DryForM

DryForM is a Google Earth Engine (GEE) workflow for dry-forest and tree-cover mapping, model training, post-processing, disturbance analysis, and interactive time-series exploration.

The repository contains the script pipeline developed in the DryForM exploratory work (2022–2024), organized as numbered modules that can be run step-by-step in GEE.

---

What this repository contains

This project is a scripted workflow (mostly JavaScript for GEE Code Editor) covering:

• creation of a sampling framework from multiple land-cover sources,
• ground-truth preparation and balancing,
• Sentinel-1/Sentinel-2 preprocessing,
• Random Forest training and inference,
• AlphaEarth embedding-based training experiments,
• optional object-based segmentation,
• post-processing and area summaries,
• disturbance detection with CCDC,
• an app-like time-series explorer.

All major modules are under DryForM_Workflow/ and are prefixed with step numbers to indicate execution order.

---

Repository structure (full map, excluding req/ internals)

.
├── .gitignore
├── README.md
└── DryForM_Workflow/
    ├── 1_SamplingFramework/
    │   ├── 1_0_LCProc_v3_2
    │   └── 1_1_Export2Drive_Optional
    ├── 2_Samples/
    │   ├── 2_0_AutoSampling
    │   ├── 2_1_GroundTruth_Visu
    │   ├── 2_2_BalanceGT
    │   └── 2_3_Separability
    ├── 3_Preprocessing/
    │   ├── 3_0_PrepS1S2_v9_(working_Allinone)
    │   └── s1_code/
    │       ├── wrapper
    │       ├── s1_ard
    │       ├── utilities
    │       ├── speckle_filter
    │       ├── terrain_flattening
    │       └── border_noise_correction
    ├── 4_Segmentation/
    │   └── 4_0_SNIC_Segmentation_ObjectOriented
    ├── 5_Training_Inference/
    │   ├── 5_1_HyperparameterTunning
    │   ├── 5_2_Training_withAutoSample
    │   ├── 5_2_Training_withGroundTruth
    │   ├── 5_3_ApplySavedModel_Optional
    │   ├── 5_Training_Inference_5_2_Training_wtihGroundTruth_Embeds
    │   └── 5_Training_Inference_5_2_Training_wtihGroundTruth_Embeds_BenchmarkMode
    ├── 6_Postprocessing/
    │   └── 6_0_PostProc_v1
    ├── 7_Disturbance/
    │   ├── 7_0_CCDC_InspectPixels_v3
    │   ├── 7_1_CCDC_CreateAutomated_v3
    │   ├── 7_2_CCDC_Extract_NonAutomated_v3
    │   └── req/
    ├── 8_Utils/
    │   ├── 8_0_Visualize_ClassificationResults
    │   ├── 8_1_recurrent_IC_removal
    │   └── 8_2_CalculateAreas
    └── 9_App/
        ├── Time Series Explorer with images_df_vd_b2v10
        └── req/

Note: scripts are stored without .js extensions, but are intended for Google Earth Engine JavaScript workflows.

---

Workflow overview (recommended order)

1) Sampling framework

• 1_0_LCProc_v3_2 builds a preliminary LC reference layer from multiple global products and AOIs.
• 1_1_Export2Drive_Optional optionally exports that layer to Google Drive.

2) Sample preparation

• 2_0_AutoSampling generates auto-sampled points from step 1 outputs.
• 2_1_GroundTruth_Visu loads/visualizes manually interpreted ground truth.
• 2_2_BalanceGT balances class distributions for training/validation.
• 2_3_Separability computes separability metrics (e.g., Jeffries–Matusita).

3) S1/S2 preprocessing

• 3_0_PrepS1S2_v9_(working_Allinone) prepares multi-temporal Sentinel-1/Sentinel-2 composites and exports tiled assets.
• Uses helper modules in 3_Preprocessing/s1_code/ for Sentinel-1 ARD processing.

4) (Optional) object-based segmentation

• 4_0_SNIC_Segmentation_ObjectOriented runs SNIC segmentation for object-oriented workflows.

