Monitoring the needs for weeding and tending operation in seedling stands are urgently required in Nordic forestry. An improved understanding of early stand development and its spatial and temporal characteristics is a prerequisite for achieving this goal. To meet these scientific objectives and the practical needs of decision-makers, we demonstrated the use of Landsat Time Series (LTS) data for monitoring and predicting the timing of regeneration in the forest stands. We evaluated how NBR and NDVI can be used to assess vegetation development in seedling stands with varying site types and to estimate the young seedling stands attributes.
5.1 NBR and NDVI values in young seedling stands
Our analysis focused on seedling stands with typical sizes of 1-3 hectares which is approximately 1-3 pixels in Landsat imagery. This limited resolution makes it challenging to capture the high variation within individual stands. Consequently, relying on a single prediction for the entire stand is not ideal. Stand-level decisions should be based on the distribution of predictions within the stand boundaries. Ultimately, the end-user needs to determine the feasible operational area based on this information.
Changes in NBR and NDVI indices provide valuable insights into the development of young trees after harvesting. Our findings, along with previous research (Schroeder et al. 2011; White et al. 2019; Chirici et al. 2020; Giannetti et al. 2020) indicate that clear-cut areas can be effectively identified using cloud-free imagery captured during suitable times of the year. For stands with excessive density (more than 3000 stems per hectare), weeding and tending are recommended treatments. These interventions are primarily necessary when unwanted vegetation, such as deciduous trees (often Betula sp.), hinders the growth of the desired conifer species.
5.2 Seedling stand forests attributes
While our study identified some trends between NDVI and NBR with seedling stand attributes (e.g., number of coniferous trees, height of deciduous trees, height difference), the indices could not fully capture the structural variation within these young stands. The R-squared values for all relationships were relatively low (0.19 to 0.26), and the RMSE values were high, indicating limited accuracy in predicting specific attributes.
For comparison, a study by Rana et al. (2023) used multispectral airborne laser scanning data and reported similar findings for broadleaved tree modeling (R²: 0.14-0.25, RMSE: 5442-7782), while conifer models achieved better accuracy (R²: 0.30-0.35, RMSE: 1292-2348). Additionally, studies have shown strong predictive ability for tree height, particularly with broadleaved trees (R²: 0.60-0.65, RMSE: 1.1-2.9) and even greater success for conifer heights (R²: 0.74-0.82, RMSE: 0.9-2.5).
These findings suggest that alternative approaches (e.g., laser scanning data and fine resolution UAV images) might be more suitable for detailed structural assessments in seedling stands. Another study by Feduck et al. (2018) found that UAV RGB imagery successfully detected 76% of coniferous seedlings, suggesting that UAV RGB imagery as a promising tool for identifying coniferous seedlings.
5.3 Classifying weeding need
The height difference between coniferous and deciduous trees serves as a valuable indicator for weeding or tending needs (Rana et al. 2023). Using this information for binary classification (weeding needed vs. not needed) resulted in a Cohen's kappa of 0.55. However, it is important to note that there is a great deal of variation within forest stands in the study area. A Landsat pixel covers reflectance information from a mixed seedling stands. This inherent inhomogeneity affects model reliability, particularly for the smaller (1-3 hectares) stands common in Finland. Despite this, the trend in the LTS-based prediction indicates that the need of silvicultural treatment in seedling stands can be mapped. Additionally, low fertility sites are less sensitive to vegetation recovery and poses difficult challenges for LTS -based prediction.
Our study achieved an accuracy of 81% (kappa = 0.55) for predicting tending needs using NBR. Notably, a study by Korhonen et al. (2013) in Joutsa, Finland, reported lower performance (kappa = 0.37-0.38, accuracy = 71-72%) using airborne laser scanning data. Similar trends were observed in other Finnish studies, with accuracy ranging from 54% to 69% and kappa values from 0.27 to 0.34 (Närhi et al., 2008; Miina et al., 2018). These comparisons suggest that our approach using NBR from LTS data offers a promising alternative for identifying stands requiring tending operations.
5.4 Implications for seedling stand monitoring method development
While temporal variations due to phenology or atmosphere can be reduced (e.g., Schroeder et al. 2007), challenges remain in monitoring young stands attributes. NBR, which captures reflectance in both the NIR and SWIR regions, is particularly sensitive to this issue. As vegetation density and canopy complexity increase with stand age, SWIR reflectance decreases due to increased shadowing (Asner and Lobell 2000; Gonsamo and Pellikka 2012). This makes it difficult to accurately track early development in seedling stand, especially considering additional uncertainties from factors like atmosphere, topography, and sun/view angles (Song and Woodcock 2003; Chen et al. 2020). Forest stand data, particularly soil characteristics, could potentially improve results.
Seasonality also plays a significant role in Nordic seedling stands. Observations made in early summer differ greatly from those in late summer due to both vegetation phenology and variations in sun angle during image acquisition. For this reason, Sentinel-2 satellites, with their superior spatial resolution, hold promise for improved prediction compared to Landsat (Korhonen et al. 2017). Despite these challenges, the current Landsat-based approach demonstrates the potential for automated procedures to predict harvest years and identify stands in need of tending. Further development of reliable, large-scale methodologies utilizing satellite technology remains an urgent need for effective seedling stand monitoring.