First I'm going to share a bit of background on modeling forest biomass using optical satellite imagery, including some of the methods and data I have used.

Second, I'm going to describe one common limitation of this approach, the greening-up before growing-up problem, and why it is important to address.

Third I'm going to discuss how we compared post-harvest recovery rates between Landsat-derived predictions and forest inventory data.

Finally, I'm going to share how we are thinking about addressing this issue in our modeling approach given the insights gained through this analysis.

So why use landsat to model forest biomass? Data sources like LiDAR are better at representing forest structure, while optical imagery like Landsat is only able to measure the spectral reflectance of the surface, or in other words the color.

Landsat offers the longest history of publicly available remote sensing data, from 1984 to present day, and has repeat observations for the same location roughly twice a month. These data are available everywhere, and Landsat missions are well supported meaning that any methods developed now can likely applied for future monitoring.

The approach co-locates Landsat-derived spectral indices and disturbance related information with field inventory plots on the ground. This combined dataset is then used to train machine learning models for aboveground biomass prediction.

So, our models are dependent on measures of spectral reflectance, or 'greenness', to predict forest biomass. The problem with this is that greenness doesn't account for structure. Vegetation close to the ground floor can and does appear quite green.

Following a harvest, a stand is likely to "green-up" relatively quickly. Within a few years. This of course depends on the type of harvest, but without some adequately established advanced regeneration, this quick greening up is likely to be the result of early-successional species, or small diameter saplings.

However, we know that it may take a harvested stand many years to "grow up", or return to pre-harvest conditions. Because of the type of remote sensing data that informs our model predictions, the models are likely to assume that a stand has snapped back to pre-harvest levels of biomass much faster than is reasonable. This is the "green-up before grow-up" problem.

Here is a time series of RGB composites derived from 3 of the spectral indices used as predictors our model. The scene is of a working forest in the Adirondack region and spans 1990 to 2019. You can see how quickly the spectral indices revert to greenness after the appearance of brown harvest patches in the early 90s and near 2010. It follows that our modeled recovery rates would be fast given these kinds of inputs.

Here are a set of predicted biomass maps for the same area and time-span. The map on the left shows pixel predictions, with LCMAP landcover classifications overlaid. Near-zero predictions are white, while larger predictions, upwards of 200 metric tons per hectare are in dark green. The bar chart on the right shows the total biomass in metric tons contained in this scene for each year. As I play this video, and the predictions progress through time we can see a few harvests. One in the early 1990s and a larger one near 2010 both of which recover to pre-harvest levels of biomass in 5-10 years. I'll play it again...

This problem is particularly relevant to carbon accounting as sequestration rates are more important than carbon stocks. Faster rates of carbon sequestration would mean faster atmospheric removals, and so rates of recovery will have an impact on how working forests are assessed as components in any greenhouse gas budgeting framework. While we would all like for forests to accumulate carbon at lightspeed, we would rather have our predictions be accurate.

We compared our modeled recovery rates to those measured through networks of repeated forest inventory plots. We used both FIA plots across the entire state, as well as CFI plots from SUNY ESF forest properties. The density of forest plots in the CFI network, and long history of measurement dating back to the 70s proved to be more useful in this analysis.

We identified harvested plots as those which lost at least 30% of their plot-level biomass between inventories. Additionally, we ensured that plots were not disturbed a second time, by excluding plots that lost more than 5% of their biomass between post-harvest inventories.

So here we can see the average post-harvest growth trajectories for our three data sources. The y-axis marks cumulative plot-level biomass gain post-harvest in metric-tons per hectare, and the x-axis represents time since harvest. There are different post-harvest time-horizons available for each dataset which is a function of data availability and our plot-selection criteria.

As expected, the landsat-derived growth-rates are much faster in the first 5-10 years post harvest. However, the rates actually slow, down, and seem to "correct themselves" in comparison to the CFI trajectory at larger time-horizons of 20-25 years.

Here we show the same data, excluding FIA, and with the addition of 95% confidence intervals around our recovery rates. For our Landsat-derived predictions, we simply don't have enough observations beyond 10-15 years post-harvest to produce a narrow range of confidence.

While the whole post-harvest time-series might not be off, the outsized early rates of carbon sequestration are significant, and could have a large impact on how working forests are assessed in greenhouse gas budgeting.

So, we have identified this problem, and have a decent way to assess the extent of it, but we are still thinking about how to address it in our modeling approach.

One solution is to produce initial model predictions as described and shown here, allowing these predictions to identify when harvests have occurred in the time-series.

Then, following a harvest we might apply a different recovery trajectory, one developed with CFI, or other forest inventory data, to more accurately represent forest growth.

Ideally we would be able to apply location or environmentally relevant trajectories as we expect the growth rates to vary widely depending on site conditions. This will require further collection of forest inventory data, as FIA plots are too sparsely distributed and our SUNY ESF CFI data is geographically limited. We hope to engage in this process with partners who might be willing to share detailed historical inventory records.

Thanks to everyone for listening! Here are a few links to where you find me or these slides later. I welcome any questions and am open to any perspectives anybody might want to share.