Spectral-temporal modelling of bamboo-dominated forest succession in the Atlantic Forest of Southern Brazil

With access to collections of continuous satellite imagery over a 40-year period, spectral-temporal patterns extracted from multi-temporal imagery offer a potential new tool to model mechanisms of forest succession and monitor changes in forested landscapes. Specifically, spectral-temporal trajectories associated with successional forest change occurring over prolonged periods of time may enhance periodic ‘snapshot’ monitoring methods, especially for species that exhibit complex and non-linear dynamics. In this paper, Landsat time-series are used to examine the spectral-temporal signatures of bamboo-dominated forest succession occurring within the critically threatened Araucaria Forest, a pine-dominated subtype of the Atlantic Forest in southern Brazil. Alteration of canopy structure through ongoing anthropogenic disturbance has increased understorey light climate and given opportunity for native invasive bamboos to flourish, resulting in drastic reduction of tree regeneration and loss of biodiversity. We aimed to evaluate how spectral-temporal signatures could be used to (1) characterize stages of bamboo-dominated forest succession, (2) identify synchrony of bamboo lifecycle dynamics and (3) classify regions of bamboo-dominated forest. Changepoint analysis was performed using an extracted sample spectral-temporal signature and trajectories were fit to the resulting segments using linear regression. Based on slope values of the fitted segments, a novel description incorporating temporal information of bamboo-dominated forest succession was developed which identified four broad phases: pioneer predominance, mature bamboo, dieback and pioneer regeneration. To determine the spatial and temporal synchrony of bamboo-dominated forest succession, a hybrid model was developed by combining the modelled segments and compared to a 32-year Landsat time-series of vegetation indices by calculating root-mean square error between each pixel in the study area. The hybrid model proficiently classified regions of bamboo-dominance, achieving between 77% and 90% accuracy, which also indicated lifecycle synchrony of bamboo populations within the study area. To further assess the performance of the hybrid model, a time-weighted dynamic time warping model approach was used to determine synchrony and classify regions of bamboo. The time-weighted dynamic time warping classifier had lower overall accuracy (68%–82%), but is still considered a useful tool for automated classification purposes that take advantage of multi-temporal imagery. To compare classification performance between ‘snapshot’ and multi-temporal imagery classifiers, a maximum-likelihood classification was performed, which attained lower overall accuracies than the hybrid model (75%–84%). Overall, the use of spectral-temporal signatures offers a novel and effective approach to both describing and modelling bamboo-dominated forest succession (and forest successional processes more generally) on a landscape-scale.