[Paper Review] Analytical Insight of Earth: A Cloud-Platform of Intelligent Computing for Geospatial Big Data

Hao Xu, Yuanbin Man+4 more|arXiv (Cornell University)|Dec 27, 2023

Scientific Computing and Data Management8 citations

TL;DR

AI Earth is a cloud-native intelligent computing platform on Alibaba Cloud for geospatial big data, offering data catalogs, AI models, and an interactive/developer workflow with a large vision segmentation model and support for third-party models.

ABSTRACT

The rapid accumulation of Earth observation data presents a formidable challenge for the processing capabilities of traditional remote sensing desktop software, particularly when it comes to analyzing expansive geographical areas and prolonged temporal sequences. Cloud computing has emerged as a transformative solution, surmounting the barriers traditionally associated with the management and computation of voluminous datasets. This paper introduces the Analytical Insight of Earth (AI Earth), an innovative remote sensing intelligent computing cloud platform, powered by the robust Alibaba Cloud infrastructure. AI Earth provides an extensive collection of publicly available remote sensing datasets, along with a suite of computational tools powered by a high-performance computing engine. Furthermore, it provides a variety of classic deep learning (DL) models and a novel remote sensing large vision segmentation model tailored to different recognition tasks. The platform enables users to upload their unique samples for model training and to deploy third-party models, thereby increasing the accessibility and openness of DL applications. This platform will facilitate researchers in leveraging remote sensing data for large-scale applied research in areas such as resources, environment, ecology, and climate.

Motivation & Objective

• Address the challenge of processing and analyzing large-scale, multi-temporal remote sensing data beyond traditional desktop tools.
• Provide an intelligent cloud platform integrating data retrieval, AI model training, and application deployment for geospatial analysis.
• Enable flexible access to public datasets, user-uploaded data, and third-party models through GUI and code interfaces.

Proposed method

• Introduce AI Earth as a cloud-native platform built on Alibaba Cloud infrastructure with ACK, MaxCompute, OSS, and related services.
• Offer data catalogs with Landsat and Sentinel datasets, plus other global and regional geospatial data, preprocessed for distributed storage.
• Provide two development modes (toolbox web UI and developer code editor) and a unified API for computations on images and time series.
• Implement AI Earth Segment Anything (AIE-SEG), a large vision model with zero-shot capability for batch segmentation across images.
• Use a DAG-based computation engine with lazy evaluation, AST optimization, and caching to optimize distributed execution on a tile-based data store.
• Support for on-the-fly and batch computation modes, dynamic resource scheduling, and parallel tile/neighbor computations for scalable processing.

Figure 1: The framework of the AI Earth platform.

Experimental results

Research questions

• RQ1 How can cloud-enabled intelligent computing platforms efficiently manage and analyze geospatial big data at scale?
• RQ2 What architectural choices (DAG, lazy evaluation, tile-based storage) enable efficient, scalable remote sensing analytics in a cloud environment?
• RQ3 How can public and private remote sensing datasets be integrated and queried through a unified data catalog and STAC-compliant interface?
• RQ4 What capabilities (GUI and code interfaces) best facilitate model training, annotation, and deployment of DL models for geospatial tasks?
• RQ5 What mechanisms (caching, data loading strategies, on-the-fly vs batch) improve performance and resource utilization for large-scale time series analyses?

Key findings

• AI Earth provides a multi-petabyte data catalog including Landsat, Sentinel, and related datasets with daily updates (T+1) compared to official sources.
• The platform supports over 440 geospatial computation functions, including ML-based classification and regression, and enables training with user-provided data or pre-annotated samples.
• AIE-SEG offers interactive annotation and zero-shot capability for batch extraction of objects across images.
• A DAG-based execution model with AST optimization and lazy evaluation reduces redundant computations and optimizes data loading for leaf-node processing.
• On-the-Fly computation limits pixel calculations to the visible map viewport for debugging, while Batch computation supports specified output resolutions and projections to balance resources.
• A tile-based, distributed computation approach enables parallel processing of individual tiles and joint computation of neighboring tiles, with caching to avoid redundant work.

(a)