Lakes and rivers are the most accessible inland water resources available for ecosystems and human consumption [1
] and they are valued for their ability to store floodwaters, protect shorelines, improve water quality, and recharge groundwater aquifers [2
]. Therefore, the information of the long-time lake water storage variations is fundamental for understanding the impact of climate change and human activities on the water resources [4
]. For these reasons, many responsible organizations operate a number of inland water level stations to collect information for water resources management [1
]. However, still many remotely located lakes have not been gaged, especially in developing countries [8
Because of the use of advance techniques (e.g., remote sensing), the number of gaging stations have decreased in recent years around the globe [5
]. Remote sensing technology for monitoring changes is widely used in different applications such as land use/cover change [12
], disaster monitoring [14
], forest and vegetation changes [16
], urban sprawl [18
], and hydrology [20
]. The knowledge about water resources can be efficiently improved by the use of remote sensing which include radar, microwave, infrared, and visible sensors. Among the mentioned remote sensing methods, microwave remote sensing provides a unique capability for mapping inundation area and delineate water boundaries over large areas of the Earth’s land surface [5
]. The exploitation of satellite data about water bodies provide reliable information for the assessment of present and future water resources, climate models, agriculture suitability, river dynamics, wetland inventory, watershed analysis, surface water survey and management, flood mapping, and environment monitoring, which are critical for sustainable management of water resources on the Earth [1
Lake surface areas (especially closed lakes) [1
] are sensitive to natural changes and thus may serve as significant proxies for variations in regional environmental and fluctuations in global climate [26
]. Changes in the areal extent of lake surface water may occur due to various factors, including the progressive unveiling of the lake basin by sediments, climate change, tectonic activity causing uplift or subsidence, and the development of drainage faults [27
Satellite remote sensing for the analysis of water volume variation has been used very often in the related literature [31
]. Birkett [32
], Frappart et al. [33
] and Crétaux et al. [11
] have used successfully satellite radar altimetry to derive water levels of water bodies [8
]. Duan et al. [8
] estimated water volume variations in Lake Mead from four satellite altimetry and imagery datasets [31
]. Recently, Baup et al. [34
] combined high-resolution satellite images and altimetry to estimate the volume changes of the lakes that are mainly used for irrigation in France [31
]. Surface water volume changes derived from the combination of altimetry and imagery were removed from the total water storage anomaly estimated using observations from the GRACE gravimetry from space mission to estimate soil water content variations [35
Yan et al. [38
] detected the dynamic changes in surface areas of Lake Qinghai using Landsat TM/ETM+ images based on the model which relies on the fact that water bodies appear dark in middle and near infrared bands [1
]. These studies provide reliable estimation of lake fluctuation in water levels and areas, which is of great significance for water resources management under the background of climate change [1
]. However, these studies have focused on the detection and analysis of variations in surface extent and water level in response to either climate change or human-environment interactions [1
Among the models proposed for water feature extraction from satellite data, multi-band ratio models are the most popular models for surface water extraction [27
]. Komeili et al. [41
] developed NDWI-PCs model based on Normalized Difference Water Index (NDWI) and Principal Component Analysis (PCA) for extracting water features from Landsat TM, ETM+, and OLI imagery over Lake Urmia. Xu [42
] developed a modified NDWI (MNDWI) in which the middle infrared (MIR) band was replaced with the near infrared (NIR) band in order to decrease false positive from built-up lands. Ouma et al. [43
] developed a Water Index (WI) model using Tasseled Cap Wetness (TCW) index and NDWI for Landsat TM and ETM+ imagery over Rift Valley lakes in Kenya [27
The other proposed models include semi-automated change detection approaches [27
], e.g., single band density slicing and Maximum Likelihood (MXL) model presented by Frazier et al. [44
], conceptual clustering technique and dynamic thresholding proposed by Soh et al. [45
], supervised classification model proposed by Alecu et al. [46
], multivariate regression method [47
] and automated spectral-shape procedure presented by Yang et al. [48
]. The last but not least proposed algorithms are automatic extraction methods, for instance, Support Vector Machine (SVM) and Spectral Angle Mapper (SAM) models presented by Jawak et al. [49
], object oriented multi-resolution segmentation model developed by Shao et al. [50
], and fuzzy intra-cluster distance within the Bayesian algorithm proposed by Jeon et al. [51
]. To the best of the authors’ knowledge, a detailed analysis of Artificial Neural Networks (ANNs) models for automatic surface water extraction has not yet been presented in the related literature, whereas ANNs models can be very competitive in terms of accuracy and speed for image classification.
Starting from these motivations, the purpose of the present paper is to demonstrate: (1) the potential of ANNs approach for a fast, robust, accurate and automated water feature extraction without using any ancillary data; and (2) the detection and analysis of variations in surface extent and water level in response to both climate change and human-environment interactions using multisensory remote sensing techniques.
