shield_lock
金大哥
← 返回 Skill 列表
extension
分类: 数据与分析需要 API Key

gee-tile-temporal-mosaic

Use when a Google Earth Engine task must build a low-cloud mosaic for an ROI across multiple Sentinel-2 or Landsat tiles while minimizing tile count, coordinating acquisition dates, and avoiding implicit cross-date per-pixel selection.

person作者: user_ae8b557chubcommunity
download下载资源包
<p align="center"> <h1 align="center">GEE Tile Temporal Mosaic Skill</h1> </p> <p align="center"> <a href="README_CN.md">中文说明</a> </p> <p align="center"> <img src="assets/gee-tile-temporal-mosaic-workflow.png" alt="Illustrated workflow for a tile-level temporally coordinated GEE mosaic"> </p>

Google Earth Engine skill for building a spatially complete, low-cloud ROI mosaic from the fewest required Sentinel-2 or Landsat tiles while keeping each tile tied to one selected acquisition.

This package is designed for OpenAI Codex and Claude Code. It is a workflow skill, not a generic qualityMosaic recipe: it selects tiles and scenes at the image level, coordinates dates across tiles, and uses fixed-priority overlap handling.

What It Solves

The skill is intended for requests such as:

  • “My city spans three Sentinel-2 MGRS tiles. Find the minimum tile set and make one low-cloud mosaic for May-September 2019-2023.”
  • “Use one scene per tile, keep the extra tiles within five days of the anchor, and calculate cloud fraction inside my ROI rather than using scene metadata.”
  • “Let the largest-coverage tile anchor the search, but evaluate several anchor candidates because the anchor date changes the other tile choices.”

It deliberately avoids silently mixing arbitrary dates per pixel. The default output is a priority_mosaic: one selected image per required tile, with the largest-contribution tile above lower-priority tiles. A lower tile can fill an upper tile's masked gap, but cannot replace a valid upper-tile pixel.

Compatibility

  • OpenAI Codex: install the repository as a skill folder under ~/.codex/skills.
  • Claude Code: install the same repository under ~/.claude/skills.
  • GEE Code Editor JavaScript and Earth Engine Python API/geemap are both supported. The agent must ask which backend to use before querying Earth Engine; it must not silently choose one.

The package contains instructions and reference templates. Earth Engine access, authentication, a Cloud Project ID, and an ROI are supplied by the user at run time; no credentials or private assets are bundled.

Installation

OpenAI Codex

git clone https://github.com/NightSensingLab/gee-tile-temporal-mosaic.git \
  ~/.codex/skills/gee-tile-temporal-mosaic

Invoke it explicitly with:

$gee-tile-temporal-mosaic

Claude Code

git clone https://github.com/NightSensingLab/gee-tile-temporal-mosaic.git \
  ~/.claude/skills/gee-tile-temporal-mosaic

Install the complete repository. The root SKILL.md, agents/openai.yaml, and references/ files are all part of the skill package.

Method In Brief

  1. Determine the minimum geometric tile set using incremental ROI coverage, not the sum of overlapping tile areas.
  2. Compute ROI-local cloud, clear, shadow, and footprint metrics for candidate scenes.
  3. Keep several anchor-scene candidates and evaluate coupled tile/date combinations instead of fixing the lowest-cloud anchor greedily.
  4. Apply hard thresholds and date-gap constraints, then rank final combinations by visible clear coverage, masked gaps, temporal spread, and target-date distance.
  5. Assemble the chosen images in explicit tile priority order and report the selected scene/date and quality diagnostics for every tile.

Every run also produces a practical handoff: a preview, a selection report, and a short copy-paste script that imports the exact selected scenes, reapplies the same mask, and builds the priority mosaic. The script exposes mosaic for clipping and downstream analysis. It never substitutes an unfiltered collection, qualityMosaic, or cross-date median.

The skill does not use qualityMosaic or a cross-date median by default. If no acceptable combination exists, it returns an explicit incomplete or no-solution state rather than silently expanding the time window.

Repository Layout

SKILL.md                         Core instructions and guardrails
agents/openai.yaml               Codex UI metadata
references/selection-design.md   Set cover, candidate search, scoring, overlap
references/sentinel2-javascript.md
                                 Sentinel-2 Code Editor pattern
references/python-geemap.md      Python/geemap pattern
references/landsat.md            Landsat Collection 2 masking notes
examples/                        Realistic prompts and expected diagnostics

Prompt Examples

Three-tile seasonal Sentinel-2 mosaic

Use $gee-tile-temporal-mosaic. My ROI spans three Sentinel-2 MGRS tiles.
Search 2019-2023 May-September. Require anchor local cloud <= 3%, other-tile
local cloud <= 20%, and a maximum five-day gap from the anchor. Keep one scene
per tile, evaluate several anchor candidates, use priority_mosaic, and output
GEE Code Editor JavaScript. Print tile dates, local cloud fractions, final
clear coverage, masked-gap fraction, and fallback state. Do not use
qualityMosaic or cross-date median.

Python/geemap output

Use $gee-tile-temporal-mosaic in Python/geemap mode. Preserve the same tile
selection, local cloud metrics, date-gap constraint, and overlap semantics.
Initialize Earth Engine with PROJECT_ID, keep server-side reducers, and make
Drive export opt-in.

Backend choice and handoff

If the request does not specify a backend, ask before running:

Use native GEE Code Editor JavaScript, Python/geemap, or both?

Native JavaScript returns a Code Editor script with Map.addLayer and print. Python/geemap returns the local script plus a geemap HTML/PNG preview. Both variants include a report with tile IDs, scene IDs, UTC dates, ROI-local cloud fractions, coverage, final visible clear/gap, overlap order, and fallback state, followed by a copy-paste import + mask + mosaic block.

Strict temporal ownership

Use $gee-tile-temporal-mosaic with overlapMode=exclusive_tile. No lower-priority
tile may fill a masked gap in another tile. Return an incomplete masked result
when a selected tile is cloudy.

Important Limitations

  • A five-year May-September search with one final output is a global best seasonal mosaic, not a representative result for each year.
  • priority_mosaic can expose a lower tile in an upper tile's masked gap; the per-tile date difference is therefore reported explicitly.
  • Exact candidate-combination search grows quickly with tile count. Use a small topN or a bounded beam search for large tile sets and state the approximation.
  • Local cloud metrics depend on the chosen cloud-probability and scene-class thresholds. Snow, bright roofs, haze, and cloud shadow require study-specific review.

License

The original skill files are released under the MIT License. Dataset terms and Earth Engine collection terms remain governed by their respective providers; see THIRD_PARTY_NOTICES.md.