Abstract
We describe in this paper Hydra, an ensemble of convolutional neural networks(CNN) for geospatial land classification. The idea behind Hydra is to create aninitial CNN that is coarsely optimized but provides a good starting pointingfor further optimization, which will serve as the Hydra's body. Then, theobtained weights are fine tuned multiple times to form an ensemble of CNNs thatrepresent the Hydra's heads. By doing so, we were able to reduce the trainingtime while maintaining the classification performance of the ensemble. Wecreated ensembles using two state-of-the-art CNN architectures, ResNet andDenseNet, to participate in the Functional Map of the World challenge. Withthis approach, we finished the competition in third place. We also applied theproposed framework to the NWPU-RESISC45 database and achieved the best reportedperformance so far. Code and CNN models are available athttps://github.com/maups/hydra-fmow