Computer Vision Methods, Limitations, and Potential Benefits for Farmers

An Exploratory Case Study from Mangaon, Maharashtra

Abstract

This case study examines the potential use of artificial intelligence (AI) and computer vision to estimate turmeric planting material from a smartphone photograph. The image, supplied during planning for turmeric cultivation at Mangaon, Maharashtra, showed three piles of harvested turmeric rhizomes. A general-purpose multimodal AI assistant produced a preliminary estimate of 12,500–19,500 pieces in total, including an estimated 1,500–2,500 mother rhizomes and 11,000–17,000 finger rhizomes. The farmer subsequently reported approximate counts of 8,500 mother rhizomes and 17,500 finger rhizomes, giving a total of approximately 26,000 pieces. On this comparison, the initial total-count range underestimated the reported total by approximately 25.0% at its upper bound and 51.9% at its lower bound. The finger-rhizome range came closer, while the mother-rhizome estimate substantially undercounted the reported quantity. The case therefore illustrates both the promise and the limitations of image-based estimation. It distinguishes the capabilities of trained computer-vision systems from the unverified reasoning used in this conversational estimate, proposes a more rigorous measurement protocol, and explains how calibrated image-based tools could help farmers with inventory, seed planning, labour allocation, and yield forecasting.

Keywords

Artificial intelligence; computer vision; turmeric; rhizome counting; finger rhizomes; mother rhizomes; agricultural inventory; image-based estimation; farmer decision support; Maharashtra.

1. Introduction

Farmers routinely make decisions using visual estimates: how much seed material is available, how many planting units can be prepared, how much labour will be required, and whether the available stock is sufficient for a planned area. Such estimates are useful but can be difficult when produce is irregularly shaped, piled in layers, partly hidden, or mixed with leaves and roots. Manual counting and weighing provide stronger evidence but require time, labour, and suitable equipment.

Computer vision offers a potential alternative or supplement. In agricultural applications, image-analysis systems have been developed to detect, segment, count, and estimate the dimensions of objects such as fruit, leaves, grains, and root crops. Research has shown that approaches based on object detection, regression, and density estimation can be applied to agricultural counting, while heavily overlapping objects and perspective distortion remain important challenges (Farjon & Edan, 2024; Dolata et al., 2021).

The present case arose from a practical question: could an AI assistant use a single photograph of turmeric seed stock to estimate the number of mother rhizomes and finger rhizomes well enough to support farm planning? The exercise was conducted in the context of planned turmeric cultivation at Mangaon, Maharashtra, using the Kokan Halad-1 variety as identified by the farmer. The focus of this paper is not to claim a validated AI product, but to document the estimation, compare it with the farmer’s later count, and outline a scientifically stronger approach.

2. Case Background and Research Question

The photograph showed three visible piles of turmeric planting material. The piles contained differently sized rhizomes and some associated plant material. The image did not include a ruler, weighing scale, measured container, depth sensor, or other calibrated reference. The farmer’s practical objective was to understand the approximate quantity of planting material and its potential use in cultivation.

The central research question was: to what extent can a general-purpose multimodal AI assistant estimate the number of turmeric mother rhizomes and finger rhizomes from a single ordinary photograph, and what additional measurements would be needed to make such an estimate dependable for farm decisions?

For this paper, ‘mother rhizome’ refers to the larger central rhizome unit, while ‘finger rhizome’ refers to the elongated lateral rhizome pieces. These categories can be operationally defined differently in farm practice, so future studies should establish a consistent counting rule before collecting reference counts.

3. Methodological Framework

3.1 What a multimodal AI assistant may infer from an image

A multimodal AI assistant processes image information alongside language and learned visual patterns. It may identify the main objects, distinguish foreground from background, compare relative dimensions, and generate plausible explanations about the scene. Such behaviour can resemble visual reasoning, but it should not automatically be interpreted as a precise physical measurement. Unless a system explicitly runs a calibrated computer-vision pipeline, its explanation of how it reached a number may be a plausible reconstruction rather than a faithful record of internal calculations.

