Skip to content

flood

Flood susceptibility and event-risk mapping. Entry point: FloodRiskUseCase.

from climate_change import FloodRiskUseCase

Use case

use_case

FloodRiskUseCase 

FloodRiskUseCase(dask_engine: DaskEngine)

Bases: BaseUseCase

Entry point for the flood risk domain.

Minimal config (all optional — defaults match the Niger flood notebook example):

{ "aoi_geojson": {"type": "Polygon", "coordinates": [...]}, "gee_project": "your-gee-project-id", "model_type": "ensemble", # "rf" | "xgboost" | "ensemble" "output_dir": "outputs", "prefix": "flood_2022_2023", "flood_event": "Niger River flood pulse Oct 2022-Jan 2023", # Main analysis period "start_date": "2022-01-01", "end_date": "2023-01-31", # Flood event windows "pre_flood_start": "2022-05-01", "pre_flood_end": "2022-07-31", "post_flood_start": "2022-10-01", "post_flood_end": "2023-01-31", # Rainfall windows "rain_7d_start": "2022-08-18", "rain_7d_end": "2022-08-25", "rain_30d_start": "2022-08-01", "rain_30d_end": "2022-08-31", # Sentinel-2 MNDWI window "mndwi_start": "2022-10-01", "mndwi_end": "2023-01-31", # JRC flood label window "flood_label_start":"2021-01-01", "flood_label_end":"2021-12-31", # Sampling "n_pixels": 3000, "scale": 90, }

Source code in flood/use_case.py 
78
79
def __init__(self, dask_engine: DaskEngine) -> None:
    super().__init__(dask_engine)

fetch_data 

fetch_data(config: AnalysisConfig) -> dict

Authenticate GEE and parse the AOI geometry (see _fetch_from_dict).

Source code in flood/use_case.py 
83
84
85
86
87
88
89
90
91
92
def fetch_data(self, config: AnalysisConfig) -> dict:
    """Authenticate GEE and parse the AOI geometry (see `_fetch_from_dict`)."""
    flat: dict = {
        "aoi_geojson": config.aoi_geojson,
        "start_date": config.start_date,
        "end_date": config.end_date,
        "country": config.country,
        **config.extra_params,
    }
    return self._fetch_from_dict(flat)

preprocess 

preprocess(raw_data: dict, config: AnalysisConfig) -> dict

Build the feature stack and sample flooded/dry pixels (see _preprocess_raw).

Source code in flood/use_case.py 
94
95
96
def preprocess(self, raw_data: dict, config: AnalysisConfig) -> dict:
    """Build the feature stack and sample flooded/dry pixels (see `_preprocess_raw`)."""
    return self._preprocess_raw(raw_data)

run_model 

run_model(features: dict, config: AnalysisConfig) -> AnalysisOutput

Train models, export the risk COG, and package an AnalysisOutput.

Source code in flood/use_case.py 
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def run_model(self, features: dict, config: AnalysisConfig) -> AnalysisOutput:
    """Train models, export the risk COG, and package an `AnalysisOutput`."""
    model_type = config.extra_params.get("model_type", "ensemble")
    dict_config = {"model_type": model_type, **config.extra_params}
    result = self._run_model_dict(features, dict_config)
    raster_paths: dict[str, str] | None = None
    raster_error: str | None = None
    raster_result: FloodCogResult | None = None
    try:
        raster_result = export_flood_cog(
            rf_model=features["_rf"],
            xgb_model=features["_xgb"],
            datasets=features["datasets"],
            output_dir=dict_config.get("output_dir", "outputs"),
            prefix=dict_config.get("prefix", "flood"),
            model_type=model_type,
            aoi_geojson=config.aoi_geojson,
        )
        raster_paths = {"flood_risk": raster_result["flood_risk"]}
    except Exception as exc:
        raster_error = str(exc)

    if raster_paths and raster_paths.get("flood_risk"):
        distribution = flood_raster_distribution(raster_paths["flood_risk"])
        labels = distribution["labels"]
        percentages = distribution["percentages"]
        counts = distribution["counts"]
        result["charts"]["risk_distribution"] = {
            "labels": labels,
            "data": percentages,
            "counts": counts,
            "colors": [
                "#E74C3C",
                "#E67E22",
                "#F1C40F",
                "#2ECC71",
            ],
        }
        pct_by_label = dict(zip(labels, percentages, strict=False))
        count_by_label = dict(zip(labels, counts, strict=False))
        result["stats"].update(
            {
                "very_high_risk_pct": pct_by_label.get("Very High", 0.0),
                "high_risk_pct": pct_by_label.get("High", 0.0),
                "medium_risk_pct": pct_by_label.get("Medium", 0.0),
                "low_risk_pct": pct_by_label.get("Low", 0.0),
                "very_high_risk_pixels": count_by_label.get("Very High", 0),
                "high_risk_pixels": count_by_label.get("High", 0),
                "medium_risk_pixels": count_by_label.get("Medium", 0),
                "low_risk_pixels": count_by_label.get("Low", 0),
                "mapped_pixel_count": distribution["valid_pixel_count"],
            }
        )
        total_area_ha = _aoi_area_ha(config.aoi_geojson)
        if total_area_ha:
            result["stats"]["total_area_ha"] = round(total_area_ha, 1)
            _attach_risk_area(result["charts"]["risk_distribution"], total_area_ha)
        _attach_population_stats(result, raster_result, labels)

    geojson_features = []
    if config.aoi_geojson:
        geojson_features.append(
            {
                "type": "Feature",
                "geometry": config.aoi_geojson,
                "properties": {"type": "boundary"},
            }
        )
    for p in result.pop("_sample_points", []):
        geojson_features.append(
            {
                "type": "Feature",
                "geometry": {"type": "Point", "coordinates": [p["lon"], p["lat"]]},
                "properties": {"risk_class": p["risk_class"]},
            }
        )
    metadata = {
        "model": model_type,
        "country": config.country,
        "start_date": config.start_date,
        "end_date": config.end_date,
        "raster": raster_paths or {},
        "windows": {
            "pre_flood": {
                "start": dict_config.get("pre_flood_start") or dict_config.get("pre_sar_start"),
                "end": dict_config.get("pre_flood_end") or dict_config.get("pre_sar_end"),
            },
            "post_flood": {
                "start": dict_config.get("post_flood_start")
                or dict_config.get("flood_sar_start"),
                "end": dict_config.get("post_flood_end") or dict_config.get("flood_sar_end"),
            },
            "rainfall_7d": {
                "start": dict_config.get("rain_7d_start"),
                "end": dict_config.get("rain_7d_end"),
            },
            "rainfall_30d": {
                "start": dict_config.get("rain_30d_start"),
                "end": dict_config.get("rain_30d_end"),
            },
            "mndwi": {
                "start": dict_config.get("mndwi_start"),
                "end": dict_config.get("mndwi_end"),
            },
        },
    }
    if raster_error:
        metadata["raster_error"] = raster_error

    return AnalysisOutput(
        module="flood",
        geojson={"type": "FeatureCollection", "features": geojson_features},
        raster_path=raster_paths,
        stats={**result["stats"], "country": config.country},
        shap=result.get("charts", {}).get("shap"),
        charts=result.get("charts", {}),
        metadata=metadata,
    )

run 

run(config: dict) -> dict

Single AOI analysis — full pipeline in one call.

Source code in flood/use_case.py 
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
def run(self, config: dict) -> dict:
    """Single AOI analysis — full pipeline in one call."""
    raw_data = self._fetch_from_dict(config)
    features = self._preprocess_raw(raw_data)
    result = self._run_model_dict(features, config)

    model_type = config.get("model_type", "ensemble")
    raster_paths: dict[str, str] | None = None
    raster_error: str | None = None
    raster_result: FloodCogResult | None = None
    try:
        raster_result = export_flood_cog(
            rf_model=features["_rf"],
            xgb_model=features["_xgb"],
            datasets=features["datasets"],
            output_dir=config.get("output_dir", "outputs"),
            prefix=config.get("prefix", "flood"),
            model_type=model_type,
            aoi_geojson=config.get("aoi_geojson"),
        )
        raster_paths = {"flood_risk": raster_result["flood_risk"]}
    except Exception as exc:
        raster_error = str(exc)

    result["raster"] = raster_paths or {}
    result["stats"].update(
        {
            "bbox": features["bbox"],
            "flood_event": config.get("flood_event", ""),
            "prefix": config.get("prefix", "flood"),
        }
    )
    if raster_error:
        result["raster_error"] = raster_error

    if raster_paths and raster_paths.get("flood_risk"):
        distribution = flood_raster_distribution(raster_paths["flood_risk"])
        labels = distribution["labels"]
        percentages = distribution["percentages"]
        counts = distribution["counts"]
        result["charts"]["risk_distribution"] = {
            "labels": labels,
            "data": percentages,
            "counts": counts,
            "colors": [
                "#E74C3C",
                "#E67E22",
                "#F1C40F",
                "#2ECC71",
            ],
        }
        pct_by_label = dict(zip(labels, percentages, strict=False))
        count_by_label = dict(zip(labels, counts, strict=False))
        result["stats"].update(
            {
                "very_high_risk_pct": pct_by_label.get("Very High", 0.0),
                "high_risk_pct": pct_by_label.get("High", 0.0),
                "medium_risk_pct": pct_by_label.get("Medium", 0.0),
                "low_risk_pct": pct_by_label.get("Low", 0.0),
                "very_high_risk_pixels": count_by_label.get("Very High", 0),
                "high_risk_pixels": count_by_label.get("High", 0),
                "medium_risk_pixels": count_by_label.get("Medium", 0),
                "low_risk_pixels": count_by_label.get("Low", 0),
                "mapped_pixel_count": distribution["valid_pixel_count"],
            }
        )
        total_area_ha = _aoi_area_ha(config.get("aoi_geojson"))
        if total_area_ha:
            result["stats"]["total_area_ha"] = round(total_area_ha, 1)
            _attach_risk_area(result["charts"]["risk_distribution"], total_area_ha)
        _attach_population_stats(result, raster_result, labels)
    return result

run_date_ranges 

run_date_ranges(config: dict, date_ranges: list[dict]) -> list[dict]

Not supported — flood is single-area only. See UseCaseInfo.single_area_only.

Source code in flood/use_case.py 
340
341
342
343
344
345
346
347
348
349
350
def run_date_ranges(self, config: dict, date_ranges: list[dict]) -> list[dict]:
    """Not supported — flood is single-area only. See UseCaseInfo.single_area_only."""
    raise ValueError(
        "Flood analysis only supports a single AOI/date-range at a time. Each "
        "run requires six independent satellite date windows (pre/post-flood "
        "SAR, 7-day and 30-day rainfall, Sentinel-2 MNDWI, JRC flood label) to "
        "all resolve to non-empty imagery — running many configs in parallel "
        "multiplies the odds that one window comes up empty, so the "
        "two-date-range comparison and multi-region batch modes are disabled "
        "for this module rather than run unreliably."
    )

run_multi_regions 

run_multi_regions(configs: list[dict]) -> list[dict]

Not supported — flood is single-area only. See UseCaseInfo.single_area_only.

Source code in flood/use_case.py 
352
353
354
355
356
357
358
359
360
361
362
def run_multi_regions(self, configs: list[dict]) -> list[dict]:
    """Not supported — flood is single-area only. See UseCaseInfo.single_area_only."""
    raise ValueError(
        "Flood analysis only supports a single AOI/date-range at a time. Each "
        "run requires six independent satellite date windows (pre/post-flood "
        "SAR, 7-day and 30-day rainfall, Sentinel-2 MNDWI, JRC flood label) to "
        "all resolve to non-empty imagery — running many configs in parallel "
        "multiplies the odds that one window comes up empty, so the "
        "two-date-range comparison and multi-region batch modes are disabled "
        "for this module rather than run unreliably."
    )

Features

features

fetch_terrain 

fetch_terrain(aoi: Geometry, scale: int = 90) -> xr.Dataset

Download SRTM elevation from GEE. Returns Dataset with variable 'elevation' and (lat, lon) coords.

Source code in flood/features.py 
162
163
164
165
166
167
168
169
def fetch_terrain(aoi: ee.Geometry, scale: int = 90) -> xr.Dataset:
    """
    Download SRTM elevation from GEE.
    Returns Dataset with variable 'elevation' and (lat, lon) coords.
    """
    dem = ee.Image("USGS/SRTMGL1_003").select("elevation").clip(aoi)
    da = _download_image(dem, aoi, scale).squeeze()
    return xr.Dataset({"elevation": da.rename({"x": "lon", "y": "lat"})})

fetch_twi 

fetch_twi(aoi: Geometry, scale: int = 500) -> xr.Dataset

Compute Topographic Wetness Index from HydroSHEDS flow accumulation + SRTM slope. Returns Dataset with variables 'twi' and 'flow_acc_log'.

Source code in flood/features.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
def fetch_twi(aoi: ee.Geometry, scale: int = 500) -> xr.Dataset:
    """
    Compute Topographic Wetness Index from HydroSHEDS flow accumulation + SRTM slope.
    Returns Dataset with variables 'twi' and 'flow_acc_log'.
    """
    dem = ee.Image("USGS/SRTMGL1_003").select("elevation").reproject("EPSG:4326", None, 500)
    slope_rad = (
        ee.Terrain.products(dem)
        .select("slope")
        .clip(aoi)
        .multiply(np.pi / 180)
        .max(ee.Image(0.001))
    )
    flow_acc = ee.Image("WWF/HydroSHEDS/30ACC").select("b1").clip(aoi)
    twi = flow_acc.multiply(810_000).divide(slope_rad.tan()).log().rename("twi")
    flow_acc_log = flow_acc.add(1).log().rename("flow_acc_log")

    raw = _download_image(
        twi.addBands(flow_acc_log), aoi, scale, band_names=["twi", "flow_acc_log"]
    )
    ds = raw.to_dataset(dim="band").rename({"x": "lon", "y": "lat"})
    return ds

fetch_sar_change 

fetch_sar_change(aoi: Geometry, pre_start: str, pre_end: str, flood_start: str, flood_end: str, scale: int = 90) -> xr.Dataset

Compute Sentinel-1 VV backscatter change (pre − flood). Positive values indicate a backscatter drop → open water / flood signal. Returns Dataset with variable 'vv_change'.

Source code in flood/features.py 
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
def fetch_sar_change(
    aoi: ee.Geometry,
    pre_start: str,
    pre_end: str,
    flood_start: str,
    flood_end: str,
    scale: int = 90,
) -> xr.Dataset:
    """
    Compute Sentinel-1 VV backscatter change (pre − flood).
    Positive values indicate a backscatter drop → open water / flood signal.
    Returns Dataset with variable 'vv_change'.
    """
    sar = (
        ee.ImageCollection("COPERNICUS/S1_GRD")
        .filterBounds(aoi)
        .filter(ee.Filter.eq("instrumentMode", "IW"))
        .filter(ee.Filter.listContains("transmitterReceiverPolarisation", "VV"))
        .select("VV")
    )
    vv_change = (
        sar.filterDate(pre_start, pre_end)
        .mean()
        .unmask(0)
        .clip(aoi)
        .subtract(sar.filterDate(flood_start, flood_end).mean().unmask(0).clip(aoi))
        .rename("vv_change")
    )
    da = _download_image(vv_change, aoi, scale).squeeze()
    return xr.Dataset({"vv_change": da.rename({"x": "lon", "y": "lat"})})

fetch_rainfall 

fetch_rainfall(aoi: Geometry, start_7d: str, end_7d: str, start_30d: str, end_30d: str, scale: int = 500) -> xr.Dataset

Download CHIRPS cumulative rainfall for two windows. Returns Dataset with variables 'rainfall_7d' and 'rainfall_30d'.

Source code in flood/features.py 
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
def fetch_rainfall(
    aoi: ee.Geometry,
    start_7d: str,
    end_7d: str,
    start_30d: str,
    end_30d: str,
    scale: int = 500,
) -> xr.Dataset:
    """
    Download CHIRPS cumulative rainfall for two windows.
    Returns Dataset with variables 'rainfall_7d' and 'rainfall_30d'.
    """
    chirps = ee.ImageCollection("UCSB-CHG/CHIRPS/DAILY").filterBounds(aoi)
    rain = (
        chirps.filterDate(start_7d, end_7d)
        .sum()
        .clip(aoi)
        .rename("rainfall_7d")
        .addBands(chirps.filterDate(start_30d, end_30d).sum().clip(aoi).rename("rainfall_30d"))
    )
    raw = _download_image(
        rain,
        aoi,
        scale,
        band_names=["rainfall_7d", "rainfall_30d"],
    )
    return raw.to_dataset(dim="band").rename({"x": "lon", "y": "lat"})

fetch_landcover 

fetch_landcover(aoi: Geometry, scale: int = 100) -> xr.Dataset

Download ESA WorldCover 2021 land cover. Returns Dataset with variable 'Map' (raw class values 10–95).

Source code in flood/features.py 
257
258
259
260
261
262
263
264
def fetch_landcover(aoi: ee.Geometry, scale: int = 100) -> xr.Dataset:
    """
    Download ESA WorldCover 2021 land cover.
    Returns Dataset with variable 'Map' (raw class values 10–95).
    """
    worldcover = ee.ImageCollection("ESA/WorldCover/v200").first().select("Map").clip(aoi)
    da = _download_image(worldcover, aoi, scale).squeeze()
    return xr.Dataset({"Map": da.rename({"x": "lon", "y": "lat"})})

fetch_dist_river 

fetch_dist_river(aoi: Geometry, scale: int = 90) -> xr.Dataset

Compute Euclidean distance to permanent water (JRC occurrence ≥ 70 %). Returns Dataset with variable 'dist_river' in metres.

Source code in flood/features.py 
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
def fetch_dist_river(aoi: ee.Geometry, scale: int = 90) -> xr.Dataset:
    """
    Compute Euclidean distance to permanent water (JRC occurrence ≥ 70 %).
    Returns Dataset with variable 'dist_river' in metres.
    """
    water_mask = ee.Image("JRC/GSW1_4/GlobalSurfaceWater").select("occurrence").gte(70).clip(aoi)
    dist_river = (
        water_mask.fastDistanceTransform(256, "pixels")
        .sqrt()
        .multiply(ee.Image.pixelArea().sqrt())
        .rename("dist_river")
        .clip(aoi)
    )
    da = _download_image(dist_river, aoi, scale).squeeze()
    return xr.Dataset({"dist_river": da.rename({"x": "lon", "y": "lat"})})

fetch_mndwi 

fetch_mndwi(aoi: Geometry, start_date: str, end_date: str, scale: int = 90) -> xr.Dataset

Compute Sentinel-2 MNDWI = (Green − SWIR1) / (Green + SWIR1). Returns Dataset with variable 'mndwi'.

Source code in flood/features.py 
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
def fetch_mndwi(
    aoi: ee.Geometry,
    start_date: str,
    end_date: str,
    scale: int = 90,
) -> xr.Dataset:
    """
    Compute Sentinel-2 MNDWI = (Green − SWIR1) / (Green + SWIR1).
    Returns Dataset with variable 'mndwi'.
    """
    s2 = (
        ee.ImageCollection("COPERNICUS/S2_SR_HARMONIZED")
        .filterBounds(aoi)
        .filterDate(start_date, end_date)
        .filter(ee.Filter.lt("CLOUDY_PIXEL_PERCENTAGE", 30))
        .select(["B3", "B11"])
        .median()
        .clip(aoi)
    )
    mndwi = s2.normalizedDifference(["B3", "B11"]).rename("mndwi")
    da = _download_image(mndwi, aoi, scale).squeeze()
    return xr.Dataset({"mndwi": da.rename({"x": "lon", "y": "lat"})})

build_feature_datasets 

build_feature_datasets(aoi: Geometry, config: dict) -> dict[str, xr.Dataset]

Download all feature bands from GEE in parallel. The seven fetch calls are independent HTTP requests; they run concurrently via DaskEngine.run_io_parallel (ThreadPoolExecutor) sharing the GEE session. The dict is keyed by band group name and consumed by both sample_training_data and cog_export.export_flood_cog.

Source code in flood/features.py 
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
def build_feature_datasets(aoi: ee.Geometry, config: dict) -> dict[str, xr.Dataset]:
    """
    Download all feature bands from GEE in parallel.
    The seven fetch calls are independent HTTP requests; they run concurrently
    via DaskEngine.run_io_parallel (ThreadPoolExecutor) sharing the GEE session.
    The dict is keyed by band group name and consumed by both sample_training_data
    and cog_export.export_flood_cog.
    """
    from climate_change.core.dask_engine import DaskEngine
    from climate_change.core.population import fetch_population_count_safe

    scale = config.get("scale", 90)
    pre_start, pre_end = _pre_flood_window(config)
    post_start, post_end = _post_flood_window(config)
    rain_7d_start, rain_7d_end = _rain_7d_window(config)
    rain_30d_start, rain_30d_end = _rain_30d_window(config)
    mndwi_start, mndwi_end = _mndwi_window(config)

    return DaskEngine.run_io_parallel(
        {
            "terrain": lambda: fetch_terrain(aoi, scale=scale),
            "twi": lambda: fetch_twi(aoi, scale=500),
            "sar": lambda: fetch_sar_change(
                aoi,
                pre_start,
                pre_end,
                post_start,
                post_end,
                scale=scale,
            ),
            "rainfall": lambda: fetch_rainfall(
                aoi,
                rain_7d_start,
                rain_7d_end,
                rain_30d_start,
                rain_30d_end,
                scale=500,
            ),
            "landcover": lambda: fetch_landcover(aoi, scale=100),
            "dist_river": lambda: fetch_dist_river(aoi, scale=scale),
            "mndwi": lambda: fetch_mndwi(
                aoi,
                mndwi_start,
                mndwi_end,
                scale=scale,
            ),
            # Raw population COUNT for exposure reporting (population within
            # medium/high/very-high risk zones). Best-effort — see
            # core.population.fetch_population_count_safe.
            "population_count": lambda: fetch_population_count_safe(aoi, scale=scale),
        }
    )

build_gee_feature_stack 

build_gee_feature_stack(aoi: Geometry, config: dict) -> ee.Image

Assemble the 10-band GEE image used for stratified sampling. Static (event-independent) and dynamic (event-specific) bands are built concurrently in background threads, then selected into FEATURE_COLS order.

Source code in flood/features.py 
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
def build_gee_feature_stack(aoi: ee.Geometry, config: dict) -> ee.Image:
    """
    Assemble the 10-band GEE image used for stratified sampling.
    Static (event-independent) and dynamic (event-specific) bands are built
    concurrently in background threads, then selected into FEATURE_COLS order.
    """
    from concurrent.futures import ThreadPoolExecutor

    with ThreadPoolExecutor(max_workers=2) as pool:
        f_static = pool.submit(_build_static_image, aoi)
        f_dynamic = pool.submit(_build_dynamic_image, aoi, config)
        static_img = f_static.result()
        dynamic_img = f_dynamic.result()

    stack = ee.Image.cat([static_img, dynamic_img]).select(FEATURE_COLS)
    # Exclude permanent water bodies (ESA WorldCover) — flood risk applies to
    # land that can flood, not water that already is water year-round.
    valid_land = exclusion_mask(aoi, [WATER_CLASS])
    return stack.updateMask(valid_land)

sample_training_data 

sample_training_data(aoi: Geometry, config: dict, n_pixels: int = 5000, scale: int = 90, seed: int = 42) -> pd.DataFrame

Build the 10-band feature stack and derive JRC flood labels concurrently in background threads, then sample n_pixels flooded + n_pixels non-flooded pixels.

Static features (elevation, TWI, dist_river, landcover, coords), dynamic features (vv_change, rainfall, mndwi), and the JRC label image are all fetched in parallel — no external feature_stack argument required.

1 = seasonal flood water not part of the permanent baseline

0 = all other land

Returns a DataFrame with columns = FEATURE_COLS + ['is_flooded'].

Source code in flood/features.py 
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
def sample_training_data(
    aoi: ee.Geometry,
    config: dict,
    n_pixels: int = 5000,
    scale: int = 90,
    seed: int = 42,
) -> pd.DataFrame:
    """
    Build the 10-band feature stack and derive JRC flood labels concurrently in
    background threads, then sample n_pixels flooded + n_pixels non-flooded pixels.

    Static features (elevation, TWI, dist_river, landcover, coords), dynamic
    features (vv_change, rainfall, mndwi), and the JRC label image are all
    fetched in parallel — no external feature_stack argument required.

    Labels: 1 = seasonal flood water not part of the permanent baseline
            0 = all other land
    Returns a DataFrame with columns = FEATURE_COLS + ['is_flooded'].
    """
    from concurrent.futures import ThreadPoolExecutor

    flood_start, flood_end = _label_window(config)

    with ThreadPoolExecutor(max_workers=3) as pool:
        f_static = pool.submit(_build_static_image, aoi)
        f_dynamic = pool.submit(_build_dynamic_image, aoi, config)
        f_label = pool.submit(_build_jrc_label, aoi, flood_start, flood_end)
        static_img = f_static.result()
        dynamic_img = f_dynamic.result()
        flood_label = f_label.result()

    feature_stack = ee.Image.cat([static_img, dynamic_img]).select(FEATURE_COLS)
    # Exclude permanent water bodies (ESA WorldCover) — flood risk applies to
    # land that can flood, not water that already is water year-round.
    valid_land = exclusion_mask(aoi, [WATER_CLASS])
    feature_stack = feature_stack.updateMask(valid_land)
    labeled = feature_stack.addBands(flood_label)
    records = _sample_labeled_pixels(
        labeled=labeled,
        flood_label=flood_label,
        aoi=aoi,
        scale=scale,
        n_pixels=n_pixels,
        seed=seed,
    )
    df = pd.DataFrame([f["properties"] for f in records]).dropna()
    df["is_flooded"] = df["is_flooded"].astype(int)
    return df

align_datasets 

align_datasets(datasets: dict[str, Dataset], ref_key: str = 'terrain', method_continuous: InterpOptions = 'linear', method_categorical: InterpOptions = 'nearest') -> dict[str, xr.Dataset]

Interpolate all datasets onto the reference grid (terrain at 90 m by default). Landcover is treated as categorical and uses nearest-neighbour interpolation. population_count is passed through entirely unchanged, at its own native ~100 m resolution — it is never interpolated onto this reference grid. Summing population within a risk zone requires the opposite: upsampling the (categorical) classification onto population's native grid — see core.population.population_exposure, which cog_export.py uses for this instead of relying on align_datasets. Interpolating population here would silently corrupt totals (verified against live GEE data — see core.population.fetch_population_count's docstring).

Each dataset is chunked before interpolation so that xarray produces Dask-backed lazy arrays; dask.compute() then materialises them all in a single parallel scheduler pass.

Source code in flood/features.py 
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
def align_datasets(
    datasets: dict[str, xr.Dataset],
    ref_key: str = "terrain",
    method_continuous: InterpOptions = "linear",
    method_categorical: InterpOptions = "nearest",
) -> dict[str, xr.Dataset]:
    """
    Interpolate all datasets onto the reference grid (terrain at 90 m by default).
    Landcover is treated as categorical and uses nearest-neighbour interpolation.
    population_count is passed through entirely unchanged, at its own native
    ~100 m resolution — it is never interpolated onto this reference grid.
    Summing population within a risk zone requires the opposite: upsampling
    the (categorical) classification onto population's native grid — see
    core.population.population_exposure, which cog_export.py uses for this
    instead of relying on align_datasets. Interpolating population here would
    silently corrupt totals (verified against live GEE data — see
    core.population.fetch_population_count's docstring).

    Each dataset is chunked before interpolation so that xarray produces
    Dask-backed lazy arrays; dask.compute() then materialises them all
    in a single parallel scheduler pass.
    """
    from dask.base import compute

    _CHUNK = {"lat": 256, "lon": 256}
    ref = datasets[ref_key]

    lazy: dict[str, xr.Dataset] = {ref_key: ref}
    for key, ds in datasets.items():
        if key == ref_key or ds is None or key == "population_count":
            continue
        method = method_categorical if key == "landcover" else method_continuous
        lazy[key] = ds.chunk(_CHUNK).interp(lat=ref.lat, lon=ref.lon, method=method)

    keys = list(lazy)
    computed = compute(*[lazy[k] for k in keys])
    result = dict(zip(keys, computed))
    pop_ds = datasets.get("population_count")
    if pop_ds is not None:
        result["population_count"] = pop_ds
    return result

Model

model

FloodModel 

FloodModel()

Orchestrates the full ML pipeline for a single flood event. Always trains both RF and XGBoost. config['model_type'] selects which probabilities drive the primary stats, risk map, and COG export: "rf" — Random Forest "xgboost" — XGBoost "ensemble" — mean of RF + XGBoost (default) Trained models are stored on self.rf / self.xgb so the use case can pass them directly to cog_export.export_flood_cog.

Source code in flood/model.py 
342
343
344
def __init__(self) -> None:
    self.rf: RandomForestClassifier | None = None
    self.xgb: Booster | None = None

predict 

predict(df: DataFrame, config: dict | None = None) -> dict
Parameters

df : DataFrame with columns = FEATURE_COLS + ['is_flooded'] config : optional dict; reads config['model_type'] (default 'ensemble')

Returns

dict with keys: stats, charts

Source code in flood/model.py 
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
def predict(self, df: pd.DataFrame, config: dict | None = None) -> dict:
    """
    Parameters
    ----------
    df     : DataFrame with columns = FEATURE_COLS + ['is_flooded']
    config : optional dict; reads config['model_type'] (default 'ensemble')

    Returns
    -------
    dict with keys: stats, charts
    """
    model_type = (config or {}).get("model_type", "ensemble")
    if model_type not in VALID_MODEL_TYPES:
        raise ValueError(f"model_type must be one of {VALID_MODEL_TYPES}, got '{model_type}'")

    X = df[FEATURE_COLS].to_numpy(dtype=np.float64)
    y = df["is_flooded"].to_numpy(dtype=np.intp)

    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42, stratify=_safe_stratify(y)
    )

    # Train RF and XGBoost concurrently when the Dask cluster is running,
    # otherwise fall back to sequential execution (e.g. during unit tests).
    from climate_change.core.dask_engine import DaskEngine

    client = DaskEngine.get_client_if_running()
    if client is not None:
        f_rf = client.submit(train_rf, X_train, y_train, pure=False)
        f_xgb = client.submit(train_xgb, X_train, y_train, X_test, y_test, pure=False)
        (self.rf, rf_meta), (self.xgb, xgb_meta) = cast(list, client.gather([f_rf, f_xgb]))
    else:
        self.rf, rf_meta = train_rf(X_train, y_train)
        self.xgb, xgb_meta = train_xgb(X_train, y_train, X_test, y_test)

    assert self.rf is not None
    assert self.xgb is not None

    eval_result = evaluate_models(self.rf, self.xgb, X_test, y_test)
    shap_payload = compute_shap_importance(self.xgb, X_test)
    uncertainty = compute_uncertainty(
        np.array(eval_result["rf"]["probabilities"]),
        np.array(eval_result["xgb"]["probabilities"]),
    )
    charts = build_flood_charts(eval_result, shap_payload, uncertainty, model_type)
    # Risk percentages from the selected model
    result_key = _MODEL_TYPE_KEY.get(model_type, "ensemble")
    primary_probs = np.array(eval_result[result_key]["probabilities"])
    risk_labels = classify_flood_risk(primary_probs)

    _ORDER = ["Very High", "High", "Medium", "Low"]
    pct_arr = (
        pd.Series(risk_labels)
        .value_counts(normalize=True)
        .mul(100)
        .round(1)
        .reindex(_ORDER)
        .fillna(0.0)
        .to_numpy(dtype=np.float64)
    )
    very_high_pct, high_pct, medium_pct, low_pct = (float(v) for v in pct_arr)

    stats = {
        "model_type": model_type,
        "n_pixels_sampled": int(len(df)),
        "flooded_pct": round(float(y.mean() * 100), 1),
        "rf_cv_f1": round(rf_meta["cv_f1_mean"], 4)
        if rf_meta["cv_f1_mean"] is not None
        else None,
        "rf_f1": eval_result["rf"]["f1"],
        "rf_auc": eval_result["rf"]["auc"],
        "xgb_cv_f1": round(xgb_meta["cv_f1_mean"], 4)
        if xgb_meta["cv_f1_mean"] is not None
        else None,
        "xgb_f1": eval_result["xgb"]["f1"],
        "xgb_auc": eval_result["xgb"]["auc"],
        "ensemble_f1": eval_result["ensemble"]["f1"],
        "ensemble_auc": eval_result["ensemble"]["auc"],
        "selected_f1": eval_result[result_key]["f1"],
        "selected_auc": eval_result[result_key]["auc"],
        "selected_threshold": eval_result[result_key]["threshold"],
        "top_flood_driver": shap_payload["features"][0],
        "very_high_risk_pct": very_high_pct,
        "high_risk_pct": high_pct,
        "medium_risk_pct": medium_pct,
        "low_risk_pct": low_pct,
        **uncertainty,
    }

    lon_idx = FEATURE_COLS.index("longitude")
    lat_idx = FEATURE_COLS.index("latitude")
    _sample_points = [
        {
            "lon": round(float(X_test[i, lon_idx]), 5),
            "lat": round(float(X_test[i, lat_idx]), 5),
            "risk_class": str(risk_labels[i]),
        }
        for i in range(len(X_test))
    ]
    return {"stats": stats, "charts": charts, "_sample_points": _sample_points}

classify_flood_risk 

classify_flood_risk(prob: ndarray) -> np.ndarray

Map flood probability array to 4-class string labels.

