Mapping Aquaculture Ponds From Satellite: Pond Count, Water Area and Status

In March, before the summer crop, a feed company's regional manager asks a simple question: how many ponds in his territory will be stocked this season? His sales team has a list from last year with about two thousand names on it. Some of those ponds have been converted to paddy, some have been merged into larger ones, some new ones have been dug where there was coconut plantation two years ago, and some are sitting dry because the owner lost the last crop to disease. Nobody has walked every bund since the list was made.
The same question, in different words, reaches a bank deciding on crop loans, an insurer pricing pond cover, a fisheries department planning extension visits and an exporter trying to show where its shrimp came from. Each of them needs a current count, an area and a status for each pond, across a district rather than a farm.
A satellite cannot tell you what is in a pond. It can tell you, every few days and without a site visit, where the ponds are, how big they are, and whether each one holds water. For a district-sized question, that is most of the answer. Free data from the European Copernicus programme is enough to do it: radar from Sentinel-1, which sees through the monsoon cloud, and optical imagery from Sentinel-2, which separates water from wet soil and vegetation more cleanly when the sky is clear.
This field guide walks through the method in the order you would build it: what to ask of the data, which sensor to use for what, how to build a water map that ignores passing floods, how to separate ponds from rice fields and rivers, how to measure them, how to classify their status on each date, and how to check the result on the ground. The pipeline at the end runs on a synthetic scene with known answers, so you can see what it gets right and what it misses.
What a Satellite Can and Cannot Tell You About a Pond
The first job is to separate the questions a satellite answers well from the ones it does not. Promising the wrong ones is the fastest way to lose a customer.
Question | Answer from free satellite data | Why |
Where are the ponds, and how many? | Yes, for ponds above about a quarter of a hectare | Ponds are regular, water-filled shapes separated by bunds |
How large is each pond? | Yes, within a few percent for ponds of a hectare or more | Area comes from the pixel count; edge pixels limit accuracy |
Is the pond filled, empty or drying? | Yes, on each clear date or radar pass | Water has a distinct signature in both sensors |
When was it filled and drained? | Approximately, to within the revisit interval | A time series shows the change |
Is it stocked, and with what species? | No | Water looks the same with or without shrimp |
What is the water quality? | Only coarse hints, such as colour and algae | Dissolved oxygen, pH and ammonia need sensors in the water |
Who owns it? | No | Ownership needs land records or a survey |
A satellite counts ponds, measures them and tells you when they hold water. It cannot see shrimp, oxygen or ownership, and a product that implies it can will fail its first field check.
The useful framing for a customer is "presence, area and status, every few days, everywhere". It is enough to answer most planning questions, and it tells them where to send people for the questions only a visit can answer.
Two Sensors, Two Strengths
Two free missions do most of the work.
Property | Sentinel-1 (radar) | Sentinel-2 (optical) |
What it measures | Microwave backscatter from the surface | Reflected sunlight in 13 bands |
Pixel size used here | 10 m | 10 m for visible and near-infrared, 20 m for short-wave infrared |
Clouds and night | Sees through both | Blocked by cloud; daytime only |
How water appears | Dark: smooth water reflects the pulse away from the satellite | Low near-infrared reflectance; water indices are high |
Common confusions | Wind-roughened water looks brighter; wet smooth soil and tarmac look dark | Cloud shadow, dark soil, shaded areas |
Best use | Every pass through the monsoon; the backbone of the time series | Clean water and vegetation separation on clear days |
Sentinel-1 lost one of its two satellites in 2021, which halved its revisit for several years. The constellation is whole again: as of 24 June 2026 the two new units, Sentinel-1C and Sentinel-1D, re-established the nominal six-day repeat cycle. Sentinel-2 offers frequent optical passes over India, but coastal Andhra Pradesh in July and August can go weeks without a clear view, which is why the radar carries the time series.
Researchers have already shown that the radar alone can do the national job. A 2019 study by the German Aerospace Center mapped land-based aquaculture at the pond level for the entire coastal zone of India from Sentinel-1 time series at 10 m resolution. Studies across Asia since then combine both missions, and that combination is the approach this guide follows.
