WHAT THE EVIDENCE SAYS
A conditional generative image-augmentation pipeline with EfficientNet-B0 reached reported classification accuracy 98.48% on its forest-scene dataset, versus 97.21% for a less optimized conditional model in the author's ablation comparison.
WHY IT MATTERS / OUR READ
Classification categories and predictive holdouts are not field ecology. Geography-separated, independently labeled imagery and on-ground verification must precede restoration or habitat-condition claims.
CAVEATS / UNCERTAINTY
Only one image dataset was used; the researchers note possible limited generalization to varied real forest and sensor conditions.
SOURCE
Lu et al. · Discover Artificial Intelligence · forest scene classifier
View original evidence ↗