本文介绍了 CARTO 首个地理空间基础模型 PlacePulse 的构建过程，阐述其 256 维嵌入所捕捉的信息特征，并评估其在 74 项预测任务中的性能表现。

How we built PlacePulse, CARTO's first geospatial foundation model: what its 256-number embeddings capture and how they perform across 74 prediction tasks.

我们的新论文提出了该模型，并配套提供了一个评估地理空间嵌入的框架，涵盖从信息保留到在新预测任务上的表现等多个维度。在本文中，我们将深入探讨该方法及主要发现。 掩码自编码器（Masked Autoencoder）能够在无标签的情况下学习压缩表示。在训练过程中，它对 30% 的输入属性进行掩码处理，并从剩余特征中重建这些被掩码的属性，从而迫使模型学习地点属性之间的关系，而非简单复制输入数据。最终结果为覆盖美国的 1,055,537 个 H3 分辨率 7 级单元格中的每一个生成了一个单独的嵌入向量。通过主成分分析（PCA）对 256 维嵌入进行着色，展示了所呈现嵌入向量的可视化效果。 PlacePulse 完全基于结构化数据进行学习。它捕捉地点的人口统计、经济和环境特征，提供了不同于基于图像或行为信号（如访问记录和搜索活动）构建的模型的视角。我们的评估旨在考察这些结构化属性能提供多少关于地点的信息，以及这些信息在不同任务中的实用性。 如何评估一种旨在支持多种用途的表示方法？单一任务的表现只能给出部分答案。我们的基准测试既考察嵌入向量捕获的内容，也考察其应用于新预测任务时的实用性。由于该框架独立于模型架构和输入数据，它为比较不同的地理空间嵌入提供了共同基础。 当 2,288 个输入特征被压缩为 256 个数值时，有多少信息得以保留？PlacePulse 以 0.887 的中位 R² 值重建其标准化输入，表明这种紧凑表示保留了原始数据中的大部分变异。每个输入特征的重建效果至少与始终预测其均值的基线模型相当。该模型还充分利用了嵌入向量的全部宽度。一项逐步移除高度相关维度的测试显示，256 个维度中有 251 个被保留，这表明很少有维度仅仅是彼此重复。综合来看，这些结果表明 PlacePulse 在实现大幅压缩的同时，仍在广泛的地点特征集合中保留了信息。 模型从未获得显式的位置信息。它看不到坐标、距离或边界；只看到地点属性和局部上下文。然而，嵌入空间却呈现出强烈的地理特性。聚类结果同样清晰可读。在 k = 10 时，嵌入向量主要根据居住密度对地点进行划分，从密集的城市核心区到稀疏的农村地区，同时也捕捉到了超越单纯密度差异的模式。在 k = 50 和 k = 100 时，更精细的聚类大多嵌套在较粗的聚类内部，而非重新洗牌，这表明存在稳定的地点类型层级结构。 我们在来自 12 个外部来源的 74 个留出目标上测试了嵌入向量作为特征的表现，涉及四种空间单元，始终使用相同的梯度提升模型和按州整体留出的交叉验证方法。这一系列测试中没有任何内容用于训练。我们将其性能与 Google 的人口动态基础模型（Population Dynamics Foundation Model）进行了比较，后者从聚合搜索趋势、兴趣点及其繁忙程度、天气和空气质量以及卫星衍生的嵌入向量中学习。结果因领域而异： 这些结果表明，结构化和行为数据为理解地点提供了互补的视角。每种表示在不同的预测任务中具有各自的优势，使得它们的组合成为下一步极具潜力的方向。结合它们是否能提高性能，仍有待进一步评估。 单个 256 维向量将每个单元格转化为地点的机器可读描述。这开启了四种利用地理信息的方式： PlacePulse 旨在使地理空间基础模型适用于现实世界分析，从寻找相似地点到预测新位置的成果。想了解它能揭示您自身足迹的哪些信息？请申请早期访问权限。 Lucía 是 CARTO 的高级数据科学家，她开发空间统计和机器学习解决方案，以发掘基于位置的数据的隐藏潜力，帮助组织最大化地理空间信息的价值。Miguel 是 CARTO 的数据科学负责人，领导一支优秀团队，致力于为用户提供空间分析能力，并利用空间统计和机器学习为客户开发定制模型。 查看 PlacePulse Embeddings 如何帮助零售选址团队将表现最佳的门店转化为可复用的“指纹”，以寻找下一个选址。CARTO 的地理空间基础模型 PlacePulse 帮助公用事业公司和充电网络发现基于采用率的模型所遗漏的电动汽车充电基础设施缺口。

