1. Introduction
While inland water bodies, which account for 2.5–2.75% of all the water on Earth, constitute a small portion of the Earth’s surface, they are nevertheless essential to sustain a wide variety of aquatic species and some human activities, especially in areas where a dry climate is dominant [
1]. For a global carbon cycle, freshwater plays an important role as a sink and a source of the carbon [
2]. The legacy use of large-scale inland water bodies spans from the areas of irrigation, energy production, fishery, and transportation to satisfying our domestic needs [
3,
4]. Harsh climatic conditions threaten the ecological functioning of these water bodies [
5,
6]. The size and distribution of most freshwater lakes present a challenge for traditional onsite watershed assessment due to time, cost, and logistic constraints. As a solution to this challenge, Landsat imagery and other remotely sensed data sets have been used as effective tools to assess water quantity and quality indicators over water bodies. Mapping inland water body distribution in space and time has proven to be a cost-effective and sustainable method of management [
7,
8]. A global gridded dataset for the freshwater is a prerequisite for a water resources management and research [
9,
10,
11,
12,
13]. In Egypt, Ghana, Malawi and Tanzania, this method has been used in the study of inland water bodies, which are inaccessible and plagued with sparse datasets [
14,
15,
16,
17]. The applicability of the satellite’s wide areal coverage, re-visits, and multiple freely accessible data channels permits researchers to carry out studies of freshwater bodies in areas where the in-situ observations are sparse. In 2008, the Landsat archive was opened to the public and since then, efforts have been made to atmospherically correct the available images, making distinguishing Earth’s surface features in varying conditions easier [
18,
19]. For the practical application of remotely sensed images, water body extraction or mapping can be performed using either a hard or soft classification method. The hard classification method comprises of techniques such as the maximum likelihood supervised classification [
20], decision tree classifications and minimum distance [
21,
22,
23]. The soft classification is generally a mixed pixel decomposition method [
24,
25]. This can be achieved by using techniques such as thematic classification [
26,
27], single band threshold [
28,
29], the spectral relationship method [
30,
31] and the water index method, which is the most commonly used method because of its specific application [
27,
32,
33,
34]. These methods make use of the reflectivity index of each band and provide water body information based on signature differences between water and other land surface features. Spectral water indices are combinations of surface reflectance at two or more wavelengths that indicate the relative abundance of features of interest. These combinations can enhance water information while constraining other land-cover types as well as eliminating noise components.
Historically, Lake Chad was ranked among the largest lakes in the world. Weather changes, precipitation frequencies and poor irrigation have caused the water level to drop to about 90% of its value in 1963 [
35,
36,
37,
38,
39]. Following this decrease, extensive information on what might have caused this phenomenon is well documented and available. Coe et al. [
36] pointed out that persistent drought was the driving force behind the lake’s continuous decrease while another study tied this tremendous decrease to an increase in irrigation activities around the lake [
40]. Studies of the lake’s hydrology were carried out using altimetric measurements of water height to estimate river discharge around the Chari confluence using imperial regression techniques [
41]. To compensate for the lack of hydrological data, some researchers reconstructed the past levels and inundated areas within the basin [
42]. Annual maximum inundation in the northern pool was observed using Landsat Multispectral; Scanner (MSS) band 7 (0.8–1.1 μm) data from 1973 to 1976 and Meteosat data in the visible channel from 1977 to 1990 [
43,
44,
45,
46]. Meteosat thermal maximum composite data was used to account for water covered by aquatic vegetation and the time series of inundated area estimates for Lake Chad from 1986–2001 [
47]. Estimates of actual evapotranspiration over the Lake Chad Basin has been thoroughly covered using satellite imagery and GRACE Terrestrial Water Storage (TWS) [
48]. An in-depth study of the hydrological variation within the basin using a combination of remotely sensed and hydrological model data-sets to characterize the spatiotemporal variability and multiscale variability of the lake’s hydrological cycle showed that rainfall intensity plays a vital role in lake water availability and identified their bimodal relationship [
49]. Using multivariate regression analysis of the time series from 1988–2012 on a sub-regional level, Erik et al. [
50] employed two hydrological systems, lake levels and rainfall data, to investigate the correlations and predictive capacities of lake water levels, rainfall and temperature variability to inter-annual variations of harvested maize and millet.
