AI 中文总结
研究将随机索引机制用于多时相图像变化检测，提出距离阈值聚类词汇表，结合空间上下文积累形成无需训练的检测管道，在多数据集评估中接近基线，发现并描述了一些问题，还进行了多种分析及讨论。
AI 中文摘要
随机索引（RI）几乎仅用于文本分析（特别是时间随机索引，TRI，用于跟踪随时间变化的词义），它用固定随机向量表示离散词汇表，并通过向量求和积累上下文。我们探索将此机制移植到多时相图像的变化检测中。使用k均值构建视觉词汇表的简单移植是不稳定的，因为小的辐射度变化会导致像素在采集日期之间频繁重新分配，破坏了RI所依赖的像素到向量的对应关系。我们提出了一种距离阈值（领导者）聚类词汇表，并给出了一个简短的形式论证——基于词汇表覆盖/填充属性的稳定半径——说明为什么它比k均值对采集噪声更具鲁棒性。结合空间上下文积累，这产生了一个无需训练的变化检测管道，在四个双时相遥感数据集（农业、河流、城市极化合成孔径雷达、野火）上针对变化向量分析基线进行评估，它始终接近但未超过该基线。多种子验证暴露并修复了概率稀疏RI变体中的退化向量故障模式，并揭示了第二个未解决的变异性来源——对领导者聚类访问顺序的敏感性——我们对其进行了描述，但尽管尝试了三次修正，仍无法消除，将其报告为主要开放问题。结果与独立的重新实现进行了交叉检查。我们报告了一次消融、超参数敏感性分析、大津法与高斯混合阈值比较，并讨论了RI增量积累以用于未来的长时间序列监测。
英文摘要
Random Indexing (RI), used almost exclusively in text analysis (notably Temporal Random Indexing, TRI, for tracking word meaning over time), represents a discrete vocabulary with fixed random vectors and accumulates context by vector summation. We explore transplanting this mechanism to change detection in multitemporal images. A naive transplant using k-means to build the visual vocabulary is unstable: pixels are frequently reassigned between acquisition dates due to small radiometric shifts, destroying the pixel-to-vector correspondence RI depends on. We propose a distance-threshold (leader) clustering vocabulary instead, and give a short formal argument -- a stability radius from the vocabulary covering/packing properties -- for why this is provably more robust to acquisition noise than k-means. Combined with spatial context accumulation, this yields a training-free change detection pipeline, evaluated on four bi-temporal remote-sensing datasets (agriculture, river, urban PolSAR, wildfire) against a Change Vector Analysis baseline, which it consistently approaches but does not surpass. Multi-seed validation exposed and let us fix a degenerate-vector failure mode in a probabilistic-sparsity RI variant, and revealed a second, unresolved source of variability -- sensitivity to the leader-clustering visitation order -- which we characterize but, despite three attempted corrections, could not eliminate, reporting it as the main open problem. Results are cross-checked against an independent re-implementation. We report an ablation, a hyperparameter sensitivity analysis, an Otsu-vs-Gaussian-mixture thresholding comparison, and discuss RI incremental accumulation for future long time-series monitoring.
