How to Map Floodwater from Radar Imagery using Semantic Segmentation - Benchmark

by Tammy Glazer, Philip Popien


Deluge in southern Louisiana. Credit: NOAA Remote Sensing Division.

Welcome to the benchmark solution tutorial for our newest competition run in partnership with Microsoft AI for Earth and Cloud to Street.

In this challenge, you are tasked with detecting the presence of floodwater in Sentinel-1 global synthetic aperture radar (SAR) imagery. While traditional optical sensors measure reflected solar light, SAR operates in the microwave band of the electromagnetic spectrum. It is therefore uniquely positioned to provide a day-and-night supply of images of the Earth's surface through the coverage of clouds, smoke, vegetation, and precipitation. With the increasing prevalence of extreme weather events due to climate change, your machine learning models and insights have the potential to strengthen flood mapping algorithms and improve disaster risk management efforts around the world.

The training data for this competition consist of 542 "chips" from flood events around the world, which each contain two polarization bands. Each chip has a corresponding label mask that indicates which pixels in a scene contain water. You can optionally supplement this training data with information on permanent water sources and elevation, available through the Microsoft AI for Earth STAC API (see the competition STAC resources page for more information on bringing in this data). To enter this challenge, you will package and submit your code to generate predictions as single-band tifs with binary pixel values.

In this post, we'll cover:

Let's get started!

In [1]:
%load_ext watermark
In [2]:
%watermark -v
Python implementation: CPython
Python version       : 3.9.6
IPython version      : 7.26.0

In [3]:
%watermark -i

Explore the Data


First we'll download the training data to our local computer:

$ tree training_data
├── flood-training-metadata.csv
├── train_features
│   ├── awc00_vh.tif
│   ├── awc00_vv.tif
│   ├── wvy31_vh.tif
│   └── wvy31_vv.tif
└── train_labels
    ├── awc00.tif
    ├── awc01.tif
    ├── wvy30.tif
    └── wvy31.tif

2 directories, 1627 files

To better understand what's in our data, we can begin by exploring trends across different locations and flood events. Let's load the metadata and look at what we have.

In [4]:
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

# This is where our downloaded images and metadata live locally
DATA_PATH = Path("training_data")
In [6]:
train_metadata = pd.read_csv(
    DATA_PATH / "flood-training-metadata.csv", parse_dates=["scene_start"]
In [7]:
image_id chip_id flood_id polarization location scene_start
0 awc00_vh awc00 awc vh Bolivia 2018-02-15
1 awc00_vv awc00 awc vv Bolivia 2018-02-15
2 awc01_vh awc01 awc vh Bolivia 2018-02-15
3 awc01_vv awc01 awc vv Bolivia 2018-02-15
4 awc02_vh awc02 awc vh Bolivia 2018-02-15

The training data for this competition consist of two polarization bands that bring out distinct physical properties in a scene: VV (vertical transmit, vertical receive) and VH (vertical transmit, horizontal receive). Each band is identified by a unique image_id, which is equivalent to {chip_id}_{polarization}. A chip_id consists of a three letter flood_id and a two digit chip number. Each unique chip has both a _vv and _vh band. For example, awc00_vh represents the vv band of chip number 00 for flood awc.

In [8]:
(1084, 6)
In [9]:

Let's take a look at how many flood events are in the training data and how many chips we have per event.

In [10]:
flood_counts = train_metadata.groupby("flood_id")["chip_id"].nunique()
count    13.000000
mean     41.692308
std      19.947367
min      15.000000
25%      28.000000
50%      32.000000
75%      65.000000
max      69.000000
Name: chip_id, dtype: float64

The training data include chips from 13 flood events. We have anywhere from 15 to 69 chips (30 to 138 images) per unique event, with half of events containing fewer than 32 chips (64 images).

We can also take a look at the distribution of chips by location.

In [11]:
location_counts = (
In [12]:
plt.figure(figsize=(12, 4))
location_counts.plot(kind="bar", color="lightgray")
plt.ylabel("Number of Chips")
plt.title("Number of Chips by Location")
Text(0.5, 1.0, 'Number of Chips by Location')

The data cover a wide geographic range. We have more than 60 chips for floods in the United States, India, Paraguay, and Slovakia.

Let's also take a look at when the images were originally captured.

