Satellite archives are too large to download, so modern Earth-observation pipelines read small windows of imagery directly from cloud storage, run a model on each window ("chip") independently, and stitch the predictions back together. This design is forced partly by the transformer's quadratic cost. Self-attention compares every image patch with every other patch, so memory and compute grow with the square of the number of patches, and a whole scene or region cannot fit on one device. The price is that the model never sees past the chip edge. Fields, rivers and deforestation fronts that cross tile boundaries get cut in half, and a pixel near the edge has less context than a pixel in the center.
This proposal uses a structural property of linear recurrent models, the family that includes state space models such as Mamba and linear recurrent units such as BLUR. Their per-layer update is a linear map, and composing linear maps is associative. Each cloud worker can therefore process its own tile and emit a tiny summary of how its tile transforms an incoming hidden state. A cheap scan over those summaries across workers then gives every tile the exact state it would have received if one machine had processed the whole region in a single pass. The claim is that this turns tile-by-tile cloud inference into exact whole-region inference. Communication per layer is proportional to the hidden-state size times the number of tiles, not to the number of pixels, and no worker ever holds more than its own chunk. We will test whether that exact long-range state actually improves maps near tile boundaries enough to matter.
Core Claim
For a diagonal linear recurrence $h_t = a_t \odot h_{t-1} + b_t$, the state at position $t$ is a function of the input sequence. Here $h_t$ is the hidden state at position $t$, $a_t$ is the (possibly input-dependent) decay, $b_t$ is the input contribution, and $\odot$ is elementwise product. A contiguous segment of positions acts on an incoming state as a single affine map $(A, B)$. Two consecutive segments compose as
$$(A_2, B_2) \circ (A_1, B_1) = (A_2 \odot A_1,\; A_2 \odot B_1 + B_2),$$
and this composition is associative. This is the property that parallel prefix scans already exploit inside one GPU, as in BLUR, which uses associative scans in both forward and backward directions. We lift the same scan one level up, from threads inside a GPU to workers reading different chunks of a cloud-optimized array.
Each recurrent layer then runs in three steps. First, each worker scans its tile locally from a zero state and records its segment map. Second, the workers perform a scan across tiles over these small maps, whose size is the state dimension times the channel count per tile. Third, each worker corrects its local states with the incoming carry, which is an elementwise update. Non-recurrent parts of the block, such as the short convolution in Mamba, need only a halo of a few tokens from neighboring tiles. The MLPs are fully local.
The result matches single-device execution up to floating-point reduction order. It costs one small synchronization per layer.
The 2D ordering problem is handled by making the scan order match the storage layout. The Vision Mamba in Remote Sensing survey identifies the flattening of 2D images into 1D scans as the main weakness of SSMs on imagery. We therefore use a tile-major hierarchical order: tiles are visited in a serpentine path, and pixels within each tile are visited row by row. Each tile is then a contiguous segment of the global sequence and coincides with a Zarr or Cloud-Optimized GeoTIFF chunk. Several directions (four serpentine orientations, run forward and backward in the style of BLUR) are separate scans that are merged per pixel.
This is offline archival mapping, not live streaming. BLUR's requirement to see the whole sequence, which rules it out for streaming, is therefore not a limitation here.
Relation to Prior Work
Three pieces of prior work motivate the gap without closing it.
TorchGeo provides lazy windowed reads and a grid sampler that fetches chips independently for inference. It is the natural host for this work, but it has no mechanism for carrying model state between chips.
The Vision Mamba survey catalogs roughly 120 SSM models for remote sensing and explicitly notes that none addresses execution over cloud chunking or lazy data graphs.
From Copernicus Big Data to Extreme Earth Analytics parallelizes deep learning next to Sentinel archives. Its parallelism is over independent tiles, whereas we make tiles dependent in a controlled way.
