INTRODUCTION
As a city-dweller residing for most parts in the tropical wetland that is Kolkata, I've always found the city to be water-positive: the South-Westerlies bring copious quantities of rainfall, the murky Hooghly river never ceases to appear menacing, the vegetation is evergreen, and the soft soil seems to emanate a distinct watery-smell throughout the year. Yet, underneath this seeming abundance lies a worrying prospect - the city's groundwater levels are receding.
HYPERLINKED SECTIONS
BACKGROUND
For all the misery that a Drought brings (it is the most costly weather-event after a Cyclone), it is rather innocuously defined as 'drier-than-normal conditions that results in water-related problems'.
While the phenomenon can largely be attributed to natural causes such as (lack of) Precipitation and (dry) Seasonality, humans are exacerbating it through Global Warming and broader Climate Change.
Having read National Geographic's succinct explanation as to why tracking the onset of drought is complicated, I can draw parallels to my study on deformation at volcanic sites as this phenomena also develops gradually.
Droughts can be categorized as-
Meteorological - When precipitation is below average for an extended period of time
Hydrological - When the water level in natural or man-made storage sites such as aquifers, lakes and reservoirs fall below a threshold
Agricultural/Ecological - When poor farming practises or excessive heat or low precipitation reduces soil moisture availability for crops
The repercussions from droughts can be wide-ranging, from destruction of crops and vegetation, famine, migration, soil erosion and resulting wildfires to loss of biodiversity, increase in water pollution, scarcity of water for irrigation & urban consumption and low power generation at dams. There is actually a category of drought known as socio-economic drought which occurs when there is a shortage of produce that is dependent on water supply i.e. when the demand for products such as grains and vegetables exceeds the supply.
'Does a Drought end when it rains?'
Not really. The continuity of rains matter. Essentially, the precipitation needs to soak the soil and replenish the groundwater for the drought conditions to subside and for vegetation to reappear.
Developing & underdeveloped economies are most vulnerable to drought due to their high dependence on agriculture, as can be ascertained from UNCCD's report. I was taken aback to know the drought conditions in India - nearly 50% of land area in India was affected by it as of 2019 (30% as of 2023) and in the five years from 2016-17 to 2021-22, an astounding 35 million hectares of farmlands have been damaged. To give you a perspective of how large this figure is, it is roughly 22% of India's total arable land! (160 million hectares as estimated in 2011). The vagaries of weather - too much rainfall at certain locations and too little at others - is particularly brutal.
All is not lost however: Gujarat, a state at the receiving end of prolonged dry weather, has taken substantial measures to monitor and combat drought and can be a role model across the country. Globally, I notice the emphasis on developing Satellite-based Early Warning Systems to detect drought conditions as well as on practical methods such as irrigation planning, rainwater harvesting, cultivating resistant crops, sustainable farming and even innovative techniques such as cloud-seeding to combat it.
DROUGHT MONITORING USING REMOTE SENSING
I have used Sentinel-2 Multispectral Satellite Imagery to monitor the (hydrological) drought conditions at Indirasagar Dam's Reservoir in Madhya Pradesh, India over the last four years (2019-2022). The reservoir is the largest in the country in terms of area and water storage capacity and the water is utilized extensively for irrigation purposes across the state which is experiencing increasing mean annual temperatures making it susceptible to drought.
The video below contains an elaborate walkthrough demonstrating the processing chain involved in Satellite Imagery Analytics for Drought Monitoring - from downloading the multispectral datasets, processing it on an imagery analytics tool, and mapping it out using GIS applications-
Prior to monitoring drought conditions at Indirasagar, I had opted to study Theewaterskloof Dam's reservoir in South Africa, the time period being the same (2019-2022). The latter was, for me, an extension to the tutorial which monitored drought conditions at this location for five years until 2020.
Theewaterskloof, which stores and supplies water to the city of Cape Town, was affected by severe drought from 2016 - 2018 and water levels were estimated to have receded by as much as 89%. I will include snippets from that study in this post as well.
a. TIMELINE OF IMAGERY DATASETS
For the Indirasagar study, I had selected two imagery datasets from each of the four years, acquired during summer and monsoon respectively.
For the Theewaterskloof study, I had selected three imagery datasets from each of the four years, two acquired during summer and one during monsoon (seasons are reversed in the Southern Hemisphere).
For some strange reason, it took me a while to acknowledge that drought conditions can exist during the monsoons (less or no rainfall).