In [13]:
year = train_metadata.scene_start.dt.year
year_counts = train_metadata.groupby(year)["flood_id"].nunique()
2016    1
2017    2
2018    6
2019    3
2020    1
Name: flood_id, dtype: int64
In [14]:
awc    1
ayt    1
coz    1
hbe    1
hxu    1
jja    1
kuo    1
pxs    1
qus    1
qxb    1
tht    1
tnp    1
wvy    1
Name: scene_start, dtype: int64

The training data cover flood events that occurred between 2016 and 2020. Images for each event were captured on the same day.


Next, we can begin exploring the image data. We'll add a feature_path column to our training metadata, which will contain a Path object with the full path to a training image. We'll also add a label_path column that points to each image's corresponding water label.

Recall that each set of two polarizations (vh and vh) correspond with a single water label.

In [15]:
# Path style access for pandas -- can be installed with `pip install pandas_path`
from pandas_path import path
In [16]:
train_metadata["feature_path"] = (
    str(DATA_PATH / "train_features")
    / train_metadata.image_id.path.with_suffix(".tif").path
In [17]:
train_metadata["label_path"] = (
    str(DATA_PATH / "train_labels")
    / train_metadata.chip_id.path.with_suffix(".tif").path

Now we can take a look at some of the flood imagery!

A GeoTIFF is a raster image file that contains geographic metadata describing the location of the image. This metadata can include bounding coordinates, an affine transform, and a coordinate reference system (CRS) projection. The package rasterio makes it easy to interact with our geospatial raster data.

In [18]:
import rasterio
In [19]:
# Examine an arbitrary image
image_path = train_metadata.feature_path[0]
In [20]:
with as img:
    metadata = img.meta
    bounds = img.bounds
    data =  # read a single band
In [21]:
{'driver': 'GTiff',
 'dtype': 'float32',
 'nodata': 0.0,
 'width': 512,
 'height': 512,
 'count': 1,
 'crs': CRS.from_epsg(32720),
 'transform': Affine(10.0, 0.0, 314030.0,
        0.0, -10.0, 8585890.0)}
In [22]:
BoundingBox(left=314030.0, bottom=8580770.0, right=319150.0, top=8585890.0)
In [23]:
array([[-16.208015 , -17.71951  , -16.281353 , ...,   0.       ,
          0.       ,   0.       ],
       [-15.3288965, -18.231857 , -16.451893 , ...,   0.       ,
          0.       ,   0.       ],
       [-15.353134 , -16.88831  , -15.585904 , ...,   0.       ,
          0.       ,   0.       ],
       [-15.741662 , -15.230668 , -13.455255 , ...,   0.       ,
          0.       ,   0.       ],
       [-15.498258 , -14.100984 , -13.11027  , ...,   0.       ,
          0.       ,   0.       ],
       [-16.055603 , -14.1121   , -14.76084  , ...,   0.       ,
          0.       ,   0.       ]], dtype=float32)

We will need to be able to identify pixels with missing data, since we will only be evaluated on predictions made for valid input pixels. The metadata tells us that a value of 0.0 represents missing data for an input image. In rasterio, you can access two different kinds of missing data masks. The first mask is a GDAL-style mask, in which non-zero elements (typically 255) indicate that the corresponding data elements are valid.

In [24]:
with as img:
    gdal_mask = img.dataset_mask()
In [25]:
array([[255, 255, 255, ...,   0,   0,   0],
       [255, 255, 255, ...,   0,   0,   0],
       [255, 255, 255, ...,   0,   0,   0],
       [255, 255, 255, ...,   0,   0,   0],
       [255, 255, 255, ...,   0,   0,   0],
       [255, 255, 255, ...,   0,   0,   0]], dtype=uint8)

The second mask is a numpy masked array, which has the inverse sense: True values indicate that the corresponding data elements are invalid. To load the data as a numpy masked array and access this type of missing data mask, simply pass a masked flag to read.

In [26]:
with as img:
    numpy_mask =, masked=True)
In [27]:
  data=[[-16.20801544189453, -17.71950912475586, -16.281352996826172,
         ..., --, --, --],
        [-15.328896522521973, -18.231857299804688, -16.451892852783203,
         ..., --, --, --],
        [-15.353134155273438, -16.888309478759766, -15.585904121398926,
         ..., --, --, --],
        [-15.74166202545166, -15.230668067932129, -13.455254554748535,
         ..., --, --, --],
        [-15.498257637023926, -14.100983619689941, -13.110269546508789,
         ..., --, --, --],
        [-16.05560302734375, -14.112099647521973, -14.76084041595459,
         ..., --, --, --]],
  mask=[[False, False, False, ...,  True,  True,  True],
        [False, False, False, ...,  True,  True,  True],
        [False, False, False, ...,  True,  True,  True],
        [False, False, False, ...,  True,  True,  True],
        [False, False, False, ...,  True,  True,  True],
        [False, False, False, ...,  True,  True,  True]],

Pixel values represent energy that was reflected back to the satellite measured in decibels. To better visualize the bands or channels of Sentinel-1 images, we will create a false color composite by treating the two bands and their ratio as red, grean, and blue channels, respectively. We will prepare a few helper functions to visualize the data.

In [28]:
import warnings

In [29]:
# Helper functions for visualizing Sentinel-1 images
def scale_img(matrix):
    Returns a scaled (H, W, D) image that is visually inspectable.
    Image is linearly scaled between min_ and max_value, by channel.

        matrix (np.array): (H, W, D) image to be scaled

        np.array: Image (H, W, 3) ready for visualization
    # Set min/max values
    min_values = np.array([-23, -28, 0.2])
    max_values = np.array([0, -5, 1])

    # Reshape matrix
    w, h, d = matrix.shape
    matrix = np.reshape(matrix, [w * h, d]).astype(np.float64)

    # Scale by min/max
    matrix = (matrix - min_values[None, :]) / (
        max_values[None, :] - min_values[None, :]
    matrix = np.reshape(matrix, [w, h, d])

    # Limit values to 0/1 interval
    return matrix.clip(0, 1)

def create_false_color_composite(path_vv, path_vh):
    Returns a S1 false color composite for visualization.

        path_vv (str): path to the VV band
        path_vh (str): path to the VH band

        np.array: image (H, W, 3) ready for visualization
    # Read VV/VH bands
    with as vv:
        vv_img =
    with as vh:
        vh_img =

    # Stack arrays along the last dimension
    s1_img = np.stack((vv_img, vh_img), axis=-1)

    # Create false color composite
    img = np.zeros((512, 512, 3), dtype=np.float32)
    img[:, :, :2] = s1_img.copy()
    img[:, :, 2] = s1_img[:, :, 0] / s1_img[:, :, 1]

    return scale_img(img)

def display_random_chip(random_state):
    Plots a 3-channel representation of VV/VH polarizations as a single chip (image 1).
    Overlays a chip's corresponding water label (image 2).

        random_state (int): random seed used to select a chip

    Returns: chip and labels plotted with pyplot
    f, ax = plt.subplots(1, 2, figsize=(9, 9))

    # Select a random chip from train_metadata
    random_chip = train_metadata.chip_id.sample(random_state=random_state).values[0]
    chip_df = train_metadata[train_metadata.chip_id == random_chip]

    # Extract paths to image files
    vv_path = chip_df[chip_df.polarization == "vv"].feature_path.values[0]
    vh_path = chip_df[chip_df.polarization == "vh"].feature_path.values[0]
    label_path = chip_df.label_path.values[0]

    # Create false color composite
    s1_img = create_false_color_composite(vv_path, vh_path)

    # Visualize features
    ax[0].set_title("S1 Chip", fontsize=14)

    # Load water mask
    with as lp:
        lp_img =

    # Mask missing data and 0s for visualization
    label = == 0) | (lp_img == 255), lp_img)

    # Visualize water label
    ax[1].imshow(label, cmap="cool", alpha=1)
    ax[1].set_title("S1 Chip with Water Label", fontsize=14)


Let's inspect a few chips and their water labels.

In [30]:
In [31]:
In [32]:

Pretty interesting!

You'll notice that some images contain high floodwater coverage, while others have little to no coverage. Water spread may be concentrated in one location or spread out, depending on a variety of surface characteristics like topography, terrain, and ground cover. Missing pixels are displayed in yellow.

Finally, let's confirm that the first few training images are the expected size of 512 x 512 pixels.

In [33]:
examples = [[x]) for x in range(5)]
for image in examples:
(512, 512)
(512, 512)
(512, 512)
(512, 512)
(512, 512)

For this benchmark, we will train a first pass model using Sentinel-1 imagery only. You are encouraged to incorporate supplementary elevation and permanent water source data into your own model training using the Planetary Computer STAC API.

Split the Data

The test set for this competition may include flood events not included in the training data. We do not want to overstate our model's performance by overfitting to one or more specific events. To be sure that our method is sufficiently generalizable, we will set aside a portion of the training imagery to validate the model during its development.

Since individual events are independent, we can randomly subset the training data into a training and validation using the flood_id field. For this benchmark, we will holdout three flood events or approximately one third of available images for our validation set. You are encouraged to think carefully about the implications of different holdouts when determining how best to validate your model.

