遥感
计算机科学
遥感应用
合成孔径雷达
图像(数学)
图像处理
计算机视觉
雷达成像
反射率
地球遥感
图像分辨率
辐射测量
人工智能
像素
地质学
地球观测
环境科学
图像分割
天基雷达
高光谱成像
波浪雷达
图像融合
作者
L L Xu,Dongyang Hou,Yuhan Wang,Peng Shen,Gaozhi Zhou,Liujue Zhang,Ziyu Li,Ji Qi,Wentao Yang,Huihui Feng,Haifeng Li
标识
DOI:10.1109/tgrs.2026.3752912
摘要
The multimodal reasoning paradigm inspired by DeepSeek-R1 has attracted wide attention: step-by-step reasoning promises not only higher accuracy in complex scenarios but also greater interpretability. On remote sensing image understanding, however, this paradigm does not reliably deliver these advantages. Further analysis shows that existing reasoning models often build correct conclusions upon lengthy or even ineffective trajectories, while the later part of a trajectory detaches from the visual evidence and keeps generating content irrelevant to the discrimination, so that the error rate rises as the trajectory grows. We trace this behavior to the penalty mechanism of reinforcement learning with verifiable rewards (RLVR): the penalty on ineffective reasoning is normalized over the full length of the trajectory, so longer trajectories receive weaker per-token gradient penalties, leaving redundant information poorly suppressed and allowing overthinking to persist. To this end, we propose RS-Reasoning, grounded in the “Concise Reasoning-then-Answering” paradigm, and introduce a Length-Adaptive Penalty (LAP) into RLVR: taking the average reasoning length measured at the end of the exploration stage as the threshold, LAP applies a penalty to trajectories that exceed this length, compensating for the pertoken gradient penalty diluted by length normalization, reducing redundant information, and alleviating the overthinking problem. On Qwen2.5-VL-7B, classification accuracy on INF improves by 2.86% over the GRPO baseline and by 21.27% over SFT, and the number of reasoning tokens on SAR decreases by 25.19% relative to GRPO. The proposed method yields more concise and more accurate reasoning for remote sensing scene classification.
科研通智能强力驱动
Strongly Powered by AbleSci AI