Entry 1: Define the Area, the Season and the Unit
What it is. Before any image is downloaded, decide where you are mapping, over what period, and what unit you will report: a pond, a farm, a village or a block.
How to do it. Draw the area of interest from administrative boundaries, such as mandal or block outlines, and mask out the sea and large reservoirs with a coastline and water-body layer. Set the period to match the crop calendar: in much of coastal Andhra Pradesh the main shrimp crop and the drying period between crops give a clear rhythm, and the reporting dates should fall where the decisions are made, such as just before stocking.
Decide the reporting unit early. A feed company thinks in farms and villages, an insurer in ponds, a fisheries department in blocks. The pipeline produces ponds; aggregation to the other units is a join with boundaries afterwards.
How it fails. Teams map a whole district at once and then discover their validation budget covers a village. Start with one block of a few hundred square kilometres, prove it, then scale.
Check before moving on. You should be able to say in one sentence what the customer will receive, for which area, on which dates.
Entry 2: Build a Water-Frequency Map
What it is. Instead of trusting any single image, compute for every pixel the fraction of observations in which it held water. A pond that holds water for most of the season has a high frequency. A rice field flooded for two or three weeks, or a low patch after rain, has a low one.
How to do it. For Sentinel-2, compute a water index per clear date. The normalised difference water index, NDWI, is green minus near-infrared divided by their sum; the modified version, MNDWI, replaces near-infrared with short-wave infrared and separates water from built-up land better. A threshold near zero, tuned on a few known ponds, turns each index image into a water mask. For Sentinel-1, water is the dark tail of the backscatter distribution; a threshold set from known ponds, or from the dip between the two peaks of the histogram, gives the mask. Mark clouds and cloud shadows as missing, not as dry.
Then count. Water frequency is the number of valid observations showing water divided by the number of valid observations. The India coastal study used a related idea, a per-pixel temporal median of all Sentinel-1 data, both to suppress radar speckle and to reduce confusion with temporarily inundated rice fields.
How it fails. Two ways recur. Counting cloudy dates as dry pulls every frequency down and makes real ponds disappear. And a season with an unusual flood can lift frequencies across a whole floodplain; this is why the shape test in the next entry matters as much as the frequency.
Check before moving on. Look at the frequency map over a handful of ponds you know. Working ponds should be bright for most of the season, with dark gaps only during known drying periods.
Treat cloudy dates as missing, not as dry. A water map that counts clouds as land quietly deletes ponds in exactly the season you care about.
Reading Radar Over a Pond
Radar is the backbone of the time series, so it is worth knowing what the numbers mean before trusting a threshold.
Sentinel-1 sends a microwave pulse sideways and measures how much comes back. A calm water surface behaves like a mirror: almost all the energy bounces away from the satellite, and the pixel is dark, typically well below the return from soil or crops. Vegetation, rough soil and buildings scatter energy in many directions, and some comes back, so they are brighter. That contrast is what makes ponds stand out, and it is why the bunds between them, which are dry and often grassed, show up as bright lines between dark rectangles.
Three effects complicate the picture. The first is speckle: radar images are grainy because each pixel adds up many small reflections that interfere with each other. A single image of a pond is mottled, and averaging over time or over neighbouring pixels is needed before thresholding. This is why the frequency and median approaches work better than any single date.
The second is wind. A breeze ruffles the water surface into small waves that scatter some energy back to the satellite, so a pond on a windy pass is less dark than on a calm one. Large, exposed ponds are affected most. Requiring two agreeing observations before changing a pond's status keeps a windy afternoon from turning filled ponds into empty ones.
The third is polarisation. Sentinel-1 over land usually records two channels, VV and VH, named for the orientation of the transmitted and received waves. VV generally gives the clearest contrast between water and land; VH responds more to vegetation and helps separate a flooded rice field, where stems poke through the water, from open water. Using both channels in the threshold, or at least checking VH where rice and ponds sit side by side, removes many confusions.
Radar sees water as darkness. Speckle, wind and floating vegetation all add light, so judge a pond on the season, not on one pass.
None of this needs specialist software to handle. The free Copernicus processing tools and the common open-source libraries calibrate, filter and terrain-correct the images to a standard product; the pond-specific work starts after that, with the frequency map.
