Methodology

For each source series the watch takes the change in depth from its May reading to its latest reading of the same year, and compares that with the median of the same mandal’s May-to-that-month change over the previous ten years. These are seasonal depth changes, not direct recharge volumes. Station composition, gaps and changes in monitoring practice can affect comparisons.

A mandal is flagged short only when it misses its own normal by at least 1 m and by at least twice its own year-to-year spread. Either test alone fails in an opposite way: metres alone flag every naturally swinging hard-rock mandal, and spread alone flags a delta mandal that moved twenty centimetres more than usual. The spread is a median absolute deviation rather than a standard deviation, because one displaced well inflates a standard deviation enough to hide the very season it should flag. Mandals with fewer than seven comparable years are left out rather than compared against a normal that is not yet established.

Rainfall context is CHIRPS at 0.05° averaged over each mandal polygon, walked back to 1981 — 45 years and thirteen El Niño events, which is what makes the rainfall composite worth reporting at all. Since October 2026 the whole record is CHIRPS v3, rebuilt from 1981 rather than spliced: v2 production ends after December 2026, and v3 corrects gauges for wind undercatch and uses about four times the station sources, so the same month reads differently. The state’s own rain gauges are shown beside it for the current water year; for June to August 2026 they read drier still. The ocean state is NOAA’s Oceanic Niño Index.

ONI groups rainfall-history composites; it does not generate local groundwater predictions. Adding it to the three-month forecast was tested on rolling origin and made the forecast worse — 1.776 m to 1.850 m mean error, and worse in every aquifer. The groundwater forecast experiment remains research-only. Historical ENSO associations do not establish a local forecast or causation.

Checked against IMD’s own rainfall grid. The India Meteorological Department’s 0.25° grid (Pai et al. 2014), the rainfall officials quote, agrees with the CHIRPS record the model reads: across 604 mandals from 2014 to 2025 the June to September totals match on average (CHIRPS at 1.00 of IMD), month by month they move together (r = 0.87), and a wet or a dry season reads the same way in 78% of mandal-seasons. As the model’s rain, IMD would cut the nowcast’s error by about 2% (1.060 m to 1.041 m over 2024 and 2025, better in 18 of 24 months). IMD publishes a year’s grid only after the year ends, while CHIRPS posts each month in about three weeks, and the weekly nowcast needs this season’s rain, so CHIRPS stays the input.

When the three-month forecast can and cannot be trusted in a dry year. Retrained every quarter since 2018 and scored on what followed, forecasts whose three months turned out at least 25% drier than normal during the monsoon expected about 1.2 m more recharge than happened (2.26 m mean error against 1.81 m for assuming no change). That happens when the forecast is made before the season fails, as from June for September: nothing measured then says the monsoon will fail. Forecasts made from October, once the season is over, carry no such bias (between −0.3 m and +0.6 m). Shrinking toward no change and correcting by past bias were both tested; neither improved every period, so the released forecast is unchanged. The model was also re-run on the State’s official mandal outlines: error moved by under 0.3% (1.797 m against 1.800 m), so its inputs are unchanged.

Each released forecast says how such forecasts have fared. The same backtest, split by the month a forecast is made and by whether the rain over the three months before it ran short, near normal or above, shows where the forecast earns its place: made in June it beat assuming no change by 11–21% whatever the rain; made in July, after a dry or ordinary spell (24% and 16%) but not after a wet one; made in August or September, after an ordinary spell (19% and 12%), and about as well as no change after a dry or wet one. Every forecast carries the verdict of its own cell, with the error behind it, and “lower confidence” where it has not beaten no change. A rule built on El Niño at the forecast date was tested first and rejected: forecasts made while the Pacific was already warm beat no change as often as any others. In October 2026 the rainfall record was rebuilt from 1981 so that mandals drawn in more than one part average rain over every part (seven mandals moved by 1.8–3.1%); retrained on it, the forecast’s error was 1.794 m. Later that month the months the State carried forward in 2021 were taken out as missing (see the caveats below): over the backtest since 2018 the released forecast’s error fell from 2.22 m to 2.17 m; over the months since 2024 it rose from 1.73 m to 1.81 m, still well under assuming no change (2.22 m).

