This model detects vessels of 26 feet or more in length within Level-1 Ground Range Detected (GRD) dual-polarization (VV+VH) Synthetic Aperture Radar (SAR) products from the Sentinel-1 satellite. The model will accept either a SAR product and JSON file defining the region of interest within the product, or a preprocessed image and JSON file specifying the geographic coordinates of the corners of the image. The model returns lat/lon bounding box coordinates and confidence scores. SAR imagery can be obtained through cloud cover, darkness, and haze, and this model can therefore be particularly helpful in these situations.
Many models are available for limited use in the free Modzy Basic account.
This model was trained on a dataset consisting of 2015-2017 New York Harbor ESA SAR GRD data, preprocessed and chipped into 416 by 416 pixel image chips. This model achieves a precision value of 0.49, a recall value of 0.40, and an F1 score of 0.44. This model is designed to identify vessels of 26 feet or more in length.
44% F1 Score – F1 Score is the harmonic mean of the precision and recall, with best value of 1. It measures the balance between the two metrics.
49% Precision – A higher precision score indicates that the majority of labels predicted by the model for different classes are accurate.
40% Recall – A higher recall score indicates that the model finds and predicts correct labels for the majority of the classes it is supposed to find.
This model is based on the YOLOv3 deep neural network, a method first published in 2018 by Joseph Redmon and Ali Farhadi from the University of Washington. The YOLOv3 architecture was chosen for this model due to its state-of-the-art performance in terms of speed and accuracy. This architecture is known for its ability to perform detection at different scales and is therefore able to accurately detect smaller objects.
This model was trained on a dataset consisting of 2015-2017 New York Harbor ESA SAR GRD dataset. This data was preprocessed using the pipeline described above and then chipped into 416 by 416 pixel image chips. 80 percent of the chips were used for training. The training took approximately 6 hours on 1 NVIDIA Tesla K80. The model was trained using Adam optimization with initial learning rate 0.0001 and early stopping with 3 consecutive epochs, without improvement in validation loss.
Validation was performed on 20 percent of the 2015-2017 New York Harbor ESA SAR GRD dataset, preprocessed image chips were used for validation.
The input(s) to this model must adhere to the following specifications:
This model will output the following:
Get a video demo and join the community of developers and customers building the future of Artificial Intelligence.