AMP: Wavelet-enhanced Foundation Model Adaptation for Weakly Supervised Urban Tree Canopy Mapping

Accepted by IEEE Transactions on Geoscience and Remote Sensing 2026

Mingnuo Teng^1*,  Jianhua Guo^2*,  Huanjing Yue¹,  Jingyu Yang^1†

¹Tianjin University, ²Aerospace Information Research Institute

🌟 Method Overview

This repository contains the official PyTorch implementation of AMP, a framework designed for high-accuracy urban tree canopy mapping using weakly supervised sparse annotations. By integrating wavelet-enhanced adaptation with foundation models (DINOv2), our method effectively captures fine-grained spatial details in complex urban environments.

---

📷 Result Display

🛠️ Installation & Training & Inference

1. Clone the repository

git clone https://github.com/Tengmingn/AMP.git
cd AMP
conda create -n amp python=3.10
conda activate amp
pip install -r requirement

2. Download the UTC-Sparse dataset

Get our dataset at https://ieee-dataport.org/documents/utc-sparse

3. Prepard third party code

cd third_party
# dinov2
git clone https://github.com/facebookresearch/dinov2.git
# dinov3 if you want
git clone https://github.com/facebookresearch/dinov3.git

Then, insert the user_forward_features function into the respective model files. The corresponding function codes are provided in ./third_party/dinov2_userfunc and ./third_party/dinov3_userfunc.

Alternatively, you can download our pre-packaged third-party folder and extract it directly into the ./third_party directory. https://pan.baidu.com/s/1Y_2E5OTQpKRoBlxD-HDMdw?pwd=57kr password: 57kr

4. Reproduce

Our pre-trained and fine-tuned models are available at: https://pan.baidu.com/s/1TzAoo50w1w1lWD5H_yIrHA?pwd=edhq password: edhq

5. Training

bash tools/train_tree.sh <num_gpu> <port>
bash tools/train_vaihingen.sh <num_gpu> <port>

6. Inference

bash tools/test_tree.sh

❤️ Acknowledgement

We sincerely acknowledge the following works for their valuable contributions and inspiration to this project:

• SparseFormer: A Credible Dual-CNN Expert-Guided Transformer for Remote Sensing Image Segmentation With Sparse Point Annotation (TGRS 2025) — for providing the implementation of the multi-branch pseudo-label generation strategy.(https://github.com/Yujia73/SparseFormer)

• Sparsely Annotated Semantic Segmentation with Adaptive Gaussian Mixtures — for providing the foundational network structure implementation.(https://github.com/Luffy03/AGMM-SASS)

• DINOv2: Learning Robust Visual Features without Supervision — for providing the pretrained visual feature backbone used in this work.(https://github.com/facebookresearch/dinov2)

We deeply appreciate the authors of these works for making their research and code publicly available to the community.

🎓 Citations

@ARTICLE{11520877, author={Teng, Mingnuo and Guo, Jianhua and Yue, Huanjing and Yang, Jingyu}, journal={IEEE Transactions on Geoscience and Remote Sensing}, title={Wavelet-Enhanced Foundation Model Adaptation for Weakly Supervised Urban Tree Canopy Mapping}, year={2026}, volume={64}, number={}, pages={4409814-4409814}, keywords={Labeling;Modeling;Vegetation;Trees (botanical);Annotations;Remote sensing;Semantic segmentation;Training;Pixel;Modules (abstract algebra);Semantic segmentation;sparse annotation;urban tree canopy (UTC);vision foundation models;weakly supervised learning}, doi={10.1109/TGRS.2026.3693826}}

📧 Contact

If you have any inquiries, please reach out us via email at tmn@tju.edu.cn