Agricultural Statistics & Data Science · Sample surveys and official agricultural statistics · also Crop Estimation Surveys, General Crop Estimation Surveys (GCES), Crop-cutting surveys
How crop production is measured: area from complete enumeration, sample surveys and area frames, yield from crop-cutting experiments on randomly located plots, the Mahalanobis and Sukhatme debate on plot size, crop insurance CCEs, and newer satellite, model and smartphone-based methods.
In the exams: AICE-JRF Ag Stat Unit 5 'crop estimation surveys' (heavy); UPSC ISS Statistics III-IV official statistics (present); state DES statistical-officer exams (present)
How much rice did a district produce this season? The answer drives procurement, food security planning, trade policy, price support and crop insurance payouts, and it comes from crop estimation surveys. This subject covers how production is built up from area and yield; how crop area is measured by complete enumeration, sample schemes and area frames; how yield is measured by crop-cutting experiments in a stratified multistage design; the practical problems of plot size, border bias, location, harvesting and driage; how district and state yields and their standard errors are estimated; the quality-check schemes that keep the system honest; the use of crop-cutting for area-yield insurance; methods for horticultural and plantation crops; the growing role of remote sensing, crop models and smartphones; crop-loss estimation; and international practice.
The Indian system is described in most detail, because India pioneered random crop-cutting surveys under Mahalanobis and Sukhatme and runs one of the largest crop estimation systems in the world, but the principles are general and the last chapter compares other systems. Scheme details (sampling fractions, plot sizes, numbers of experiments, insurance rules) change over time and differ between states and countries; this manual states the principles and well-established features and points to the current official instructions for the rest.
In examinations, ICAR AICE-JRF/SRF Agricultural Statistics Unit 5 names the agricultural statistics system and crop estimation surveys, and official-statistics papers of the Indian Statistical Service and state statistical services ask about land-use classification, crop estimation methodology and the schemes that support it. The sampling theory behind the subject is in the Sampling Theory I and II subjects; the institutional system is in the Agricultural Statistics System subject.
Almost every national system estimates crop production as the product of two separately estimated quantities: the area under the crop and the average yield per unit area. Area is a characteristic of land and can be measured or recorded field by field; yield can only be measured by harvesting, so it is estimated from a sample of plots. Production of a crop in a district is then the district area multiplied by the estimated district yield, and state or national production is the sum of district productions. The two estimates usually come from different operations with different staff, frames and timing: area from land records, area surveys or remote sensing, and yield from crop-cutting experiments. Their errors are therefore roughly independent, and the precision of production depends on both. If the area estimate A and the yield estimate Y are independent, the relative variance (squared coefficient of variation) of production P = AY is approximately the sum of the relative variances: CV(P) squared is about CV(A) squared + CV(Y) squared. A district area known within 2% and a yield estimated with a 4% CV give a production CV of about the square root of (4 + 16), or 4.5%.
Area records in countries with cadastral land records are close to a complete enumeration, with sampling error near zero and non-sampling errors that depend on the diligence of the village official. Yield, by contrast, varies greatly from field to field because of soils, sowing dates, varieties, pests, irrigation and management, and it is measured on small plots, so its sampling error is the main component of the error of production. Most of the methodology of crop estimation surveys is therefore about designing, conducting and supervising the yield survey well, and most quality checks focus on it. Non-sampling errors matter in both parts: area may be recorded for the crop that was planned rather than the crop that was sown, mixed crops may be apportioned arbitrarily, and crop-cutting plots may be located or harvested carelessly. The total error of production is the sum of sampling and non-sampling errors, and in practice the latter often dominate.
Governments need production figures long before crop-cutting is complete, so estimates are released in stages. In India, advance estimates of production for each season are issued several times during the agricultural year, starting from area reported early in the season (through TRS and remote sensing) combined with expected yields based on weather, crop condition reports and past trends, and moving towards final estimates as crop-cutting results and complete area figures arrive. Each revision replaces assumptions with measurements. Users should treat early estimates as forecasts with larger errors and expect revisions, sometimes large in years of unusual weather. A good statistical system publishes its revision history so that the reliability of early figures can be judged. Forecasting methods (weather-based regression, remote-sensing indices, crop models) are covered in the crop forecasting subject; the crop-cutting survey supplies the benchmark against which those forecasts are calibrated.
In countries with detailed cadastral records, the village accountant (called the patwari, talati, karnam or village administrative officer in different Indian states) inspects every field each season and records the crop sown, the area and whether it is irrigated. This crop inspection, known in north India as girdawari, produces area statistics by complete enumeration from the field level upward. It also produces land-use statistics, classified in India into nine categories: forest; area under non-agricultural uses; barren and unculturable land; permanent pastures and other grazing land; land under miscellaneous tree crops; culturable waste; other fallow; current fallow; and net area sown. Complete enumeration gives figures for every village, which local administration and crop insurance need, but its quality depends on the field official's workload and supervision, and final figures can arrive late.
India uses two sample-based schemes alongside the land records. Under the Timely Reporting Scheme (TRS), a random sample of about one fifth of the villages in land-record states is selected each season, and the village official completes the crop inspection in those villages first, on a priority basis, so that state-level area estimates are available quickly for the advance estimates of crop production. In states without a village land-record system (historically the permanently settled areas such as West Bengal, Odisha and Kerala), the scheme for Establishment of an Agency for Reporting Agricultural Statistics (EARAS) set up a staff to enumerate area in a sample of villages, again about a fifth of the villages, so that sample-based estimates replace the missing complete enumeration. Both schemes use the sample villages as first-stage units of an estimation procedure: the sample total of area is expanded to the stratum by the inverse of the sampling fraction or by ratio methods using land-use totals. The exact sampling fractions and administrative arrangements have changed over time and differ between states, so the current scheme documents of the Directorate of Economics and Statistics should be consulted for details.
Where no reliable list of fields or holdings exists, area frames are used. The territory is stratified by land use (for example intensively cropped, mixed, pasture, urban) usually from satellite images, divided into segments with identifiable boundaries, and a sample of segments is visited; every field in a sampled segment is measured or recorded with its crop. The total area of a crop in a stratum is estimated by expanding the sample segment totals by the inverse of the sampling fraction. The United States June Agricultural Survey is the best-known area-frame survey, and many countries have adopted the approach with FAO support. Satellite classification of crops gives complete coverage but with classification errors; combining it with ground segments through regression or calibration estimators corrects the bias and reduces the variance. India's national crop forecasting programmes use remote sensing for area of major crops in this spirit.
Mixed cropping (several crops sown together in the same field, such as sorghum with pigeonpea, or cotton with groundnut) complicates area statistics, because a field's area must be shared among its crops. Land-record systems apportion the area of a mixture among its component crops using standard ratios fixed for each mixture and region, or from the seed rates or row ratios observed. The resulting crop areas are therefore partly conventional, and different apportionment rules can change reported areas substantially. For yield, a crop-cutting plot in a mixed field yields more than one crop; the produce of each is weighed and converted to yield per unit of apportioned area, so that yield times apportioned area gives the crop's production in the field. Errors in apportionment then pass into both area and yield, though they partly cancel in production. FAO's methodological work under the Global Strategy gave particular attention to mixed and continuous cropping because it is common in Africa and South Asia.
Yield in India is estimated by the General Crop Estimation Surveys (GCES), which follow the methodology developed by the Indian Council of Agricultural Research and its statistical wing (now ICAR-IASRI) under P. V. Sukhatme from the 1940s, and the parallel work of P. C. Mahalanobis at the Indian Statistical Institute. The design is stratified multistage random sampling. Strata are administrative units, usually tehsils, taluks or development blocks within a district. Within each stratum, the first-stage units are villages, selected at random; the second-stage units are fields (survey numbers) growing the crop in each selected village, selected at random from the list prepared during crop inspection; and the third-stage unit is a plot of specified size and shape located at random within each selected field, which is harvested to measure yield. The number of crop-cutting experiments allotted to a district depends on the crop's importance and the precision required; typically a small number of fields (often two) is selected per village and one plot per field. Selecting few fields per village and more villages spreads the sample and reduces the between-village component of variance, which is usually large.