Maintaining consistency in terms of the timeline felt appropriate - not just in the selection of the month but also in the selection of the day from the month. For example, one can reasonably expect May 15th - 31st to be reasonably warmer that May 1st - 15th in an Indian summer. To compare a May 28th image to a May 2nd image from the previous year wouldn't be a like-for-like comparison: the water body would be expected to be drier (more evaporation), all else remaining constant.
Also, while Sentinel-2 datasets are available cyclically throughout the year, the tricky aspect is the presence of clouds in the imagery - it is very important that the cloud cover over the study area in particular (reservoir) is negligible as they hinder the analysis of surface features due to lack of adequate reflectance data. I had to be mindful of this and had to move to another date range if there were no cloud-free images in the desired date-range. For the Indirasagar study, most of the datasets which I had selected were relatively cloud-free (<1% cloud cover for most and 2-3% occasionally).
Just to clarify, free-to-use earth observation satellites such as Sentinel-2 travel on a fixed path and offer standard-sized coverage (based on swath width). If the study area is very large or does not lie within a single pass of the satellite, one may need to stitch two or more imagery datasets together prior to processing. In contrast, if the study area lies within the imagery, one prefers to clip the geographic extent of the image as it helps in reducing the size of the imagery (Sentinel-2 dataset is ~1 GB per image) allowing subsequent processing steps to run quicker.
Technically known as subset, clipping is not restricted to geographic extent alone - multispectral datasets capture surface reflectances across multiple ranges of wavelengths (called spectral bands) and one can choose to remove the bands that are not relevant to the study. Band subset has the same effect as Geographic subset in that the size of the image is reduced which saves up on processing time.
The polygon on the left in Figure 6 depicts the geographic extent of a Sentinel-2 multispectral dataset - you can observe the water body which lies within - it is the Indirasagar Reservoir, our area of interest. Notice the clouds that lie south-west in the satellite view on the right of Figure 6. While the cloud cover percentage in the image is miniscule (<1 % as mentioned in the product page), had that chunk of clouds been directly over the reservoir, I would be hesitant to select this dataset for the study.
This product (Figure 7) is not useful at all though-
b. PROCESSING THE IMAGERY ON SNAP
I was fortunate that I could use Level-2A Sentinel-2 Imagery products for my study instead of Level-1C products. This is because Level-2A products are being generated relatively recently (December 2018) and hence, were available for my entire study period from 2019 to 2022. These products are atmospherically corrected i.e. they contain Bottom of Atmosphere (BoA) reflectance values instead of Top of Atmosphere (ToA) reflectance values.
The algorithm generated BoA reflectance eliminates atmospheric noise which reaches the satellite sensor i.e. the image resembles more like how the sunlight has been reflected just off the earth's surface instead of resembling how it appears after the reflected sunlight has passed through the atmosphere.
BoA reflectance is useful because we are measuring Ground surface properties here (water extent) and not Atmospheric properties (for example, pollution). Converting Level-1C product to Level-2A product would normally entail deploying the Sen2Cor processing algorithm which is a time-consuming process. Having Level-2A directly from the imagery distribution service helped me save considerable processing time. I suspect if one were to monitor drought using the 'Evotranspiration' method i.e. measuring the water vapour content in the atmosphere, then Level-1C products from the applicable satellite service provider would have been sufficient to do the analysis.
To analyse and visualize drought monitoring from the Level-2A products, I have used three applications - SNAP, QGIS & ArcGIS Pro. SNAP imagery processing software was predominantly used for the main analysis (identifying water pixels) whereas the latter two GIS applications (Geographic Information System) helped to refine, organize and chart the extracted imagery output onto a map.
Much of what is done in the post-processing phase is elaborately explained in my video walkthrough. I'd recommend you to watch it if you are interested to see the workings involved. Read on to get a detailed summary of the steps and the final output.
Exploring the Product on SNAP
I visualized the raw satellite imagery product in different color-modes to begin with-
An optical imagery product such as the one i am using contains several spectral 'bands' of information - the manipulation of which helps us to ascertain, study, monitor or measure the characteristics of the surface which we had intended to discover.
Spectral bands contains reflectance information in varying resolution and central wavelengths. Understand the basics of spectral bands here.
Besides spectral bands, Atmospheric correction information such as cloud pixels is also stored in separate masks within the imagery product.
Processing on SNAP
There are three broad processing steps which I've performed on SNAP for drought monitoring purposes-
In Phase 1 (Refining), the imagery product needs to be 'Resampled' first so that all the bands have the same resolution. This is because the software is only able to run certain processing commands if there is an equivalence in resolution across all the bands used during calculation.
