MTP: Advancing Remote Sensing Foundation Model via Multi-Task Pretraining
Foundation models have reshaped the landscape of Remote Sensing (RS) by enhancing various image interpretation tasks. Pretraining is an active research topic, encompassing supervised and self-supervised learning methods to initialize model weights effectively. However, transferring the pretrained models to downstream tasks may encounter task discrepancy due to their formulation of pretraining as image classification or object discrimination tasks. In this study, we explore the Multi-Task Pretraining (MTP) paradigm for RS foundation models to address this issue. Using a shared encoder and task-specific decoder architecture, we conduct multi-task supervised pretraining on the SAMRS dataset, encompassing semantic segmentation, instance segmentation, and rotated object detection. MTP supports both convolutional neural networks and vision transformer foundation models with over 300 million parameters. The pretrained models are finetuned on various RS downstream tasks, such as scene classification, horizontal and rotated object detection, semantic segmentation, and change detection. Extensive experiments across 14 datasets demonstrate the superiority of our models over existing ones of similar size and their competitive performance compared to larger state-of-the-art models, thus validating the effectiveness of MTP.
PDF AbstractCode
Tasks
Aerial Scene Classification
Building change detection for remote sensing images
Change Detection
Change detection for remote sensing images
Decoder
Image Classification
Instance Segmentation
Object
object-detection
Object Detection
Object Detection In Aerial Images
Oriented Object Detection
Scene Classification
Segmentation
Self-Supervised Learning
Semantic Segmentation
ImageNet
EuroSAT
DOTA
RESISC45
xView
LoveDA
Million-AID
Satlas
DOTA 2.0
SAMRS
SpaceNet 1
Results from the Paper
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Uses Extra Training Data |
Benchmark |
|---|---|---|---|---|---|---|---|
| Change detection for remote sensing images | CDD Dataset (season-varying) | MAE+MTP(ViT-L+RVSA) | F1-Score | 0.9798 | # 3 | ||
| Change Detection | CDD Dataset (season-varying) | MAE+MTP(ViT-B+RVSA) | F1-Score | 97.87 | # 5 | ||
| Change Detection | CDD Dataset (season-varying) | MAE+MTP(ViT-L+RVSA) | F1-Score | 97.98 | # 3 | ||
| Change Detection | CDD Dataset (season-varying) | IMP+MTP(InternImage-XL) | F1-Score | 98.33 | # 2 | ||
| Change detection for remote sensing images | CDD Dataset (season-varying) | IMP+MTP(InternImage-XL) | F1-Score | 0.9833 | # 2 | ||
| Change detection for remote sensing images | CDD Dataset (season-varying) | MAE+MTP(ViT-B+RVSA) | F1-Score | 0.9787 | # 4 | ||
| Object Detection In Aerial Images | DIOR | IMP+MTP(InternImage-XL) | AP50 | 78.0 | # 3 | ||
| Object Detection In Aerial Images | DIOR | MAE+MTP(ViT-B+RVSA) | AP50 | 79.4 | # 2 | ||
| Object Detection In Aerial Images | DIOR | MAE+MTP(ViT-L+RVSA) | AP50 | 81.1 | # 1 | ||
| Object Detection In Aerial Images | DIOR-R | IMP+MTP(InternImage-XL) | mAP | 72.17 | # 3 | ||
| Object Detection In Aerial Images | DIOR-R | MAE+MTP(ViT-L+RVSA) | mAP | 74.54 | # 1 | ||
| Object Detection In Aerial Images | DIOR-R | MAE+MTP(ViT-B+RVSA) | mAP | 71.29 | # 4 | ||
| Object Detection In Aerial Images | DOTA | IMP+MTP(InternImage-XL) | mAP | 80.77% | # 13 | ||
| Object Detection In Aerial Images | DOTA | MAE+MTP(ViT-B+RVSA) | mAP | 80.67% | # 15 | ||
