Forest and agroecosystem biomonitoring

Landscape-scale biomonitoring produces maps of quantities no satellite or sensor ever measures directly: biomass, species occurrences, water held at root depth. Every product is the last link of a proxy chain, and the error budget is dominated not by instrument precision but by calibration — how tightly the proxy was tied to the ground, and how well that tie survives elsewhere and later.

The proxy chains#

Optical sensing exploits plant physiology: chlorophyll absorbs red light, leaf structure scatters near-infrared, so a spectral index stands in for green cover and photosynthetic activity. The proxy saturates in dense canopy and is confounded by atmosphere, phenology and background soil. LiDAR is more direct physics: the return time of a laser pulse gives distance, so canopy height and vertical structure are measured, not inferred — but biomass still requires an allometric model that converts height and structure into mass, and that model is fitted on harvested trees and field plots. Soil-moisture probes read the dielectric permittivity of the substrate, which depends strongly on water content — but the permittivity-to-water conversion shifts with texture and organic matter, so a probe is calibrated per soil type, not per product line. Environmental DNA works by a different chain altogether: organisms shed tissue and cells into water, soil and air; primers amplify the fragments; sequences are matched against a barcode reference database. What comes out is detection weighted by the probability of detection — shaped by primer match, library completeness, DNA degradation and transport — not a census.

Why ground truth dominates the error budget#

The absolute accuracy of every mapped quantity is set by the field data used to train and validate its retrieval. Those field data carry their own errors and scale mismatches: a tree measured by tape, a plot that covers a fraction of a pixel, a pixel that averages many crowns and soil patches. Representativeness is the sharper failure — a retrieval calibrated in one forest type or one soil and applied in another inherits a bias no amount of down-stream statistics removes. Drift is the error’s time dimension: sensors age and their response shifts, atmospheric state differs between scenes, and an eDNA reference database keeps growing, so the same sample processed two years apart can return a different species list — the measurement changed although the forest did not. Sensor fusion compounds the accounting: when sparse LiDAR footprints are extended using optical and radar imagery, each fused layer contributes its own calibration error, and the total error is only ever as good as the worst-calibrated layer.

The limiting quantity#

The scarce resource is fresh, representative ground truth. Everything else in the stack is engineering; a cheaper or higher-resolution sensor does not repair a biased calibration, and no algorithm does. This is worth holding onto wherever monitoring acquires consequences — carbon inventories and compliance reporting convert the calibration error directly into an error in tonnes and in obligations, which is why verification costs concentrate in field campaigns rather than in computing. What those measurements are verifying — whether a restoration trajectory is actually converging on its reference — is the subject of ecosystem restoration; the detection-versus-abundance limits of eDNA reappear on the population side in biodiversity genomics.