reviewed input data and excluded tiles with annotation errors.slight modifications for road network extraction For flood prediction, train a siamese U-Net with ResNet-50 encoder.For buildings and roads, train a U-Net with EffecientNet-V2 encoder.use additional training data from spacenet-2 (buildings), spacenet-3 (roads), spacenet-5 (roads), and xBD dataset (for flood).same postprocessing as baseline - hyperparameters were adjusted. UPerNet for the building and flood network decoder. swin-transformer backbone (initialized with ImageNet-22K weights).three independent models were used (building segmentation, road segmentation, and flood segmentation model).spacenet-2 buildings, spacenet-3 roads, massachusetts buildings and roads, inria aerial image labeling dataset buildings. pretrain with additional training data.Binary cross entropy loss to classify each whole image as 'flooded' or 'not-flooded'. use focal and dice loss for segmentating flood and non-flood.if 50% of buildings in image are flooded, all instances of building are flooded. heuristic for flood attribution: if 30% of roads in image are flooded, all instances are assigned flood.siamese U-Net (initialized using the pretrained U-Net weights) and train on spacenet-8 for flood and non-flood data.finetune the pretrained U-Net on spacenet-8 pre-event imagery and reference data.pretrain U-Net (ResNet-50 encoder, initialized with ImageNet weights) on additional training data from spacenet-2 (buildings), -3 (roads), and -5 (roads) to predict roads and buildings.binary cross entropy loss for building segmentation, regional mutual information loss for road speed and flood segementation.data augmentation during training (motion blur, gaussian noise, hue and saturation shift, random brighness and contrast, random gamma).You can find more detailed writeup in each of the subdirectories The goal of SpaceNet 8 is to leverage the existing repository of datasets and algorithms from SpaceNet Challenges 1-7 ( ) and apply them to a real-world disaster response scenario, expanding to multiclass feature extraction and characterization for flooded roads and buildings and predicting road speed. To help address this need, the SpaceNet 8 Flood Detection Challenge will focus on infrastructure and flood mapping related to hurricanes and heavy rains that cause route obstructions and significant damage. As these events become more frequent and severe, there is an increasing need to rapidly develop maps and analyze the scale of destruction to better direct resources and first responders. Winning Solutions and Baseline for SpaceNet 8 Flood Detection ChallengeĮach year, natural disasters such as hurricanes, tornadoes, earthquakes and floods significantly damage infrastructure and result in loss of life, property and billions of dollars.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |