fix: remove gf3 water priors from production workflow

.env.example

@@ -119,8 +119,6 @@ IDL_DINSAR_DEM_BASE_FILE=D:\SRTM30m\SRTMDEM_RSP_SARscape
 SRTM_DEM_DIR=D:\SRTM30m
 GF3_GEO_DEM_PATH=D:\DEM\gf3_dem.jp2
 WATER_RESULTS_DIR=D:\WaterResult
-GF3_WATER_USE_DLTB=false
-GF3_WATER_DLTB_CACHE_DIR=
 GF3_WATER_DEM_PATH=
 GF3_WATER_DEFAULT_CARTOGRAPHIC=true
 GF3_WATER_DEFAULT_OUT_VECTOR=true

backend/app/config.py

@@ -215,8 +215,6 @@ class Settings(BaseSettings):
     RADAR_PREVIEW_BUILD_ON_DEMAND: bool = True
 
     WATER_RESULTS_DIR: str = ""
-    GF3_WATER_USE_DLTB: bool = False
-    GF3_WATER_DLTB_CACHE_DIR: str = ""
     GF3_WATER_DEM_PATH: str = ""
     GF3_WATER_DEFAULT_CARTOGRAPHIC: bool = True
     GF3_WATER_DEFAULT_OUT_VECTOR: bool = True

backend/app/services/gf3_water_extraction_service.py

@@ -134,15 +134,8 @@ def run_gf3_hh_hv_water_extraction(
 
     hh = _existing_path(hh_path, label="HH input")
     hv = _existing_path(hv_path, label="HV input")
-    use_dltb = _as_bool(params.get("use_dltb"), bool(settings.GF3_WATER_USE_DLTB))
-    dltb_cache_dir = (
-        _optional_existing_path(
-            params.get("dltb_cache_dir") or settings.GF3_WATER_DLTB_CACHE_DIR,
-            label="GF3_WATER_DLTB_CACHE_DIR",
-        )
-        if use_dltb
-        else None
-    )
+    use_dltb = False
+    dltb_cache_dir = None
     dem = _optional_existing_path(
         params.get("dem") or params.get("dem_path") or settings.GF3_WATER_DEM_PATH,
         label="GF3_WATER_DEM_PATH",
@@ -162,9 +155,9 @@ def run_gf3_hh_hv_water_extraction(
         out_dir=out_dir,
         dem=dem,
         dltb_cache_dir=dltb_cache_dir,
-        dltb_mode=str(params.get("dltb_mode") or "soft"),
-        water_vector=_path_param_list(params, "water_vector"),
-        paddy_vector=_path_param_list(params, "paddy_vector"),
+        dltb_mode="off",
+        water_vector=[],
+        paddy_vector=[],
         threshold_method=str(params.get("threshold_method") or "percentile"),
         score_percentile=_numeric_param(params, "score_percentile", 95.0, float),
         hv_percentile=_numeric_param(params, "hv_percentile", 20.0, float),
@@ -227,9 +220,13 @@ def run_gf3_hh_hv_water_extraction(
             "hv": str(hv),
             "dem": str(dem) if dem else None,
             "dltb_cache_dir": str(dltb_cache_dir) if dltb_cache_dir else None,
+            "water_vector": [],
+            "paddy_vector": [],
         },
         "runtime": {
-            "dltb_enabled": use_dltb and dltb_cache_dir is not None,
+            "prior_inputs_enabled": False,
+            "dltb_enabled": False,
+            "deep_learning_enabled": False,
             "vector_requested": out_vector,
             "vector_runtime_available": vector_runtime_available,
             "vector_output_enabled": vector_output_enabled,

docs/GF3_WATER_EXTRACTION_INTEGRATION_20260615.md

@@ -15,12 +15,18 @@ The goal is integration, not bulk import. Code that is reusable should be copied
 - `gf3_water/`: active Python package for GF-3 HH/HV water extraction.
 - `scripts/`: thin CLI/data-preparation wrappers.
 - `docs/`: algorithm and integration notes.
-- `data/`: local scenes and prior data.
+- `data/`: local scenes and prior data from the experiment workspace.
 - `outputs/`: generated products.
 - `gf3-water-ai4g-unet/`: legacy deep-learning experiment/checkpoints.
 
 The active production-facing implementation is the non-DL `gf3_water` package. It consumes SARscape ENVI HH/HV geocoded assets, for example `_hh_geo` and `_hv_geo`, and writes raster, preview, vector and `metadata.json` outputs.
 
+Production policy:
+
+- Do not use DLTB, hydro, water-vector or paddy-vector priors.
+- Do not use deep-learning checkpoints or U-Net inference.
+- Use the current HH/HV machine-learning-style threshold, candidate and morphology workflow.
+
 ## Current System Entry Points
 
 The existing system already has a suitable production chain:
@@ -88,16 +94,15 @@ Do not store runtime data in Git.
 
 Current integration does not use DLTB priors. The GF-3 HH/HV processor runs from SAR backscatter, morphology and optional vector/DEM inputs only.
 
-Optional future runtime assets:
+Not transferred:
 
 - `D:\Code\Water\data\priors\dltb_cache\heilongjiang`
   - Current size observed: 25 files, about 8.5 GB.
-  - Do not transfer for the current workflow.
-  - If DLTB is re-enabled later, transfer to a managed runtime asset path, for example `D:\production_assets\gf3_water\priors\dltb_cache\heilongjiang`, then set `GF3_WATER_USE_DLTB=true` and `GF3_WATER_DLTB_CACHE_DIR`.
-
-Optional runtime assets:
-
 - Hydro prior vectors under `data/priors/hydro`.
+- Water/paddy vector priors.
+
+Optional runtime asset:
+
 - DEM path if slope filtering should be enabled.
 
 Build-only assets:
@@ -118,8 +123,6 @@ These are large model artifacts and should stay outside the application until a
 
 Add configuration keys:
 
-- `GF3_WATER_USE_DLTB=false`
-- `GF3_WATER_DLTB_CACHE_DIR`
 - `GF3_WATER_DEM_PATH`
 - `GF3_WATER_DEFAULT_CARTOGRAPHIC=true`
 - `GF3_WATER_DEFAULT_OUT_VECTOR=true`
@@ -156,5 +159,5 @@ flowchart TD
 ## Open Decisions
 
 - Whether runtime execution should be in-process Python API first or always subprocess CLI. Initial integration should use in-process API because it fits the existing worker model and keeps job accounting simple.
-- DLTB priors are disabled for the current workflow. Re-enabling them later should be a deployment decision because the cache is about 8.5 GB.
-- Whether legacy AI4G U-Net should be supported later. It should not block the current HH/HV production chain.
+- DLTB, hydro, water-vector and paddy-vector priors are not part of the current workflow.
+- Legacy AI4G U-Net/deep-learning checkpoints are not part of the current workflow.