In [34]:
import random
In [35]:
random.seed(9)  # set a seed for reproducibility

# Sample 3 random floods for validation set
flood_ids = train_metadata.flood_id.unique().tolist()
val_flood_ids = random.sample(flood_ids, 3)
['pxs', 'qxb', 'jja']
In [36]:
val = train_metadata[train_metadata.flood_id.isin(val_flood_ids)]
train = train_metadata[~train_metadata.flood_id.isin(val_flood_ids)]

Since a single input to our model will contain both VV and VH bands, let's identify our input paths by chip_id.

In [37]:
# Helper function for pivoting out paths by chip
def get_paths_by_chip(image_level_df):
    Returns a chip-level dataframe with pivoted columns
    for vv_path and vh_path.

        image_level_df (pd.DataFrame): image-level dataframe

        chip_level_df (pd.DataFrame): chip-level dataframe
    paths = []
    for chip, group in image_level_df.groupby("chip_id"):
        vv_path = group[group.polarization == "vv"]["feature_path"].values[0]
        vh_path = group[group.polarization == "vh"]["feature_path"].values[0]
        paths.append([chip, vv_path, vh_path])
    return pd.DataFrame(paths, columns=["chip_id", "vv_path", "vh_path"])
In [38]:
# Separate features from labels
val_x = get_paths_by_chip(val)
val_y = val[["chip_id", "label_path"]].drop_duplicates().reset_index(drop=True)

train_x = get_paths_by_chip(train)
train_y = train[["chip_id", "label_path"]].drop_duplicates().reset_index(drop=True)
In [39]:
# Confirm approx. 1/3 of images are in the validation set
len(val_x) / (len(val_x) + len(train_x)) * 100

Build the Model

The goal of our first pass is to build a relatively simple model that takes radar imagery as input and outputs binary masks that indicate which pixels in a scene contain floodwater. Once we test this basic approach, we can attempt to improve our model by adding additional sophistication and complexity. We will use a lightweight PyTorch wrapper called PyTorch Lightning for this benchmark solution.

Rather than train an entire convolutional neural network (CNN) from scratch, we will tune a U-Net model for semantic segmentation. U-Net was first designed in 2015 to localize abnormalities in biomedical images. By applying a variety of data augmentation techniques, it can achieve high performance using relatively small training datasets. U-Net can be broadly thought of as an encoder network followed by a decoder network, where the encoder can be initialized using a pretrained backbone. For this exercise, we will initialize our encoder using ResNet34. ResNet was prepared by Microsoft Research Asia in 2015 and is pretrained on the ImageNet dataset.

First, we will need to read the training data into memory, convert the data to PyTorch tensors, and serve the data to our model in batches. Luckily, the PyTorch Dataset and DataLoader classes make implementing these complex tasks relatively straightforward. A Dataset object allows us to define custom methods for working with the data, and a DataLoader object parallelizes data loading. If you haven't worked with these classes before, we highly recommend this short tutorial.

Our custom dataset will inherit an abstract class called and override the following methods:

  • __len__(): returns the length of the dataset, measured as number of samples
  • __getitem__(): provided as index, returns a sample from the dataset

The dataset object will return samples as dictionaries with keys for:

  • chip_id: the chip id
  • chip: a two-band image tensor (VV and VH)
  • label: the label mask, if it exists
In [40]:
import torch
In [41]:
class FloodDataset(
    """Reads in images, transforms pixel values, and serves a
    dictionary containing chip ids, image tensors, and
    label masks (where available).

    def __init__(self, x_paths, y_paths=None, transforms=None): = x_paths
        self.label = y_paths
        self.transforms = transforms

    def __len__(self):
        return len(

    def __getitem__(self, idx):
        # Loads a 2-channel image from a chip-level dataframe
        img =[idx]
        with as vv:
            vv_path =
        with as vh:
            vh_path =
        x_arr = np.stack([vv_path, vh_path], axis=-1)

        # Min-max normalization
        min_norm = -77
        max_norm = 26
        x_arr = np.clip(x_arr, min_norm, max_norm)
        x_arr = (x_arr - min_norm) / (max_norm - min_norm)

        # Apply data augmentations, if provided
        if self.transforms:
            x_arr = self.transforms(image=x_arr)["image"]
        x_arr = np.transpose(x_arr, [2, 0, 1])

        # Prepare sample dictionary
        sample = {"chip_id": img.chip_id, "chip": x_arr}

        # Load label if available - training only
        if self.label is not None:
            label_path = self.label.loc[idx].label_path
            with as lp:
                y_arr =
            # Apply same data augmentations to label
            if self.transforms:
                y_arr = self.transforms(image=y_arr)["image"]
            sample["label"] = y_arr

        return sample

Our custom dataset normalizes input pixel values by applying min-max normalization with a minimum of -77 and a maximum of 26.