The use of Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+) and Operational Land Imager (OLI) sensor in delineating the lake’s open water features and providing its area estimates are yet to be covered. Recent surface area estimates are critical to understanding the desiccation story of the basin in recent times [
51]. Cadre Harmonisé, a food security analysis tool unique to West Africa, demonstrated that this crisis has contributed to soaring food insecurity, with more than 6 million affected in the region of the four countries surrounding the lake [
52]. In the FAO’s response strategy for 2017–2019, aimed at addressing the Lake Chad Basin crisis, part of the strategic framework is to manage the natural resources as well as resource-based conflict reduction by promoting sustainable management of land use, pasture and water resources at the community level [
53]. This involves monitoring these resources at cross-border levels. The purpose of this study, therefore, is to determine the area changes of available surface water within the Lake Chad basin from 2003 and 2016 using Landsat TM, 7 ETM+ and 8 OLI imagery. GRACE-derived TWSA changes and some meteorological data from the Lake Chad Basin Commission (hereafter LCBC) were used to complement our area estimates. The resulting series of Landsat satellite data can be used to delineate water features information and to map surface water area changes in the future. In terms of shoreline protection strategies, the information gathered from the extraction of water features using remotely sensed data will be beneficial to management authorities not only in terms of finances but also consistency. Better knowledge of the water coverage duration, beginning and ending dates for the vast range of marshlands surrounding the lake is important for the measurement, modeling and management of marshland ecosystems in this area.
The specific objectives of this study are to analyze the changes in areal surface water coverage in Lake Chad using the appropriate water index derived from Landsat imagery and to investigate the link between hydroclimatic parameters and surface water extent.
We examined the aerial changes in the Lake Chad area over the 13-year period, 2003–2016. A series of 416 Landsat images acquired between 2003 and 2016 were used to investigate these changes and they were compared against GRACE TWSA and rainfall data. We tested four reflectance indices: AWEI, NDWI, MNDWI and NDVI for extraction of the lake surface area from our Landsat data. For consistency reasons, the years 2007 and 2015 were chosen as our test periods. Both years had high quality data. In terms of accuracy, MNDWI gave us the best results and was therefore used to calculate the areal extent during our study period. The changes were then statistically analyzed. This method has been covered extensively with an abundance of techniques which made use of band ratios and individual bands [
33,
54,
55].
2. Study Area
The Lake Chad Basin, whose interconnected water system includes Lake Chad (
Figure 1a), spans through Cameroon, Niger, Nigeria, Chad and the Central African Republic, providing natural resources for over 30 million people in the region [
56]. The basin is situated at an altitude higher than 300 m above the mean sea level where the Sahara Desert meets the Savannah lands, in the middle of the Sahel (
Figure 1b(i)). Between 1963 and 2013, Lake Chad, which is a freshwater lake that lies at the center of the basin, has lost about 90% of its water mass, shrinking from ~25,000 km
2 to 2500 km
2 [
57,
58]. This significant loss of water within the basin is evident in the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) in
Figure 1b(i). The lake shoreline receded from point A to point B. This reduction, which is mainly because of climate change and human factors, threatens the resources and livelihood of the more than 30 million people in the region. Numerous studies have been conducted with the aim of understanding these fluctuations and providing some sort of solution to ameliorate the Lake Chad situation. Lake levels have fluctuated considerably over the past decades [
36,
37]. Climate variability and environmental degradation were implicated as the main causes of lake level fluctuations. Previous study shows that from 1983 to 1994, water used for irrigation in this area increased fourfold compared to the levels during 1953–1979 [
36]. Another study used a hydrological model to simulate the effects of bathymetry, human interactions, and climate variability on the lake level, surface area, and water storage. They found that lake siltation and severe droughts caused the lake to split in 1972. Until the 1990s, the lake failed to merge back into a single lake during the wet seasons. The authors also highlighted irrigation withdrawals as the main cause of water depletion in this area [
38]. It is still challenging to evaluate an exact amount of existing open water within the Lake Chad Basin.
There is a crucial need to properly manage whatever amount of open water is left within the basin.
