Learn AI Create the Future — with Pranay
Main site

Satellite and Agricultural AI

AI for smallholder farms

Most of the world's farms are smaller than a single satellite pixel and their farmers may not own a smartphone, so techniques built for large mechanised farms often do not transfer, and that gap is the actual engineering problem.

Read these first

On this page 6
  1. Why this needs its own lesson
  2. How the mismatch actually looks
  3. Where this shows up
  4. An honest warning
  5. Remember this
  6. What to learn next

One lesson, three depths. Pick the one that fits you today — you can switch any time.

Beginner — No maths. Plain English.

Most of the techniques earlier in this section were built for large farms, and most of the world's farms are not large.

Think of watching a film on your own phone, with a full data plan. Now compare that to watching one on a neighbour's shared phone, which only gets signal near the edge of the village. Both are "using a smartphone". The second changes what you can realistically build for that person.

Farming has the same split. A mechanised farm spanning many hectares, with reliable internet and money for equipment, is one world. A family farming under a hectare, sharing one phone with patchy signal, is a different world. It is also the more common one globally.

Why this needs its own lesson

Everything covered so far in this section — reading bands, computing NDVI, detecting change, predicting yield, spot spraying — assumes conditions that hold for large farms. Those conditions quietly break for small ones.

Free satellite pixels are often bigger than the whole field. A public satellite pixel covers 10 to 30 metres on a side. Many smallholder plots are smaller than that. One "pixel" of NDVI can average across five different farmers' fields, growing five different crops. Nobody can tell whose field is whose from that one blurred number.

Precision equipment costs more than the farm is worth. A spot-spraying rig, covered in the previous lesson, is built for large mechanised fields. There, the herbicide saved pays for the machine. On a tiny plot, the maths never works out.

Models trained on the wrong crops. Public datasets and pretrained models lean heavily toward globally traded staple crops grown in large monoculture fields — wheat, maize, soy. A smallholder farm intercropping several vegetables in one small plot looks nothing like that training data.

How the mismatch actually looks

Large farm            Smallholder farm
one crop,             several crops in
many satellite        one plot, smaller
pixels across it      than a single pixel
     |                        |
     v                        v
 clean per-field         one blurred number
 NDVI signal             covering several
                         different farms at once

Where this shows up

  • Agricultural extension programmes in India and across Africa deliberately choose SMS and voice-based advisory tools over apps. A basic phone with no internet is what many farmers actually have.
  • Government and NGO efforts to build crop-type maps for smallholder regions. These need far higher-resolution imagery — often expensive commercial imagery — than free public satellites give.
  • Microfinance and crop insurance products designed for small landholdings face exactly the resolution problem above, when trying to verify individual small claims from satellite data.

An honest warning

A tool built for a thousand-hectare farm is not automatically useful once shrunk down for a half-hectare one. It may not even be cheap or fair at that scale. Sometimes the right answer is a genuinely different, lower-tech design — an SMS advisory service instead of a satellite dashboard.

Building AI for smallholder agriculture is as much a design and access problem as a modelling problem. A more accurate model that nobody can afford or reach helps nobody.

Remember this

  • Free satellite pixels are frequently larger than an entire smallholder field, which breaks field-level analysis by default.
  • Equipment and connectivity costs that a large farm can absorb are often out of reach for a small one.
  • The right solution for smallholder farms is sometimes a different kind of tool entirely, not a scaled-down version of the large-farm one.

What to learn next

Developer — Code and libraries.

Setup

bash
pip install numpy

Minimal runnable code

This demonstrates the "mixed pixel" problem directly: several small, differently-planted plots sitting inside the footprint of one or two satellite pixels.

mixed_pixel_demo.py
import numpy as np

# Ground truth: a 30m x 30m stretch of land at 1m resolution, containing
# five separate smallholder plots side by side, each growing something different.
# NDVI values are illustrative, not measured.
ndvi_1m = np.zeros((30, 30))
plot_bounds = [(0, 6), (6, 11), (11, 18), (18, 24), (24, 30)]   # column ranges, in metres
plot_ndvi =  [0.75,     0.20,     0.55,      0.10,      0.68]   # true NDVI per plot

for (start, end), value in zip(plot_bounds, plot_ndvi):
    ndvi_1m[:, start:end] = value

print("ground truth, one row across all five plots (1m resolution):")
print(np.round(ndvi_1m[0], 2))

# What a single 10m-resolution satellite pixel reports: the average over its 10x10m footprint
pixel_10m = ndvi_1m[0:10, 0:10].mean()
print(f"\none 10m satellite pixel covering the first ~1.5 plots: {pixel_10m:.2f}")
print("(true values underneath that single pixel: okra-like plot 0.75, fallow patch 0.20)")

# A single 30m-resolution pixel (e.g. Landsat) covering this entire stretch
pixel_30m = ndvi_1m.mean()
print(f"\none 30m satellite pixel covering all five plots at once: {pixel_30m:.2f}")
print("(that one number gets reported as 'the' NDVI for five different farmers' fields)")
Output
ground truth, one row across all five plots (1m resolution):
[0.75 0.75 0.75 0.75 0.75 0.75 0.2  0.2  0.2  0.2  0.2  0.55 0.55 0.55
 0.55 0.55 0.55 0.55 0.1  0.1  0.1  0.1  0.1  0.1  0.68 0.68 0.68 0.68
 0.68 0.68]

one 10m satellite pixel covering the first ~1.5 plots: 0.53
(true values underneath that single pixel: okra-like plot 0.75, fallow patch 0.20)

one 30m satellite pixel covering all five plots at once: 0.47
(that one number gets reported as 'the' NDVI for five different farmers' fields)

What actually happened

No single farmer's field in this simulation actually has an NDVI of 0.53 or 0.47. Those numbers are averages the satellite reports because its pixel footprint straddles several fields at once — none of the true values in plot_ndvi are anywhere near either reported number.

