WAIMLAp 2026 PUCP

Workshop on Artificial Intelligence and Machine Learning Applications

Cross-regional deforestation mapping in the Amazon

Evaluating DINOv3 representations with Sentinel‑1 and Sentinel‑2

Christian Carhuancho1 · Jorge Chamorro1 · Jorge Fatama1

1Pontificia Universidad Católica del Perú, Lima, Perú

WAIMLAp 2026 · November 18–20, 2026 · Lima, Perú

Research question

Can an optimized DINOv3 trained in Pará map deforestation in Perú with no local labels, and how many local labels does it need?

Motivation & research problem

  • Monitoring Amazon deforestation still depends on expert interpretation.
  • A model trained in one region loses accuracy when land cover, sensors or clearing practices change.
  • Perú lost 3.4 Mha of forest in 2000–2020. This study targets Ucayali and San Martín.
  • Vision transformers have shown less generalization capability compared to CNNs in tropical regions, which makes it crucial to study the transferability of vision transformers.
  • No cross-country Amazon study with vision foundation models was found in our research of the current state of the art.

Objectives / contribution

Evaluate the cross-country transferability of an optimized DINOv3 model trained in Pará, Brazil, for deforestation mapping in Ucayali and San Martín, Peru, while assessing zero-shot performance and local label requirements.

  1. Evaluate source-domain performance: Compare DINOv3 against conventional CNN architectures, namely U-Net and SegNet, for deforestation mapping in Pará, Brazil.
  2. Assess cross-country transferability: Evaluate the ability of the Brazil-trained DINOv3 model to generalize to Ucayali and San Martín, Peru, without using local labeled data (zero-shot transfer).
  3. Quantify local label requirements: Determine how many labeled samples from Peru are needed to improve the transferred model’s performance and adapt it to the target regions.

Methodology

Brazil baselines U-Net and SegNet Optical image and radar
DINOv3 on Pará Foundation model Adapted on the Brazil data
Transfer to Peru Ucayali and San Martín Local Label First

Multimodal architecture

Multimodal architecture combining DINOv3 optical features and CNN-based radar features through a fusion decoder to generate deforestation masks.

Results & discussion

0.900 Pará test IoU DINOv3 · F1 0.947. Above SegNet (0.878) and U-Net (0.858).
0.469 Peru zero-shot IoU F1 0.638. False positives reach 50.4% with no local labels.
0.633 Peru few-shot IoU F1 0.775 after one labeled tile per department.
Pará test set results
Model False positives (FP) False negatives (FN) IoU F1
U-Net 10.7% 4.3% 0.858 0.924
SegNet 7.4% 5.6% 0.878 0.935
DINOv3 6.5% 4.1% 0.900 0.947
True colors (S2) PRODES (mask) Prediction Disagreement
Full study area in Pará

Forest Correct prediction False positives False negatives

Peru test set results
Model Region FP FN IoU F1
Zero-shot San Martín 55.5% 3.6% 0.438 0.609
Ucayali 43.5% 16.8% 0.507 0.673
Peru 50.4% 10.4% 0.469 0.638
Few-shot San Martín 24.7% 13.1% 0.677 0.807
Ucayali 16.9% 33.4% 0.587 0.740
Peru 21.4% 23.6% 0.633 0.775
True colors (S2) MapBiomas (mask) Zero-shot Few-shot
Lowest zero-shot test tile in San Martín (t3)

Correct deforestation Correct forest False positives False negatives

The zero-shot algorithm detects the larger clearing, hence its primary mistake is to generate false alarms within the closed forest as well as at the clearing boundaries. The few-shot algorithm eliminates the false alarm in the interior of the forest but creates more false negatives by missing some of the clearing areas.

Conclusions

Conclusions: DINOv3 proved effective for satellite imagery, outperforming U-Net and SegNet in Pará. The frozen backbone performed similarly to the fine-tuned model, while zero-shot transfer to Peru remained limited. However, few-shot adaptation using just one labeled tile per department substantially improved transfer performance, highlighting the reusability of frozen features and the potential of lightweight domain adaptation.

Limitations: The evaluation covers few tiles, two departments, one image year (2024), and a single run. Training in Pará used cloud-free imagery, so performance under cloudy conditions remains untested, potentially affecting results in Peru. MapBiomas’ labels represent accumulated clearing, while spatial proximity between adaptation and test tiles and the uncalibrated threshold may also affect results.

Future work: Explore Domain-Adversarial Neural Networks (DANN) to improve cross-domain alignment using unlabeled Peruvian imagery. Further steps include validation with GeoBosques on independent sites, threshold recalibration, larger patches, and evaluation under cloudy conditions.

Key references

  1. Chhibber, Deep learning based multi-source data fusion to map deforested areas in Amazon rain forest, M.Sc. thesis, Univ. Twente, 2022.
  2. Siméoni et al., DINOv3, arXiv:2508.10104, 2025.
  3. Soto Vega et al., Assessing the generalization capacity of CNNs and vision transformers for deforestation detection in tropical biomes, ISPRS Archives XLVIII-3, 2024.
  4. Móstiga et al., Two decades of accelerated deforestation in Peruvian forests (2000–2020), Reg. Environ. Change 24(2), 2024.
  5. MINAM/PNCBMCC, Protocolo metodológico para la detección de la pérdida de bosque (2001–2019), GeoBosques.

Acknowledgments / funding

Data: Rainforest-Sentinel (Chhibber, 2022) with PRODES labels (INPE) for Pará; Copernicus Sentinel-1/2 imagery (2024) and MapBiomas land-cover maps (2022–2025) as the reference mask for Ucayali and San Martín, Peru.

Models: DINOv3 via timm.

Compute: One NVIDIA RTX 2060 (6 GB), provided by Grupo IA PUCP through Dr. Edwin Villanueva.

Continue the conversation

Corresponding author

Christian Carhuancho

a20172665@pucp.edu.pe

Pontificia Universidad Católica del Perú

Scan the QR code to open the project link from the poster.

QR code from the poster, linking to the project page.