Thereafter, we 'Subset' the product. This process entails discarding unnecessary information such as extra geographic area from the product and extra bands of information. The former is very pertinent in Theewaterskloof processing as you'll realize from Figure 10 that the reservoir forms just a tiny portion of the total imagery product. To process the entire 1GB imagery product would be a waste of computing resources. Also, if we are sure which bands are to be used during our processing (as we are in this particular case), we can remove the undesirable bands to make the product size even lower.
Phase 2 (Analysis) is the most important step in SNAP, across all the three applications rather.
In this phase, we manipulate the bands of information to isolate the pixels which are 'water' from the imagery.
In hydrological drought monitoring, what we are essentially setting out to do is to see how the reservoir area changes over a period of time. If the reservoir shrinks by a significant portion, it is indicative of drought. Therefore, in this step, we use the band combination methods (water radiometric indices) proposed by researchers to identify the water pixels.
The four radiometric indices which we've used for this study are:
a) NDWI - Normalized Difference Water Index
b) MNDWI - Modified Normalized Water Index
c) MDWI+5 - Revised Modified Normalized Vegetation Index
d) AWEI - Automated Water Extraction Index
These indices are either newly devised by researchers or they are an advanced version of an earlier method or sometimes even a method suited during certain scenarios: eg. in a specific terrain type - urban environment or a vegetated land.
The way to input these indices on SNAP is like a formula in mathematics. For example NDWI is 'B3 band minus B8 band divided by B3 band plus B8 band'. (B3 has a Central Wavelength of 560 nanometers whilst B8 has a Central Wavelength of 842 nanometers)
As you'll infer from the figure above, the reflectance percentage of water drastically diminishes from the wavelength 600 nanometers (0.6 micrometer * 1000) and beyond while soil & green vegetation's reflectance tends to increase and move further away from that of water at this juncture. Thus, the surface reflectance properties of these materials show a clear divergence at this particular wavelength and beyond and NDWI formula exploits this peculiarity to isolate water pixels from non-water pixels. Cool, isn't it?
In Phase 3 (Extraction), we take the output of all the four radiometric indices and ask SNAP to select only those pixels which are 'water pixels across all the indices' i.e. select the best of all worlds. This helps us to be assured that all the shortlisted pixels (10m each in size) are very likely to be water.
Thereafter, we create a separate 'Water Mask' layer from our shortlisted pixel selection and export it to our computer as a GeoTIFF file. These GeoTIFFs will be inputted in QGIS application later for further processing.
SNAP is a powerful, free-to-use application. For those who may not want to refine the imagery product on a GIS platform, they can still do quite a lot of interesting analysis and comparison from within SNAP application itself. For example, we can compute the area of the water pixels in the imagery by creating a Water Mask and computing the Mask's Area.
Or, we can do some Time Series Analysis as well. Refer to the video below where we are are comparing the pixel values of the NDWI band across all the 8 images on a graph simultaneously upon hovering the cursor over a product.
Video 2: Demo of Time Series Analysis tool on SNAP
Question for your to ponder: Why is there is a clear cyclicity observed in the readings on the graph on the bottom-left of the video?
Processing on QGIS
I'd have preferred to do the ensuing processing steps, which I had performed on QGIS, on SNAP itself. Moving data from one platform to the other is a little tedious and one has to know the basics of operating different applications as well. That doesn't do away with the fact that QGIS is a very powerful GIS application - suitable to do a variety of mapping workflows. The best aspect is that it is open-source and free-to-use making it a favorite among beginners as well as professionals alike.
In QGIS, I performed data transformation and extracted water pixels simultaneously for all the images in the timeline.
By data transformation, I mean converting Raster form of imagery output into a Vector form (Vector vs Raster explained). In a Raster image, each pixel has a data value but then there are no shapes acting as containers for clusters of similar data values. Converting the imagery it into a vector format helps address our requirement of having distinct shapes in the image containing similar pixels within: a reservoir, in geospatial terms, is shaped like a polygon within which lie only water pixels.
So what we are doing in QGIS with this data transformation is to create two distinct types of shapes from the raster imagery data - that of clusters of water pixels and of clusters of non-water pixels.
Thereafter, we discard the non-water clusters because they is not useful for us in the analysis-
Notice that there are water shapes outside the reservoir area as well.