Entry 3: Separate Ponds From Everything Else That Holds Water
What it is. A coastal district has many surfaces that hold water: rivers, canals, irrigation tanks, salt pans, mangrove creeks, flooded rice and temporary puddles. Ponds differ from all of them in shape.
How to do it. Turn the water-frequency map into a binary mask with a threshold such as 0.5, clean it with a small morphological opening to cut thin connections across bunds, and label each connected region. For each region compute its area, how completely it fills its bounding rectangle, and how elongated it is. Aquaculture ponds are compact, close to rectangular and separated from their neighbours by bunds a few metres wide.
Water body | Frequency | Shape | How to reject it |
Aquaculture pond | High, with drying gaps | Compact rectangle, 0.25 to 5 ha typical | Keep |
River or canal | Very high | Long and thin | Elongation above a limit |
Flooded rice field | Low, a few weeks | Irregular, large | Frequency below the threshold |
Irrigation tank or reservoir | High | Large, irregular outline | Area and rectangularity |
Salt pan | Variable | Rectangular, often very regular grids | Spectral brightness when dry; local knowledge |
Mangrove creek | High | Branching | Rectangularity |
Salt pans are the hardest confuser, because they are also rectangular, man-made water bodies. They tend to dry to a bright, white crust that optical bands pick up, and their locations are usually well known locally. Keep a separate class for them rather than forcing them into ponds.
How it fails. Ponds that share a thin bund merge into one region at 10 m resolution, and a group of small ponds can come out as one large, oddly shaped one. The opening step reduces this; a final check of very large regions against higher-resolution imagery catches the rest.
Check before moving on. Overlay the accepted regions on a recent high-resolution basemap for a few villages and count the misses and merges by eye.
Entry 4: Count and Measure
What it is. With ponds isolated, the count is the number of accepted regions and the area is the pixel count times the pixel area, 100 square metres at 10 m.
How to do it. Report area to one decimal place of a hectare and say how it was measured. Expect a small underestimate, because the pixels along each bund are a mixture of water and soil and many fall below the threshold.
The pixel size sets a floor on what can be mapped at all. The table assumes a square pond and loses one pixel all round to the bund.
Pond size | Side length in 10 m pixels | Clean interior pixels |
0.1 ha | 3.2 | About 1 |
0.25 ha | 5.0 | About 9 |
0.5 ha | 7.1 | About 26 |
1.0 ha | 10.0 | About 64 |
2.0 ha | 14.1 | About 147 |
A pond of a tenth of a hectare has about one clean pixel; no method will map it reliably at 10 m. At a quarter of a hectare there are enough pixels to detect it, though its area will be uncertain. From half a hectare up, detection and area are dependable. Small-farm districts with many ponds under half a hectare need commercial imagery at 3 m or finer for the smallest ones, and the report should say which size class each figure covers.
How it fails. The usual error is reporting a district total without a size limit, so the customer compares a satellite count of ponds above 0.25 ha with a survey list that includes every backyard tank.
Check before moving on. State the minimum pond size in every report, and compare areas with a sample of surveyed ponds.
Entry 5: Classify Status on Each Date
What it is. For each mapped pond and each date, decide whether it is filled, empty or in between. The pond outline comes from the whole season; the status comes from that day's image.
How to do it. Compute the fraction of the pond's pixels that show water on the date. Above a high threshold it is filled, below a low one it is empty, and in between it is filling, draining or partly obscured. Optical indices also separate an empty pond with wet mud from one with dry, cracked soil, and from one where the bottom is being prepared. A study in the coastal region of Guntur district evaluated spectral indices on Landsat-8 imagery for this job and found that NDWI, a water ratio index and a combination of indices separated active and empty ponds with near-perfect agreement, with a kappa above 0.9.
The sequence of statuses is where the value lies. A pond that fills in March and stays filled is likely stocked. One that drains in May after a short time filled may have lost its crop early. One that has been empty for two seasons is probably out of production.
How it fails. Shallow water with dense algae can look like vegetation in optical indices, and wind-roughened water looks brighter to radar than calm water. Both make a filled pond look empty on a single date, which is why status should change only when two consecutive observations agree.
Check before moving on. Pick ten ponds whose owners can tell you their filling and draining dates, and compare.