Source code in flood/model.py 
83
84
85
86
87
88
89
def classify_flood_risk(prob: np.ndarray) -> np.ndarray:
    """Map flood probability array to 4-class string labels."""
    risk = np.full(prob.shape, "Low", dtype=object)
    risk[(prob >= 0.25) & (prob < 0.50)] = "Medium"
    risk[(prob >= 0.50) & (prob < 0.75)] = "High"
    risk[prob >= 0.75] = "Very High"
    return risk

train_rf 

train_rf(X_train: ndarray, y_train: ndarray, cv_folds: int = 5) -> tuple[RandomForestClassifier, dict]

Fit a balanced Random Forest and report CV F1. Returns (model, metadata).

Source code in flood/model.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
def train_rf(
    X_train: np.ndarray,
    y_train: np.ndarray,
    cv_folds: int = 5,
) -> tuple[RandomForestClassifier, dict]:
    """Fit a balanced Random Forest and report CV F1. Returns (model, metadata)."""
    rf = RandomForestClassifier(
        n_estimators=RF_N_ESTIMATORS,
        max_depth=RF_MAX_DEPTH,
        min_samples_leaf=RF_MIN_SAMPLES_LEAF,
        class_weight="balanced",
        random_state=42,
        n_jobs=-1,
    )
    rf.fit(X_train, y_train)
    folds = _safe_cv_folds(y_train, cv_folds)
    if folds is None:
        return rf, {"cv_f1_mean": None, "cv_f1_std": None}
    cv_f1 = cross_val_score(rf, X_train, y_train, cv=folds, scoring="f1")
    return rf, {"cv_f1_mean": float(cv_f1.mean()), "cv_f1_std": float(cv_f1.std())}

train_xgb 

train_xgb(X_train: ndarray, y_train: ndarray, X_val: ndarray, y_val: ndarray, cv_folds: int = 5) -> tuple[Booster, dict]

Fit XGBoost binary flood classifier via the low-level Booster API with scale_pos_weight and eval-set logging. Returns (model, metadata).

Uses xgboost.train() rather than the XGBClassifier sklearn wrapper because the wrapper's fit() rejects a y whose sorted unique values aren't exactly [0, 1] — an AOI/time window where every sampled pixel is flooded (or none are) is a legitimate data state, not invalid input, and would otherwise crash training.

Source code in flood/model.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
def train_xgb(
    X_train: np.ndarray,
    y_train: np.ndarray,
    X_val: np.ndarray,
    y_val: np.ndarray,
    cv_folds: int = 5,
) -> tuple[Booster, dict]:
    """
    Fit XGBoost binary flood classifier via the low-level Booster API with
    scale_pos_weight and eval-set logging. Returns (model, metadata).

    Uses xgboost.train() rather than the XGBClassifier sklearn wrapper because
    the wrapper's fit() rejects a y whose sorted unique values aren't exactly
    [0, 1] — an AOI/time window where every sampled pixel is flooded (or none
    are) is a legitimate data state, not invalid input, and would otherwise
    crash training.
    """
    scale_pos_weight = float((y_train == 0).sum() / max((y_train == 1).sum(), 1))
    params = {**XGB_PARAMS, "scale_pos_weight": scale_pos_weight}
    dtrain = DMatrix(X_train, label=y_train)
    dval = DMatrix(X_val, label=y_val)
    booster = xgb_train(
        params,
        dtrain,
        num_boost_round=XGB_N_ESTIMATORS,
        evals=[(dval, "validation")],
        verbose_eval=False,
    )

    folds = _safe_cv_folds(y_train, cv_folds)
    if folds is None:
        return booster, {"cv_f1_mean": None, "cv_f1_std": None}
    cv = StratifiedKFold(n_splits=folds)
    cv_f1_scores = []
    for train_idx, val_idx in cv.split(X_train, y_train):
        fold_scale_pos_weight = float(
            (y_train[train_idx] == 0).sum() / max((y_train[train_idx] == 1).sum(), 1)
        )
        fold_booster = xgb_train(
            {**XGB_PARAMS, "scale_pos_weight": fold_scale_pos_weight},
            DMatrix(X_train[train_idx], label=y_train[train_idx]),
            num_boost_round=XGB_N_ESTIMATORS,
        )
        fold_pred = (fold_booster.predict(DMatrix(X_train[val_idx])) >= 0.5).astype(int)
        cv_f1_scores.append(f1_score(y_train[val_idx], fold_pred))

    cv_f1 = np.array(cv_f1_scores)
    return booster, {"cv_f1_mean": float(cv_f1.mean()), "cv_f1_std": float(cv_f1.std())}

positive_class_proba 

positive_class_proba(clf: RandomForestClassifier, proba: ndarray) -> np.ndarray

Return P(y=1) from a binary sklearn classifier's predict_proba output.

predict_proba only has one column when the classifier saw a single class during training (e.g. every sampled pixel in this AOI/time window was flooded, or none were) — proba[:, 1] would then raise IndexError. Falls back to reading clf.classes_ to know whether that lone column represents class 0 or class 1.

Source code in flood/model.py 
164
165
166
167
168
169
170
171
172
173
174
175
176
def positive_class_proba(clf: RandomForestClassifier, proba: np.ndarray) -> np.ndarray:
    """
    Return P(y=1) from a binary sklearn classifier's predict_proba output.

    predict_proba only has one column when the classifier saw a single class
    during training (e.g. every sampled pixel in this AOI/time window was
    flooded, or none were) — proba[:, 1] would then raise IndexError. Falls
    back to reading clf.classes_ to know whether that lone column represents
    class 0 or class 1.
    """
    if proba.shape[1] == 2:
        return proba[:, 1]
    return proba[:, 0] if clf.classes_[0] == 1 else np.zeros(proba.shape[0])

find_best_threshold 

find_best_threshold(probs: ndarray, y_true: ndarray) -> tuple[float, float]

Return (threshold, best_f1) that maximises F1 on the precision-recall curve.

