EvaNet

Abstract

Accurate and timely mapping of flood extent from high-resolution satellite imagery plays a crucial role in disaster management such as damage assess- ment and relief activities. However, current state- of-the-art solutions are based on U-Net, which can- not segment the flood pixels accurately due to the ambiguous pixels (e.g., tree canopies, clouds) that prevent a direct judgement from only the spectral features. Thanks to the digital elevation model (DEM) data readily available from sources such as United States Geological Survey (USGS), this work explores the use of an elevation map to im- prove flood extent mapping. We propose, EvaNet, an elevation-guided segmentation model based on the encoder-decoder architecture with two novel techniques: (1) a loss function encoding the phys- ical law of gravity that if a location is flooded (resp. dry), then its adjacent locations with a lower (resp. higher) elevation must also be flooded (resp. dry); (2) a new (de)convolution operation that in- tegrates the elevation map by a location-sensitive gating mechanism to regulate how much spectral features flow through adjacent layers. Extensive experiments show that EvaNet significantly outper- forms the U-Net baselines, and works as a perfect drop-in replacement for U-Net in existing solutions to flood extent mapping. Full paper with an extended appendix is available at: (https://github.com/MTSami/EvaNet/blob/master/full_version_with_appendix.pdf)

Quantitative Results

Note: 4-Channel input refers to (RGB+Elevation) data provided as input to model, whereas 3-Channel input refers to just RGB data.

	Elevation 4-Channel Input	Elevation 4-Channel Input	U-Net 3-Channel Input	U-Net 3-Channel Input	U-Net 4-Channel Input	U-Net 4-Channel Input
	Dry	Flood	Dry	Flood	Dry	Flood
Accuracy	95.805	95.805	88.076	88.076	94.927	94.927
Precision	95.018	97.102	92.682	82.075	95.452	94.116
Recall	98.182	92.208	87.073	89.594	96.156	93.066
F1-Score	96.574	94.592	89.789	85.669	95.803	93.588

Qualitative Results

Below, the predicted flood region (red) for each model is illustrated.

Installation

Requirements

• Linux
• Anaconda Environment with Python>=3.7
• PyTorch>=1.12.1, torchvision>=0.13.1 and CUDA>=11.3

Conda environment can be created from environment.yml file:

conda env create -f environment.yml

To activate:

conda activate elev

Usage

Data Preparation

Please download dataset from https://uab.box.com/s/rjglwac0qswq13tlvi9axztqicc31lgp

Data must be organized in the following manner for all models:

model/
  --data/
    --repo/
      --FloodNetData/
        --Region_1_Features7Channel.npy
        --Region_1_labels.npy
        --...

To process data for EvaNet, run the following:

cd ./Eva_Net_4_channel/data
python data_maker.py

Train and test models

The run.ipynb (jupyter notebook) file can be used to train and test the different models.

Alternatively

Training

Please change directory to the model root directory. For example, to train EvaNet cd ./Eva_Net_4_channel then run:

python main.py --mode training
               --train_region Region_1_3_TRAIN
               --epochs 350
               --lr 1e-6
               --val_freq 1
               --saved_model_epoch 0

Inference

From model root directory run:

python main.py --mode testing
               --test_region Region_2_TEST
               --saved_model_epoch 135
               --out_dir /output
               --saved_model_dir /saved_models

NOTE:

• --out_dir is where the model predictions are saved.
• --saved_model_dir is the folder where model weights are saved.