Multistage sampling avoids the need for a complete list of all fields in a district: only the selected villages need field lists, which the crop inspection provides. Clustering the fieldwork in villages saves travel. Stratification by tehsil or block ensures geographical spread and gives estimates for administrative units. The price of clustering is that fields in the same village are more alike than fields from different villages, so the variance of the estimate depends heavily on the number of villages. For two-stage sampling with n villages and m fields per village (ignoring finite population corrections), the variance of the mean yield per plot is about S_b squared / n + S_w squared / (nm), where S_b squared is the between-village variance and S_w squared the within-village variance. With costs c1 per village and c2 per field, the most efficient number of fields per village is about the square root of [c1 S_w squared / (c2 S_b squared)].
A district's total number of crop-cutting experiments is shared among strata (tehsils or blocks), usually in proportion to the area under the crop in each stratum, so that the sample is roughly self-weighting by area. Where some strata are known to be more variable, Neyman allocation, with experiments proportional to area times the standard deviation of yield, gives a more precise district estimate for the same total; in practice proportional allocation is common because standard deviations are uncertain and administrative fairness matters. Within a stratum, villages are selected at random, either with equal probability or with probability proportional to the crop area in the village, and fields are selected at random from those growing the crop. When villages are selected with probability proportional to crop area, the plot mean is close to self-weighting for yield per hectare, which simplifies estimation. The number of experiments across crops is also balanced against the workload of field staff, since each experiment requires a visit at harvest.
An early methodological controversy in Indian crop surveys concerned the size of the crop-cutting plot. Mahalanobis's surveys in Bengal used very small plots, while Sukhatme's experiments showed that small plots tended to overestimate yield. The cause is border bias: when an investigator marks a small plot, plants on the boundary are more often included than excluded, and the proportion of boundary plants is larger for small plots. Larger plots reduce the relative effect of border errors, at the cost of more harvesting work and a smaller number of plots for the same budget. The resolution adopted in GCES was a plot large enough to keep border bias small but small enough to be harvested by the field staff, with the size fixed by crop. Rectangular or square plots of a few tens of square metres are used for most field crops; the specific dimensions vary by crop and have been revised over time, so the current instructions of the state agricultural statistics authority should be followed.
For a given area, a square or circle has the shortest boundary and therefore the smallest border problem; long narrow plots have more boundary per unit area. Circular plots are hard to mark accurately in the field and were largely abandoned in favour of squares and rectangles marked with measuring tapes and corner pegs. Clear rules for boundary plants (for example, include a plant if more than half of its base lies inside the plot line) reduce investigator discretion. Equipment such as rigid frames or pre-measured ropes improves accuracy. Plot size also affects the variance per plot: as with experimental plots, larger plots average over more variation, but the effect is modest compared with the effect of the number of villages and fields sampled, because most yield variation is between fields and villages, not within fields.
Real fields have gaps where seeds failed, internal bunds and irrigation channels, trees and irregular shapes. Instructions must say whether such features inside a randomly located plot are included (they usually are, because they are part of the field's productive reality and excluding them would bias yield upward) and how to handle plots falling across a field boundary (usually by shifting the plot inward along a stated rule). Consistent handling matters more than any single choice, because inconsistent rules between investigators and years create spurious differences in yield.
The plot must be located at random within the selected field, never chosen by eye, because investigators and farmers tend to choose better-looking patches. The standard procedure is to measure the length and breadth of the field from a fixed reference corner, draw two random numbers within those dimensions, and place a corner of the plot at the point so defined, with rules for what to do if the plot would fall outside the field (shift it inward). The farmer is informed and the plot is harvested when the crop is ready, with the farmer's harvest as the reference date. Experiments are lost when the farmer harvests before the investigator arrives, when the field is mixed-cropped in a way the instructions do not cover, or when the crop has failed; the instructions specify how to record such cases, including treating failed crops as zero yield rather than replacing the field, to avoid bias.
The produce of the plot is harvested, threshed and cleaned in the investigator's presence, and the grain (or other economic produce, such as seed cotton or tubers) is weighed with a calibrated balance. The weight, plot dimensions, date, variety, irrigation and other ancillary information are recorded on the schedule. With mobile applications now widely used, the plot location is geotagged and photographs of the plot, the harvest and the weighing are attached, which strengthens supervision and audit.
