Assessing Deep Learning Models Trained on Public versus Commercial Data using FloodPlanet, a High-Resolution Commercial Imagery Flood Dataset, for Inundation Detection
Abstract
Floods cause more damage than any other disaster. Mapping floods from public and commercial, optical, and radar satellites improve response, relief, and mitigation efforts to save money, lives, and property. The advancement in deep learning and increased spatial resolution of commercial satellites like Planetscope provide an opportunity to improve flood detection. However, the lack of training and validation data is a major obstacle to deep learning in flood detection.
The newly proposed FloodPlanet dataset reduces this barrier. FloodPlanet is a multi-sensor, co-located, labeled, spatial-temporal, commercial SmallSat based dataset to test, train, and validate deep learning algorithms for surface water detection. The dataset contains high-resolution labels of surface water based on PlanetScope commercial data for 18 flood events across the globe, along with associated Harmonized Landsat Sentinel-2 (HLS) or Sentinel-1 imagery. To provide a baseline for this dataset, we modified U-Net and proposed an attention-boosted convolutional neural network model for water segmentation, which was trained and tested on the dataset. To demonstrate the advantage of adopting deep learning in flood detection, we compared the proposed model's ability to outperform the current NDWI thresholding-based method to extract surface water from Planetscope data. The experiment result shows that deep learning outperforms Otsu on NDWI by 16% in Intersection Over Union (85% vs 69%). Two extra experiments were designed to assess the value of the commercial SmallSat dataset to train deep learning algorithms for surface water detection. We first compare the performance difference of the proposed model trained on labeled public data with their associated public label and FloodPlanet. We then train the proposed model using FloodPlanet's high-resolution label on public data to see the performance difference from using FloodPlanet only. All assessments are based on common performance metrics such as Intersection Over Union, F1 score, and Total Operating Characteristic. We took advantage of the increased spatial and temporal resolution of commercial satellites to produce high-resolution labels, aiming to improve surface water detection using publicly available data during flood events through advanced deep learning approaches.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMGC13A..05Z