High-Resolution Global Inland Surface Water Monitoring using PlanetScope
Data and Supervised Learning with Bootstrapped Noisy Labels
Abstract
High-resolution mapping and monitoring of global inland surface water
bodies are critical to address challenges in sustainable water
management practices. Planet currently operates the largest
constellation of Earth Observation satellites and collects images at
very high spatial (0.5 m - 5 m) and temporal (near-daily) resolutions.
Here, we use PlanetScope data (resampled to 3 m) to develop a holistic
and fully automated pipeline running on the Google Cloud Platform for
monitoring global inland surface water. We incorporate the
openly-available Global Reservoir and Dam (GRanD) data set into a
three-stage supervised learning approach which initiates with an
unsupervised label-generation step consisting of k-means clustering and
NIR-based thresholding. We then rank the labels generated from these
steps and the water labels extracted from the latest ESRI 10 m land
cover data based on image contours. The best (noisy) labels having the
least number of contours from this unsupervised learning stage are
bootstrapped to train a deep-learning based semantic segmentation model
(U-Net) on a KubeFlow pipeline. We subsequently create a new refined
dataset by using these model predictions as labels which are passed to a
Stochastic Gradient Descent (SGD)-based multi-temporal supervised label
refinement stage (SGD classifier running on the same label for multiple
input scenes). Finally, we iterate over the SGD based-supervised and
U-Net-based label refinement steps to successively denoise the
bootstrapped data until we obtain an acceptable test accuracy
(F1 score > 0.9). Visual inspection of the
results obtained over different climatic regions, terrains, and seasons
across the globe shows that our approach works quite well. We also
aggregate these predictions to detect temporal changes in surface water
area. However, the model predictions exhibit high uncertainty in
agricultural areas and complex terrains characterized by hill shadows
and clouds. This issue could potentially be mitigated using
hard-negative mining. Nevertheless, with the nearly-daily imaging
capability of Planet, the high-fidelity surface water maps developed
using this proposed supervised learning approach could be beneficial to
the global water community for dealing with water security issues as
part of the UN sustainable development goals.