The Mann-Kendall Trend Test, Explained for Remote Sensing

Fit a straight line through ten years of NDVI values and you'll get a slope — positive, negative, or flat. But a slope alone doesn't tell you whether that trend is real or just the kind of up-and-down wobble you'd see in any noisy time series, even one with no underlying trend at all. That's the specific gap the Mann-Kendall test fills.

The problem with slope alone

Imagine ten years of NDVI values for a single pixel. A linear regression will always produce some slope — it's a mathematical guarantee, not a finding. If the values happened to drift upward in years 1–4 purely by chance, then downward in years 5–10, the regression line might still show a mild positive or negative slope depending on the exact numbers, without there being any real underlying process driving the change.

This matters a lot in remote sensing specifically, because year-to-year satellite measurements carry real noise — from atmospheric conditions, sensor calibration drift, slightly different acquisition dates or angles, and genuine short-term weather variation. A slope computed from ten data points, several of which are noisy, needs a way to say: is this pattern distinguishable from randomness, or not?

What Mann-Kendall actually tests

The Mann-Kendall test is a non-parametric method — it doesn't assume the data follows a normal distribution, which matters for satellite time series that often don't. Instead of fitting a line, it looks at every pair of values in the sequence and asks a simple question for each pair: is the later value higher or lower than the earlier one?

Each pair contributes +1 if the later value is higher, −1 if lower, and 0 if tied. Sum these up across all pairs to get the statistic S. If the data has a genuine upward trend, most pairs will show later-higher-than-earlier, and S will be strongly positive. If there's no trend, positive and negative pairs roughly cancel out, and S stays close to zero.

The test then compares S to what you'd expect from pure chance, using the known statistical distribution of S for a dataset of that size, producing a Z-score. A common threshold is |Z| ≥ 1.96, corresponding to the standard 95% confidence level (p < 0.05) — if the Z-score clears that bar, the trend is considered statistically significant, not just a slope that happened to come out non-zero.

Why this matters for a map, not just a single pixel

In a spatial trend map — showing, say, vegetation trend across an entire district over 15 years — every pixel gets its own trend line and its own Mann-Kendall test. This produces two layers of information that are easy to conflate but need to stay separate:

• Magnitude — how steep the trend is (the slope), typically shown as a red-to-blue color gradient.
• Significance — whether that particular pixel's trend clears the Mann-Kendall threshold, typically shown as a mask or overlay.

A pixel can show a large slope that is not statistically significant (common with short time series or highly variable land cover, like cropland with irregular irrigation), and a pixel can show a small but statistically significant slope (common with a long, consistent time series and low year-to-year noise, like stable forest).

Practical thresholds

Mann-Kendall's statistical power depends heavily on how many time points you feed it. As a rough guide:

• Fewer than 8 years: the test has low power — real trends may fail to reach significance even when they exist.
• 10–15 years: a reasonable minimum for most vegetation and land-surface temperature trend work.
• 20+ years: where the test performs best, which is part of why harmonized multi-sensor archives (stitching together Landsat 5, 7, 8, and 9) matter — they extend usable time series back to the 1980s instead of starting at 2013 or 2015.