To prevent overfitting during training, we'll increase the size of our training data by applying a set of data augmentations to our input, including random cropping, random 90 degree rotations, and horizontal and vertical flipping. The image augmentation library albumentations is a helpful resource for this task.

In [42]:
import albumentations
In [43]:
# These transformations will be passed to our model class
training_transformations = albumentations.Compose(
        albumentations.RandomCrop(256, 256),

Next, we will create a custom class to define our training loss function and a helper function to calculate our validation metric, Intersection over Union (IOU).

For training, we will use a standard mixture of 50% cross-entropy loss and 50% dice loss, which improves learning when there are unbalanced classes. Since our images tend to contain more non-water than water pixels, this metric should be a good fit. Broadly speaking, cross-entropy loss evaluates differences between predicted and ground truth pixels and averages over all pixels, while dice loss measures overlap between predicted and ground truth pixels and divides a function of this value by the total number of pixels in both images. This custom class will inherit torch.nn.Module, which is the base class for all neural network modules in PyTorch. A lower XEDiceLoss score indicates better performance.

For validation, we will calculate IOU. As a reminder, IOU represents the size of the intersection divided by the size of the union of pixels. A higher IOU indicates better performance.

In [44]:
class XEDiceLoss(torch.nn.Module):
    Computes (0.5 * CrossEntropyLoss) + (0.5 * DiceLoss).

    def __init__(self):
        self.xe = torch.nn.CrossEntropyLoss(reduction="none")

    def forward(self, pred, true):
        valid_pixel_mask =  # valid pixel mask

        # Cross-entropy loss
        temp_true = torch.where((true == 255), 0, true)  # cast 255 to 0 temporarily
        xe_loss = self.xe(pred, temp_true)
        xe_loss = xe_loss.masked_select(valid_pixel_mask).mean()

        # Dice loss
        pred = torch.softmax(pred, dim=1)[:, 1]
        pred = pred.masked_select(valid_pixel_mask)
        true = true.masked_select(valid_pixel_mask)
        dice_loss = 1 - (2.0 * torch.sum(pred * true)) / (torch.sum(pred + true) + 1e-7)

        return (0.5 * xe_loss) + (0.5 * dice_loss)
In [45]:
def intersection_and_union(pred, true):
    Calculates intersection and union for a batch of images.

        pred (torch.Tensor): a tensor of predictions
        true (torc.Tensor): a tensor of labels

        intersection (int): total intersection of pixels
        union (int): total union of pixels
    valid_pixel_mask =  # valid pixel mask
    true = true.masked_select(valid_pixel_mask).to("cpu")
    pred = pred.masked_select(valid_pixel_mask).to("cpu")

    # Intersection and union totals
    intersection = np.logical_and(true, pred)
    union = np.logical_or(true, pred)
    return intersection.sum(), union.sum()

Given our new FloodDataset, PyTorch Lightning will allow us to train a model using minimal for loops. This library keeps the flexibility of PyTorch but removes most of the boilerplate training code, making it less error prone, cleaner to read, and easier to update. By subclassing pl.LightningModule, most of the training logic will happen for us behind the scenes. We can simply define what a forward call and training_step are, and provide our model with a train_dataloader to serve the data to our model. Behavior like checkpoint saving and early stopping are easily parameterized.

We define the following methods within our model:

  • forward: define the forward pass for an image
  • training_step: switch the model to train mode, implement the forward pass, and calculate training loss for a batch
  • validation_step: switch the model to eval mode and calculate the validation metric for a batch
  • train_dataloader: call an iterable over the training dataset for automatic batching
  • val_dataloader: call an iterable over the validation dataset for automatic batching
  • configure_optimizers: configure an optimizer and a scheduler to dynamically adjust the learning rate based on the number of epochs; this step automatically calls backward and step in each epoch
  • validation_epoch_end: calculate the validation metric for an epoch with the outputs of validation_step and reset the global tracker

We'll also add a couple of additional, optional, helper methods:

  • _prepare_model: initialize a pretrained U-Net model
  • _get_trainer_params: prepare checkpoint and early stopping behavior
  • fit: fit a pl.Trainer object for training automation
In [46]:
import pytorch_lightning as pl
import segmentation_models_pytorch as smp
In [47]:
class FloodModel(pl.LightningModule):
    def __init__(self, hparams):
        super(FloodModel, self).__init__()
        self.backbone = self.hparams.get("backbone", "resnet34")
        self.weights = self.hparams.get("weights", "imagenet")
        self.learning_rate = self.hparams.get("lr", 1e-3)
        self.max_epochs = self.hparams.get("max_epochs", 1000)
        self.min_epochs = self.hparams.get("min_epochs", 6)
        self.patience = self.hparams.get("patience", 4)
        self.num_workers = self.hparams.get("num_workers", 2)
        self.batch_size = self.hparams.get("batch_size", 32)
        self.x_train = self.hparams.get("x_train")
        self.y_train = self.hparams.get("y_train")
        self.x_val = self.hparams.get("x_val")
        self.y_val = self.hparams.get("y_val")
        self.output_path = self.hparams.get("output_path", "model-outputs")
        self.gpu = self.hparams.get("gpu", False)
        self.transform = training_transformations

        # Where final model will be saved
        self.output_path = Path.cwd() / self.output_path

        # Track validation IOU globally (reset each epoch)
        self.intersection = 0
        self.union = 0

        # Instantiate datasets, model, and trainer params
        self.train_dataset = FloodDataset(
            self.x_train, self.y_train, transforms=self.transform
        self.val_dataset = FloodDataset(self.x_val, self.y_val, transforms=None)
        self.model = self._prepare_model()
        self.trainer_params = self._get_trainer_params()

    ## Required LightningModule methods ##

    def forward(self, image):
        # Forward pass
        return self.model(image)

    def training_step(self, batch, batch_idx):
        # Switch on training mode

        # Load images and labels
        x = batch["chip"]
        y = batch["label"].long()
        if self.gpu:
            x, y = x.cuda(non_blocking=True), y.cuda(non_blocking=True)

        # Forward pass
        preds = self.forward(x)

        # Calculate training loss
        criterion = XEDiceLoss()
        xe_dice_loss = criterion(preds, y)

        # Log batch xe_dice_loss
        return xe_dice_loss

    def validation_step(self, batch, batch_idx):
        # Switch on validation mode

        # Load images and labels
        x = batch["chip"]
        y = batch["label"].long()
        if self.gpu:
            x, y = x.cuda(non_blocking=True), y.cuda(non_blocking=True)

        # Forward pass & softmax
        preds = self.forward(x)
        preds = torch.softmax(preds, dim=1)[:, 1]
        preds = (preds > 0.5) * 1

        # Calculate validation IOU (global)
        intersection, union = intersection_and_union(preds, y)
        self.intersection += intersection
        self.union += union

        # Log batch IOU
        batch_iou = intersection / union
            "iou", batch_iou, on_step=True, on_epoch=True, prog_bar=True, logger=True
        return batch_iou

    def train_dataloader(self):
        # DataLoader class for training

    def val_dataloader(self):
        # DataLoader class for validation

    def configure_optimizers(self):
        # Define optimizer
        optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learning_rate)

        # Define scheduler
        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
            optimizer, mode="max", factor=0.5, patience=self.patience
        scheduler = {
            "scheduler": scheduler,
            "interval": "epoch",
            "monitor": "val_loss",
        }  # logged value to monitor
        return [optimizer], [scheduler]

    def validation_epoch_end(self, outputs):
        # Calculate IOU at end of epoch
        epoch_iou = self.intersection / self.union

        # Reset metrics before next epoch
        self.intersection = 0
        self.union = 0

        # Log epoch validation IOU
        self.log("val_loss", epoch_iou, on_epoch=True, prog_bar=True, logger=True)
        return epoch_iou

    ## Convenience Methods ##

    def _prepare_model(self):
        unet_model = smp.Unet(
        if self.gpu:
        return unet_model

    def _get_trainer_params(self):
        # Define callback behavior
        checkpoint_callback = pl.callbacks.ModelCheckpoint(
        early_stop_callback = pl.callbacks.early_stopping.EarlyStopping(
            patience=(self.patience * 3),