In this case, the assistant initially separated the scene into three piles and proposed ranges for their contents. It then discussed possible methods such as pile-volume estimation, bulk-density assumptions, average piece weight, and extrapolation from visible rhizomes. Those methods are reasonable components of a future estimation workflow, but the original conversation did not establish that the assistant measured the pile dimensions, applied a verified turmeric-specific density value, or ran a detector that individually segmented every rhizome. The estimates must therefore be treated as exploratory visual judgments, not as instrument-grade measurements.

3.2 Computer-vision methods that could be used in a dedicated system

A purpose-built system could begin with image preprocessing to improve contrast and reduce distractions caused by uneven lighting, shadows, soil, leaves, and mesh. Object detection could then locate visible rhizomes, while instance segmentation could outline each visible piece separately. Detection-based counting is suitable when individual objects are sufficiently distinct; density-estimation or regression approaches may be more appropriate when objects overlap so heavily that drawing a separate outline around every item is impractical. The agricultural counting literature describes these as distinct families of methods rather than a single universal solution (Farjon & Edan, 2024).

A dedicated model would need labelled turmeric images in which human annotators mark or count mother rhizomes and finger rhizomes separately. Training data should represent different varieties, sizes, soil backgrounds, moisture levels, camera angles, pile densities, lighting conditions, and degrees of overlap. The model should be evaluated on photographs from farms not used in training to test whether it generalises beyond familiar examples.

Scale calibration is a separate problem. A ruler, calibration board, known-size tray, or measured container in the same plane as the rhizomes can convert pixel dimensions into approximate physical dimensions. A single image generally cannot uniquely reveal hidden depth or the exact size of an object when perspective and scale are unknown. Multiple views, depth cameras, or photographs of material spread in a single layer can improve the measurement basis. Research on root-crop imaging likewise identifies occlusion and perspective distortion as important difficulties in estimating object dimensions (Dolata et al., 2021).

3.3 From image counts to weight and planting potential

If the visible pieces can be counted reliably, the next step is to estimate the hidden population. This is difficult because the surface of a pile may not represent the distribution of sizes and types inside it. A stronger method is to weigh a representative sample of mother rhizomes and finger rhizomes separately, count the pieces in each sample, and calculate the average weight per piece for each category. The total mass can then be measured directly with a scale, or estimated from a calibrated volume and a locally measured bulk density.

For a weight-based estimate, the basic relationship is: estimated number of pieces in a category = total mass of that category ÷ mean mass per piece. This approach requires category-specific weights. Dividing total mixed stock weight by one average rhizome weight would be unreliable if mother and finger rhizomes differ substantially in size. Where possible, direct counting from spread-out batches and weighing the same batches is preferable to inferring the whole stock from the apparent volume of a pile.

4. Initial Estimate and Farmer-Reported Count

The assistant’s initial image-based estimate was a total of approximately 12,500–19,500 rhizome pieces. Within that total, it estimated 1,500–2,500 mother rhizomes and 11,000–17,000 finger rhizomes. The farmer later reported approximate actual counts of 8,500 mother rhizomes and 17,500 finger rhizomes, or about 26,000 pieces altogether. These farmer-reported counts are treated here as the reference figures for the case comparison; they were not independently audited by the authors of this paper.

Category	Initial AI estimate	Farmer-reported count	Comparison
Mother rhizomes	1,500–2,500	≈ 8,500	Substantial underestimation
Finger rhizomes	11,000–17,000	≈ 17,500	Upper bound short by ≈ 500
Total pieces	12,500–19,500	≈ 26,000	Total undercounted

5. Error Analysis

The farmer-reported total of approximately 26,000 pieces lies above the initial estimate range. Compared with the upper bound of 19,500, the difference is 6,500 pieces, equivalent to an underestimation of 25.0% relative to the reported count. Compared with the lower bound of 12,500, the difference is 13,500 pieces, or approximately 51.9% of the reported count. The total-count estimate therefore should not be described as highly accurate on the evidence available.