We'll proceed to refine our output and perform analysis and visualization steps next on ArcGIS Pro. This could have been done on QGIS itself, however, I am more comfortable using ArcGIS Pro to perform these type of activities.
Processing on ArcGIS Pro
Esri's ArcGIS Pro is a premium, paid application and is useful to perform a wide variety GIS mapping, location analytics and imagery analytics workflows. It is user-friendly and a trial version is also available. The software also has several dedicated extensions for performing specialized workflows specific to certain sectors and industries. For example: solar, airports, military, urban planning among others. The workflows performed on SNAP earlier can also be performed on ArcGIS Pro itself by using its Image Analyst extension.
Watching my video walkthrough would be the ideal if you'd like to know the processing steps on ArcGIS in detail as there are several minor steps which are difficult to outline in the article itself.
Explore the slider below and try to Q) guess this Processing step done on ArcGIS Pro-
Slider 1: Before - After comparison of an important processing step in ArcGIS Pro
A: We've refined the output to do away with several small water pools surrounding the reservoir from our original imagery product which we had exported from QGIS as a vector shapefile.
Why have we done so?
A: Because we do not want to overestimate the area computation of the reservoir - we'd like to restrict it to the main reservoir as much as we can.
So what's the way to do so?
A: Ideally, we can locate the shapefile for the reservoir from within ArcGIS' repository (from ArcGIS Online or from Living Atlas) or from other online sources on the Web.
Or we can create one ourselves using a handy technique, as I've done for this Indirasagar Drought Monitoring case - Because I had converted the raster product into vector format, hundreds of water shapes were created across all the eight sets of imagery. It is easy to know that the largest shape in the attribute table belongs to that of the reservoir. Therefore, using the Buffer tool I created an outline around the largest shape (from across all the eight years) and then used it as a base in the Clip tool to eliminate noise from all the eight imagery layers. The output of a single layer clipped is depicted as the image on the right in the slider above.
Besides cleaning up and organizing the data further, deriving 'Geometry' attributes (Area & Perimeter) using the Merge tool, converting Area in Sq. m to Sq. km using 'Calculate field' function on the attribute table and adding new fields such as Date & Season, ArcGIS Pro is particularly useful when it comes to 'Symbolizing' the layer.
We can visualize the output on the map in multiple ways using shapes, transparency, colors etc. In a way, that's the essence of map-making - how to present the information so that the desired message is highlighted and conveyed to to the reader.
The map output above shows the four outlines of Theewaterskloof's reservoir during one summer day from 2019 to 2022. It is a nice way to visually observe whether there was drought and if so, which portions of the reservoir receded the most.
In Theewaterskloof's case, it becomes evident that it was not affected by drought as the years went by, rather it was emerging from the drought it had faced between 2016-18 and was showing improvement in water levels almost year-on-year.
Besides the visual component, one can also analyze the map's data using dynamic charts, statistic generators and infographs-
From Indirasagar Reservoir's chart above, it becomes very clear that there was a significant decline in water levels in the period between 2021 monsoon and 2022 summer when compared to the three previous years, thereby signalling a period of drought to me.
The charts below validate and further our understanding on the drought conditions for both our study areas.
ArcGIS Pro finally helps me to combine all the useful data points into a beautiful map which conveys the main message / learning / inference which I derived from the study-
I hope you found this article to be interesting and appreciated the importance of water as a precious resource. Feel free to drop in your feedback on projects@mapmyops.com or on the YouTube video's comment section.
Much Thanks to RUS Copernicus for the tutorial & ArcGIS Pro for the trial software.
ABOUT US
Intelloc Mapping Services | Mapmyops.com is based in Kolkata, India and engages in providing Mapping solutions that can be integrated with Operations Planning, Design and Audit workflows. These include but are not limited to - Drone Services, Subsurface Mapping Services, Location Analytics & App Development, Supply Chain Services & Remote Sensing Services. The services can be rendered pan-India, some even globally, and will aid an organization to meet its stated objectives especially pertaining to Operational Excellence, Cost Reduction, Sustainability and Growth.
Broadly, our area of expertise can be split into two categories - Geographic Mapping and Operations Mapping. The Infographic below highlights our capabilities.
Our 'Mapping for Operations'-themed workflow demonstrations can be accessed from the firm's Website / YouTube Channel and an overview can be obtained from this flyer. Happy to address queries and respond to documented requirements. Custom Demonstration, Training & Trials are facilitated only on a paid-basis. Looking forward to being of service.
Regards,