5) Training and inference

• 5_1_HyperparameterTunning tunes Random Forest parameters.
• 5_2_Training_withGroundTruth runs supervised RF training with visual GT.
• 5_2_Training_withAutoSample alternative training with auto-sampled GT.
• 5_3_ApplySavedModel_Optional applies previously exported model assets.
• AlphaEarth/Satellite-Embeddings scripts currently present in this repository root of step 5:
  • 5_Training_Inference_5_2_Training_wtihGroundTruth_Embeds
  • 5_Training_Inference_5_2_Training_wtihGroundTruth_Embeds_BenchmarkMode

6) Post-processing

• 6_0_PostProc_v1 mitigates no-data artifacts in classification outputs.

7) Disturbance analysis (CCDC)

• 7_1_CCDC_CreateAutomated_v3 creates CCDC outputs.
• 7_2_CCDC_Extract_NonAutomated_v3 extracts disturbance-relevant bands.
• 7_0_CCDC_InspectPixels_v3 interactive per-pixel inspection.
• Uses reusable helper module under 7_Disturbance/req/.

8) Utilities

• 8_0_Visualize_ClassificationResults for visual QA/QC.
• 8_1_recurrent_IC_removal for recurrent deletion of images in image collections.
• 8_2_CalculateAreas calculates Forest / Non-forest area totals by zone (World_UNctry_grid) using pixel-area summaries and exports table results to Drive.

9) App / integration

• Time Series Explorer with images_df_vd_b2v10 integrates classification, imagery chips, temporal profiles, and disturbance layers.
• Uses helper modules under 9_App/req/.

---

Helper and support files

3_Preprocessing/s1_code/

These files implement the Sentinel-1 ARD chain used by step 3:

• wrapper: orchestrates the S1 preprocessing workflow and calls the other helper modules.
• s1_ard: core ARD routine and parameter-driven S1 preprocessing logic.
• utilities: shared math/conversion functions (e.g., dB-linear conversions).
• speckle_filter: mono/multi-temporal speckle filtering implementations.
• terrain_flattening: slope/radiometric terrain correction functions.
• border_noise_correction: additional Sentinel-1 border noise masking.

5_Training_Inference/ (AlphaEarth note)

The currently tracked repository structure contains two embedding-oriented scripts in 5_Training_Inference/.

If additional files now exist in another branch/path (for example an AlphaEarth/ subfolder or *_Embeds_v3.js variants), they are not present in this checked-out tree and should be added/committed so they can be documented here precisely.

req/ folders

Both 7_Disturbance/req/ and 9_App/req/ contain reusable helper modules required by the main scripts in those directories.

.gitignore

Includes .DS_Store ignore rule for macOS metadata files.

---

Prerequisites

• A Google Earth Engine account with permission to run and export tasks.
• Access to required GEE assets referenced in scripts (AOIs, GT, intermediate outputs).
• (Optional) OEEex plugin or task management strategy for large export batches.

---

How to use

1. Open Google Earth Engine Code Editor.
2. Copy a script from this repository into a new GEE script.
3. Update parameters in each script (AOI, year, asset paths, class fields, export paths).
4. Run modules in numbered order (1 → 9), using optional modules as needed.
5. Validate outputs after each stage before proceeding.

Because many scripts reference private/project assets, adaptation to your own GEE assets is expected.

---

Areas of interest and classes

The workflow is configured around multiple AOIs (e.g., Chaco, Caatinga, Cerrado, Tanzania/Tanganyika in different script versions) and supports reclassified LC targets for model training.

Class schemas and remapping arrays are defined directly in training scripts; adjust them consistently across balancing, training, validation, and visualization modules.

---

Outputs

Depending on the executed path and options, outputs may include:

• LC proxy/sampling framework layers,
• training and validation sample assets,
• quarterly (or alternative interval) S1/S2 composites,
• trained classifier assets and classified rasters,
• post-processed classification maps,
• CCDC disturbance layers,
• area-estimation tables,
• summary tables and visualization layers.

---

Important notes

• This repo captures working research scripts; variable names, comments, and assumptions are tuned to DryForM processing context.
• Some modules are explicitly marked as optional or experimental.
• Several scripts note known limitations (e.g., Sentinel-1 overlap artifacts, non-automated extraction steps) and ongoing optimization ideas.
• Some helper modules were adapted from external community/academic code and are integrated here for workflow reproducibility.

---

Contributing

Contributions are welcome for:

• documentation improvements,
• script refactoring and parameterization,
• reproducibility improvements (config-driven assets/AOI setup),
• automation of currently manual steps.

When submitting changes, please keep the numbered workflow logic intact and document any breaking changes in step dependencies.