Lakhasly

Water, as one of the most essential resources on the Earth, not only sustains human life and ecological cycles, but also plays a vital role in economic development and mineral exploration (Vorosmarty et al., 2000).Therefore, constructing large-scale to over all water conditions, combining multi-modal and multi-temporal RS data and domain knowledge to improve the interpretation generalization in other areas and long-term perception ability would be the future research directions of water body classification.In this context, water identification aims at classifying if a pixel from RS images is water or not, which has gradually evolved into a hot research topic in RS, and scholars have carried out many studies to extract water bodies from various RS images (Hollstein et al., 2016; Dang and Li, 2021).Focusing on multiple-sensor images processing, Li et al. (2021c) presented an encoder-decoder-based dense-localfeature-compression (DLFC) network to extract valuable spatial and spectral details.How to enhance the interpretation effect with limited samples has always been a hot topic for scholars in RS. Li et al. (2021b) used a region of interest (ROI) to build water labels and then proposed a pixel-based CNN to synchronously combine spectral and texture information for water segmentation.To fully use microwave images, Xue et al. (2021) introduced a densecoordinate-feature concatenate network (DCFCN) for merging water body features from dual-polarimetric SAR images and thus address the ground interference in single-polarization SAR images. A portion of this is surface water and mainly involves rivers, lakes, canals, and ponds; the oceans are always excluded from this category due to their large extent and salty characteristic (Huang et al., 2018).Later, many improved water indexes were developed, such as the modified NDWI (MNDWI) (Xu, 2006), EWI (Yan et al., 2007), NWI (Feng, 2012), HRWI (Yao et al., 2015) and two-step TSUWI (Wu et al., 2018).Next, a restricted receptive field DeconvNet (RRF DeconvNet) was presented to compresses redundant layers and then uses an edge weighting loss to extract accurate water edges (Miao et al., 2018).With the growth of a severe global water shortage and increasing flood disasters, the efficient and accurate assessment of available surface water has become an essential part of ecological protection, urban planning, and industrial production.Over the past decades, Earth observation techniques have advanced significantly and many kinds of high-resolution RS images are accessible for various real-world applications, including efficiently identifying available water bodies.Therefore, data-driven ML- and DL based methods can adaptively leverage spectral and spatial information to build discriminative features for efficient and accurate water classification.Due to the simplicity of DT, Hollstein et al. (2016) developed several DTbased classification methods for simultaneously extracting water, snow, clouds, and so on from Sentinel 2 images.Differently, (Morsy et al., 2018) designed a unsupervised land/water classification method to automatically monitor the changes of coastal areas from multi-spectral airborne LiDAR data.Isikdogan et al. (2017) demonstrated an early application of a data-driven FCN for surface water mapping of Landsat 7 images, which was different from the threshold methods in adopting different regions and imaging conditions.

Water, as one of the most essential resources on the Earth, not only
sustains human life and ecological cycles, but also plays a vital role
in economic development and mineral exploration (Vorosmarty et al.,
2000). A portion of this is surface water and mainly involves rivers,
lakes, canals, and ponds; the oceans are always excluded from this
category due to their large extent and salty characteristic (Huang et al.,
2018). Generally, surface water bodies shrink and expand periodically
under the influence of natural or human factors (Li et al., 2022a). With
the growth of a severe global water shortage and increasing flood disasters, the efficient and accurate assessment of available surface water has
become an essential part of ecological protection, urban planning, and
industrial production. Over the past decades, Earth observation techniques have advanced significantly and many kinds of high-resolution
RS images are accessible for various real-world applications, including
efficiently identifying available water bodies. In this context, water
identification aims at classifying if a pixel from RS images is water or
not, which has gradually evolved into a hot research topic in RS, and
scholars have carried out many studies to extract water bodies from
various RS images (Hollstein et al., 2016; Dang and Li, 2021).
A water body’s reflectance is greatly lower than other GERS elements in the range from visible light to near-infrared, which is the main
mechanism for extracting water bodies from optical RS images. On this
theory basis, water indexes (or so-called threshold-based method in Li
et al. (2022a)) were proposed. They first selected some single or several
spectral bands and a suitable threshold according to the water’s spectral
curves and then identify water and non-water pixels via predefined
threshold rules. In 1996, a normalized difference water index (NDWI)
was the indicator for the assessment of water resources (McFeeters,
1996). A single infrared band was also used for producing a classification map (Frazier et al., 2000). Later, many improved water indexes
were developed, such as the modified NDWI (MNDWI) (Xu, 2006),
EWI (Yan et al., 2007), NWI (Feng, 2012), HRWI (Yao et al., 2015)
and two-step TSUWI (Wu et al., 2018). These indexes enhance the
extraction performance with various scenes and data types to a certain
extent. However, these methods, rely totally on spectral information
and are inherently rule-based and inefficient to geospatial objects’
high variations. This shortcoming of the water indexes causes terrible
boundary extraction and generalization when they are transferred to
other remote areas. Therefore, data-driven ML- and DL based methods can adaptively leverage spectral and spatial information to build
discriminative features for efficient and accurate water classification.
