Wetlands are important for water control, carbon storage, biodiversity and climate resilience, but they are difficult to map because their edges are often gradual and mixed with forest, shrubland or upland. The main problem is that many wetland maps use government polygon inventories as if every 10 m pixel inside or outside a polygon were perfectly correct. In practice, pixels near polygon boundaries can be uncertain, and this can affect both model training and accuracy assessment. It should be noted that another practical problem is data preparation: conventional wetland mapping often needs long satellite time series, cloud masking and many processing steps before a model can be trained. In this study, we address these two problems by using a boundary aware 10 m wetland mapping workflow for the 2024 GeoNB LiDAR covered part of New Brunswick, Canada. We used 64 AlphaEarth satellite embedding bands as the main predictor group, together with LiDAR terrain variables and Sentinel 2 spectral bands. We also removed a 20 m zone around wetland polygon edges from training and testing so that uncertain boundary pixels did not control the results. Two models were tested: an ExtraTrees classifier and an Attention U Net segmentation model. The ExtraTrees model reached a wetland F1 score of 75.17% and an overall accuracy of 96.51%. The Attention U Net model reached a wetland F1 score of 75.78% and an overall accuracy of 96.92%. These scores should be read within the evaluation setup used for each model, since the ExtraTrees metrics were computed on the full mapped domain and the Attention U Net metrics were computed on the held out test area. AlphaEarth variables appeared in 19 of the 20 most important ExtraTrees predictors, although these latent variables should not be read as direct physical measurements. The AlphaEarth only Attention U Net result was also close to the LiDAR only result. Both final maps were conservative in area: ExtraTrees mapped 6.75% wetland and Attention U Net mapped 6.38%, compared with the 7.29% buffered reference. On the basis of these results, AlphaEarth embeddings appear to carry useful wetland information while reducing the need for long satellite image collection, cloud masking and compositing. The study also shows that removing uncertain polygon edge pixels gives a clearer and more defensible way to evaluate wetland maps made from government inventory data.
It may be argued that forest monitoring based only on land-cover conversion can miss functional decline inside areas that remain spectrally forest-like. This study presents a label-free remote-sensing framework for mapping candidate hidden forest degradation from annual satellite data. The method combines physically interpretable Sentinel-2 vegetation, moisture and burn-sensitive indices with Landsat land surface temperature and year-to-year AlphaEarth embedding drift. These signals are normalised and fused into an annual hidden functional degradation score, then separated into two pathways: candidate spectral cover-loss/conversion pixels and candidate hidden degradation within pixels that still satisfy relaxed forest-like spectral conditions. Dynamic World probabilities are excluded from model construction and used only afterwards as a public land-cover transition reference, which avoids direct label leakage while retaining the limitation that both Dynamic World and part of the model input use Sentinel-2 information. It should be noted that, across six annual transitions from 2019 to 2025, AlphaEarth drift and the combined score separated Dynamic World transition pixels from stable-forest reference pixels with AUC values of 0.948–0.999 and 0.939–0.999, respectively. At the top 2% threshold for 2024–2025, F1 reached 0.937 for AlphaEarth drift and 0.916 for the combined score. These validation results apply to the full-domain score layers rather than to direct confirmation of the hidden stable-forest pathway itself. The combined score does not always outperform AlphaEarth drift on F1, but it provides a more physically interpretable candidate early-warning layer for forest-condition assessment and follow-up validation.