The Pipeline in Code
The script below runs the whole chain on a synthetic scene with known answers: 40 rectangular ponds between about 0.4 and 2.5 hectares on a 3 km square of 10 m pixels, a quarter of them drained for part of the season, a river, a rice field flooded on three of twelve dates, three percent random noise on every image and one date half covered by cloud. The core steps fit in a few lines.
import numpy as np
from scipy import ndimage
# obs: water masks per date (T, H, W); valid: False where cloud
freq = (obs & valid).sum(0) / valid.sum(0)
mask = ndimage.binary_opening(freq >= 0.5, structure=np.ones((2, 2)))
lab, n = ndimage.label(mask)
ponds = []
for i, sl in enumerate(ndimage.find_objects(lab), start=1):
region = lab[sl] == i
area_ha = region.sum() * 100 / 10000
bh, bw = region.shape
rect = region.sum() / (bh * bw) # 1.0 = perfect rectangle
elong = max(bh, bw) / max(1, min(bh, bw))
if 0.3 <= area_ha <= 5 and rect >= 0.75 and elong <= 4:
ponds.append((sl, area_ha))
# status on one date: share of each pond's pixels that hold water
filled = [sl for sl, _ in ponds if (obs[t][sl] & valid[t][sl]).mean() > 0.6]
The results on the synthetic scene are a fair picture of what to expect on real data.
Measure | Truth | Pipeline |
Ponds | 40 | 38 accepted |
Total pond area | 50.2 ha | 47.5 ha |
Other water bodies | River and rice field | Both rejected |
Ponds drained on the test date | 12 | 10 classified empty |
The pipeline found 38 of the 40 ponds and rejected the river and the rice field. It underestimated the total area by about 5 percent, almost all of it from the edge pixels that the cleaning step removes, and it missed two of the twelve drained ponds on the test date, both among the ponds it had not mapped at all. On real imagery the numbers will be worse in places with small ponds, dense algae or frequent cloud, and the validation step exists to measure by how much.
Checking the Map on the Ground
A map without an accuracy figure is an opinion. The check does not need to be expensive, but it needs to be planned.
Draw a random sample of mapped ponds and a random sample of points that the map calls "not pond", stratified by village or by size class. Visit them, or check them on recent high-resolution imagery where a visit is not practical, and record what is there. A few hundred points across a block is usually enough to report a confusion matrix: how many mapped ponds are real, how many real ponds the map missed, and how often the status was right.
Report three numbers with every map: the share of mapped ponds that were confirmed, the share of real ponds that were found, and the status agreement on the sampled dates. Add the minimum pond size and the dates of the imagery. A customer can act on those numbers; a single overall accuracy figure hides the errors that matter to them.
Publish the accuracy with the map: confirmed, found and status agreement, with the minimum pond size. A count without them invites the first field visit to discredit the whole product.
Who Uses This and What Each Needs
The same pond layer serves several customers, but each needs a different cut of it.
User | Question | What they need from the map | Update rate |
Feed and seed companies | Where will demand be this crop? | Count and area of ponds filled before stocking, by village | Monthly in season |
Banks and lenders | Is the pond real and in production? | Pond outline, area and status history for a named location | On application, then per crop |
Insurers | What is at risk and did the loss happen? | Area, filling date and draining date | Per crop, plus event checks |
Fisheries departments | Where is aquaculture expanding or declining? | Count and area by block, year on year | Seasonal |
Exporters and certifiers | Where did this lot come from? | Pond outlines and activity dates for supplier farms | Per crop |
Environmental regulators | Are ponds expanding into mangroves or wetlands? | New ponds against protected-area boundaries | Annual |
The last row matters in India's deltas. The India coastal study notes that the east coast lost significant mangrove area to pond conversion, in places such as the Godavari delta and the Sundarbans. A time series of pond outlines is the simplest evidence of where and when new ponds appeared.
Where the Method Breaks
The method is robust for ordinary earthen ponds in open country. It weakens in predictable places.
Very small ponds. Below about a quarter of a hectare, 10 m pixels cannot resolve them; use finer commercial imagery or accept a size limit.
Lined, covered or indoor systems. Biofloc tanks, polyhouses and covered ponds do not look like open water from space.
