search

weecology / cross_scale_structural_diversity

git diary for investigating how scale affects ecological drivers implications on scaling rules and forest structure
MIT License
0 stars 0 forks source link

Evaluate the usability of current state of the art DeepForest boxes #4

Open MarconiS opened 3 years ago

MarconiS commented 3 years ago

DeepForest is impressively good approach, but may not be enough in reliably use crown estimates at the individual level. For example, extracting boxes for NEON VST points taller than 3m, we can notice uncertainty rising as the crown radius gets bigger. In particular, there seems to be an under-representation of the crown size (Fig 1). This error propagates in the DBH-CR relationships (Fig2). However, it seems that using 20x20 plots average, the effect of this error may be reduced, but it requires a more quantitative backup.

If that is confirmed, makes probably sense to approach the study using the FIA subplot as an elementary unit, also because it would back up the logic behind graph descriptors (proxies of plot level interactions/demography)

MarconiS commented 3 years ago
Screen Shot 2021-01-20 at 10 01 06 AM

Fig 1: relationship between observed - delineated crowns. No chm filter at this stage, despite a 4m delta chm filter would probably be required (to avoid misattributed crowns).

MarconiS commented 3 years ago
4M_DBH_CR_AVE

Fig 2 correlation between Crown Diameter (CD) and DBH from field (first row) and delineated (second row). Values are averages at the plot level.