An Active Learning Model for Seagrass Detection in Remote Senseing Imagery

Abstract

In this paper, we developed an active learning model for seagrass detection in Landsat-8 satellite images. We first trained a patch-based deep convolutional neural network (DCNN) model using domain experts labeled region of interest (ROIs) in the images as ground truth. We then applied the trained DCNN model to a whole image to produce a pixelwise classification. Finally, we sampled large patches in the classified image as ground truth to train high-resolution network (HRNet), a semantic segmentation model, for robust seagrass detection. In addition, we optimized the hyperparameters in the HRNet model to achieve better performance. Experiments on the Landsat-8 images show that HRNet is more robust than the DCNN for seagrass detection since the larger input patches used in the HRNet contain more context information than the smaller input patches used in the DCNN. The context information enables a trained HRNet to directly be applied to time series images for seagrass detection collected at the same location without model finetuning.