In order to examine the robustness of the algorithm, the result of the proposed model has been compared with the results of different satellite-derived indexes including Normalized Difference Water Index (NDWI) [41
], Modified Normalized Difference Water Index (MNDWI) [42
], Water Ratio Index (WRI) [41
], Normalized Difference Vegetation Index (NDVI) [41
], Automated Water Extraction Index (AWEI) [41
], and Normalized Difference Water Index (NDWI) and Principal Component Analysis (NDWI-PCs) [41
] models, which are the latest models published in the related literature about extraction of surface water from satellite imagery. The datasets (for the period 1975–2015) prepared over Urmia Lake, Lake Sevan, and Van Lake (situated in similar geographical regions) have been used in the Experimental Section because the mentioned lakes are under intensive natural and human driving forces.
The paper starts with a description of our study areas in Section 2
. Section 3
describes complementary datasets that have been used in this paper. In Section 4
, our approach and algorithms to extract and analyse the surface water extent of the lakes using satellite imagery has been shown. The proposed model, which has been used for deriving the lakes water levels and lakes surface areas from satellite data, is presented in Section 5
. The results from satellite altimetry and satellite imagery, together with local climate data, are used in the same section for discussing about the impact of factors such as climate change and anthropogenic activities on the drying up of the studied lakes. Section 6
will summarize and conclude the results of this study.
5. Results and Discussion
In order to detect the surface area changes of the lakes in the period 1975–2015, the water surface of each lake in each temporal image was extracted using MLP classifier. In the MLP classifier, 100,000 pixels (extracted from the Landsat dataset) have been used for training/testing the net. The training sets contain 60% of the data, and the test sets contain the remaining 40%. Pixel selection for the training/test set has been performed randomly and repeated six times.
After the several attempts to properly select the number of units in the hidden layers, architecture 4-10-4 has been finally chosen for its good performance in terms of classification accuracy, Root-Mean Square Error (RMSE), and training time. In total, 35,000 training cycles were sufficient to train the network. The inputs of the net consist of Landsat data in spectral bands 480, 560, 660, and 825 nm, and the output provides the pixel classification in terms of water body, urban area, bare lands, and green lands. One MLP classifier has been used in classifying all the images (Figure 5
An accuracy assessment has been carried out in order to assess the classifiers more appropriately. For each of the 27 mosaicked images, 20,000 pixels (which is 6% of each image) have randomly been selected and then labelled by visual interpretation. The same procedure has been used to calculate the accuracy of the MNDWI, AWEI, NDVI, NDWI, WRI, and NDWI-PCs classifiers for each of the 27 mosaicked images.
For visual interpretation, Landsat ETM+ data in spectral bands 560 nm, 660 nm, and 825 nm have been used in RGB format. The main reason for using this combination is the high contrast of water and dry/land areas (due to the high absorption and reflectance of NIR (825 nm) by water and the terrestrial vegetation and dry soil, respectively) in NIR band [31
The accuracy of the whole dataset classified by MLP ANNs is 95.52% with a standard deviation of 2.00%, 3.88% and 4.47% commission (the samples which are committed to the wrong class) and omission (the samples which are omitted from the right class) errors and average, respectively. The average accuracy computed for the NDWI-PCs (which generates the best results after MLP NNs) is 86.10% with a standard deviation of 2.47%, 7.67% and 13.90% commission and omission errors and average, respectively. An improvement of 9.42% in accuracy has been obtained on the dataset classified by MLP NNs with respect to the same dataset classified by the NDWI-PCs model. Moreover, MLP classifier generated less Commission error with respect to the NDWI-PCs model. The results of accuracy assessment applied to different models are displayed in Table 1
and Table 2
The outputs of the MLP NNs classifier have been overlaid to produce the surface water changes (for each five year) in time-series started from 1975 to 2015 (Figure 6
). The results show that the Urmia Lake surface area was ~4724.69 km2
, ~4111.12 km2
, ~3184.73 km2
, and ~1642.71 km2
in 2000, 2005, 2010, and 2015, respectively (Table 3
The Urmia Lake surface area has decreased ~613.57 km2 between 2000 and 2005, ~926.39 km2 between 2005 and 2010, and ~1542.02 km2 between 2010 and 2015, while the Lake Sevan and Van Lake surface areas have increased ~14.69 km2 and ~16.15 km2 between 2000 and 2005, 9.93 km2 and ~5.99 km2 between 2005 and 2010, and 3.57 km2 and ~1.37 km2 between 2010 and 2015, respectively. The most intense changes in Urmia Lake is detected between 2010 and 2015, during which the lake lost ~65.23% of its surface area in comparison with the year 2000 and 48.41% of its surface area in comparison with the year 2010.
