Multiple Testing in Remote Sensing: Addressing the Elephant in the Room.

Abstract

In environmental remote sensing, statistical tests are often conducted at the pixel level, generating thousands of p-values and substantially increasing the risk of Type I errors. Traditional multiple testing corrections aim to control the probability of any false positives, but often at the cost of drastically reduced statistical power. In contrast, false discovery rate (FDR) control limits the proportion of false positives among significant results while preserving power. It is widely used in image-based disciplines such as neuroimaging and, more broadly, medical imaging, where error control is critical. However, its adoption in remote sensing remains unclear. We conducted a Scopus-based review of twenty years of literature (2004–2023) in remote sensing journals to assess the use of FDR control in spatiotemporal trend analyses, with particular attention to studies employing the Mann-Kendall test, a non-parametric method widely used for detecting monotonic trends at the pixel level. Our results reveal that only 0.03% of remote sensing articles cited the seminal Benjamini–Hochberg (1995) method, and none of the studies using pixel-wise Mann-Kendall testing explicitly applied FDR control. This striking absence highlights a critical risk of overestimating significance in large-scale remote sensing analyses, especially as increasing data precision demands more rigorous error control. We call for routinely incorporating FDR procedures into remote sensing workflows to control the inflation of Type I error in large-scale multiple testing.