Grain weighed at harvest contains more moisture than grain in storage, so its weight is converted to a dry-weight basis. Driage experiments keep a sample of the harvested produce (for example a kilogram of paddy), dry it under standard conditions and re-weigh it after a fixed period; the ratio of dry to fresh weight is the driage factor, applied to the plot weights. Some crops also need conversion between forms, such as paddy to rice (with a milling outturn ratio), seed cotton to lint, or green to dry chillies. The yield per hectare from a single experiment is: plot produce (kg) x driage factor x 10,000 / plot area (square metres). For example, 10 kg of paddy from a 25 square metre plot, with a driage factor of 0.85, gives 10 x 0.85 x 10,000 / 25 = 3,400 kg/ha of dry paddy.
Within a stratum (tehsil or block), the estimate of average yield is the mean of the plot yields, each converted to kg/ha, when villages are selected with equal probability and the same number of plots is taken in each; otherwise appropriate weights are used. The district average yield is the weighted mean of stratum yields, with weights proportional to the area under the crop in each stratum, so that strata with more crop area count more. District production is district area times district yield, or equivalently the sum over strata of stratum area times stratum yield; state production is the sum of district productions and the state yield is state production divided by state area. Because weights are areas, the state yield is not the simple average of district yields; it is an area-weighted average, which is how official yield figures should be interpreted.
For a stratified two-stage design with villages and fields, the variance of the stratum mean yield is estimated from the between-village and within-village mean squares. If villages are selected by simple random sampling and m fields are taken in each of n villages, an unbiased estimator of the variance of the stratum mean (ignoring finite population corrections at both stages) is s_b squared / n, where s_b squared is the variance between village means: the within-village variation is included automatically in the variation of village means. The district variance combines stratum variances with squared area weights: V(district yield) is the sum of W_h squared V(stratum mean h), where W_h is the share of stratum h in district crop area. The percentage standard error (the CV of the estimate) is the usual reporting measure. National systems aim for percentage standard errors of a few per cent at state level for major crops; district-level standard errors are larger because the sample per district is smaller. A reliability target can be turned into a sample size: if the CV of yield between plots is C and the target CV of the mean is e, the number of independent plots needed is about (C/e) squared, inflated by a design effect for clustering in villages.
Suppose a district has three tehsils with rice areas of 30,000, 20,000 and 50,000 ha and estimated mean yields of 2.8, 3.4 and 2.5 t/ha. District production is 30,000 x 2.8 + 20,000 x 3.4 + 50,000 x 2.5 = 84,000 + 68,000 + 125,000 = 277,000 t, and the district yield is 277,000 / 100,000 = 2.77 t/ha. The simple average of the three tehsil yields, 2.90 t/ha, would overstate the district yield because the largest tehsil has the lowest yield. If the standard errors of the tehsil means are 0.10, 0.15 and 0.08 t/ha, the variance of the district yield is 0.3 squared x 0.01 + 0.2 squared x 0.0225 + 0.5 squared x 0.0064 = 0.0009 + 0.0009 + 0.0016 = 0.0034, so its standard error is about 0.058 t/ha and its percentage standard error about 2.1%. The example shows that the largest stratum often dominates the variance even when its own standard error is the smallest.
Because both area enumeration and crop-cutting depend on very large numbers of field staff, India operates a central scheme for Improvement of Crop Statistics (ICS). Under it, supervisors from the central National Statistical Office field operations and from the states carry out sample checks: they verify the crop inspection (area enumeration) in a sub-sample of the TRS villages, comparing the officially recorded crops and areas with what is found in the field, and they supervise a sub-sample of crop-cutting experiments, observing whether plots are located at random, marked correctly, harvested and weighed properly. The results are used to estimate the extent of discrepancy between recorded and observed area figures, to identify districts and practices with problems, and to recommend corrective action. ICS findings over the years have repeatedly highlighted issues such as delayed crop inspection, recording of crops without field visits and deviations from the crop-cutting protocol.
Quality in crop estimation is built in at several points: training and manuals for field staff; random selection of villages, fields and plot locations by the statistical office rather than by field staff; supervision of a proportion of experiments by senior officers; scrutiny of schedules for impossible values (such as yields far beyond the plausible range for the crop and area); comparison with the previous year, with neighbouring districts and with independent indicators such as satellite vegetation indices and market arrivals; and audit trails with geotags and photographs. Non-response through lost experiments is monitored, because lost experiments are not a random sample: experiments are more often lost in early-harvested or poor fields. Large losses should trigger investigation, not silent replacement.