Our new paper presents the model alongside a framework for evaluating geospatial embeddings, from the information they preserve to their performance on new prediction tasks. In this post, we take a closer look at the approach and some of the main findings. A masked autoencoder learns compressed representations without labels. During training, it masks 30% of the input attributes and reconstructs them from the remaining features, forcing the model to learn relationships among place attributes rather than simply copy the input. The result is a single embedding for each of the 1,055,537 H3 resolution 7 cells covering the United States. Visualizations of the presented embeddings colored by principal component analysis (PCA) of the 256-dimensional embeddings. PlacePulse learns entirely from structural data. It captures the demographic, economic, and environmental characteristics of places, offering a different perspective from models built on imagery or behavioral signals such as visits and search activity. Our evaluation examines how much these structural attributes can tell us about a place and how useful that information is across different tasks. How do we assess a representation designed to support many different uses? Performance on a single task gives us only part of the answer. Our benchmark examines both what the embeddings capture and how useful they are when applied to new prediction tasks. Because the framework is independent of the model architecture and input data, it provides a common basis for comparing different geospatial embeddings. How much information survives when 2,288 input features are compressed into 256 numbers? PlacePulse reconstructs its standardized inputs with a median R² of 0.887, indicating that the compact representation preserves much of the variation in the original data. Every input feature is reconstructed at least as well as a baseline that always predicts its mean. The model also makes use of the embedding’s full width. A test that progressively removes highly correlated dimensions retains 251 of the 256, suggesting that few dimensions simply duplicate one another. Together, these results show that PlacePulse achieves substantial compression while retaining information across a broad set of place characteristics. The model is never given an explicit location. It sees no coordinates, distances, or boundaries; only place attributes and local context. Yet the embedding space becomes strongly geographic. The clusters are readable too. At k = 10, the embedding partitions places mainly by settlement density, from dense urban cores to sparse rural areas, while also capturing patterns beyond differences in density alone. At k = 50 and k = 100, the finer clusters mostly nest inside the coarser ones rather than reshuffling, which suggests a stable hierarchy of place types. We tested the embeddings as features on 74 held-out targets from 12 outside sources, across four spatial units, always with the same gradient-boosted model and cross-validation that holds out whole states. Nothing in that battery was used in training. We compared its performance with Google’s Population Dynamics Foundation Model, which learns from aggregated search trends, points of interest and their busyness, weather and air quality, and satellite-derived embeddings. The results varied by domain: These results suggest that structural and behavioral data offer complementary perspectives on place. Each representation has strengths across different prediction tasks, making their combination a promising next step. Whether combining them improves performance is a question for further evaluation. A single 256-dimensional vector turns every cell into a machine-readable description of place. That opens up four ways to work with geography: PlacePulse is built to make geospatial foundation models practical for real-world analysis, from finding similar places to predicting outcomes in new locations. Want to see what it can reveal about your own footprint? Request early access. Lucía is a Senior Data Scientist at CARTO, where she develops spatial statistics and machine learning solutions that unveil the hidden potential of location-based data, enabling organizations to maximize the value of geospatial information. Miguel is Head of Data Science at CARTO, leading an amazing team working to provide spatial analytics capabilities to our users as well as to develop custom models for our customers using spatial statistics and machine learning. See how PlacePulse Embeddings help retail site selection teams turn top-performing stores into a reusable fingerprint for finding the next location. PlacePulse, CARTO's geospatial foundation model, helps utilities and charging networks spot EV charging infrastructure gaps that adoption-based models miss.