Figure 1b(ii) shows that on an average, about 55% of the lake area is covered by vegetation and marshes. Open water occupies about 10% of the lake area. A brief investigation using a false color composite image showed the barrier of land described by Gao et al. [
38] was present until the early 80s when vegetation began sprouting over the area (
Figure 1b(iii)). From the late 90s to recent times, we can see major changes in terms of vegetation spread. The two horizontal bars on the January-1987 false color composite image represent a large area mainly covered by sand which separated the Northern from the Southern parts of Lake Chad. With clogged irrigation channels and river channels blocked by siltation, water could barely reach the Northern part of Lake Chad. Bdliya et al. [
59] analyzed the causes and impacts of invasive species and sedimentation along the Yobe River and found that some of the river channels dried out and as a result, these channels which serve as the lake intake, are filled by sediments.
Large strips of land colonized by invasive species contributed to the diversion of flow away from the Lake. In the January-1999 false color composite image in
Figure 1b(iii), we can still see this split, but this time with some sparse vegetation. In January-2016, the split was covered with more vegetation and marshes. The southern part of the lake, which is mainly fed by the Chari-Logone River, is the most hydrologically active part of the lake. Several rivers in north-eastern Nigeria, including the Yobe River and its tributaries, flow into Lake Chad (
Figure 1a(i)).
The climate is dry and hot in the North but mild and humid in the South. During the dry seasons, in the past, the lake’s depth varied from about 2–3 m in the Southern pool, to about 3–6 m in the Northern pool, with an average of about 3.5 m [
58]. A recent study showed that during extreme conditions, the mean depth varies between 0.5–2 m in the Southern pool, and from 0–1.8 m in the Northern pool [
42]. A fairly linear relationship exists between the water depth and water surface area [
58]. As mentioned previously, the lake surface area dropped from ~25,000 km
2 to its current extent of ~2000 km
2. In situ lake levels measured from the Bol village station (provided by the LCBC) and altimetry lake levels from Hydroweb can be seen in
Figure 2.
Hydroweb is developed by Laboratoire d’Etudes en Geophysique et Oceanographie Spatiale (Toulouse, France). They use multiple altimetry data to provide water levels for lakes around the world (
http://hydroweb.theia-land.fr). We could not obtain lake level data from 2013 onward from the LCBC. As such, altimetry lake levels from Hydroweb were used to see what the trend has been like from that period. Despite the multiple rises and falls as seen in
Figure 2a, the lake experienced an upward trend throughout the study period. Inter-and intra-annual variations can be seen during this period, with intra-annual variations reaching ~250 for the in-situ lake levels.
Figure 2b presents the monthly in-situ lake levels. In the first half of the year, with little or no precipitation in the lake basin, the average lake level steadily decreases from January-July to about 100 cm. It then increases to about 300 cm from August-November when rainfall increases. In 2009, October, November and December had no record of lake level. A temporal linear interpolation was done between them.
6. Discussion
Lake Chad plays an important role in supporting urban development and serving millions of people across the countries it borders. Therefore, proper management of this water resource is necessary. Given its size and location, monitoring its area using freely available remote-sensing observations will help in proper management of the lake at a lesser cost.
A recent record of the Lake Chad’s area dynamics (2003–2016) was investigated using high spatial resolution (30 m) Landsat data. Due to cloud cover, some downloaded data set collected mostly during the rainy season was disregarded for this study (
Figure 3). Insufficient data may lead to estimated area uncertainties. However, we believe the observation frequency of at least two years covering the dry and raining season could help in this study. As such 2007 and 2015 were chosen based on cloud free data and reference data availability. AWEI, MNDWI, NDVI and NDWI indices were used to extract the lake area during this period. The performance of these water index can vary depending on the water type, atmospheric and land surface conditions [
27]. The results of the four indices show that the overall accuracy of water features delineation was greater than 80% and the overall kappa coefficient, greater than 0.8. In terms of accuracy, MNDWI stood out with an averaged overall accuracy of 96.3% and kappa coefficient of 0.9. MNDWI was therefore used to delineate water from non-water features for our study area. MNDWI overcomes the shortcomings of its predecessor, the NDWI by using Shortwave Infrared band to replace the Near Infrared band used in NDWI. The MNDWI maximizes the difference in spectral features. It enhances water features by restraining the non-water features within a Landsat image. Here, we presented our results along with: (1) Threshold analysis, (2) Aerial changes in the Lake Chad, and (3) Error analysis.
