Research map / Time series / Foundation models for time series
Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Exact, a weakly supervised (image-level label) framework for satellite image time series crop segmentation that uses spatial prototype clues with contrastive learning and temporal-aware affinity propagation to produce clue-based CAM pseudo labels.
Problem
Pixel-level annotation for crop parcel segmentation in satellite image time series is costly; standard CAM-based weakly supervised methods from natural images suffer spatial over-activation from homogeneous neighboring parcels and temporal semantic bias from anomalous periods.
Approach
Builds class-wise prototypes via optimal-transport clustering on temporal dense embeddings, uses them in a clue-based contrastive loss, and propagates CAMs with affinities derived from temporal-to-class attention (temporal-aware affinity), then generates CB-CAMs from prototype similarity as pseudo labels.
Key results
- Exact pseudo labels on PASTIS train reach 84.1% OA and 75.6% mIoU, versus 81.2% OA and 69.5% mIoU for the baseline CAM.
- A TSViT segmenter trained on Exact pseudo labels reaches 80.3% OA and 61.8% mIoU on PASTIS test, about 95% of its fully supervised mIoU (65.1%).
- Exact pseudo labels on the Germany train set reach 88.3% OA and 80.6% mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Exact-generated pseudo labels instead of the baseline labels | TSViT segmentation model · PASTIS test | mIoU: 57.7 -> 61.8 | Helped |
| Add clue-based contrastive learning Lcbl (ablation) | Exact classification network · PASTIS train | mIoU: 69.5 -> 72.3 | Helped |
| Add temporal-aware affinity propagation loss Ltap to the Lcbl variant (ablation) | Exact classification network · PASTIS train | mIoU: 72.3 -> 73.4 | Helped |
| Generate CAMs using the learned space-time clues (ablation) | Exact · PASTIS train | mIoU: 73.4 -> 75.6 | Helped |
| Use temporal dense embeddings rather than spatial dense embeddings to generate CAMs (ablation) | TSViT baseline · PASTIS train | mIoU: 66.7 -> 68.3 | Helped |
| Fuse spatial and temporal embedding CAMs (ablation) | TSViT baseline · PASTIS train | mIoU: 68.3 -> 69.5 | Helped |
| Use temporal-aware affinity instead of low-level affinity cues (ablation) | Exact · PASTIS train | mIoU: 69.9 -> 72.3 | Helped |
| Combine low-level cues with temporal-aware affinity (ablation) | Exact · PASTIS train | mIoU: 72.3 -> 71.8 | Hurt |
| Use only 2 class-specific prototypes rather than increasing the prototype count (ablation) | Exact · PASTIS train | mIoU | Helped |
| Increase the similarity temperature beyond the reported selected setting (ablation) | Exact · PASTIS train | mIoU | Mixed |
Surprising findings
- Off-the-shelf natural-image WSSS modules (e.g., PAMR, TS-CAM, SIPE, FPR) adapted to SITS often fail to improve or even degrade performance relative to the baseline.
- Low-level cue affinity (color, intensity) gives only limited gains on SITS, while temporal-aware high-level affinity helps more.
Limitations
- Exact still shows limitations in edge processing of crop parcel boundaries.
- Evaluated only on two SITS crop datasets (PASTIS and Germany).
Experimental setup
Datasets: PASTIS Germany crop dataset
Benchmarks: PASTIS Germany
Metrics: mIoU OA recall precision
Baselines: MCTFormer ViT-PCM LP-CAM TSCD DuPL SeCo PAMR TS-CAM SIPE FPR
Builds on: TSViT class activation map PAMR SIPE FPR MCTformer
Closest papers by meaning
- PARCS: A Deployment-Oriented AI System for Robust Parcel-Level Cropland Segmentation of Satellite Images (AAAI 2023 · similarity 0.88)
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.86)
- Fields of The World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Segmentation (AAAI 2025 · similarity 0.85)
- Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.85)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.85)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.