        # Specify where TensorBoard logs will be saved
        self.log_path = Path.cwd() / self.hparams.get("log_path", "tensorboard-logs")
        logger = pl.loggers.TensorBoardLogger(self.log_path, name="benchmark-model")

        trainer_params = {
            "callbacks": [checkpoint_callback, early_stop_callback],
            "max_epochs": self.max_epochs,
            "min_epochs": self.min_epochs,
            "default_root_dir": self.output_path,
            "logger": logger,
            "gpus": None if not self.gpu else 1,
            "fast_dev_run": self.hparams.get("fast_dev_run", False),
            "num_sanity_val_steps": self.hparams.get("val_sanity_checks", 0),
        return trainer_params

    def fit(self):
        # Set up and fit Trainer object
        self.trainer = pl.Trainer(**self.trainer_params)

Fit the Model

Finally, it's time to fit our model. A FloodModel can be instantiated using a dictionary of hparams. The only required hparams are the training and validation data, but there are several additional hyperparameters we can specify to explore modeling strategies, including the learning rate, patience, and batch size. Consider experimenting with different models and using hyperparameter tuning to find the best combination of parameters to improve performance.

Once we specify our hparams, we can simply call the fit method to begin training!

In [48]:
hparams = {
    # Required hparams
    "x_train": train_x,
    "x_val": val_x,
    "y_train": train_y,
    "y_val": val_y,
    # Optional hparams
    "backbone": "resnet34",
    "weights": "imagenet",
    "lr": 1e-3,
    "min_epochs": 6,
    "max_epochs": 1000,
    "patience": 4,
    "batch_size": 32,
    "num_workers": 0,
    "val_sanity_checks": 0,
    "fast_dev_run": False,
    "output_path": "model-outputs",
    "log_path": "tensorboard_logs",
    "gpu": torch.cuda.is_available(),
In [ ]:
flood_model = FloodModel(hparams=hparams)

PyTorch Lightning lets you log PyTorch models and metrics into a directory for visualization within the TensorBoard UI. TensorBoard is a machine learning visualization toolkit that's helpful for tracking metrics across batches, epochs, and models. You can install TensorBoard through the command line to visualize your logged data by specifying the root log directory and running --logdir=tensorboard_logs/benchmark-model.

In [50]:
tensor(0.3069, device='cuda:0')

Our best validation IOU is 0.3069.

Prepare for Submission

Now that our model is trained, we can finally prepare our submission. For this competition, your inference code will be executed to generate a score. This means that we'll need to save our model, write inference code, package everything up, and upload to the competition submissions page.

Let's walk through the process. Below we will:

  1. Create a local benchmark-pytorch/ directory with an benchmark-pytorch/assets subdirectory to store our trained model and weights
  2. Write a file to load our model and weights, perform inference on images stored in data/test_features, generate chip-level predictions, and save predictions to submission/ (see an example file here)
  3. zip the contents of benchmark-pytorch/ (not the directory itself) into a file called
  4. Upload to the competition submissions page; the file will be unzipped and will be run in a containerized execution environment to obtain our model's IOU

Note: If you'd like to look at this benchmark submission already completed, you can see the end product at the runtime repo.

First, let's make the benchmark-pytorch/ directory and its benchmark-pytorch/assets subdirectory.

In [51]:
submission_path = Path("benchmark-pytorch")
submission_assets_path = submission_path / "assets"

Next, let's save our model and weights.

In Pytorch, the learnable parameters of a model including its weights and biases are contained in it's parameters. A state_dict is a Python dictionary object that maps each layer to its parameter tensor. PyTorch provides a handy method for serializing our model's state_dict and, in turn, saving our model's weights.

In [52]:
weight_path = submission_assets_path / "", weight_path)

To load our trained model in the competition runtime environment, we'll save a simplified FloodModel class to benchmark-pytorch/ This model class will instantiate a generic U-Net model and will contain a helper method to perform inference for a given chip.

In [53]:
%%file benchmark-pytorch/
import numpy as np
import pytorch_lightning as pl
import rasterio
import segmentation_models_pytorch as smp
import torch

class FloodModel(pl.LightningModule):
    def __init__(self):
        self.model = smp.Unet(

    def forward(self, image):
        # Forward pass
        return self.model(image)

    def predict(self, vv_path, vh_path):
        # Switch on evaluation mode

        # Create a 2-channel image
        with as vv:
            vv_img =
        with as vh:
            vh_img =
        x_arr = np.stack([vv_img, vh_img], axis=-1)

        # Min-max normalization
        min_norm = -77
        max_norm = 26
        x_arr = np.clip(x_arr, min_norm, max_norm)
        x_arr = (x_arr - min_norm) / (max_norm - min_norm)