For each prototype mandal the water watch asks three questions and counts the yes answers. Is its groundwater flagged short of its own seasonal normal (the Monsoon Watch rule above, from measured wells)? Is its gauge rainfall for the water year 20% or more below the department’s normal (measured)? Is its soil moisture at 30 cm among the driest quarter of years for the same calendar day (modelled)? A source with no usable value for the mandal is left out of the count, never counted as a yes.

Three yeses mean three sources point the same way, which is a reason to check that mandal first. It is not a drought declaration, a crop-loss estimate or a ranking of need. The soil model is driven by rainfall, so the rain and soil answers are not independent; groundwater is. Each source is joined to its boundary only when the join is unique: a boundary two source rows claim gets no value rather than the first or the worst.

The projection. For each mandal with at least four past winters in the APWRIMS monthly series (2014on), the May depth is its latest reading plus its own drawdown from the same month to May: the median of its past winters for a typical winter, and the largest for a dry one. Set against the deepest May the same series has recorded, a mandal is “beyond its record” when the typical projection is deeper, “in a dry winter” when only the dry one is. Against its own record, because the depth at which a well fails depends on how deep it was drilled, which no public record gives; beside it, whether the May depth would be more than 10m down, a class boundary on CGWB’s depth-to-water maps, since a delta mandal can break its own record with water three metres down. How many people live in those mandals comes from WorldPop’s 2020 population grid (1 km, UN-adjusted, CC BY 4.0), summed over each outline: a modelled estimate that counts everyone in the mandal, towns with piped supply included, not a census and not a count of people on wells.

How it has fared. Each past year was projected from the other years’ winters alone, from its August reading (4,710 mandal-years). The typical projection was off by a median 1.6 m, against 1.9 m for assuming no change. Mandals called beyond their record went past it in 33% of years, against 12% of all mandal-years and 6% of those called within it. A ranking of where to look first, not a forecast of this winter’s rain.

The crop water check is an FAO-56 Chapter 8 root-zone water balance run for every mandal, for a crop and growth stage the reader chooses. It starts from the soil moisture APWRIMS publishes (NRSC’s VIC model: plant-available water to 5, 30, 100 and 150 cm as a share of what the soil holds) and steps forward a day at a time to a week ahead with ECMWF’s open IFS forecast of reference evapotranspiration (Penman-Monteith) and rain, via Open-Meteo. How much the soil holds in millimetres comes from ISRIC SoilGrids 2.0 (field capacity minus wilting point, by depth). Rain under a fifth of the day’s reference evapotranspiration is treated as evaporated; the crop uses Ks × Kc × ETo; stress begins once it has used the share p of the root zone’s available water that Table 22 gives it, adjusted for its rate of use, and below that its water use falls in proportion (Eq. 84). Roots are 0.2 m at the initial stage and afterwards the larger Table 22 depth, the one its footnote gives for rainfed crops, capped at 1.5 m, the deepest the soil model reports. The pipeline runs its own copy of the calculation for every crop and stage and the page must reproduce its counts. Crop and stage are chosen, not observed; no irrigation, runoff or capillary rise is modelled; it is a screening view, not a watering instruction.

Crop vegetation is NOAA STAR’s weekly Vegetation Condition Index, the same 4 km files Drought Watch reads: this week’s greenness against the same week in every year on record. Here each 4 km cell is weighted by its cropland share from ESA WorldCover 2021 (10 m, read at about 330 m), so forest on the Eastern Ghats does not stand in for fields; a mandal with less than half a cell of cropland keeps the plain value and says so. The headline is the mean of the last four weeks, classed by the drought manual’s Table 3.4 (60–100 normal, 40–60 moderate, below 40 severe). WorldCover is a 2021 snapshot, not this season’s sowing, and the index measures plant vigour, not yield.