Supervision checks are themselves a sample and can be analysed as one. If supervisors re-enumerate crop areas in a random sub-sample of villages, the mean difference between recorded and observed areas, with its standard error, estimates the bias of the recorded figures, and a ratio of observed to recorded area can be used to adjust estimates or to target retraining. For crop-cutting, the proportion of supervised experiments with protocol deviations estimates the deviation rate in the whole survey. Supervision is most useful when the sub-sample is selected at random and the supervised experiments cannot be predicted by field staff; announced or purposively chosen checks measure what staff do when watched, not what they do otherwise.
Area-yield index insurance pays all insured farmers in an insurance unit when the measured yield of that unit falls below a threshold. In India's Pradhan Mantri Fasal Bima Yojana (PMFBY), the threshold yield of a crop in an insurance unit is the average yield of the best five of the last seven years multiplied by the indemnity level (for example 70%, 80% or 90%), and the claim, as a proportion of the sum insured, is (threshold yield - actual yield) / threshold yield when the actual yield is below the threshold. The actual yield is estimated from crop-cutting experiments conducted in the insurance unit. This turns crop-cutting from a statistical exercise into the basis for payments, which raises the stakes for accuracy and timeliness and requires many more experiments, because insurance units (such as a village or village panchayat for major crops) are much smaller than districts.
Scheme guidelines prescribe a minimum number of crop-cutting experiments per insurance unit, for example four per village or village panchayat for major crops under the PMFBY operational guidelines, and larger numbers for larger units. With only a few experiments per unit, the sampling error of the unit yield is large: if the CV of yield between plots within a village is 30%, the mean of four plots has a CV of about 15%, so two neighbouring villages with the same true yield can easily receive different payouts. This is a basic statistical limitation of small-area yield estimation by crop-cutting alone, and it motivates combining CCEs with other information.
Technology is used in three ways. Mobile applications record crop-cutting experiments with geotagged, time-stamped photographs and upload the data directly, reducing manipulation and delay. Remote sensing and crop models are used to select the locations of experiments (smart sampling, which places experiments to represent the variation in crop condition) and to provide independent yield estimates for checking. And technology-based yield estimates combining satellite data, weather data and models with a reduced number of CCEs have been introduced for some crops and states, in which the model yield and the CCE yield are combined with stated weights. The details of such schemes change frequently, so the current guidelines of the implementing ministry should be consulted; the statistical principles are those of regression and small-area estimation.
Suppose the yields of a crop in an insurance unit over the last seven years were 2.0, 2.4, 1.8, 2.6, 2.2, 2.5 and 2.3 t/ha, the indemnity level is 80% and this year's crop-cutting gives an average yield of 1.5 t/ha. The best five years average (2.6 + 2.5 + 2.4 + 2.3 + 2.2)/5 = 2.4 t/ha, so the threshold yield is 0.8 x 2.4 = 1.92 t/ha. The shortfall is (1.92 - 1.5)/1.92 = 21.9% of the threshold, and each insured farmer in the unit receives about 21.9% of the sum insured, whatever the yield on their own field. Two statistical points follow. The payout depends on the sampling error of the unit's yield estimate, which with few CCEs can be large. And the threshold depends on past yield estimates, which themselves have errors and may be revised, so the quality of historical crop statistics directly affects insurance fairness.
Fruit and vegetable crops do not fit the crop-cutting model well. Many are harvested in several pickings over weeks or months (tomato, chilli, okra), perennial fruit trees vary in age and bearing, intercropping and kitchen gardens are common, and the economic produce is perishable and sold in many forms. Area records for horticultural crops are also less reliable than for field crops. Methods developed by ICAR-IASRI and others therefore use different units: for orchards, the number of bearing trees and the yield per tree; for vegetables, the yield from repeated pickings on sample plots over the season. A typical orchard survey selects villages, then orchards, then sample trees, and records the number of bearing trees per hectare and the produce harvested from sample trees, often by enumerating fruits and weighing a sample. Production is then estimated as area x bearing trees per hectare x yield per tree, or as number of bearing trees x yield per tree. For example, 150 bearing trees per hectare yielding 40 kg each gives 6 t/ha.
