This repo contains the code and data to reproduce the work published in the paper "Giezendanner et al. , Inferring the past: a combined CNN--LSTM deep learning framework to fuse satellites for historical inundation mapping, CVPR 23 Earthvision workshop", as well as the inferred dataset.

Use the following citation when these data or model are used:

Giezendanner, J.; Mukherjee, R.; Purri, M.; Thomas, M.; Mauerman, M.; Islam, A. K. M. S.; Tellman, B. Inferring the Past: A Combined CNN-LSTM Deep Learning Framework to Fuse Satellites for Historical Inundation Mapping. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June, 2023. https://openaccess.thecvf.com/content/CVPR2023W/EarthVision/html/Giezendanner_Inferring_the_Past_A_Combined_CNN-LSTM_Deep_Learning_Framework_To_CVPRW_2023_paper.html

Inferred dataset

Fractional Inundated Area for Bangladesh, 2001-2022, 500 meters resolution, every 8 days

The model output dataset contains 985 .tiff files covering most of Bangladesh every 8 days, at 500 meters resolution, from 2001 to 2022. The dataset can be found at 10.25739/2edm-jh03.

Model and data

Data

The data for training (cross-validation) and inference can be found at 10.25739/be28-vs34.

The following data is available:

• Fractional inundated area generated from Sentinel-1 binary map
• MODIS Terra 8-days composite
• DEM and slope derived from FABDEM upscaled to 500 meters
• HAND upscaled to 500 meters (MERIT Hydro)

The data is expressed in 32x32 500 meter resolution chips.

Code

The code is organised as follow:

• Source/0_ModelTraining contains the code to train and cross-validate the model.
• Source/1_Inference contains the code to run inference on the trained model.
• Source/Helpers contains diverse helper functions
• Source/ModelClasses contains the classes of the model

Both the code for the CNN-LSTM proposed in the paper, as well as the trained weights (see release) are provided.