The groundwater category is the official Dynamic Ground Water Resources assessment, made each year by the Central Ground Water Board and the State Ground Water Department under the GEC-2015 method, read from the public dashboard of INGRES. In Andhra Pradesh the assessment unit is the mandal; its stage of extraction is the year’s groundwater draft over the annual extractable resource (safe up to 70%, semi-critical to 90%, critical to 100%, over-exploited above; saline where the water is too salty to use). Units are matched to our boundaries by district and name, then by the district a boundary’s neighbours give it (the prototype map labels some same-named mandals with another district), then by spelling; city wards assessed separately stay unmatched, and every match records how it was made. The statewide stage is INGRES’s own figure, not a sum we recompute.

How the crop water check has fared. The check is re-run on every week from early July to early September of each kharif season from 2024, with the soil moisture APWRIMS held on each date and the weather that actually followed (ERA5 reanalysis, so “within seven days” is tested as if the forecast were perfect). The outcome is crop vegetation three weeks later: the satellite index over cropland, which the soil model does not use. Two comparisons were run first and are kept on record, because each misleads for a reason that is not the check. Pooled across mandals, the calls do not separate vegetation at all: the places called short most often are the chronically dry ones, the index compares each place with its own past, and irrigated fields stay green whatever the soil model says. Inside one mandal but pooled across seasons, the gap is large, mostly because “short” calls come in drier seasons: the check tells a dry season from a wet one, as rainfall alone would. The verdict holds both fixed: the same mandal in the same season, vegetation after a “short” call against after a “comfortable” one. And because the check is a rainfed water balance, it is read over rainfed fields: the 437 mandals where less than 50% of the cropland was mapped irrigated (ESA WorldCereal), with the vegetation index weighted to rainfed cropland only. “Backed by its record” needs it at least 5index points lower, overall and in every season; “weak” at least 1 point lower; “not borne out” otherwise. On 2024–26, of 21 crop-stage pairs, 0 are backed, 5 weak, 12 not borne out and 4 untested over rainfed fields (over all cropland, irrigated fields included: 0 backed, 7 weak, 12 not borne out, 2 untested). The index covers whatever is growing, not the chosen crop, so this tests whether the check finds real water shortage, not whether a particular crop suffered.

At field scale. The 4 km index is one value for some 1,600 hectares, irrigated and rainfed, crop and scrub. Sentinel-2 (ESA) sees the same ground at 10 m every two to three days; the record reads it at 160m over rainfed cropland pixels only (WorldCover cropland that WorldCereal does not map as irrigated), takes the greenest clear view of each week, and compares the three-week change in NDVI, less the State’s change that week, inside the same mandal and season. The bars were set before the outcomes were read: “backed” at 0.02less NDVI change after “short”, in every season of at least two; “weak” at 0.005. Kharif is cloudy: 35.6% of rainfed mandal-weeks had a clear view in both weeks. A greenest-view composite can read a little lower in a week with fewer clear passes; taking out the State’s change that week removes most of that, not all. Correction: the first build of this reading, published on 4 October 2026, took the archive’s reflectance offset off a second time, which pushed every pixel’s NDVI towards 1; its verdicts are withdrawn, and the figures here are rebuilt from the corrected reads. On 2024–26: 0 backed, 8 weak, 9 not borne out and 4 untested.

Where field teams would learn most. This Week lists the mandals where four or more of six published signals point to stress at once: groundwater short of its seasonal normal, gauge rain 20% or more below normal, soil among the driest quarter of years, four or more of the seven reference crops short of water at mid-season, crop vegetation severely below normal, and a semi-critical or worse official category. It is a count of stated tests, shown signal by signal, to direct verification visits: not a ranking of need, an allocation or a declaration. The signals are not all independent.

The live scorecard. The track record re-runs the check with the weather that happened. The stricter test is the calls the page actually made: each Monday the refresh writes down every mandal’s call for every crop and stage, on that week’s ECMWF forecast, and never changes it (phase3_levels/data/field_calls). Three weeks on, when NOAA has published that week’s vegetation index, the calls are scored the way the track record scores them, over rainfed fields. 5 weeks frozen so far, 0 scored.