For finger rhizomes, the upper end of the initial estimate was 17,000, only about 500 pieces below the farmer’s approximate count of 17,500. This is an underestimation of approximately 2.9% relative to the reported count. However, the lower end of the range was 11,000, which is about 37.1% below the reported count. The apparent closeness of the upper bound is useful but should not be confused with a precise point estimate.

The mother-rhizome estimate was the weakest component. The reported count of approximately 8,500 was 6,000 pieces above the initial upper bound of 2,500 and 7,000 above the lower bound of 1,500. This corresponds to underestimation of approximately 70.6% to 82.4% relative to the farmer-reported count. The result shows that a visually plausible explanation can still produce a large category-specific error.

Metric	Calculation	Result
Total undercount vs upper bound	(26,000 − 19,500) ÷ 26,000	25.0%
Finger undercount vs upper bound	(17,500 − 17,000) ÷ 17,500	2.9%
Mother undercount vs upper bound	(8,500 − 2,500) ÷ 8,500	70.6%

6. Interpretation and Limitations

The largest limitation was occlusion: many rhizomes were hidden beneath the visible surface. The piles also contained pieces of different sizes and shapes, and the photograph did not provide a calibrated scale or a second view. A single image can suggest the outer shape of a pile, but cannot reliably reveal its internal packing or the number of concealed objects. Any method that converts pile volume to count must account for air gaps, variable piece size, and the proportion of leaves or other material.

A second limitation is that the initial estimates were produced in a conversational setting, not by a trained turmeric-specific counting model. It would be scientifically inaccurate to claim that the system definitively reconstructed three-dimensional geometry, identified a standard-sized bowl, calculated a verified bulk density, or individually counted the surface pieces unless those operations were actually performed and documented. These are plausible techniques for a designed system, not confirmed steps in this particular interaction.

A third limitation concerns the reference count. The farmer described the counts as approximate. The case does not document a formal counting protocol, independent repeat count, weighing record, or uncertainty range for the reference values. As a result, the comparison is informative but preliminary. Future research should preserve photographs, original estimates, exact counts, measured weights, and a transparent record of the calculation process.

Finally, the number of seed pieces alone does not determine production. Yield depends on the weight and health of planting material, variety, planting density, soil, water management, drainage, nutrition, disease pressure, weather, and harvest practices. A count estimate can support inventory planning, but it should not be converted directly into a confident yield forecast without locally validated agronomic data.

7. Proposed Farmer-Centred Workflow

A practical farmer-facing system should combine image analysis with a small amount of physical calibration. First, the farmer should photograph the rhizomes from directly above and from two or more side angles, using good lighting and avoiding extreme perspective. A ruler or known-size board should be included in the same plane as the objects. If feasible, spreading the stock into shallow trays or a single layer will reduce overlap and make direct object detection more reliable.

Second, the system should distinguish mother rhizomes from finger rhizomes and display visible-object detections so the farmer can correct missed or duplicate counts. Third, the farmer should weigh and count a few random samples from each category. The sample size and selection method should be recorded, and more than one sample should be used because piece weights vary. Fourth, the total stock should be weighed on a scale whenever possible. The image-based count can then be cross-checked against the measured mass and sample mean weights.

Fifth, the application should display a range and explain the main sources of uncertainty rather than presenting a false impression of exactness. It should indicate whether its result is based on directly detected objects, a calibrated mass-to-count calculation, or a visual estimate. Over time, with permission and appropriate data handling, anonymised image-count-weight records from multiple farms could be used to train and validate a turmeric-specific model.

