Tree Health Monitoring Using Multispectral Images

Abstract

Background: Walking an orchard or a forest plot tree-by-tree to check for stress is slow, and two people rarely agree on borderline cases. Multispectral cameras, mounted on drones or carried by satellites, pick up reflectance outside what the eye can see, which makes it possible to catch stress in a canopy before the leaves actually start to show it. Methods: We put together a pipeline for tree health monitoring built around five-band UAV imagery (RGB, red-edge and near-infrared), standard preprocessing into an orthomosaic, a set of vegetation indices (NDVI, GNDVI, NDRE, SAVI) computed per tree crown, and two classifiers — a Random Forest trained on the indices alone, and a CNN given the full stacked band-plus-index image. The design choices were informed by a review of publicly available multispectral orchard and forest datasets rather than a single in-house collection effort. Results: NDVI and NDRE turned out to be the two indices that separated healthy from stressed crowns most cleanly, and the CNN working on the full band stack did better than the index-only Random Forest — a pattern that lines up with what recent multispectral orchard studies have reported. Conclusion: Multispectral imaging paired with a learned classifier is a workable, non-destructive way to catch tree stress early at scale, though the field still lacks standardized labelled datasets and models tend not to transfer cleanly across species.

Keywords