| Object Detection In Aerial Images | DOTA | MAE+MTP(ViT-L+RVSA) | mAP | 81.66% | # 7 | ||
| Oriented Object Detection | DOTA 1.0 | MAE+MTP(ViT-L+RVSA) | mAP | 81.66 | # 4 | ||
| Oriented Object Detection | DOTA 2.0 | MAE+MTP(ViT-L+RVSA) | mAP | 58.41 | # 3 | ||
| Oriented Object Detection | DOTA 2.0 | MAE+MTP(ViT-B+RVSA) | mAP | 56.08 | # 5 | ||
| Oriented Object Detection | DOTA 2.0 | IMP+MTP(InternImage-XL) | mAP | 55.13 | # 6 | ||
| Image Classification | EuroSAT | MAE+MTP(ViT-B+RVSA) | Accuracy (%) | 98.76 | # 9 | ||
| Image Classification | EuroSAT | IMP+MTP(IntenImage-XL) | Accuracy (%) | 99.24 | # 1 | ||
| Image Classification | EuroSAT | MAE+MTP(ViT-L+RVSA) | Accuracy (%) | 98.78 | # 7 | ||
| Object Detection In Aerial Images | FAIR1M-2.0 | MAE+MTP(ViT-B+RVSA) | mAP | 51.92 | # 2 | ||
| Object Detection In Aerial Images | FAIR1M-2.0 | IMP+MTP(InternImage-XL) | mAP | 50.93 | # 3 | ||
| Object Detection In Aerial Images | FAIR1M-2.0 | MAE+MTP(ViT-L+RVSA) | mAP | 53.00 | # 1 | ||
| Building change detection for remote sensing images | LEVIR-CD | MAE+MTP(ViT-L+RVSA) | F1 | 92.67 | # 1 | ||
| Params(M) | 305 | # 9 | |||||
| Building change detection for remote sensing images | LEVIR-CD | MAE+MTP(ViT-B+RVSA) | F1 | 92.22 | # 5 | ||
| Params(M) | 86 | # 8 | |||||
| Building change detection for remote sensing images | LEVIR-CD | IMP+MTP(InternImage-XL) | F1 | 92.54 | # 2 | ||
| Params(M) | 335 | # 10 | |||||
| Change Detection | LEVIR-CD | IMP+MTP(InternImage-XL) | F1 | 92.54 | # 2 | ||
| Change Detection | LEVIR-CD | MAE+MTP(ViT-L+RVSA) | F1 | 92.67 | # 1 | ||
| Change Detection | LEVIR-CD | MAE+MTP(ViT-B+RVSA) | F1 | 92.22 | # 6 | ||
| Semantic Segmentation | LoveDA | IMP+MTP(InternImage-XL) | Category mIoU | 54.17 | # 5 | ||
| Semantic Segmentation | LoveDA | MAE+MTP(ViT-L+RVSA) | Category mIoU | 54.17 | # 5 | ||
| Semantic Segmentation | LoveDA | MAE+MTP(ViT-B+RVSA) | Category mIoU | 52.39 | # 15 | ||
| Aerial Scene Classification | NWPU (20% as trainset) | IMP+MTP(InternImage-XL) | Accuracy | 96.27 | # 1 | ||
| Aerial Scene Classification | NWPU (20% as trainset) | MAE+MTP(ViT-L+RVSA) | Accuracy | 95.88 | # 2 | ||
| Aerial Scene Classification | NWPU (20% as trainset) | MAE+MTP(ViT-B+RVSA) | Accuracy | 95.57 | # 6 | ||
| Change Detection | OSCD - 3ch | IMP+MTP(InternImage-XL) | F1 | 55.61 | # 2 | ||
| Change Detection | OSCD - 3ch | MAE+MTP(ViT-B+RVSA) | F1 | 53.36 | # 3 | ||
| Change Detection | OSCD - 3ch | MAE+MTP(ViT-L+RVSA) | F1 | 55.92 | # 1 | ||
| Semantic Segmentation | SpaceNet 1 | IMP+MTP(InternImage-XL) | Mean IoU | 79.16 | # 5 | ||
| Semantic Segmentation | SpaceNet 1 | MAE+MTP(ViT-L) | Mean IoU | 79.69 | # 1 | ||
| Semantic Segmentation | SpaceNet 1 | MAE+MTP(ViT-B+RVSA) | Mean IoU | 79.63 | # 2 | ||
| Semantic Segmentation | SpaceNet 1 | MAE+MTP(ViT-L+RVSA) | Mean IoU | 79.54 | # 3 | ||
| Change Detection | WHU Building Dataset | IMP+MTP(InternImage-XL) | F1-score | 0.9559 | # 1 | ||
| Change Detection | WHU Building Dataset | MAE+MTP(ViT-L+RVSA) | F1-score | 0.9475 | # 3 | ||
| Change Detection | WHU Building Dataset | MAE+MTP(ViT-B+RVSA) | F1-score | 0.9432 | # 5 | ||
| Object Detection In Aerial Images | xView | MAE+MTP(ViT-L+RVSA) | AP50 | 19.4 | # 1 | ||
| Object Detection In Aerial Images | xView | IMP+MTP(InternImage-XL) | AP50 | 18.2 | # 2 | ||
| Object Detection In Aerial Images | xView | MAE+MTP(ViT-B+RVSA) | AP50 | 16.4 | # 3 |