In order to analyse the time series of height above reference surface variations of lakes extracted from radar altimetry data, two models have been used. In the first model, the time series were treated as a whole under the hypothesis that the time series has a decreasing (blue lines in Figure 7
) trend in Urmia Lake and an increasing trend in Sevan and Van Lakes (Figure 7
). In the second model, by applying PELT algorithm, the time series have been divided into segments (black lines in Figure 7
) with its own statistical characteristics that are similar within each subseries and different between subseries. It seems that the increasing mono-trend (red lines in Figure 7
) fitted to the whole time series can have different behaviour when multiple inner trends are taken into account.
The generated results revealed a significant change in the surface area of Urmia Lake. These changes confirm different natural and human-made external driving forces in the watershed area. Investigations on the Urmia Lake water level breakpoints on 2000 and 2008 shows that the lake had experienced rapid changes in its history. These rapid changes could refer to intensive dam construction on one the hand and intensive and extensive cultivation activities by increasing the irrigation land on the other hand (Figure 8
and Figure 9
). The above-mentioned changes from 2000 to 2015 have been confirmed by the surface area generated using MLP NNs classifier.
Moreover, monitoring the lake’s long term changes (Figure 6
) and comparing them with long term anthropogenic activities (e.g., land use change, land over use, dam construction and urbanization as it has been shown in Figure 8
) in the watershed area of the lake indicate that the unsustainable land management has been significantly impact in drying up the Urmia Lake as well.
The results also show that with increasing the intensive human activities in different periods, the drying trend increased. For instance, there are 51 dams in Urmia Lake basin which have been constructed to supply irrigation demands (Figure 9
). Moreover, based on the authors knowledge, there are 224 projects (72 reservoir dams, 124 weirs and conduction facilities, 17 pumping stations and 10 flood controlling and artificial feeding) under study, 231 of which were assessed to be constructed in the near future [62
]. The under construction projects regulate 1499.9 MCM water (under study projects regulate 657.2 MCM water). Therefore, total regulated water volume will be 3869.1 MCM within approximately 20 years [62
shows the most important projects in this area. Long term climate change and perturbation enhanced the negative effects of mismanagement and caused critical condition for the Urmia Lake in the resent years. As it has been shown in Figure 10
, the annual average precipitation on Urmia Lake basin was 382.4 mm from 1966 to 1990, which has been decreased to 315.5 mm from 1991 to 2015. The temperature average over Urmia Lake basin was nearly 11.6 °C from 1966 to 1990 that has reached an average of 12 °C from 1991 to 2015.
In this study, a multisensory, multitemporal remote sensing approach has been used to monitor water level and storage variations of Urmia Lake, Lake Sevan, and Van Lake. In order to examine the proposed model, these three study areas have been selected because they are under intensive natural and human driving forces. Landsat TM, ETM+ and OLI multitemporal images and Topex/Poseidon, Jason-1 and 2, and GFO satellite data have been used together with climate data to identify lake parameters (water level variations) and separate them from other land cover types (surface water extraction). The results show that, the Urmia Lake’s water level and area decrease at a significant rate, which is dramatically high in comparison with Lake Sevan and Van Lake.
Furthermore, an approach based on multilayer perceptron neural networks algorithm has been introduced for surface water change detection, which shows high performance in simultaneously detecting the surface water changes in comparison with the other state of the art models presented in the related literature. In conclusion, the proposed model (as a global model which can be applied to another datasets just with some parameter adjustments based on the type of data) has been proven to be effective in detecting the water surface changes in different lakes. The construction of the NN model can be a time consuming process (which is considered as the weakness of the model in comparison with the other algorithms) since building up the NN architecture is synonymous to a strenuous activity involving trial and error.
Then, the impact of factors such as climate change (e.g., rain variation) and anthropogenic activities (e.g., dam construction and water overuse) have been demonstrated. The results show that despite the long-term transformation of the environment by human activities as well as climate change in the watershed area of the three lakes, Lake Urmia is in critical situation and urgent action is needed for the lake to survive. It can be concluded that construction of dams in Urmia Lake basin was not the main factor in declining the lake level but the drying up of Urmia Lake has been occurred due to a chain of reasons, which are highly influenced by anthropogenic activities and climate change.
Managing water supply and irrigation, strict water rights, and modifying farming to conserve water and averting new dam construction in the basin are suggested to help Urmia Lake to make a recovery. However, continuous time series of temperature, precipitation and other meteorological observations and estimations (e.g., evaporation and aridification) on the lake and in the watershed, as well as of the discharge of surface water to the lake would help to better constrain the water balance.