| Model | Modalities | Any-to-Any Generation | Multi-Scale Features |
| RemoteCLIP | optical, text | X | X |
| CROMA | optical, radar | X | X |
| AnySat | aerial, optical, radar, NAIP | X | X |
| DeCUR | optical, radar | X | X |
| DOFA | optical, radar, hyperspectral, NAIP | X | X |
| MetaEarth | optical (unimodal) | X | ✓ |
| Galileo | optical, radar, elevation, weather, location, population, ... | X | ✓ |
| TerraMind | optical, radar, land use, elevation, vegetation index, location, text | ✓ | ✓ |
",
+ "image_path": "e99accac7b0552322e0145e2b418e3c374d12d577dbd6d217a88426840df1874.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 483,
+ 555,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 483,
+ 555,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 483,
+ 555,
+ 507
+ ],
+ "type": "text",
+ "content": "Table 1. Comparison of TerraMind to other model architectures. TerraMind represents a first-of-its-kind generative, multimodal model."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 532,
+ 555,
+ 592
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 532,
+ 555,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 532,
+ 555,
+ 592
+ ],
+ "type": "text",
+ "content": "be open-sourced to the community. TerraMesh builds on existing datasets, which we expand by adding modalities from external data sources or by applying pseudo-labeling. We provide an overview of the aligned image modalities and a detailed dataset description in the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 313,
+ 594,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 594,
+ 556,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 594,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "content": "Base datasets. TerraMesh is based on SSL4EO-S12 [6, 65] and MajorTOM-Core [23], two unlabeled remote sensing datasets containing co-aligned radar and optical imagery from Sentinel-1 and Sentinel-2 satellites. SSL4EO-S12 has lower geographic coverage but is multi-seasonal. MajorTOM-Core covers most of the Earth's land surface at a single timestamp. For MajorTOM-Core, we apply a subsampling scheme based on LULC classes and ecoregions. TerraMesh includes a total of approximately 9 million globally distributed samples from both Sentinel-1 and Sentinel-2,"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "7385"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "text",
+ "content": "each measuring "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "inline_equation",
+ "content": "264 \\times 264"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "text",
+ "content": " pixels at "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "inline_equation",
+ "content": "10\\mathrm{m}"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 262,
+ 83
+ ],
+ "type": "text",
+ "content": " resolution."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 55,
+ 84,
+ 297,
+ 277
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 84,
+ 297,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 84,
+ 297,
+ 277
+ ],
+ "type": "text",
+ "content": "Additional modalities. We obtain co-aligned yearly LULC maps by ESRI with nine land use classes. Additionally, we leverage SEnSeI v2 [22] as a cloud and ice annotation model and update the ESRI LULC classes for better spatiotemporal alignment. NDVI maps are computed using the corresponding spectral bands from Sentinel-2. DEM is extracted from the Copernicus DEM 30m dataset [2], which provides global coverage of the Earth's elevation at a 30m resolution. Captions are generated synthetically by constructing RGB images from Sentinel-2 patches using the corresponding spectral bands and processing them with LLaVANext [37]. A tailored prompt guides the model to describe the content of each image as described in [47]. For geolocations, we round latitude and longitude from the center of each patch to the nearest quarter degree and store the discretized coordinates as strings in a pre-defined format."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 288,
+ 117,
+ 300
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 288,
+ 117,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 288,
+ 117,
+ 300
+ ],
+ "type": "text",
+ "content": "4. Methods"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 55,
+ 308,
+ 296,
+ 369
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 308,
+ 296,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 308,
+ 296,
+ 369
+ ],
+ "type": "text",
+ "content": "TerraMind pretraining is two-staged following [52]. We first pretrain unimodal tokenizer models, tokenize the modalities, and then leverage token-level and pixel-level input to pretrain the TerraMind encoder-decoder architecture. We describe those individual stages in the following."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 55,
+ 376,
+ 139,
+ 388
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 376,
+ 139,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 376,
+ 139,
+ 388
+ ],
+ "type": "text",
+ "content": "4.1. Tokenization"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 55,
+ 394,
+ 296,
+ 478
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 394,
+ 296,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 394,
+ 296,
+ 478
+ ],
+ "type": "text",
+ "content": "We develop modality-specific tokenizers to encode each modality into a sequence of discrete tokens for pretraining and decode token sequences back to images. Thus, TerraMind is in principle compatible with any modality, as long as it can be tokenized and aligned with other modalities. For reasons of space, we delegate most experiments related to the tokenizer performances to the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "text",
+ "content": "Image-like modalities. We train autoencoder-based architectures with a quantization step in the bottleneck for image-like modalities such as S-1, S-2, LULC, NDVI, and DEM. Tokenizer encoders process an input image and generate a latent representation for each "
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "inline_equation",
+ "content": "16 \\times 16"
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "text",
+ "content": " patch, which is then discretized with finite-scalar-quantization (FSQ) [51] into one of "
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "text",
+ "content": " codewords. All tokenizers use a vocabulary size of 16K besides the simpler LULC modality for which we use 4K. These codewords are then used by the diffusion decoder to reconstruct the original image. The benefit of leveraging diffusion decoders lies in facilitating cross-modal generation in TerraMind by transforming tokens back into images. By mapping each codeword to a unique integer in "
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "inline_equation",
+ "content": "\\{0, 1, \\dots, N - 1\\}"
+ },
+ {
+ "bbox": [
+ 55,
+ 479,
+ 296,
+ 694
+ ],
+ "type": "text",
+ "content": ", we obtain discrete tokens for each image patch. We pretrain the tokenizers in a self-supervised setting. FSQ as quantization method enhances training stability [51] compared to vector quantization [63] by eliminating the need for codebook-related loss terms. Notably, FSQ is"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 72,
+ 556,
+ 240
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 72,
+ 556,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 72,
+ 556,
+ 240
+ ],
+ "type": "text",
+ "content": "heavily influenced by ideas of neural compression [27]. For example, on 12-band S-2 images, we achieve compression rates of over "
+ },
+ {
+ "bbox": [
+ 313,
+ 72,
+ 556,
+ 240
+ ],
+ "type": "inline_equation",
+ "content": "3000\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 313,
+ 72,
+ 556,
+ 240
+ ],
+ "type": "text",
+ "content": " by applying quantization. We summarize the architecture of our tokenizers in Figure 4. The main objective of the overall tokenizer is to encode image patches consistently into discrete tokens based on semantic similarity to enable cross-modal correlation learning. Therefore, the loss of some details is an expected trade-off, since the focus is on grouping similar patches rather than preserving all fine-grained features. Naturally, more accurate reconstructions facilitate cross-modal generation, however the main focus of the pretraining lies on consistent cross-modal correlation learning. We provided further details on the pretraining of the tokenizers in the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 318,
+ 252,
+ 555,
+ 330
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 318,
+ 252,
+ 555,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 252,
+ 555,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 252,
+ 555,
+ 330
+ ],
+ "type": "image",
+ "image_path": "71da8b1dfb813f0f4d5e280206571aa38f59c2eb458f021352749f9465058d54.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 313,
+ 342,
+ 556,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 342,
+ 556,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 342,
+ 556,
+ 365
+ ],
+ "type": "text",
+ "content": "Figure 4. Tokenizer for image-like modalities combining finite-scalar quantization [51] with diffusion decoding."