Rainfed or irrigated. ESA WorldCereal maps actively irrigated cropland at 10 m from Sentinel-1 and -2; over Andhra Pradesh its season is rabi 2020–21. Counted per 4 km vegetation cell and per mandal against ESA WorldCover’s cropland, it marks 437 of 670 mandals as mostly rainfed (23.9% of the State’s cropland mapped irrigated). It is one season’s satellite detection, not the irrigation census, and it cannot say which crop is grown: no open map does for groundnut, cotton, chilli or the pulses.

Rabi Outlook, the digest and field reports. The Rabi Outlook puts reservoir storage by basin beside soil moisture against its own past in the mostly rainfed mandals, and the northeast monsoon beside what past El Niño years brought; it recommends no crop, release or sowing date. The weekly digest is the same figures on one A4 sheet, printed to PDF by each deploy. Field reports are written on the phone and shared as text; nothing is sent anywhere until a person shares it, and a collector’s page reads the shared messages back into a table beside what the site called that week.

The Drought Watch follows the Manual for Drought Management (2020), the procedure a State must use to declare drought. Step 1, Trigger 1 (Table 3.11): a dry spell of four weeks in a row with under half the week’s normal rain sets it; without one, only large-deficient rainfall (60% or more below normal) does. Weekly totals come from the AP DES gauges via APWRIMS; SPI from CHIRPS v3 since 1981 is reported as the second rainfall route.

Step 2, Trigger 2 (Table 3.12), uses three of the manual’s four impact indicators per mandal, each with its own table: the Vegetation Condition Index (NOAA’s 4 km product; the manual prefers 56–500 m), the Percent Available Soil Moisture (the NRSC model at 30 cm, a four-week mean) and the Groundwater Drought Index(APWRIMS wells, ten years or more). Severe needs two severe and the third at least moderate; moderate needs two at moderate or worse. Area sown and reservoir storage are shown where they exist, by district and by reservoir. Step 3, field verification of crop loss (33% qualifies, over 50% for severe), and the notification by 31 October are the State’s. Each judgement the manual leaves open is listed on the page.

Open the Drought Watch

Boundaries. The State’s mandal geography gives every boundary vertex with the official area, LGD codes, revenue division and assembly and parliamentary constituency, but not the order the vertices run in. The outline is rebuilt as the shortest closed path through them (or, failing that, a concave hull tuned to the official area) and is drawn only where it lands within 3% of the official area, does not lie over its official neighbours, and touches the mandal it replaces (names repeat across districts, so a record that lands far from its mandal is treated as someone else’s and withheld, codes and all); elsewhere the public prototype stays. 595 of 670 outlines are official today. Compared with the prototypes, about a third were the wrong shape (area off by more than 15%, or overlapping less than their coarseness explains). The model still computes its inputs on the prototype outlines until it is re-tested on these.

Constituencies. Each mandal is counted in the one assembly constituency its State record names, or, without one, the constituency its centre falls in; a constituency’s figures are its mandals’ counts and medians. Its outline is the State’s own, rebuilt the same way and kept within 3% of the official area, else the union of its mandals.

The State network’s latest reading. One recent reading per location (early September), with the pre-monsoon, post-monsoon and year-ago levels. Its pre-monsoon value equals our APWRIMS May reading for 585 of 587 mandals, so it is the same department series, a month ahead. It is shown beside our monthly series, never merged into it, and the model does not use it.

Open Constituencies

Visits are counted with GoatCounter, which sets no cookies, stores no IP address and builds no profile across sites. What is recorded is the page opened, and — on the Monsoon Watch page — whether a visit used the film, the Pacific map, the mandal map or the download, and which of the page’s sections came into view. Each of those is counted at most once per visit, so a figure means the number of visits that did something rather than the number of clicks one reader made.

It exists to answer one question honestly: which of this evidence is actually reaching anyone. Nothing is recorded for a browser that sends Global Privacy Control, and any reader can switch counting off for this site for good by opening a page with #toggle-goatcounter on the end of the address.