Two papers supply methods we reuse or compare against. BLUR supplies the recurrence we build on: diagonalized linear recurrent units with eigenvalues constrained inside the unit circle, run bidirectionally with parallel scans. Its benchmarks are 1D forecasting (ETT, Weather, ECL) and sequential image classification on a single device, while we add cross-worker scans and a storage-aligned 2D order. Chunked TabPFN shows that exact attention can be computed chunk by chunk with an online softmax, keeping memory linear while compute stays quadratic. It is the closest rival, since it too gives exact whole-context results from chunks. The difference is what must move between chunks: attention needs every chunk's keys and values at every layer, while recurrence needs only a fixed-size state.
One line of work constrains the design. DynamicEarthNet found that feeding denser daily sequences hurt U-TAE and ConvLSTM, and attributes this to redundancy, atmospheric artifacts and training instability, not to sequence cost. We read this as a warning: longer context is not automatically useful, so we must measure how far useful information actually travels. For the same reason we deliberately do not make "fix the daily paradox with Mamba" the core claim. Attention over 30 frames is cheap, so quadratic cost is not what is failing there.
Research Program
We will build a TorchGeo-compatible sampler that emits tiles in the hierarchical scan order, together with a distributed executor that runs the three-step layer protocol across workers, using a Dask- or Ray-style scheduler over Zarr stores.
The model is a U-Net-shaped segmentation network whose bottleneck and decoder stages use bidirectional multi-direction linear recurrent blocks. We use two variants, BLUR-style units with fixed decays and selective Mamba-style units with input-dependent decays, because the survey reports that selective scans dominate remote-sensing results.
Models are trained on regions of moderate size (for example 4×4 tiles) that fit on one GPU, using the identical tile-major order. At inference the model runs on regions tens to hundreds of tiles across. This raises a length-generalization question: does a model trained on 16-tile sequences behave well on 1,000-tile sequences? We will address it by training with randomized region sizes and by measuring the decay spectrum of learned states.
Three analyses accompany the system. The first is an exactness check comparing distributed outputs with single-device outputs on regions small enough to run both. The BFASTmonitor port to Earth Engine documents numerical divergence when solvers move between runtimes, so this check is necessary rather than cosmetic. The second is a communication and wall-clock model as a function of tile count and layer count. The third is an influence analysis, in which we measure the gradient of an output pixel with respect to inputs k tiles away, giving the effective context radius the learned models actually use.
Baselines
The naive reference is the same recurrent network run as it would be today: each chip is processed independently from a zero initial state, and predictions are stitched. This is also the ablation that isolates the key ingredient. Weights, data and compute are identical, and only the cross-tile state carry is removed. Beating it means higher intersection-over-union overall and, more specifically, closing the gap between accuracy on pixels within a few tokens of a tile edge and accuracy in tile interiors.
The standard practical baseline is overlapped tiling with blended stitching, where each chip is read with a margin and only its center is kept. We will run it with overlap margins of 0, 12.5, 25 and 50 percent, charging the extra reads and compute to its budget. The carry-over method must beat it at equal bytes read, or match it while reading fewer bytes.
The convolutional reference is TorchGeo's DeepLabV3 with a ResNet-50 backbone on independent chips. On the Earth Surface Water dataset the tutorial reports overall accuracy 0.977 and IoU 0.824 after 10 epochs (TorchGeo tutorial). We rerun it under our tuning budget rather than quoting it as a target.
The strongest competitor on accuracy is a ViT segmentation model whose attention spans the whole region, computed chunk by chunk with the online-softmax scheme of Chunked TabPFN. This gives exact global attention and is the natural "just keep attention" answer. It gets the same training data, the same number of tuning trials, and the same GPU-hour budget. We report its accuracy, FLOPs and bytes moved between workers. No Lacuna paper reports a result for chunked exact attention on geospatial segmentation, so there is no published number to quote.
Beating this baseline does not require higher accuracy. It means matching its accuracy within noise at a clearly lower cost in compute and communication, or exceeding it at equal cost.
For the temporal extension, the reference is U-TAE on DynamicEarthNet. It reports 37.9% mIoU with purely supervised monthly training and 43.6% with context-aware consistency semi-supervision (DynamicEarthNet).