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 313,
+ 377,
+ 556,
+ 475
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 377,
+ 556,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 377,
+ 556,
+ 475
+ ],
+ "type": "text",
+ "content": "Sequence-like modalities. We treat both captions and geolocations as text and use a single text tokenizer to process both modalities. By discretizing the geographic coordinates and representing them as strings, we introduce special coordinate tokens into the vocabulary. This allows us to encode geolocations as a sequence of discrete tokens, beginning with a latitude token followed by a longitude token. For textual data, we modify the existing WordPiece tokenizer [33]."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 313,
+ 480,
+ 395,
+ 493
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 480,
+ 395,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 480,
+ 395,
+ 493
+ ],
+ "type": "text",
+ "content": "4.2. Pre-training"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 313,
+ 498,
+ 556,
+ 594
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 498,
+ 556,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 498,
+ 556,
+ 594
+ ],
+ "type": "text",
+ "content": "Architecture. TerraMind uses a symmetric Transformer-based encoder-decoder architecture proposed in [52], which is designed to process sequences of multimodal tokens. In addition to discrete tokens, TerraMind accepts pixel-level inputs, specifically satellite imagery and digital elevation maps. For pixel-level inputs, we apply learnable patch-wise linear projections to generate patch embeddings for each "
+ },
+ {
+ "bbox": [
+ 313,
+ 498,
+ 556,
+ 594
+ ],
+ "type": "inline_equation",
+ "content": "16 \\times 16"
+ },
+ {
+ "bbox": [
+ 313,
+ 498,
+ 556,
+ 594
+ ],
+ "type": "text",
+ "content": " patch, similar to the approach used in ViT [17]."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 313,
+ 594,
+ 557,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 594,
+ 557,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 594,
+ 557,
+ 713
+ ],
+ "type": "text",
+ "content": "Dual-scale early fusion. In contrast to [52], we not only embed token-level data but additionally leverage pixel-level data across a range of input modalities to introduce a dual-scale feature representation to support the structuring of the embedding space. Both tokens and patches represent a 16x16 pixel area. Tokens represent this area via a single discrete integer value, while the image patches describe the same area with the actual floating point sensor data. Thus, during pretraining, the model not only learns a correlation between modalities (i.e., cross-modal learning) but also between dif"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 56,
+ 704,
+ 259,
+ 712
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 704,
+ 259,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 704,
+ 259,
+ 712
+ ],
+ "type": "text",
+ "content": "https://planetarycomputer.microsoft.com/dataset/io-lulc-annual-v02"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 295,
+ 748,
+ 316,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 316,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 316,
+ 757
+ ],
+ "type": "text",
+ "content": "7386"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 193
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 193
+ ],
+ "type": "text",
+ "content": "ferent levels of abstraction within the same modality. The low-level token information enables cross-modal correlation learning, while adding pixel level input accounts for spatial nuances. Based on dual-scale features the model further learns to better structure pixel-level data in the embedding space via the corresponding information from the discrete token. We illustrate the pretraining paradigm in Figure 5. The model is agnostic to processing tokens or patches in the input space, while the target is generally token-level data. We use six pixel-level modalities and eight token-level modalities."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 56,
+ 203,
+ 289,
+ 290
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 56,
+ 203,
+ 289,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 203,
+ 289,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 203,
+ 289,
+ 290
+ ],
+ "type": "image",
+ "image_path": "e0986ca258c6aa8b5d527085e8ce35c951950eed3bc2982fada04e93f0661922.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 55,
+ 303,
+ 296,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 303,
+ 296,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 303,
+ 296,
+ 336
+ ],
+ "type": "text",
+ "content": "Figure 5. Illustration of the pre-training task. Given an encoded multimodal sample of random subsets of patches and input tokens, the decoder predicts target tokens for the masked input."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 347,
+ 296,
+ 418
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 347,
+ 296,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 347,
+ 296,
+ 418
+ ],
+ "type": "text",
+ "content": "Masking strategy. TerraMind applies a masked modeling approach in the token space following [52]. The model leverages a set of randomly selected target tokens that have to be reconstructed from a randomly selected set of input tokens and pixel-level data. During pre-training, we sample input and target data from a Dirichlet distribution."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 55,
+ 419,
+ 296,
+ 598
