Water
is the vital natural resource for human survival and development, as well as an
important restriction factor of the Eco-environment. Accurate quantitative
water body recognition is crucial to many applications including environmental
monitoring, resource survey, flood assessment, and drought detection. The
presented research addresses a pervasive crucial concern, volumetric detection
of water bodies.
Introduction
Accurate
information on surface water is important for assessing the role it plays in
urban ecosystem services in the context of human survival, climate change, and
hydrological cycle. Surface water refers to water on the surface of the Earth,
such as a river, lake, wetland, and the ocean. Usually, the ocean is excluded
in the definition due to its size and because it is salty, though smaller
saline water bodies are usually included. Surface water bodies are critical
freshwater resources, for both human and ecological systems. They are of paramount
importance in sustaining all forms of life. Water helps preserve the
biodiversity in riparian or wetland ecosystems by providing habitats to a
plethora of flora and fauna. It is not only critical to the ecosystems as a key
component of the hydrologic cycle but also touches every aspect of our lives,
such as drinking water, agriculture, electricity production, transportation,
and industrial purposes.
Surface
water bodies are dynamic as they shrink, expand, or change their appearance or
course of flow with time, owing to different natural and human-induced factors.
Variations in water bodies impact others’ natural resources and human assets
and further influence the environment. Change in surface water volume usually
causes serious consequences. The changes in urban water bodies make a huge
difference to human lives and may cause disasters, such as surface subsidence,
urban inland inundation, and health problems. In extreme cases, a rapid
increase in surface water can result in flooding. Therefore, it is crucial to
know about urban water distribution and changes in the water area to
efficiently detect the existence of surface water, to extract its extent, to
quantify its volume, and to monitor its dynamics. The spatial and temporal
change pattern of the surface water has important practical significance and
scientific value for water resources management, biodiversity, emergency
response, and global climate change. Also, the precise extraction of surface
water bodies is of great significance for urban planning, socio-economic
development, urban environmental testing, urban heat-island effects, and urban
ecosystem maintenance.
Overview of the
Methodology
In
recent years, satellite remote-sensing technology has developed rapidly and has
the characteristics of a wide observation range, short return period, and so
on. It has been widely used in many fields such as military reconnaissance,
environmental protection, mapping, and geography. Among current urban
water-extraction technologies, a mainstream method uses remote-sensing imagery
to gather urban water information in a timely and accurate way. Previous urban
water-resource surveys have been based on low- and medium-resolution images.
However, small water bodies such as small ponds and narrow rivers cannot be
extracted due to the limited spatial resolution of these remote-sensing images.
With the improvement of the spatial resolution of remote-sensing images, many
remote-sensing satellites (such as WorldView-2, IKONOS, and rapid eye) can
provide high-resolution images. Most high-resolution remote-sensing images only
have four bands (blue, green, red, and near-infrared), lacking the short-wave
infrared (SWIR) data necessary to compute the modified normalized difference
water index (MNDWI) and the automated water extraction index (AWEI) indices.
A
high-resolution spatial multi-spectral image has more detailed spatial features
information, which can greatly improve the accuracy of urban water body
extraction. Many algorithms have been proposed for identifying water bodies
with remote-sensing imagery including single-band threshold and multi-band
threshold methods, water body index methods, sub-pixel water mapping methods,
and supervised and unsupervised classification methods. The water body index
method has the characteristics of fast calculation and high precision, so it is
widely used in practical applications. Modified Normalized Difference Water
Index (MNDWI) which uses mid-infrared bands for normalization instead of
near-infrared and green bands, and has better results for urban water body
extraction. These improvements in the water index are generally difficult to
apply in low-resolution remote-sensing images due to limited spectral
resolution. Image classification methods such as supervised or machine learning
are often used to extract water bodies from remote-sensing images.
Generally,
machine-learning methods include neural network and support vector machines,
and unsupervised classification methods include k-means clustering and ISODATA
clustering methods. The above algorithms are mainly used on low spatial
resolution remote-sensing images. The existing algorithms have undergone less
research for urban water body extraction in high-resolution satellite images. At present, the main
problem for extracting an urban water body by low spectral resolution
remote-sensing images is the ability to distinguish between the building
shadows and the water bodies which is one of the most difficult tasks. Deep
learning is the learning process that simulates the human brain. It can
automatically extract high-level features from low-level features of the input
image. For this project, a novel method for the extraction of urban water
bodies based on deep learning is considered for high spatial resolution
multispectral images. A new convolutional neural network (CNN)architecture is
designed that can extract water and detect building shadows effectively even in
complex circumstances and predict the superpixel as one of two classes
including water and no water.
The
major contributions are:
1. A novel extraction method
for urban water bodies based on deep learning is proposed for
remote-sensing images. The proposed method combines the superpixel method with
deep learning to extract urban water bodies and distinguish shadow from water.
2. A new CNN architecture is
designed, which can learn the characteristics of water bodies from the input
data.
3. To reduce the loss of image
features during the process of pooling, we included self-adaptive pooling (SAP)
Technical
Approach
Using
satellite images with resolutions better than 0.5m (GSD), an image segmentation
model can be built. Image segmentation using ‘DeepConvolutional Neural
Networks’ (D-CNN) is the method of classification of each pixel into predefined
categories. For water body identification, an image-segmentation model can be
developed, having two categories for each pixel, i.e. “Water” & “Others”
classes. Finally, a pixel count and grouping mechanism can be developed which
would count the ‘Water’ class pixels and hence calculate the total surface area
on earth covered by water. An estimate of the water quantity can be guessed by
combining the information related to the water body, such as average depth, the
total volume of water present can be estimated.
Conclusion
In
this project, a novel water body extraction method based on deep learning is
proposed for high-resolution remote-sensing images. The proposed method
combines an enhanced superpixel method with deep learning to extract urban
water bodies and distinguishes between shadow pixels and water pixels. This
study concludes that the proposed deep-learning methods can significantly
improve urban surface water detection accuracy for the high-resolution
remote-sensing imager.