8. Potential Benefits for Farmers

If developed and validated properly, image-based rhizome estimation could reduce the time required for inventory assessment and help farmers organise planting material before the season begins. Estimating mother and finger rhizomes separately could improve seed sorting, batch preparation, labour planning, and decisions about whether additional seed stock is needed. When paired with actual weights and locally measured planting rates, it could help estimate the area that can be planted. For producer groups, nurseries, and turmeric processors, the same approach could support stock records and more consistent procurement estimates.

The technology is potentially accessible because a smartphone camera is widely available and image analysis can be delivered through a mobile or web application. However, low-cost access must not be confused with guaranteed accuracy. Farmers should be able to inspect and correct the output, and important decisions should remain anchored in direct weighing, sampling, and local agricultural advice. The greatest value may come from combining AI estimates with farmers’ knowledge, rather than replacing field judgment.

9. Recommendations for Further Study

A pilot study should collect a structured dataset of turmeric piles across farms, varieties, seasons, and moisture conditions. Each pile should be photographed from multiple angles and then separated into mother rhizomes and finger rhizomes for verified counts and weights. A subset of images should be independently annotated by two people to assess consistency. The dataset should be split by farm or collection session, not merely by individual image, so near-duplicate views do not leak into both training and testing sets.

Evaluation should report mean absolute error in piece counts, percentage error, precision and recall for visible-object detection, and separate performance for mother and finger rhizomes. Because percentage error becomes unstable for small counts, absolute error should also be reported. The system should be compared against simple baselines such as sample-weight extrapolation and manual counting. Confidence intervals should be estimated from repeated samples or bootstrap procedures. A model should only be described as validated after testing on independent farm images and publishing its error metrics.

A subsequent phase could integrate count estimates with planting recommendations for Kokan Halad-1, but only after obtaining locally appropriate guidance on planting material quantity, spacing, seed health, and expected yield. This would transform a visual counting prototype into a more complete decision-support tool.

10. Conclusion

The Mangaon case demonstrates both the potential and the current limitations of AI-assisted agricultural estimation. From one photograph, a general-purpose AI assistant generated a preliminary range for turmeric rhizome quantities. The farmer’s later approximate count—8,500 mother rhizomes and 17,500 finger rhizomes—showed that the initial total range was too low, with a particularly large error for mother rhizomes. The near match between the finger-rhizome count and the upper bound is encouraging, but it does not establish overall accuracy.

The central lesson is methodological: a persuasive AI explanation is not the same as a validated measurement. Reliable farmer-facing systems should combine computer vision with scale calibration, category-specific sample weights, direct weighing where possible, transparent uncertainty, and independent testing. With these safeguards, image-based tools could help farmers make quicker and better-informed decisions about turmeric inventory and planting plans. The Mangaon exercise provides a useful starting point for such a research and development programme, rather than proof that a single photograph can already deliver a precise inventory.

References

Farjon, G., & Edan, Y. (2024). AgroCounters—A repository for counting objects in images in the agricultural domain by using deep-learning algorithms: Framework and evaluation. Computers and Electronics in Agriculture, 222, 108988. https://doi.org/10.1016/j.compag.2024.108988

Dolata, P., Wróblewski, P., Mrzygłód, M., & Reiner, J. (2021). Instance segmentation of root crops and simulation-based learning to estimate their physical dimensions for on-line machine vision yield monitoring. Computers and Electronics in Agriculture, 190, 106451. https://doi.org/10.1016/j.compag.2021.106451

Farjon, G., Huijun, L., & Edan, Y. (2023). Deep-Learning-based Counting Methods, Datasets, and Applications in Agriculture—A Review. arXiv. https://arxiv.org/abs/2303.02632

Li, H. H., Irgau, M., Janmohamed, N., Rieckmann, K. S., & Lobell, D. B. (2026). Scalable Vision-Guided Crop Yield Estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 40(1), 615–622. https://doi.org/10.1609/aaai.v40i1.37026