Source code in flood/model.py 
180
181
182
183
184
185
def find_best_threshold(probs: np.ndarray, y_true: np.ndarray) -> tuple[float, float]:
    """Return (threshold, best_f1) that maximises F1 on the precision-recall curve."""
    precisions, recalls, thresholds = precision_recall_curve(y_true, probs)
    f1s = 2 * precisions * recalls / (precisions + recalls + 1e-9)
    best_idx = int(f1s.argmax())
    return float(thresholds[best_idx]), float(f1s[best_idx])

evaluate_models 

evaluate_models(rf: RandomForestClassifier, xgb: Booster, X_test: ndarray, y_test: ndarray) -> dict

Evaluate RF, XGBoost, and their ensemble on the held-out test set. Thresholds are maximised per-model via the precision-recall curve.

Source code in flood/model.py 
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
def evaluate_models(
    rf: RandomForestClassifier,
    xgb: Booster,
    X_test: np.ndarray,
    y_test: np.ndarray,
) -> dict:
    """
    Evaluate RF, XGBoost, and their ensemble on the held-out test set.
    Thresholds are maximised per-model via the precision-recall curve.
    """
    rf_prob = positive_class_proba(rf, rf.predict_proba(X_test))
    xgb_prob = xgb.predict(DMatrix(X_test))
    ens_prob = (rf_prob + xgb_prob) / 2.0

    def _metrics(prob: np.ndarray, label: str) -> dict:
        thresh, _ = find_best_threshold(prob, y_test)
        pred = (prob >= thresh).astype(int)
        # roc_auc_score is undefined (returns NaN, which isn't valid JSON) when
        # y_test ends up with only one class — a real possibility for a scarce
        # class routed through the non-stratified split fallback in predict().
        has_both_classes = len(np.unique(y_test)) > 1
        auc = round(float(roc_auc_score(y_test, prob)), 4) if has_both_classes else None
        return {
            "label": label,
            "f1": round(float(f1_score(y_test, pred)), 4),
            "auc": auc,
            "threshold": round(thresh, 4),
            "predictions": pred.tolist(),
            "probabilities": prob.round(4).tolist(),
        }

    return {
        "rf": _metrics(rf_prob, "Random Forest"),
        "xgb": _metrics(xgb_prob, "XGBoost"),
        "ensemble": _metrics(ens_prob, "Ensemble (mean prob)"),
        "actuals": y_test.tolist(),
    }

compute_shap_importance 

compute_shap_importance(xgb: Booster, X_test: ndarray) -> dict

TreeExplainer SHAP values for XGBoost, sorted by descending mean |SHAP|. XGBoost is always used for SHAP regardless of model_type — it provides the most interpretable tree-based explanations.

Source code in flood/model.py 
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
def compute_shap_importance(xgb: Booster, X_test: np.ndarray) -> dict:
    """
    TreeExplainer SHAP values for XGBoost, sorted by descending mean |SHAP|.
    XGBoost is always used for SHAP regardless of model_type — it provides
    the most interpretable tree-based explanations.
    """
    import shap

    explainer = shap.TreeExplainer(xgb)
    shap_vals = explainer.shap_values(X_test)
    mean_abs = np.abs(shap_vals).mean(axis=0)
    rank_idx = np.argsort(mean_abs)[::-1]
    return {
        "features": [FEATURE_COLS[i] for i in rank_idx],
        "mean_abs_shap": mean_abs[rank_idx].round(4).tolist(),
    }

compute_uncertainty 

compute_uncertainty(rf_prob: ndarray, xgb_prob: ndarray) -> dict

Epistemic uncertainty from RF–XGBoost probability spread. Pixels with spread > 0.20 are flagged for field validation. Always computed regardless of model_type so the UI can display it.

Source code in flood/model.py 
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
def compute_uncertainty(rf_prob: np.ndarray, xgb_prob: np.ndarray) -> dict:
    """
    Epistemic uncertainty from RF–XGBoost probability spread.
    Pixels with spread > 0.20 are flagged for field validation.
    Always computed regardless of model_type so the UI can display it.
    """
    spread = np.abs(rf_prob - xgb_prob)
    return {
        "mean_spread": round(float(spread.mean()), 4),
        "high_uncertainty_pct": round(float((spread > 0.20).mean() * 100), 1),
        "spread_stats": {
            "min": round(float(spread.min()), 4),
            "p25": round(float(np.percentile(spread, 25)), 4),
            "p75": round(float(np.percentile(spread, 75)), 4),
            "max": round(float(spread.max()), 4),
            "mean": round(float(spread.mean()), 4),
        },
    }

build_flood_charts 

build_flood_charts(eval_result: dict, shap_payload: dict, uncertainty_payload: dict, model_type: str = 'ensemble') -> dict

Assemble frontend-ready chart payloads.

Risk distribution is derived from the selected model_type's probabilities. model_performance always includes all three models for comparison.

Source code in flood/model.py 
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
def build_flood_charts(
    eval_result: dict,
    shap_payload: dict,
    uncertainty_payload: dict,
    model_type: str = "ensemble",
) -> dict:
    """
    Assemble frontend-ready chart payloads.

    Risk distribution is derived from the selected model_type's probabilities.
    model_performance always includes all three models for comparison.
    """
    result_key = _MODEL_TYPE_KEY.get(model_type, "ensemble")
    selected_probs = np.array(eval_result[result_key]["probabilities"])

    _RISK_ORDER = ["Very High", "High", "Medium", "Low"]

    risk_labels = classify_flood_risk(selected_probs)
    counts = (
        pd.Series(risk_labels)
        .value_counts()
        .reindex(_RISK_ORDER)
        .fillna(0.0)
        .to_numpy(dtype=np.float64)
    )
    risk_pct = (counts / counts.sum() * 100).round(1)

    risk_chart = {
        "labels": _RISK_ORDER,
        "data": risk_pct.tolist(),
        "colors": [RISK_COLORS[c] for c in _RISK_ORDER],
    }

    # All three models always shown for comparison
    model_performance = {
        "rf": {"f1": eval_result["rf"]["f1"], "auc": eval_result["rf"]["auc"]},
        "xgb": {"f1": eval_result["xgb"]["f1"], "auc": eval_result["xgb"]["auc"]},
        "ensemble": {
            "f1": eval_result["ensemble"]["f1"],
            "auc": eval_result["ensemble"]["auc"],
        },
        "selected": model_type,
    }

    return {
        "risk_distribution": risk_chart,
        "shap": shap_payload,
        "uncertainty": uncertainty_payload,
        "model_performance": model_performance,
    }

COG export

cog_export

flood_raster_distribution 

flood_raster_distribution(path: str | Path) -> dict

Read an already-exported flood-risk COG and tally pixel counts/percentages per risk class (Very High/High/Medium/Low), skipping nodata (<= 0) pixels.

