Currently, accurate detection of natural phenomena, such as red tide, that adversely affect wildlife and human, using satellite images has been increasingly utilized. However, red tide detection on satellite images still remains a very hard task due to unpredictable nature of red tide occurrence, extreme sparsity of red tide samples, difficulties in accurate groundtruthing, etc. In this paper, we aim to tackle both the data sparsity and groundtruthing issues by primarily addressing two challenges: i) extreme data imbalance between red tide and non-red tide examples and ii) significant lack of hard examples of non-red tide that can enhance detection performance. In the proposed work, we devise a 9-layer fully convolutional network jointly optimized with two plug-in modules tailored to overcoming the two challenges: i) cascaded online hard example mining (cOHEM) to ease the data imbalance and ii) a hard negative example generator (HNG) to supplement the hard negative (non-red tide) examples. Our proposed network jointly trained with cOHEM and HNG provides state-of-the-art red tide detection accuracy on GOCI satellite images.