Plantation crops such as tea, coffee, rubber, coconut and arecanut are often estimated from returns of estates and processing factories and through commodity boards, supplemented by sample surveys of smallholders. In India, a national programme (CHAMAN) used remote sensing and geographic information systems with ground surveys to estimate area and production of major horticultural crops in selected districts. Remote sensing works best for orchards and plantations with distinctive canopies and is weaker for small, mixed vegetable plots.
For vegetables harvested repeatedly, the survey selects sample plots early in the season and records the produce of every picking until the crop is removed, so that the season's yield is the sum of pickings. Because this requires many visits, the number of plots is smaller than for field crops and farmers' records of pickings are sometimes used, checked on a sub-sample. Estimates of the share of produce lost after harvest are collected separately.
Satellite images provide repeated, complete coverage of cropland. For area, crops are classified from their spectral signatures and seasonal profiles, using optical sensors and, under clouds during the monsoon, synthetic aperture radar. Classification accuracy is assessed against ground truth with confusion matrices, and pixel counts are converted to unbiased area estimates by regression or calibration with ground samples, not taken at face value. For yield, vegetation indices such as the normalised difference vegetation index (NDVI) at critical growth stages are related statistically to measured yields; the relation is calibrated with crop-cutting data and then applied across the area. India's FASAL programme (Forecasting Agricultural output using Space, Agro-meteorology and Land based observations) combines remote sensing, weather-based models and field data to forecast production of major crops before harvest. Similar systems operate elsewhere, including the European Union's crop monitoring and yield forecasting system and national systems in the United States and China.
Process-based crop models simulate growth from weather, soil, variety and management, and can predict yield where no measurements exist. Used for estimation, their parameters must be calibrated to local conditions and their outputs treated as auxiliary information with error, not as observations. Assimilating satellite observations into crop models, adjusting the simulated leaf area to match observed vegetation indices, is an active research area.
Smartphones now capture CCE data, geotags and images, and computer-vision methods estimate yield components (grain counts, ear sizes) from photographs. Farmer-reported yields collected by phone or in household surveys are cheap but are subject to recall and reporting biases, which studies have found to vary with farm size and crop; they should be calibrated against objective measurements on a sub-sample. The statistical frame for using all these sources together is the same: an expensive accurate measurement on a sample (crop-cutting) combined with cheap auxiliary information on all units (satellite, model, farmer report) through ratio, regression or small-area estimators.
A satellite crop map is assessed by comparing its classes with ground truth at randomly selected reference points, summarised in a confusion (error) matrix. Overall accuracy is the proportion of reference points correctly classified; user's accuracy for a class is the proportion of points mapped as that class that are truly that class (the complement of commission error); producer's accuracy is the proportion of true points of that class that are mapped correctly (the complement of omission error). The kappa coefficient compares agreement with chance agreement, though its use is now debated. Map pixel counts are biased whenever omission and commission errors do not balance. Good practice, recommended in the remote-sensing literature, estimates class areas from the reference sample using the error matrix (a stratified estimator with map classes as strata), which gives unbiased area estimates with confidence intervals, rather than reporting pixel counts.
Crop loss is the reduction in yield, quality or value caused by pests, diseases, weeds or adverse weather, relative to an attainable yield. Loss is usually expressed as a percentage: loss = (attainable yield - actual yield) / attainable yield x 100. The attainable (or potential) yield is that obtained with protection, so the measured loss is the avoidable loss under the protection used. For example, if protected plots yield 3.2 t/ha and unprotected plots 2.4 t/ha, the avoidable loss is 0.8/3.2 = 25%.
Three approaches are common. Paired-plot experiments compare protected and unprotected plots side by side on the same fields, giving direct estimates of avoidable loss; they are the most reliable but costly. Survey methods record pest or disease incidence and severity on sampled fields and relate them to yield by regression, so that losses can be estimated from incidence surveys over large areas. Damage-yield models, developed from experiments, convert observed damage (such as percentage of dead hearts or infected leaf area) into expected yield loss. ICAR-IASRI and national plant-protection research have used these methods to estimate losses in major crops. Loss surveys use the same multistage sampling as crop estimation surveys, often in the same sample villages, with additional observations on pest incidence. Their main difficulties are separating the effect of one pest from others, from weather and from management, and the year-to-year variability of pest outbreaks, which makes average loss figures unstable.
