Agricultural Analysis

Agricultural analysis is a seasonal, time-series workflow: use remote sensing and supporting field data to describe crop condition, within-field variability, and seasonal development at scales a grower or agronomist can act on. It shares its engine with NDVI monitoring — a vegetation index tracked through time — but the question is different. Monitoring asks where has greenness changed; agricultural analysis asks is this crop developing normally, which parts of this field are underperforming, and when should someone go look. That reframing pulls in the crop calendar, field boundaries, and management context, so this page builds on the NDVI workflow rather than restating the index. As always the reasoning lives in the imagery and the decisions you make about it, so it holds on any platform: a browser environment like Google Earth Engine is one convenient way to try each step, and a Python stack (rasterio, xarray, geopandas) is another, but neither changes what the workflow is doing or why.

Frame the agronomic question and the crop calendar

An agricultural workflow is only as useful as the decision it feeds, and in farming that decision is timed. Fix the unit you report on (a field, a management zone, a set of trial plots), the crop and its calendar (planting, emergence, canopy closure, peak, senescence, harvest), and the decision window — scouting during rapid growth, a nitrogen top-dress before a phenological stage, a harvest-readiness call. The crop calendar matters more here than in generic monitoring because “low vigor” only means something relative to where the crop should be on its curve: the same index value is healthy at emergence and alarming at peak canopy. Bring field boundaries in from the start so every later summary is computed inside the parcel rather than smearing across roads and margins. Writing these down turns “check the crops” into a specification with a unit, a stage, and a threshold — the difference between a map that triggers a scouting trip and a pretty picture.

Assemble seasonal imagery and supporting data

Because crop signals move week to week, temporal cadence is as important as pixel size when choosing imagery. The tradeoffs are the ones the Resolution concept lays out: a finer sensor resolves within-field zones and small trial plots but often revisits less often, while a coarser, more frequent sensor tracks the growth curve but blurs parcel detail — and for a crop that changes over a fortnight, a monthly clear-sky record can miss the stage you care about. Confirm the source delivers the red and near-infrared Spectral Bands that vegetation indices need, and prefer calibrated surface reflectance so a value means the same thing across the season, as the Digital Imagery concept explains. For a season sampled densely enough to read a growth curve, Sentinel-2 is the usual optical basis, with NAIP as the occasional sharper look you check a field boundary against or use to see what a within-field anomaly actually is. Where persistent cloud would gut an optical time series — a wet growing season, a tropical region — radar is the complement, because SAR sees through cloud and responds to canopy structure and moisture. Gather the non-image layers now too: weather (growing-degree days, rainfall), soil or irrigation information, crop-type references, and any ground observations, and align every layer to one grid following the Coordinate Systems concept so a field boundary lands on the pixels it actually contains.

Build cloud-screened seasonal metrics

This is where a stack of dates becomes a growth curve. Screen every optical date first — clouds, shadows, and haze corrupt an index and an unmasked cloud edge reads as a spurious vigor swing — following the Cloud Masking concept, then compute the vegetation index the same way on every date exactly as the NDVI workflow insists, so a change reflects the crop and not a change of method. From the screened series derive the metrics agronomy actually uses: a smoothed seasonal trajectory per unit, phenology dates (green-up, peak, senescence), season-integrated greenness as a biomass proxy, and — where you have the bands or radar — a moisture proxy. This temporal reasoning is the Time Series concept applied to a managed crop: the shape of the curve, not any single date, is the measurement, because a field can pass through the same index value on the way up and on the way down.

Compare against the right reference and separate signal from noise

A per-field curve is data, not yet a recommendation, so compare it against an explicit reference: the same field in prior seasons, neighboring fields under similar management, a treatment-versus-control contrast, or the expected trajectory for the crop and planting date. Deciding against what is a real agronomic choice — a field can look poor against a wet-year benchmark and normal against a long-term one, and only the question decides which is honest. Then rule out the confounders, which in agriculture are many and specific: mixed pixels at field edges and in small parcels, crop rotation that makes this year’s field non-comparable to last year’s, residue and bare soil early in the season that depress an index before canopy develops, irrigation and management timing that shift the curve legitimately, cloud or shadow the mask missed, and inconsistent or outdated field boundaries. The habit is the one the NDVI page uses: treat an apparent underperformance as a hypothesis and check whether it survives these ordinary explanations — and whether it persists across several dates — before it becomes a scouting call.

Summarize, validate, and produce actionable outputs

Reduce the screened series to the unit the question named — mean or median index and phenology metrics by field, management zone, crop type, or planting window — so the result is a number that can be ranked and acted on. Where the workflow labels anything (a crop-type map, a vigor-zone classification), validate it with reference samples and a confusion matrix, the accuracy discipline the Classification Basics concept applies to any labelled output; overlaying the zoned raster against vector field records is the Raster vs Vector step that turns pixels into a per-field table. Where such a label is trained rather than thresholded, which fields go to training and which are held back is an agronomic decision as much as a statistical one — neighbouring plots under one operator share variety, planting date, and management, so an accuracy figure earned on a trial block can fall apart on the next grower’s fields; the ML pipeline note works through why that happens and how to draw the split. The deliverables of a run are usually a set: within-field vigor and phenology maps, seasonal charts per field, a ranked list of fields or zones that crossed a threshold and warrant scouting, and a QA note stating which dates were thin, which fields had few clear observations, and where the mask or the boundaries were doing heavy work. Finally, record the run so it repeats: the fields and boundaries used, the crop and calendar, the exact dates and sensors, the masking and reflectance products, the index and phenology definitions, the summary unit, and the reference. With those written down the season can be re-analysed next year, audited by an agronomist, or ported between platforms without quietly changing what it measures — which is the whole point of a concept-first workflow.