Dense algae or weed. Optical indices can classify a filled pond as vegetation; use radar and require two agreeing observations.
Wind. Strong wind roughens the surface and brightens the radar return, so a filled pond can look less like water on a windy pass.
Tree shade and steep bunds. Shadows at the pond edge reduce the clean area and bias status on small ponds.
Salt pans and newly dug ponds. Both are rectangular and may hold water briefly; local knowledge and a longer time series separate them.
None of these is a reason not to map. Each is a reason to state the limits in the report and to choose validation points where the method is weakest.
Satellite for Many Ponds, Sensors for One
The satellite layer and the pond sensors answer different questions, and the useful products combine them. From space you see every pond in a block every few days. In the water, a sensor sees dissolved oxygen, temperature and pH in one pond every few minutes, which is what a farmer needs to decide when to run aerators at three in the morning.
A practical arrangement uses the satellite layer to find and track ponds at scale, and places sensors on a sample of ponds chosen from that layer: some large, some small, some in each village cluster. The sensor ponds give ground truth for the status classification, and the satellite layer tells the agronomy team where to look when a cluster of ponds drains early.
What a Pond Layer Costs to Run
The data is free, and the compute for a block is modest: a season of Sentinel-1 and Sentinel-2 scenes for a few hundred square kilometres fits comfortably on an ordinary workstation, and cloud processing services can do the same without downloading anything. The costs are elsewhere, in people and in checking.
The table gives a rough split of effort for one block once the pipeline exists. The figures are planning estimates for a small team, not quotations.
Task | First season for a new block | Each later update | Who does it |
Area, masks and customer brief | 2 to 3 days | None | Analyst and account lead |
Data download, cloud masks, water maps | 2 days, mostly automated | A few hours, automated | Pipeline |
Threshold tuning on known ponds | 1 to 2 days | A check, under an hour | Analyst |
Validation sample and field or image check | 4 to 6 days | 1 to 2 days per season | Field staff or analyst |
Report, maps and village tables | 2 days | Half a day | Analyst |
Two lessons follow from the split. First, most of the effort is in validation and in turning the layer into something the customer reads, not in image processing. A pricing model that charges for processing and treats validation as free will lose money on every new block. Second, the cost per block falls sharply after the first season, because thresholds, masks and the validation sample carry over. A customer who commits to a full season is cheaper to serve than one who wants a single snapshot.
A First Block in Four Weeks
A first deliverable for one block can be produced in about a month by a small team.
Week 1. Agree the area, the reporting unit, the minimum pond size and the dates with the customer. Assemble boundaries, a coastline mask and a list of ponds with known filling and draining dates.
Week 2. Download the season's Sentinel-1 and Sentinel-2 data for the block, build cloud masks, compute the water masks and the frequency map, and tune the thresholds on the known ponds.
Week 3. Extract ponds, classify status on each date, and produce maps and a table by village. Draw the validation sample.
Week 4. Check the sample in the field or on high-resolution imagery, compute the three accuracy figures, adjust thresholds once, and deliver the layer with its limits written down.
After the first block, the same pipeline runs on new dates automatically, and the cost of each update is mostly the validation.
Back to the Two Thousand Names
The regional manager's list had about two thousand names. The satellite layer for his territory will not tell him which pond belongs to which customer, and it will not tell him what anyone plans to stock. It will tell him how many ponds above a quarter of a hectare exist this year, how much water area they hold, which ones filled in the last fortnight and which have sat dry for two seasons. That turns a stale list into a current map, and it sends his field staff to the villages where ponds are filling, a week before the stocking orders arrive.
The layer does not replace the field team. It tells them where to go first, and it shows the manager how his territory changes between visits.
Related service: PondWatch Aquaculture Monitoring
Related case study: Intelligent Aquaculture Hub: Edge-First GeoAI and Control for Pond Automation
Related reading: NDVI From Free Satellite Data: A Working Pipeline for Indian Crops, InSAR Explained: What Satellite Radar Can and Cannot Measure on the Ground and Which Shrimp-Pond Water-Quality Sensors Are Worth Installing
EurthTech is an engineering-first product design and embedded systems company with a development center in Hyderabad. To discuss a project, write to connect@eurthtech.com or see https://www.eurthtech.com/case-studies.




Comments