The framework, called Systematic Vegetation Index Discovery, or SVID, takes a deliberately simple approach. Instead of relying on expert intuition to design an index, it exhaustively combines the blue, green, red and near-infrared bands of the Sentinel-2A satellite using basic arithmetic operations: addition, subtraction, multiplication and division. The expressions were constrained to a single arithmetic statement with no nested operations, keeping every candidate physically interpretable and computationally cheap. Of 1,328 candidate expressions written down before any analysis began, 1,144 produced numerically valid rasters when applied pixel by pixel to real imagery.

The test bed was a 323.9-hectare agricultural landscape near Andong City in North Gyeongsang Province, a site chosen because it contains three spectrally distinct land-cover types: flooded rice paddies, pepper fields and plastic-covered greenhouses. The researchers drew on 32 cloud-free Sentinel-2A images acquired during 2022, all at the satellite’s sharpest 10-meter resolution, to capture the full phenological arc of the growing season. Rice paddies, for example, show an NDVI near zero in mid-May while still flooded, then climb steeply to a peak of 0.61 in late August. Pepper follows a later, sharper trajectory, while greenhouses show no seasonal cycle at all, because the plastic cover, not the plants, dominates the reflected signal.

Each of the 1,144 indices was then fed to its own Random Forest classifier, trained on the index’s 32-date temporal profile and evaluated against 500 independent reference points that played no role in training. The authors were careful about spatial autocorrelation: training pixels formed contiguous blocks within 49 field parcels, and the median distance from a validation point to the nearest training pixel was 85 meters, an arrangement designed to produce an honest estimate of accuracy rather than the inflated scores that random cross-validation can give on spatially structured data.

The headline result is striking. The best-performing index, (NIR − Red)/(Green × Blue), achieved 97.0 percent overall accuracy and a Kappa coefficient of 0.873, compared with 94.4 percent and 0.734 for NDVI. A McNemar test on the paired predictions confirmed the difference was significant at p = 0.002. NDVI, remarkably, ranked 89th out of 401 distinct features, landing in only the 22nd percentile of the pool the researchers had generated. Several families of near-infrared-to-visible ratios outperformed the classic index, with the top four ranked features all significantly better than NDVI.

Where did the improvement come from? Almost entirely from one class. For rice, which dominates the validation set, every index tested exceeded 96 percent producer’s accuracy. But for greenhouses, the class NDVI handled worst, producer’s accuracy jumped from 58.6 percent to 84.5 percent with the best index. The explanation lies in the physics of the scene. NDVI rests on a single red-versus-near-infrared contrast, which separates greenhouses poorly from sparsely vegetated paddies before transplanting. Indices that also weight the blue and green bands exploit the distinctive brightness of the plastic cover across the visible spectrum, sharpening the distinction. On the classified maps, the NDVI-based product assigned 25.5 hectares to greenhouses while the new index assigned 36.4 hectares, with the difference drawn almost entirely from pixels NDVI had mislabeled as rice.

Perhaps the most sobering finding concerns redundancy. Before any ranking, the researchers discovered that 388 of their expressions were algebraically identical to another expression, and a further 355 were strictly monotone transformations, such as reciprocals or sign reversals, that a rank-based classifier like Random Forest cannot distinguish. The 1,144 indices therefore encoded just 401 genuinely different features, meaning roughly two-thirds of the apparent search space carried no additional information. The largest equivalence class contained 18 different ways of writing the same quantity.

This redundancy is not merely cosmetic. Two algebraically identical formulations of the same feature scored 97.60 and 95.60 percent overall accuracy respectively, a 2.0 percentage-point gap that cannot reflect any difference in spectral information and instead arises from finite-precision arithmetic and tie-breaking inside the classifier. Across all equivalence classes containing multiple formulations, the median within-class spread was 0.6 percentage points and the maximum reached 3.4 points. Refitting the same index with different random seeds shifted accuracy by up to 1.1 points, the same order of magnitude as the differences separating the top ten features. Ten of the 401 distinct features were statistically tied with the best.

The study also clarifies what automated index discovery can and cannot deliver. A reference model trained on all four raw bands across all 32 dates, 128 features in total, achieved 96.0 percent accuracy, statistically indistinguishable from the best single index. Index discovery therefore does not recover information absent from the reflectance data; its practical contribution is a four-fold reduction in dimensionality, from 128 band-date features down to 32 temporal features of a single index, at no measurable cost in accuracy, together with an explicit algebraic form that can be inspected, transferred between sensors and implemented where computing power is limited. Notably, the raw-band model achieved the highest producer’s accuracy for both minority classes, indicating that compressing a multi-band time series into one index trades balanced class performance for parsimony.

The authors draw three practical lessons for anyone tempted to automate index design. First, expression pools should be canonicalized before evaluation, collapsing algebraic duplicates and monotone equivalents, a step that would have removed 65 percent of this pool. Second, accuracy estimates should always come with confidence intervals and paired significance tests against a named baseline, because with 500 reference points, differences smaller than about two percentage points cannot be resolved. Third, results should be reported as index families with rank stability rather than as a single winner. The study’s limitations are acknowledged: one site, one year, four spectral bands, and a validation set so imbalanced that the pepper class, with just 13 points, cannot distinguish any index from another. Still, the message is clear and likely to resonate across the field: NDVI is not the most informative index for every landscape, systematically discovered near-infrared-to-visible ratios can do measurably better, and the era of letting machines search the spectral design space, carefully and with honest uncertainty, has arrived.

Subject of Research: Systematic discovery of vegetation indices for phenology-based crop classification with Sentinel-2 imagery

Article Title: Systematic vegetation index discovery for phenology-based crop mapping

Article References: Rahimi, E., & Jung, C. (2026). Systematic vegetation index discovery for phenology-based crop mapping. Frontiers in Remote Sensing, 7, Article 1911652. https://doi.org/10.3389/frsen.2026.1911652

Image Credits: AI Generated

DOI: 10.3389/frsen.2026.1911652

Keywords: vegetation indices, NDVI, Sentinel-2, crop mapping, phenology, remote sensing, Random Forest, greenhouse detection, rice paddies, feature redundancy, accuracy assessment, machine learning