Why not just…?
Why not just use overlapping tiles? Overlap fixes seams only out to the overlap margin, and its cost grows with the square of the margin, since the whole border band is re-read and recomputed. It cannot deliver context from several tiles away, such as the upstream course of a river or the rest of a large field. This is a real threat if useful context rarely exceeds a fraction of a tile. The influence-radius analysis and the margin sweep settle it: if a 25% overlap closes most of the edge-versus-interior gap, the carry-over mechanism is not worth its synchronization cost, and we will say so.
Why not just stream exact attention chunk by chunk? It gives exact global context and reuses pretrained vision transformers, which is a genuine advantage. Two costs count against it. Its compute stays quadratic in region size, as the Chunked TabPFN authors themselves note. In a distributed setting, every worker must also receive every other worker's keys and values at every layer, so communication grows with the number of pixels rather than with a fixed state size. Our cost curves over increasing region size will show where, if anywhere, the two approaches cross.
Why not just add a downsampled overview branch for context? A coarse overview supplies long-range context cheaply and is a strong, simple competitor. It loses fine detail at a distance, for example a narrow canal or a field boundary two tiles away. We will include a two-branch variant, a chip plus a 4×-downsampled region overview, as an additional baseline under the same budget. If it matches carry-over on Fields of The World boundaries, the exactness of our method is buying little.
Why not just subsample time or use harmonic models for the temporal axis? Our core claim is spatial. We apply the same scan over Zarr time-chunks only as a secondary test on the full two-year daily DynamicEarthNet series of about 730 frames. Harmonic change tests such as BFASTmonitor work on a single spectral index and do not produce multi-class maps. Given the daily-paradox finding, though, we expect temporal gains to hinge on cloud-aware input handling more than on sequence length.
Proposed Experiments
For spatial segmentation we use two datasets. The first is Fields of The World, whose field polygons regularly cross tile edges. The only number the file reports for it is a macro F1 of 0.65–0.75 for crop classification with MOSAIKS embeddings on Iowa, which is a different task, so no reported field-boundary segmentation result was found on Lacuna. The second is the Earth Surface Water dataset from the TorchGeo tutorial, where long rivers make distant context plausibly useful.
We evaluate each model on regions of 16, 64, 256 and 1,024 tiles. Metrics are IoU, object-level F1, and recall for objects spanning two or more tiles. We also report the edge-band accuracy gap, defined as interior accuracy minus accuracy within a fixed pixel distance of a tile edge. Cost is reported as GPU-hours, peak memory per worker, bytes read from object storage, and bytes exchanged between workers.
Four checks decide the outcome. First, distributed and single-device outputs must agree to within float32 tolerance; otherwise the exactness claim fails. Second, carry-over must shrink the edge-band gap relative to zero-state chips and to the best overlap setting at equal bytes read. Third, the influence radius must exceed one tile on at least one task; if it does not, the idea reduces to overlap tiling. Fourth, against chunked exact attention, carry-over must stay within a small accuracy margin while its compute and communication grow roughly linearly with region size, where attention's grow quadratically.
The temporal run on DynamicEarthNet is reported separately against U-TAE's 37.9% supervised figure, with and without cloud-masked inputs, so that any gain can be attributed either to sequence modeling or to cleaner data.
References
- Bidirectional linear recurrent models for sequence-level multisource fusion
- Vision Mamba in Remote Sensing: A Comprehensive Survey of Techniques, Applications and Outlook
- Advancing Earth Observation Through Machine Learning: A TorchGeo Tutorial
- From Copernicus Big Data to Extreme Earth Analytics
- Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data
- DynamicEarthNet: Daily Multi-Spectral Satellite Dataset for Semantic Change Segmentation
- Implementation of BFASTmonitor Algorithm on Google Earth Engine to Support Large-Area and Sub-Annual Change Monitoring Using Earth Observation Data
- Fields of The World: A Field Guide for Extracting Agricultural Field Boundaries