.plan-26-10: Streaming TESSERA working, biodiversity action papers, and FPL takes off

Key findings from this study

• The authors demonstrate a fully browser-based streaming interface for TESSERA satellite embeddings using JavaScript, WebGPU, and WebAssembly with Zarr chunking and HTTP range requests.
• The work identifies a tension between forward programming, backwards programming with learned embeddings, and causal inference programming as three distinct styles requiring integration in geospatial machine learning workflows.
• The authors report that functional programming specialists proposed multi-domain-specific language architectures and effects-based interpretation as potential frameworks for unifying these programming paradigms.

Overview

This work presents technical progress on TESSERA, a browser-based streaming interface for satellite embeddings, and explores programming paradigm challenges associated with geospatial machine learning workflows. The author demonstrates a client-side implementation using JavaScript, WebGPU, and WebAssembly that streams Zarr-formatted embeddings via HTTP range requests, enabling real-time visualization and classification workflows including solar panel detection. The development is motivated by integration goals with existing dashboard infrastructure and addresses practical computational constraints inherent to browser-based geospatial analysis.

The work situates these technical developments within broader discussions of programming language design for scientific computing. Presentations at the WG2.8 functional programming working group framed a design problem concerning three distinct programming styles: forward programming for conventional computational workflows, backwards programming for machine learning with high-dimensional embeddings, and causal relationship establishment for scientific theory testing. The author queries how these styles might be unified within a single programming framework, receiving suggestions ranging from multi-domain-specific language approaches to effects-based interpretations. Parallel server-side implementations and ongoing transcoding efforts from numpy to Zarr formats are described as complementary production-oriented pathways.

Methods and approach

The streaming interface implements client-side processing architecture utilizing JavaScript, WebGPU, and WebAssembly to perform geospatial analysis within web browsers. TESSERA embeddings are stored in chunked Zarr format and served over HTTP, with range requests minimizing data transfer volumes. The implementation supports both classification workflows and patch-based solar panel detection using existing embedding models. A caching proxy implementation enables offline functionality through remote tile caching on local systems.

Development of production infrastructure involves an OxCaml pipeline for transcoding petabyte-scale embeddings from numpy to Zarr format, requiring solutions to parallelization challenges. A from-scratch OxCaml-Zarr implementation was initiated collaboratively, with plans to import existing ocaml-tessera pipelines to enable model training and tile inference entirely within OCaml. The work proceeds through independent parallel implementations with iterative exchange of interface design concepts and analytical approaches, distinguishing browser-oriented dynamic workflows from server-based production deployments serving downstream task projects.

Results

The streaming interface achieves functional real-time visualization of global TESSERA embeddings through browser-based architecture. Client-side processing successfully executes classification workflows and solar panel detection using patch-based embeddings, with demonstrated offline capability via caching mechanisms. The implementation provides interactive false-color visualization at global scale with responsive zooming functionality. Integration pathways with existing dashboard infrastructure and areas-of-habitat datasets are established, with browser-based limitations explicitly recognized for computationally intensive analyses requiring server infrastructure.

Presentation of the planetary programming design problem to functional programming specialists generated multiple conceptual frameworks. Respondents suggested multi-domain-specific language implementations within OCaml to separately handle forward programming, backwards programming with embeddings, and causal relationship testing. Effects-based interpretation emerged as an alternative approach, potentially enabling single program representations that could be interpreted differently using effects for sampling, stochastic simulation, and causal hypothesis testing through data structure invariant checking. Analogies to hardware programming, property-based testing frameworks, array programming for spatial ecology, and bidirectional lenses connecting synthetic models with observational samples were proposed as relevant paradigms. Practical guidance on OxCaml development emphasized current limitations in layout polymorphism while identifying feasible implementations for features such as float16 support relevant to embedding workflows.

Implications

The demonstration of browser-based streaming for satellite embeddings establishes technical feasibility for client-side geospatial machine learning workflows, with implications for deployment architectures that minimize server dependencies and enable interactive visualization at planetary scale. The explicit recognition of computational limitations suggests a stratified deployment model where browser interfaces serve exploratory and visualization functions while server infrastructure handles production-scale analyses. Integration with dashboard systems and habitat datasets indicates pathways toward operational environmental monitoring applications accessible through standard web technologies.

The programming paradigm inquiry raises fundamental questions about language design for scientific computing that combines conventional computation, machine learning inference on learned representations, and causal inference workflows. The multi-domain-specific language approach suggests architecture where distinct computational styles are implemented as separate but interoperable languages, while effects-based approaches propose unified representations with interpretation-time differentiation. The synthesis of perspectives from hardware design, property testing, spatial ecology computing, and bidirectional programming suggests cross-domain analogies that may inform development of programming frameworks for Earth observation workflows. The connection to ongoing transcoding efforts and production system deployments indicates practical constraints shaping theoretical programming language research in this domain.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.