FAO has long promoted crop-cutting as the objective method of yield measurement, and its guidelines on crop yield estimation and the World Programme for the Census of Agriculture set international standards for area and yield statistics. Under the Global Strategy to improve Agricultural and Rural Statistics, methodological research compared crop-cutting, full-plot harvest, farmer recall and crop cards (diaries of harvests) in several countries. It found that full-plot harvest is the most accurate but costliest, that crop-cutting can be biased when plots are small or poorly located, and that farmer recall can be biased in ways that depend on plot size, which has implications for research on farm size and productivity. The 50x2030 Initiative supports countries in building integrated agricultural survey systems that combine these methods.
The United States Department of Agriculture's National Agricultural Statistics Service runs objective yield surveys for major crops such as corn, soybeans, wheat and cotton. Sample fields are drawn from the June Agricultural Survey area frame, and enumerators lay out small sample units in each field during the season, counting plants, ears, pods or bolls and measuring their components. Before harvest these counts feed forecast models; at harvest the sample units are harvested and weighed. The objective measurements are combined with farmer-reported yield surveys to produce monthly forecasts and final estimates, with published reliability measures.
Despite differences, successful systems share features: probability selection of fields and plots by the statistical office; objective measurement where feasible; clear rules for awkward cases (mixed crops, failed crops, early harvest); independent supervision; combination of expensive measurements on a sample with cheap auxiliary data on everything; and publication of methods and standard errors. Where systems fail, the causes are usually non-sampling: delayed or fictitious field work, inadequate staff, and pressure on reported figures. For students and practitioners, crop estimation is the place where sampling theory meets the practical reality of fieldwork.
In a stratified area frame, if a stratum contains N segments and n are sampled at random, and the sampled segments contain a total of a hectares of a crop, the estimated crop area in the stratum is (N/n) x a. For example, with 1,000 segments, 50 sampled and 400 ha of maize found in them, the estimate is 20 x 400 = 8,000 ha. Its variance follows from simple random sampling of segments: N squared (1 - n/N) s squared / n, where s squared is the variance of maize area between sampled segments. Segments may be selected with probability proportional to size where sizes differ, and the estimate is improved by a regression estimator using the satellite-classified crop area of every segment as the auxiliary variable, which is how many systems combine remote sensing with ground surveys.
Students need the structure of crop estimation for examinations (multistage design, TRS, EARAS, ICS, land-use classes, driage, area-weighted yields) and for theses that use official crop statistics or design their own yield surveys. Knowing how official yields are produced, and their standard errors, stops them from over-interpreting small differences between districts or years.
A five-year statistician in a state statistics directorate or an insurance company allocates crop-cutting experiments, trains and supervises field staff, scrutinises schedules, computes district yields and standard errors, handles disputes over insurance yields, and works with remote-sensing teams to select CCE locations and check estimates. They decide when a lost or doubtful experiment needs re-examination and how to treat mixed and failed crops.
A senior official responsible for agricultural statistics in a state or country sets the sample sizes and allocation, approves methodological changes such as technology-assisted yield estimation, responds to ICS findings, and defends published estimates to policy makers and the public. They balance timeliness against accuracy for advance estimates and manage the shift from manual to digital field systems.
A PhD with long experience develops improved estimators (small-area models combining CCEs with satellite and weather data), designs evaluations of new methods against crop-cutting, and advises national and international bodies on crop statistics methodology. They carry the history of the subject, from the Mahalanobis-Sukhatme debates to today's model-assisted systems, into training and reform.
Exact current sampling fractions, plot dimensions by crop, numbers of experiments per insurance unit and the weights used in technology-based yield estimation change over time and differ between states; this manual gives principles and well-established features and does not quote current figures it could not verify. The detailed estimation formulas for three-stage designs with unequal probabilities and the mathematics of small-area models are in other subjects. Country systems other than India and the United States are described only in outline.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions. Every numerical answer was computed with a script. Scheme details that change over time are asked only at the level of principle.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.