6.1. Threshold Analysis
The potentials of a suitable threshold method to delineate water from non-water features with high accuracy over time was tested for this study. Thresholding significantly affects the efficiency of mapping water features and it is highly influenced by the judgement of the user [
94]. With the availability of high quality reference data for accuracy assessment, an optimal repeatable threshold method can be established and verified. Water features within the lake were delineated from non-water features using the Otsu method. The generality of the Otsu method was tested by comparing its outputs to those obtained manually with the help of the gray image histogram for the year 2015. In doing so, we also tested the performance of the widely used 0-threshold and a 0.2-threshold in our study area (
Figure 7a).
A threshold of 0 will normally allow for rapid water feature delineation from other components. However, the classification accuracy at subpixel level based from Worldview-3 high resolution image showed that the Otsu and Manual threshold performed better than the 0 threshold in this area (
Figure 8). The threshold of 0 misclassified a significant amount of non-water features as water features (
Figure 8). It had an overall accuracy of 80% and kappa coefficient of 0.78. Lowest amongst all the threshold compared (
Table 6). The Otsu’s threshold of 0.41 and the manual threshold of 0.52 had an overall accuracy of 96.8% and 98.6% and a kappa coefficient of 0.93 and 0.96 respectively. Judging from
Figure 7b, there is a huge leap from a threshold of 0 to 0.41. We decided to test a threshold of 0.2 which lies between them to see if it had any significant changes in delineating water features. The 0.2 threshold had an accuracy of 96% and kappa coefficient of 0.91. The difference in accuracy between the manual threshold and all those used in this study was in its ability to identify some turbid water pixels as seen in
Figure 8. However, when using MNDWI for delineating water features for our study area, a threshold value of 0.2–0.5 will get a high degree a separability in delineating water from non-water features. The root mean square deviation between the Otsu and manual threshold was ~0.1. These experimental results show that the Otsu algorithm could be effectively used in place of manual threshold for feature segmentation of our study area.
6.2. Estimated Area and Surface Changes
Lake area estimates for our study area were obtained by applying the optimal threshold to MNDWI. As
Figure 8 shows, MNDWI was highly accurate in delineating clear water in Lake Chad and achieved an extraction accuracy greater than 95%. However, we encountered some difficulties in extracting few turbid water pixels after applying the Otsu method of segmentation. Most of these turbid water pixels were misclassified (
Figure 6 and
Figure 8). Nonetheless, this occupied less than 5% of the extracted area of the affected images.
For this study, a total of 78 monthly area covering the entire lake was generated. Generally, there is a slight uptrend in estimated lake area during our study period. The slight increase in lake area is in line with a study which showed improving wet conditions in the last two decades with 2002–2014 described as the wettest period. The authors attributed this to increasing rainfall in the area and the lack of any major long-term drought besides the usual seasonal fluctuations [
51]. Lake Chad’s average area from 2003–2016 was estimated to be 1694 km
2 with a standard deviation of about 233. The largest areas of 2087, 2182 and 2231 km
2 were recorded in October 2013, September 2014 and July 2015 respectively. This corresponds to the rainy season. The smallest areas of 1242, 1325 and 1379 km
2 were recorded in March 2003, December 2006 and February 2009 respectively. This period corresponds to the dry season. From
Figure 10a, we see that the lake area has a phase from 2003–2012 where the area changes fluctuated within the 1000 km
2 range. The average area from 2003–2012 was about 1563 km
2. From 2013–2016, area estimates entered the 2000 km
2 range with higher seasonal fluctuations. An averaged area of about 1876 km
2 was recorded for that period. Increase in lake area could be because of no major ongoing irrigation scheme going on in the Northern section of the lake as reported by the LCBC [
95]. Increase in estimated lake area could also be partially explained by the slight increase in altimetric lake levels (
Figure 2a). Between 2008 and 2014, there was a yearly increase of about 0.4 m/year in altimetry lake levels [
51]. Lake area estimates also match with an increasing agricultural productivity. This was reported by a recent study where the authors revealed that locals had experienced an increase in maize yields from 2010 upward compared to a mild harvest during the early 2000s [
50]. A possible reason for such an increase in agricultural productivity is more surface water. An increasing trend in GRACE TWSA around this area also backs the slight increase in lake area (
Figure 10a). Summarily, increased rainfall, lake levels and a recent halt in irrigation schemes on the lake are all consistent with the recent growth of the lake surface area from during our study period. Even though we downloaded images with less than 5% cloud cover, during image processing, we noticed some images were still contaminated by clouds. As such, they were disregarded for our analysis. This explains the gaps in
Figure 10a. Spatially, water maps show a permanent water body in the southern part of the lake with patches of water appearing in the northern section of the lake depending on the season (
Figure 9a).