        # Transpose
        x_arr = np.transpose(x_arr, [2, 0, 1])
        x_arr = np.expand_dims(x_arr, axis=0)

        # Perform inference
        preds = self.forward(torch.from_numpy(x_arr))
        preds = torch.softmax(preds, dim=1)[:, 1]
        preds = (preds > 0.5) * 1
        return preds.detach().numpy().squeeze().squeeze()
Overwriting benchmark-pytorch/

The next step is to generate our script in the root of our benchmark-pytorch/ directory. The code submission format page contains helpful guidance for creating This file assumes that FloodModel can be imported from

In [54]:
%%file benchmark-pytorch/
import os
from pathlib import Path

from loguru import logger
import numpy as np
from tifffile import imwrite
from tqdm import tqdm
import torch
import typer

from flood_model import FloodModel

ROOT_DIRECTORY = Path("/codeexecution")

# Make sure the smp loader can find our torch assets because we don't have internet!
os.environ["TORCH_HOME"] = str(ASSETS_DIRECTORY / "torch")

def make_prediction(chip_id, model):
    Given a chip_id, read in the vv/vh bands and predict a water mask.

        chip_id (str): test chip id

        output_prediction (arr): prediction as a numpy array
    """"Starting inference.")
        vv_path = INPUT_IMAGES_DIRECTORY / f"{chip_id}_vv.tif"
        vh_path = INPUT_IMAGES_DIRECTORY / f"{chip_id}_vh.tif"
        output_prediction = model.predict(vv_path, vh_path)
    except Exception as e:
        logger.error(f"No bands found for {chip_id}. {e}")
    return output_prediction

def get_expected_chip_ids():
    Use the test features directory to see which images are expected.
    paths = INPUT_IMAGES_DIRECTORY.glob("*.tif")
    # Return one chip id per two bands (VV/VH)
    ids = list(sorted(set(path.stem.split("_")[0] for path in paths)))
    return ids

def main():
    For each set of two input bands, generate an output file
    using the `make_predictions` function.
    """"Loading model")
    # Explicitly set where we expect smp to load the saved resnet from just to be sure
    torch.hub.set_dir(ASSETS_DIRECTORY / "torch/hub")
    model = FloodModel()
    model.load_state_dict(torch.load(ASSETS_DIRECTORY / ""))"Finding chip IDs")
    chip_ids = get_expected_chip_ids()
    if not chip_ids:
        typer.echo("No input images found!")
        raise typer.Exit(code=1)"Found {len(chip_ids)} test chip_ids. Generating predictions.")
    for chip_id in tqdm(chip_ids, miniters=25):
        output_path = SUBMISSION_DIRECTORY / f"{chip_id}.tif"
        output_data = make_prediction(chip_id, model).astype(np.uint8)
        imwrite(output_path, output_data, dtype=np.uint8)

    logger.success(f"Inference complete.")

if __name__ == "__main__":
Overwriting benchmark-pytorch/

We also need to copy the model that segmentation-models-pytorch saved to our local computer. The container doesn't have access to the Internet, so you are responsible for packaging up any local assets and pointing your code to them. In the below, you will see that we set the TORCH_HOME environment variable to look at where our submission gets decompressed instead of its default (usually something like ~/.cache/torch). See the pytorch.hub docs for more on this.

In [55]:
!cp -R ~/.cache/torch benchmark-pytorch/assets/

Now our submission directory looks like this:

In [56]:
!tree benchmark-pytorch/
├── assets
│   ├──
│   └── torch
│       └── hub
│           └── checkpoints
│               └── resnet34-333f7ec4.pth

4 directories, 4 files

Finally, let's zip everything up, submit it, and get our score!

In [57]:
# Remember to avoid including the inference dir itself
!cd benchmark-pytorch && zip -r ../ *
updating: assets/ (stored 0%)
updating: assets/torch/ (stored 0%)
updating: assets/torch/hub/ (stored 0%)
updating: assets/torch/hub/checkpoints/ (stored 0%)
updating: assets/torch/hub/checkpoints/resnet34-333f7ec4.pth (deflated 7%)
updating: assets/ (deflated 7%)
updating: (deflated 58%)
updating: (deflated 57%)
In [58]:
!du -h

Upload Submission

We can now head over to the competition submissions page to submit our code and obtain our model's IOU!

Our submission takes about 20 minutes to execute. We should see an IOU of 0.6255 on the leaderboard. That's a great start, but now it's to you to improve this score. We hope this benchmark solution provides a helpful framework for exploring the data, designing a machine learning pipeline, training a deep learning model, and preparing your code for submission.

Head over to the STAC Overflow challenge homepage to get started. We can't wait to see what you build!