  • The 1m grid stands in for "the truth on the ground" — what you would measure if you could walk the field with a handheld sensor.
  • ndvi_1m[0:10, 0:10].mean() reproduces what a 10m satellite pixel actually measures: the average reflectance over its whole footprint, not a value for any one plot inside it.
  • This is not a bug in the satellite or the NDVI formula. It is an unavoidable consequence of pixel size being larger than the thing you want to measure.

Common mistakes

Applying large-farm NDVI thresholds directly to smallholder pixels. A threshold like "NDVI above 0.3 means healthy crop", from NDVI and vegetation indices, assumes one crop fills the pixel. A mixed pixel can sit right at a misleading middle value even when every individual plot inside it is either very healthy or completely bare.

Assuming higher-resolution imagery always exists for free. Some commercial satellites offer sub-metre resolution, but at a real cost per image, which is why most smallholder-focused programmes either pool budget across many farmers or use drones for the final, small-scale layer instead of relying on free satellites alone.

Designing the interface for a smartphone user by default. A dashboard app is the wrong assumption when the intended user shares one basic phone with a family, and connects to the internet rarely. This is a product decision, not a modelling one, but it matters as much to the system actually working.

Try it yourself

Shrink the plots: change plot_bounds so all five plots fit inside (0, 15) instead of (0, 30), keeping their relative widths. Recompute the 10m pixel average and see how much further from any true plot value it drifts as plots get smaller relative to the pixel.

What to learn next

Researcher — Mathematics and papers.

The mixed-pixel problem, formally

A satellite pixel's recorded reflectance is the area-weighted average of the true reflectance of everything within its footprint, weighted further by the sensor's point spread function (PSF) — the sensor does not sample a sharp square, it integrates light with a smooth, roughly Gaussian-shaped weighting that extends somewhat beyond the nominal pixel boundary:

R_pixel = (1 / A) * integral over the footprint of  R_true(x, y) * PSF(x, y) dA
  • R_pixel — the reflectance value the sensor reports for one pixel.
  • R_true(x, y) — the true reflectance at each point on the ground.
  • PSF(x, y) — the sensor's point spread function, describing how much each ground location contributes to the recorded value.
  • A — the effective integration area.

When field size is comparable to or smaller than pixel size, R_pixel is a mixture dominated by whichever land cover occupies the largest share of the footprint, and no per-field signal can be recovered from a single band's reflectance alone.

Sub-pixel and super-resolution approaches

Spectral unmixing treats R_pixel as a linear combination of a small number of pure "endmember" spectra (e.g. bare soil, healthy crop, water) with unknown mixing fractions, and solves for those fractions via constrained least squares. This recovers approximate land-cover proportions within a pixel, not the spatial arrangement — it cannot say where within the pixel each fraction sits.

Pansharpening and super-resolution fuse a high-resolution panchromatic band with lower-resolution multispectral bands (classical pansharpening, e.g. Landsat's 15m panchromatic band with its 30m multispectral bands), or apply learned super-resolution models trained to predict fine detail from coarse multispectral input. Both approaches improve apparent spatial detail but do not create true new ground information the sensor never captured, and can introduce artefacts that look plausible but are not measured reality.

Very high resolution commercial imagery (Planet Labs at 3-5m, Maxar at sub-metre) resolves individual smallholder plots directly, at real per-image cost that has historically limited its use to funded pilot programmes rather than routine free access.

Field-size distributions and scale of the problem

Lowder, Sánchez and Bertini (2021), analysing agricultural census data across 167 countries, estimated that farms under 2 hectares account for roughly 84% of all farms worldwide by count, while occupying a much smaller share of total farmed area — a small share of hectares, but the majority of the world's farming households. This is the scale at which the mixed-pixel problem is not an edge case but the typical case globally.

Access and connectivity constraints

Beyond imagery resolution, deployment constraints documented in agricultural-development literature include: intermittent or absent mobile data connectivity in rural regions, shared rather than individually owned devices, literacy and language barriers to app-based interfaces, and limited capital for any equipment beyond what is already owned. These motivate the widespread use of SMS and interactive voice response (IVR) systems — such as India's Kisan Call Centre network and various NGO-run voice advisory lines — as the actual deployed interface, rather than smartphone apps, for many large-scale agricultural-AI programmes targeting smallholders.

Papers and sources

  • Lowder, S. K., Sánchez, M. V., Bertini, R. (2021). Which Farms Feed the World and Has Farmland Become More Concentrated? World Development.
  • Jain, M. et al. (2016). Mapping Smallholder Wheat Yields and Field Sizes in a Semi-Arid Region in India Using Multi-Temporal Landsat and LiDAR Data. Remote Sensing of Environment — a direct case study of the resolution problem in Indian smallholder agriculture.
  • Debats, S. R. et al. (2016). A Generalized Computer Vision Approach to Mapping Crop Fields in Heterogeneous Agricultural Landscapes. Remote Sensing of Environment — field-boundary delineation specifically for small, irregular fields.
  • FAO (2021). The State of Food and Agriculture — annual reporting on smallholder access to digital agricultural tools globally.

Current state

High-resolution commercial imagery, cheaper small drones, and mobile-first low-bandwidth interfaces are narrowing this gap but have not closed it. Deployed systems at meaningful scale for smallholder farmers still lean on non-satellite signals — farmer-submitted photos, SMS-based reporting, community extension workers — as much as on remote sensing, precisely because the imagery-resolution and connectivity constraints described above remain real engineering limits, not solved problems.

What to learn next