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 419,
+ 296,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 419,
+ 296,
+ 598
+ ],
+ "type": "text",
+ "content": "We opt for masked token reconstruction to familiarize the model with the absence of entire modalities, which is crucial for a high usability of a multimodal model in Earth observation. During pre-training, the model learns an internal representation of unseen modalities which is expected to benefit a range of downstream applications. In addition, sampling input and target tokens improves the computational efficiency of the pre-training, as each token is a compressed representation of a patch with compression factors of between 250x and 3000x depending on the modality. Finally, without tokenized representations of the image-like modalities, it is challenging to learn the correlation to sequence-like modalities. The overall training objective of TerraMind boils down to a cross-modal patch-level classification problem optimized via a cross entropy loss:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 125,
+ 605,
+ 296,
+ 637
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 125,
+ 605,
+ 296,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 605,
+ 296,
+ 637
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(p _ {i}\\right), \\tag {1}",
+ "image_path": "5c964073da44bbe5bd363bfad93b2c657fb846153d402f0c333cc6ceb4c0c0de.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "y_{i}"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": " is the one-hot encoded true class of token "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "p_{i}"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": " is the predicted probability for token "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": " is the total number of possible tokens. Interestingly, we can infer an upper bound loss for a random model where the cross entropy loss will collapse to the natural logarithm of the vocabulary size "
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}_{\\mathrm{CE,random}} = -\\sum_{i=1}^{N} y_{i} \\log \\left( \\frac{1}{N} \\right) = \\log(N)"
+ },
+ {
+ "bbox": [
+ 55,
+ 642,
+ 296,
+ 715
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 72,
+ 555,
+ 348
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 72,
+ 555,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 72,
+ 555,
+ 348
+ ],
+ "type": "text",
+ "content": "Scaling. We trained three versions of TerraMind scaling across model size, compute, and data. In addition, we pretrain different versions of TerraMind with respect to the number of dual-scale features. TerraMindv1-B is pre-trained on 500B tokens for 6 days on 32 NVIDIA A100 GPUs. The model uses dual-scale features from both token-level and pixel-level. During initial experiments, we observed significant improvements from scaling model size when switching from a tiny backbone to a small backbone to a base backbone. Therefore, we pre-trained TerraMindv1-L on a large backbone with 500B tokens on 32 NVIDIA A100 GPUs trained for 10 days. Finally, to better understand the effect of scaling across the dual-scale feature representation, we pre-train TerraMindv1-B-single as a single-scale model on primarily token-level data with optical S-2 L2A data as only pixel-level input (compared to pixel-level S-2 L1C, S-2 RGB, S-1 GRD, S-1 RTC, and DEM in TerraMindv1-B and -L). TerraMindv1-B-single is pretrained on 500B tokens from over one million samples for 6 days on 32 NVIDIA A100 GPUs. We summarize the scaling behavior in model size, compute, and data in Figure 9 of the supplementary material. We additionally provide final validation losses in Table 9 comparing v1-B and v1-L with the theoretical random loss."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 313,
+ 357,
+ 390,
+ 369
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 357,
+ 390,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 357,
+ 390,
+ 369
+ ],
+ "type": "text",
+ "content": "4.3. Generation"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 313,
+ 374,
+ 556,
+ 578
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 374,
+ 556,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 374,
+ 556,
+ 578
+ ],
+ "type": "text",
+ "content": "Once pretrained, TerraMind can generate tokens for any modality, conditioned on any subset of input modalities. The generative capabilities unlock various zero-shot tasks, such as water body segmentation. For the generation of image-like modalities, the decoder receives mask tokens for the modality to be generated and predicts the corresponding tokens based on the encoded input. For sequence-like modalities, the decoder generates the output autoregressively. After generating tokens from the target modality, the corresponding tokenizer decoder allows to map from token-space to image or text space. TerraMind further supports chained generation which ensures consistency across generated modalities. The chained generation represents a conditional probability distribution where the prior probability distribution is determined by the input modality, and all subsequent modalities are generated conditioned on the input modality and potentially other generated modalities."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 313,
+ 599,
+ 446,
+ 612
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 599,
+ 446,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 599,
+ 446,
+ 612
+ ],
+ "type": "text",
