Uncovering forest dynamics using historical forest inventory data and Landsat time series

[Display omitted] •First use of imputation technique to support forest monitoring in Ukraine.•Landscape-level (21.65 km2) aggregated prediction accuracy was high.•Forest area increased (2%-4%), though local losses were common.•Greatest growing stock volume increases before 2000.•This study promotes the use of limited historical plot data in forest management. While many countries regularly produce timely and accurate estimates for a range of forest attributes through national forest inventory programs, many have access to no such inventory or their inventory is discontinuous in space and time, such as a sample-based approach in Ukraine limited to two regional (Ivano-Frankivsk and Sumy) inventories conducted in 2008–2015 at a total area of 37,700 ha. This study addresses the extent to which limited historical forest inventory data may facilitate forest mapping efforts using dense Landsat time series (LTS) and mapping techniques (i.e., classification and imputation) to predict multiyear forest dynamics. We used the Continuous Change Detection and Classification (CCDC) segmentation approach to extract inter- and intra-annual trends in LTS and utilized these fitted trends as covariates in further modeling. We developed random forest (RF) classification models for forest mapping based on field data collection and LTS, and then generated yearly forest maps for 1990–2020. The RF model accuracies were high for both producer’s and user’s accuracy (>0.93 ± 0.04). Based on yearly forest dynamics, we detected an increasing trend in forest loss during 2000–2010 and 2010–2020 in our regions. We used the gradient nearest neighbor (GNN) method to generate species presence (basal area (BA) ≥ 1.0 m2 ha−1) and growing stock volume (GSV) maps at the 30-m pixel level. Our nearest neighbor imputation models achieved better accuracies (Cohen’s kappa > 0.4) for species that are prevalent within regions, occupy distinct geographical areas, and have higher plot BA abundance. To obtain estimates of species abundance with higher R2, we used the GNN model with k = 3 nearest neighbors. Aggregating species based on similar environmental niche resulted in greater R2 values of BA that varied within two regions from low (0.129 and 0.416) for a group of hardwood deciduous species to high (0.531 and 0.712) for coniferous species. The accuracy was systematically better at 5-km hexagonal level, thus we recommend using the maps at coarser aggregations. The study showed that forests accumulated