MGRL

MGRL is a multi-reward reinforcement learning framework for remote sensing change captioning on LEVIR-CC. It generates natural language descriptions of changes between paired images using cross-entropy training and self-critical sequence training (SCST).

The training objective combines CIDEr, BLEU, change-state, detail, and knowledge-graph rewards, with different rewards activated at different training stages.

• MGRL-RSCC: Multi-Granularity Reward Reinforcement Learning for Fine-Grained Remote Sensing Change Captioning, Futian Wang, Mengqi Wang, Xiao Wang*, Wentao Wu, Haowen Wang, Zhicheng Zhao, Jin Tang, arXiv:2609.26166 [Paper]

Remote Sensing Change Captioning (RSCC), which aims to generate accurate and detailed linguistic descriptions of ground object variations from bi-temporal remote sensing images, is a critical and challenging task in intelligent remote sensing interpretation. The mainstream autoregressive training paradigm faces severe exposure bias and train-test distribution mismatch, resulting in cumulative generation errors. They tend to produce conservative and template-fixed captions while ignoring subtle scene change details. To address these challenges, this paper proposes a novel multi-granularity reward reinforcement learning paradigm, termed MGRL-RSCC. Specifically, we first leverage a CNN and hierarchical self-attention module to extract and enhance visual features from bi-temporal remote sensing images. A Transformer decoder is then utilized to complete visual-to-linguistic translation. Different from existing methods, we design a dual-decoding strategy and a two-stage joint optimization scheme, which combines token-level supervised learning via greedy decoding and multi-granularity reward-driven self-critical reinforcement learning via sampling decoding. We further construct three complementary reward functions covering linguistic fluency, change state consistency, and structural-semantic relevance to comprehensively optimize caption quality and alleviate false and missing change descriptions. Extensive experiments on multiple public RSCC benchmark datasets demonstrate that the proposed MGRL-RSCC effectively mitigates exposure bias and conservative generation problems in traditional autoregressive methods.

Setup

Prepare a Python environment with compatible PyTorch and torchvision installations, then install the dependencies:

python -m pip install -r requirements.txt

METEOR evaluation also requires Java and the following resources, which are not included:

eval_func/meteor/meteor-1.5.jar
eval_func/meteor/data/paraphrase-en.gz

Alternatively, specify their locations with JAVA_BIN, METEOR_JAR, and METEOR_PARAPHRASE.

Data Preparation

Organize the prepared dataset as follows:

datasets/LEVIR_CC/
├── images/
│   ├── train/A/  train/B/
│   └── val/A/    val/B/
├── train.txt
├── val.txt
├── vocab.json
├── tokens/
└── levir_kg4_three_cleaned.csv

Training uses the train split, while validation and checkpoint selection use the val split. The knowledge-graph CSV must use GBK encoding. Datasets and model weights are not included in this repository.

Training

bash run_mgrl.sh

The default configuration uses 150 epochs, a batch size of 32, top-k sampling with k=3, and a sampling temperature of 0.8. SCST starts at epoch 8, with loss weights of 0.9 for cross-entropy and 0.1 for reinforcement learning. Epoch indices start at zero.

Outputs are saved to checkpoints/levir_mgrl/. To select a Python interpreter or GPU:

PYTHON_BIN=python3 GPU_ID=0 bash run_mgrl.sh

Custom paths can be supplied through DATA_FOLDER, LIST_PATH, TOKEN_FOLDER, KG_CSV, and SAVEPATH. Default paths are resolved relative to the project directory.

Main Files

• work2_train_unified_reward_lrboost_once.py: multi-reward SCST training and validation.
• model/: image encoders and the caption decoder.
• kg_reward_utils.py: knowledge-graph processing and matching utilities.
• run_mgrl.sh: main experiment configuration.

Citation

If you find this work useful for your research, please cite the following papers.

@misc{wang2026MGRLRSCC,
      title={MGRL-RSCC: Multi-Granularity Reward Reinforcement Learning for Fine-Grained Remote Sensing Change Captioning}, 
      author={Futian Wang and Mengqi Wang and Xiao Wang and Wentao Wu and Haowen Wang and Zhicheng Zhao and Jin Tang},
      year={2026},
      eprint={2609.26166},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.26166}, 
}