite-based radiometers. The proposed diatom prediction based on convolutional neural network with remote sensing satellite and meteorological data is validated. Through the experiments at Ariake bay area, Kyushu, Japan with gathered time series of remote sensing data of Moderate resolution of Imaging Spectroradiometer (MODIS) derived turbidity as well as chlorophyll-a data estimated for the winter seasons (from January to March) during from 2010 to 2018 together with measured and acquired meteorological data for the same winter seasons, the proposed method is validated.

Highlights

One of the problems of DLM: Deep Learning Method (Convolutional Neural Network: NN, Recurrent Convolutional Neural Network: RCNN, etc.) is that DLM cannot consider time and spatial relations among data for input nodes
It is found that the proposed prediction method for large diatom appearance is validated with the meteorological data and Moderate resolution of Imaging Spectroradiometer (MODIS) derived turbidity as well as chlorophyll-a data estimated for the winter in 2012 and 2015 [1]
Proposed diatom prediction based on neural network with remote sensing satellite and meteorological data is validated

Summary

Introduction

One of the problems of DLM: Deep Learning Method (Convolutional Neural Network: NN, Recurrent Convolutional Neural Network: RCNN, etc.) is that DLM cannot consider time and spatial relations among data for input nodes. Another problem is that DLM is not supported physical processes physical meaning is significantly important. As truth data of the diatom appearance are obtained from the shipment data of the number of cells of diatoms which are acquired with research vessel provided by the Saga Prefectural Institute of Ariake Fisheries Promotion: SPIAFP.

Objectives

Methods

Findings

Conclusion