Due to their advantages, they have gradually become the mainstream
methods of water extraction in recent years. This part summarizes the
representative ML and DL methods in the Table 6.
Similar to lithology classification, ML classifiers were used as the
main models in the early days. Aiming at identifying urban water
bodies, a two-stage ML-based workflow was developed by Huang et al.
(2015b). It first uses several indexes to classify water bodies at the
pixel level, and then spatially geometrical and textural information is further utilized for object-level identification via ML classifiers. Due
to the simplicity of DT, Hollstein et al. (2016) developed several DTbased classification methods for simultaneously extracting water, snow,
clouds, and so on from Sentinel 2 images. Li et al. (2021a) and Bangira
et al. (2019) jointly compared several ML classifiers and threshold
methods. Their experimental results showed that different ML models
have high requirements of suitable training samples, and their overall
performances were influenced by the imaging time and regions of the
test samples. Differently, (Morsy et al., 2018) designed a unsupervised
land/water classification method to automatically monitor the changes
of coastal areas from multi-spectral airborne LiDAR data. Prosek et al.
(2020) applied SVM and K-means to compare hyper-spectral, LiDAR
and their integration data for extracting small water bodies. These
works demonstrate the effectiveness of hyper-spectral and LiDAR data.
DL plays a more critical role than ML methods in water interpretation than lithology mappings, and enormous DL studies have been
published. Isikdogan et al. (2017) demonstrated an early application
of a data-driven FCN for surface water mapping of Landsat 7 images,
which was different from the threshold methods in adopting different regions and imaging conditions. Next, a restricted receptive field
DeconvNet (RRF DeconvNet) was presented to compresses redundant
layers and then uses an edge weighting loss to extract accurate water
edges (Miao et al., 2018). Focusing on multiple-sensor images processing, Li et al. (2021c) presented an encoder–decoder-based dense-localfeature-compression (DLFC) network to extract valuable spatial and
spectral details. The DLFC models were verified on Gaofen 2, Gaofen
6, Sentinel 2, and ZY 3 images, and the overall accuracy achieved
satisfactory performance. Microwave RS is another main type of data
source for mapping surface water. The microwave sensors recording
long wavelength radiation can penetrate clouds and shallow vegetation
coverage. Therefore, the microwave sensors can work on images independent of solar radiation and thus overcome some terrible weather.
To fully use microwave images, Xue et al. (2021) introduced a densecoordinate-feature concatenate network (DCFCN) for merging water
body features from dual-polarimetric SAR images and thus address the
ground interference in single-polarization SAR images. A large number
of labeled samples have always been necessary for training DL methods,
but the labeling work is high-cost and time-consuming. How to enhance
the interpretation effect with limited samples has always been a hot
topic for scholars in RS. Li et al. (2021b) used a region of interest
(ROI) to build water labels and then proposed a pixel-based CNN
to synchronously combine spectral and texture information for water
segmentation. Dang and Li (2021) presented a multi-scale residual
network (MSResNet) in the manner of self-supervised learning (SSL)
that addressed the problems of irregular shapes of water bodies and
small annotation samples. In contrast, Abid et al. (2021) transformed
the task as an unsupervised learning process that totally eliminates
the requirements for annotation data. The main idea is to consider
the identification as a water and non-water clustering task, and the
clustering features are from the pre-trained DL models. Compared with
unsupervised learning, weakly supervised methods achieve a certain
degree of performance improvement. Lu et al. (2022a) utilize point
labels to classify water bodies via a weakly supervised neighbor feature
aggregation network (NFANet). It aggregates water’s adjacent pixels
to learn more representative features. Point labels also significantly
reduce the label cost.
Overall, water body recognition is a relatively mature interpretation
task than lithology mapping. Li et al. (2022a) compared ten methods
for water extraction in the GID data (Tong et al., 2018) and even water
index methods can achieve more than 90% pixel classification accuracy
in the challenging data. Furthermore, the data-driven approaches also
have significant boost of the overall performance, when they train
converges. However, due to the instability and limited interpretability,
especially for DL models, they do not have the remarkably transferable
performance in other areas. In the aspect of data, the existing data
are mainly in some specific regions and cannot involve the water
bodies’ characteristics in the world. Additionally, monitoring water
change for a longtime is gradually significant in water management
and flood monitoring. Therefore, constructing large-scale to over all
water conditions, combining multi-modal and multi-temporal RS data
and domain knowledge to improve the interpretation generalization
in other areas and long-term perception ability would be the future
research directions of water body classification.