Source code in flood/cog_export.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def flood_raster_distribution(path: str | Path) -> dict:
    """
    Read an already-exported flood-risk COG and tally pixel counts/percentages
    per risk class (Very High/High/Medium/Low), skipping nodata (<= 0) pixels.
    """
    with rasterio.open(path) as src:
        data = src.read(1)

    valid_values = data[data > 0]
    counts = {
        label: int((valid_values == RISK_VALUE_BY_LABEL[label]).sum()) for label in RISK_ORDER
    }
    total = int(valid_values.size)
    percentages = {
        label: round((count / total * 100), 1) if total else 0.0 for label, count in counts.items()
    }
    return {
        "labels": list(RISK_ORDER),
        "counts": [counts[label] for label in RISK_ORDER],
        "percentages": [percentages[label] for label in RISK_ORDER],
        "valid_pixel_count": total,
    }

export_flood_cog 

export_flood_cog(rf_model: RandomForestClassifier, xgb_model: Booster, datasets: dict[str, Dataset], output_dir: str, prefix: str, model_type: str = 'ensemble', aoi_geojson: dict | None = None) -> FloodCogResult

Build the full-AOI feature matrix from in-memory xarray Datasets, run inference with the selected model, classify into 4 risk classes, and write a Cloud Optimised GeoTIFF.

Source code in flood/cog_export.py 
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
def export_flood_cog(
    rf_model: RandomForestClassifier,
    xgb_model: Booster,
    datasets: dict[str, xr.Dataset],
    output_dir: str,
    prefix: str,
    model_type: str = "ensemble",
    aoi_geojson: dict | None = None,
) -> FloodCogResult:
    """
    Build the full-AOI feature matrix from in-memory xarray Datasets,
    run inference with the selected model, classify into 4 risk classes,
    and write a Cloud Optimised GeoTIFF.
    """
    if model_type not in VALID_MODEL_TYPES:
        raise ValueError(f"model_type must be one of {VALID_MODEL_TYPES}, got '{model_type}'")

    output_dir_path = Path(output_dir)
    output_dir_path.mkdir(parents=True, exist_ok=True)
    cog_path = output_dir_path / f"{prefix}_flood_risk.tif"

    # Align all bands to the terrain (90 m) reference grid
    aligned = align_datasets(datasets, ref_key="terrain")
    terrain = aligned["terrain"]
    ref_lat = terrain.lat
    ref_lon = terrain.lon
    n_lat, n_lon = len(ref_lat), len(ref_lon)

    # Extract feature arrays (all float32)
    elevation = terrain["elevation"].values.astype(np.float32)
    twi_ = aligned["twi"]["twi"].values.astype(np.float32)
    dist_river_ = aligned["dist_river"]["dist_river"].values.astype(np.float32)
    vv_change_ = aligned["sar"]["vv_change"].values.astype(np.float32)
    rain7_ = aligned["rainfall"]["rainfall_7d"].values.astype(np.float32)
    rain30_ = aligned["rainfall"]["rainfall_30d"].values.astype(np.float32)
    mndwi_ = aligned["mndwi"]["mndwi"].values.astype(np.float32)
    landcover_ = aligned["landcover"]["Map"].values.astype(np.float32)
    landcover_ = (landcover_ / 10.0) - 1.0  # normalise: match §8

    lon_grid = np.broadcast_to(ref_lon.values[np.newaxis, :], (n_lat, n_lon)).astype(np.float32)
    lat_grid = np.broadcast_to(ref_lat.values[:, np.newaxis], (n_lat, n_lon)).astype(np.float32)

    # Stack → (n_pixels, n_features) and mask nodata
    bands = np.stack(
        [
            elevation,
            twi_,
            dist_river_,
            vv_change_,
            rain7_,
            rain30_,
            mndwi_,
            landcover_,
            lon_grid,
            lat_grid,
        ],
        axis=0,
    )  # (10, n_lat, n_lon)
    X_full = bands.reshape(len(FEATURE_COLS), -1).T  # (n_pixels, 10)
    valid_mask = ~np.isnan(X_full).any(axis=1)
    # Permanent water bodies (ESA WorldCover) are excluded from flood-risk output.
    water = np.isclose(landcover_.reshape(-1), _WATER_LANDCOVER)
    valid_mask &= ~water

    transform = terrain["elevation"].rio.transform()
    crs = terrain["elevation"].rio.crs or "EPSG:4326"
    geometries = _aoi_geometries(aoi_geojson)
    if geometries:
        inside_aoi = geometry_mask(
            geometries,
            out_shape=(n_lat, n_lon),
            transform=transform,
            invert=True,
        ).reshape(-1)
        valid_mask = valid_mask & inside_aoi

    X_valid = X_full[valid_mask]

    # Inference — route by model_type
    risk_grid: np.ndarray = np.zeros(n_lat * n_lon, dtype=np.uint8)
    if X_valid.size:
        prob_rf = positive_class_proba(rf_model, rf_model.predict_proba(X_valid))
        prob_xgb = xgb_model.predict(DMatrix(X_valid))
        if model_type == "rf":
            selected_prob = prob_rf
        elif model_type == "xgboost":
            selected_prob = prob_xgb
        else:  # ensemble
            selected_prob = (prob_rf + prob_xgb) / 2.0

        # Classify and reconstruct spatial grid
        risk_labels = classify_flood_risk(selected_prob)
        risk_valid = np.array([RISK_INT[r] for r in risk_labels], dtype=np.uint8)
        risk_grid[valid_mask] = risk_valid
    risk_grid = risk_grid.reshape(n_lat, n_lon)

    population_by_class: dict[str, float] = {}
    total_population: float | None = None
    pop_ds = aligned.get("population_count")
    if pop_ds is not None:
        by_int = population_exposure(
            risk_grid, transform, pop_ds, classes=list(RISK_LABEL_BY_VALUE)
        )
        population_by_class = {RISK_LABEL_BY_VALUE[v]: by_int[str(v)] for v in RISK_LABEL_BY_VALUE}
        total_population = round(sum(population_by_class.values()), 1)

    with rasterio.open(
        cog_path,
        "w",
        driver="COG",
        height=n_lat,
        width=n_lon,
        count=1,
        dtype=np.uint8,
        crs=crs,
        transform=transform,
        nodata=0,
        compress="lzw",
    ) as dst:
        dst.write(risk_grid, 1)
        dst.update_tags(
            RISK_CLASSES="0=nodata,1=Low,2=Medium,3=High,4=Very High",
            MODEL_TYPE=model_type,
        )

    uncompressed_mb = risk_grid.nbytes / 1e6
    _log.info(
        "COG exported: %s  (%d×%d px, %.1f MB uncompressed)",
        cog_path,
        n_lat,
        n_lon,
        uncompressed_mb,
    )
    return {
        "flood_risk": str(cog_path),
        "population_by_class": population_by_class,
        "total_population": total_population,
    }