Range and AGC normalization in airborne discrete-return LiDAR intensity data for forest canopies

Recently, the intensity characteristics of discrete-return LiDAR sensors were studied for vegetation classification. We examined two normalization procedures affecting LiDAR intensity through the scanning geometry and the system settings, namely, range normalization and the effects of the automatic gain control (AGC) in the Optech ALTM3100 and Leica ALS50-II sensors. Range normalization corresponds to weighting of the observed intensities with the term (R/RRef)a, where RR is the range, RRef is a mean reference range, and a∈[2,4]a∈[2,4] is the exponent that is, according to theory, dependent on the target geometry. LiDAR points belonging to individual tree crowns were extracted for 13 887 trees in southern Finland. The coefficient of variation (CV) of the intensity was analyzed for a range of values of exponent aa. The tree species classification performance using 13 intensity variables was also used for sensitivity analysis of the effect of aa. The results were in line with the established theory, since the optimal level of aa was lower (a≈2)(a≈2) for trees with large or clumped leaves and higher (a≈3)(a≈3) for diffuse coniferous crowns. Different echo groups also showed varying responses. Single-return pulses that represented strong reflections had a lower optimal value of aa than the first and all echoes in a pulse. The gain in classification accuracy from the optimal selection of the exponent was 2%–3%, and the optimum for classification was different from that obtained using the CV analysis. In the ALS50-II sensor, the combined and optimized AGC and RR normalizations had a notably larger effect (6%–9%) on classification accuracy. Our study demonstrates the ambiguity of RR normalization in vegetation canopies.