We found that flooding events are limited to once a year during our 13-year study period. This is often during or after the rainy season when we have patches of water forming in the northern section. These water patches were considered as water features in this study. During the dry season, most of the water bodies in the northern section of the lake dries up. We recorded no direct link in terms of water out flow from the southern pool into the northern pool of the lake. Hence, the northern side of the lake still solely relies on direct rainfall or an inflow from the Yobe River. The lake bathymetry enhances water loss in this area. With water not being able to flow from the south to the north, this increases chances of evapotranspiration and seepage of the water pool in the southern part [
38]. Even though there was a slight increase in lake area during our study period, the large seasonal fluctuations reported to have occurred in the 80s [
47], still happens in recent times (
Figure 9a).
6.3. Error Analysis
An uncertainty encountered during this study was cloud cover. It reduces our data set and makes distinguishing water features difficult. Images with more than 10% cloud cover were disregarded for this study. Only water extraction derived from MNDWI was adopted for error analysis. Errors in extracting our area estimates could be from sensor properties, the lake water characteristics, and data processing. For this study, Landsat 7 ETM+ and OLI 8 images were used for water area extraction.
For the ETM+ image, a high-resolution Google Earth aerial image was used to select a reference image. This was mostly done by careful visual interpretation. This was done by comparing their water boundaries. For the OLI image, a very high-resolution multispectral image from Worldview-3 obtained from DigitalGlobe served as our reference image. Coupled with proper processing techniques, we reduced our error sources to the lakes characteristics.
The water extraction results were all acquired using the Otsu threshold. As
Table 5 and
Table 6 show, the extraction process was not flawless. Some omission errors were recorded.
Figure 5,
Figure 6 and
Figure 8 show MNDWI properly delineated water from non-water features. However, in the recognition of patches of turbid water around the lake, some omission errors are recorded. This is because, the reflectance of shallow water is usually affected by what is underneath hence affecting its spectral features. This is probably why those areas were missed during the extraction process. Even after we applied a manual threshold with reference to the Otsu threshold, we still observed some omissive errors (
Figure 8). Omissive errors around the edges were also identified. Edge pixels are likely to consist of water and non-water features [
96]. This can be addressed using spectral unmixing [
20,
97].
7. Conclusions
In this study, we estimated changes in Lake Chad’s area from Landsat 7 and 8 satellite images. MNDWI performed better than other indices for mapping the water extent from acquired Landsat images. When used with the “proper” threshold value, MNDWI can identify open water, turbid water and small water bodies which all makes part of the larger Lake Chad. This study results demonstrated that using a conventional threshold of value 0 could results in high commission errors in this area. However, when provided with reference data, optimized threshold values could be generated for large data set in a reproducible manner using the Otsu algorithm. This helped in producing high accuracy lake area extraction results in this area. Due to the difficulty in delineating some turbid water in our area, when applying MNDWI for future studies in this area, a thorough investigation of the quality of image is required. Estimated lake area generally increased from 2003–2016. From 2011–2016, the lake’s area increased by a further ~314 km2 mainly due to increasing wet conditions and less irrigation in the area. GRACE TWSA was used as a representative metric to help explain the lake area changes both seasonally and inter-annually. It could be used as a hydrological indicator for future designs of the Lake Chad’s management strategies. With the constant improvement of the Landsat and GRACE missions, the problem of limited availability of healthy data sets could be addressed by proper use of the free data set in this area.
The FAO proposed a response strategy to the water crisis in this area for the period 2017–2019. A key point they noted towards achieving this goal is to provide mapping an analysis of existing resourceful features around the Lake Chad basin and a generational analysis of land use and water utilization at cross-border levels [
53]. Water coverage durations estimated from freely available Landsat 8 imageries could be a step towards achieving this goal. Given the advancements in the OLI sensor, a thorough study using Landsat 8 images will help determine with much accuracy the water coverage seasonal duration as well as other surface features around the lake in 2017–2019.