An Elegant End-to-End Fully Convolutional Network (E3FCN) for Green Tide Detection Using MODIS Data

Using remote sensing (RS) data to monitor the onset, proliferation and decline of green tide (GT) has great significance for disaster warning, trend prediction and decision-making support. However, remote sensing images vary under different observing conditions, which bring big challenges to detection missions. This paper proposes an accurate green tide detection method based on an Elegant End-to-End Fully Convolutional Network (E3FCN) using Moderate Resolution Imaging Spectroradiometer (MODIS) data. In preprocessing, RS images are firstly separated into subimages by a sliding window. To detect GT pixels more efficiently, the original Fully Convolutional Neural Network (FCN) architecture is modified into E3FCN, which can be trained end-to-end. The E3FCN model can be divided into two parts, contracting path and expanding path. The contracting path aims to extract high-level features and the expanding path aims to provide a pixel-level prediction by using a skip technique. The prediction result of whole image is generated by merging the prediction results of subimages, which can also improve the final performance. Experiment results show that the average precision of E3FCN on the whole data sets is 98.06%, compared to 73.27% of Support Vector Regression (SVR), 71.75% of Normalized Difference Vegetation Index (NDVI), and 64.41% of Enhanced Vegetation Index (EVI).