+ "content": "4.4. Thinking-in-Modalities"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "content": "Thinking in Modalities (TiM) is a recursive fine-tuning and inference technique designed to enhance multimodal learning by leveraging the generative capabilities of the model itself. Given an input "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "x \\in \\mathcal{X}"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "content": " (e.g., an optical satellite image), the model first generates additional synthetic modalities "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\tilde{x} = f_{\\mathrm{gen}}(x)"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "content": " on a token-level using a learned generative function "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{gen}}"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 713
+ ],
+ "type": "text",
+ "content": ". These generated tokens are then concatenated with the original input and jointly processed by the downstream"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "7387"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "text",
+ "content": "model "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "f"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "text",
+ "content": " (e.g., TerraMind encoder with a segmentation head), yielding the final output "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "y = f(x, f_{\\mathrm{gen}}(x))"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "text",
+ "content": ". This formulation allows the model to reason over both observed and inferred modalities, effectively enriching the input space. TiM can leverage multiple generated modalities which are then generated in a chained approach. For example, for "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 296,
+ 156
+ ],
+ "type": "text",
+ "content": " modalities, the input is augmented with newly generated modalities:"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 163,
+ 296,
+ 178
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 163,
+ 296,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 163,
+ 296,
+ 178
+ ],
+ "type": "interline_equation",
+ "content": "\\tilde {x} ^ {(k + 1)} = \\tilde {x} ^ {(k)} \\cup f _ {\\text {g e n}} (\\tilde {x} ^ {(k)}), \\tag {2}",
+ "image_path": "60b6eb23b3dd22e279e951a2d476553e023a1c300c717e47891d033a65012a8c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 186,
+ 228,
+ 198
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 186,
+ 228,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 186,
+ 228,
+ 198
+ ],
+ "type": "text",
+ "content": "and the final model output is described by:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 146,
+ 205,
+ 296,
+ 220
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 205,
+ 296,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 205,
+ 296,
+ 220
+ ],
+ "type": "interline_equation",
+ "content": "y = f \\left(\\tilde {x} ^ {(K)}\\right). \\tag {3}",
+ "image_path": "0ccb2554b06baa9d62c6a74f70c7943a8df9804907704db9b3a3219bda8f6be8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 55,
+ 227,
+ 297,
+ 264
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 227,
+ 297,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 227,
+ 297,
+ 264
+ ],
+ "type": "text",
+ "content": "This recursive augmentation mimics a chain-of-thought process, enabling the model to iteratively refine its internal representation, particularly in scenarios with missing modalities."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 55,
+ 276,
+ 137,
+ 289
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 276,
+ 137,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 276,
+ 137,
+ 289
+ ],
+ "type": "text",
+ "content": "5. Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 55,
+ 296,
+ 296,
+ 333
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 296,
+ 296,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 296,
+ 296,
+ 333
+ ],
+ "type": "text",
+ "content": "In this section, we describe the performance gains resulting from TerraMind and experiment with the unlocked capabilities of any-to-any generation and Thinking-in-Modalities."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 55,
+ 339,
+ 201,
+ 352
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 339,
+ 201,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 339,
+ 201,
+ 352
+ ],
+ "type": "text",
+ "content": "5.1. Foundational experiments"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 55,
+ 356,
+ 296,
+ 487
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 356,
+ 296,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 356,
+ 296,
+ 487
+ ],
+ "type": "text",
+ "content": "Multimodality vs. unimodality. As a first motivational experiment, we outline the benefit of using multimodal data in Earth observation at the example of water body mapping. Specifically, we leverage the ViT-B encoders from the unimodal tokenizer models for S-1, S-2, and LULC, concatenate their embeddings, and train a segmentation head with four ConvNeXt [43] blocks as a late fusion approach. The results in Table 2 (left) suggest that regardless of which modalities we combine, the combination of two modalities always outperforms each unimodal model. Combining all three modalities achieves the best overall performance."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 78,
+ 497,
+ 273,
+ 591
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 78,
+ 497,
+ 273,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 497,
+ 273,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 497,
+ 273,
+ 591
+ ],
+ "type": "table",
+ "html": "