#!/usr/bin/env python3 """ AlgaeGuard — Autonomous HAB Detection from Orbit HuggingFace Spaces Gradio Application Tabs: 1. Live Demo — preset NDCI colormaps + optional VLM inference 2. HAB Timeline — historical bloom severity charts for 6 water bodies 3. Custom Inference — upload image or fetch live from SimSat by coordinates 4. About — project summary and links """ from __future__ import annotations import base64 import csv import json import os import re from datetime import datetime from pathlib import Path from typing import Optional import cv2 import gradio as gr import numpy as np import plotly.graph_objects as go import requests from PIL import Image # ── ZeroGPU (HF Spaces) — optional ──────────────────────────────────────────── try: import spaces HAS_SPACES = True except ImportError: HAS_SPACES = False # ── Paths ────────────────────────────────────────────────────────────────────── ROOT = Path(__file__).parent DATA_DIR = ROOT / "data" EXAMPLES_DIR = ROOT / "examples" # ── Constants ────────────────────────────────────────────────────────────────── MODEL_ID = "debrajsingha/algaeguard-lfm2-5-vl-450m" SYSTEM_PROMPT = ( "You are AlgaeGuard, an autonomous on-board satellite AI for Harmful Algal Bloom " "(HAB) early-warning using Sentinel-2 NDCI imagery. You analyze NDCI colormap images " "and produce structured bloom assessments for water utility operators. Your assessments " "guide 6–12 hour treatment protocol decisions — be precise, actionable, and structured." ) HAB_LOCATIONS = [ {"name": "lake_erie", "display": "Lake Erie (USA)", "lat": 41.66, "lon": -83.55}, {"name": "lake_taihu", "display": "Lake Taihu (China)", "lat": 31.20, "lon": 120.00}, {"name": "chesapeake_bay", "display": "Chesapeake Bay (USA)", "lat": 38.50, "lon": -76.40}, {"name": "okeechobee", "display": "Lake Okeechobee (USA)", "lat": 26.90, "lon": -80.80}, {"name": "curonian", "display": "Curonian Lagoon (Lithuania)","lat": 55.40, "lon": 21.10}, {"name": "murray_darling", "display": "Murray-Darling (Australia)", "lat": -34.10, "lon": 141.90}, ] LOC_DISPLAY = {l["name"]: l["display"] for l in HAB_LOCATIONS} LOC_BY_KEY = {l["name"]: l for l in HAB_LOCATIONS} SEVERITY_CFG = { "CLEAR": {"color": "#4ade80", "bg": "#052e16", "emoji": "✅", "action": "No action required"}, "LOW": {"color": "#a3e635", "bg": "#1a2e05", "emoji": "🟡", "action": "Monitor — rescan next pass"}, "MEDIUM": {"color": "#facc15", "bg": "#2d2000", "emoji": "🟠", "action": "Issue caution advisory to water utility"}, "HIGH": {"color": "#f97316", "bg": "#2d0c00", "emoji": "🔴", "action": "Alert water utility — activate response"}, "CRITICAL": {"color": "#ef4444", "bg": "#1c0000", "emoji": "🚨", "action": "IMMEDIATE ACTION — emergency protocol"}, } SEVERITY_ORDER = ["CLEAR", "LOW", "MEDIUM", "HIGH", "CRITICAL"] REAL_WORLD_EVENTS = [ {"location": "lake_erie", "date": "2014-08-02", "label": "Toledo Crisis 2014", "severity": "CRITICAL", "desc": "400,000 residents lost safe water for 3 days. Bloom visible 2 weeks earlier in Sentinel-2 data."}, {"location": "lake_erie", "date": "2019-07-15", "label": "Erie Bloom 2019", "severity": "HIGH", "desc": "620 sq mile bloom — largest ever recorded on Lake Erie at the time."}, {"location": "lake_taihu", "date": "2007-05-29", "label": "Taihu Crisis 2007", "severity": "CRITICAL", "desc": "Tap water cut for 2M residents in Wuxi, China. Cyanobacteria toxin levels 1000× safe limit."}, {"location": "okeechobee", "date": "2018-08-01", "label": "Florida Emergency", "severity": "CRITICAL", "desc": "Governor declared state of emergency. Toxic algae spread to Atlantic and Gulf coasts."}, {"location": "chesapeake_bay", "date": "2011-07-01", "label": "Chesapeake Dead Zone","severity": "HIGH", "desc": "Record dead zone — 1.91 cubic miles of hypoxic water. Massive fish and shellfish kills."}, {"location": "murray_darling", "date": "2010-01-10", "label": "Murray-Darling 2010","severity": "CRITICAL", "desc": "Largest bloom ever recorded — cyanobacteria stretched 1,000 km along the river system."}, ] # ── Load timeseries ──────────────────────────────────────────────────────────── def _load_timeseries() -> dict: ts_path = DATA_DIR / "timeseries.json" if ts_path.exists(): return json.loads(ts_path.read_text()) # fallback: build from CSV csv_path = DATA_DIR / "processed_index.csv" if not csv_path.exists(): return {} from collections import defaultdict ts: dict = defaultdict(list) for row in csv.DictReader(open(csv_path)): stem = row["stem"] m = re.search(r"(\d{4}-\d{2}-\d{2})$", stem) if not m: continue date = m.group(1) loc = stem[: m.start()].rstrip("_") ts[loc].append({ "date": date, "severity": row["severity"], "ndci_mean": round(float(row["ndci_mean"]), 4), "ndci_max": round(float(row["ndci_max"]), 4), "bloom_pct": round(float(row["bloom_pct"]), 2), "severe_pct":round(float(row["severe_pct"]), 2), "image": f"{stem}_cmap.png", }) for loc in ts: ts[loc].sort(key=lambda x: x["date"]) return dict(ts) TIMESERIES = _load_timeseries() # ── Model (lazy-loaded) ──────────────────────────────────────────────────────── _processor = None _model = None def _load_model(): global _processor, _model if _model is not None: return _processor, _model import torch from transformers import AutoModelForImageTextToText, AutoProcessor print(f"Loading {MODEL_ID} ...") _processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True) dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32 _model = AutoModelForImageTextToText.from_pretrained( MODEL_ID, torch_dtype=dtype, device_map="auto", trust_remote_code=True ) _model.eval() print(f"✅ Model ready on {next(_model.parameters()).device}") return _processor, _model # ── Spectral helpers ─────────────────────────────────────────────────────────── def _ndci_from_bands(bands: np.ndarray) -> np.ndarray: """bands: (H, W, ≥2) float32 — channel 0 = B04 red, channel 1 = B05 rededge""" B04, B05 = bands[:, :, 0], bands[:, :, 1] return (B05 - B04) / (B05 + B04 + 1e-8) def _ndci_to_colormap(ndci_map: np.ndarray, size: int = 512) -> np.ndarray: """Returns BGR uint8 (H, W, 3)""" clipped = np.clip(np.nan_to_num(ndci_map, nan=0.0), -1.0, 1.0) norm = ((clipped + 1) / 2 * 255).astype(np.uint8) cmap = cv2.applyColorMap(norm, cv2.COLORMAP_JET) if cmap.shape[0] != size: cmap = cv2.resize(cmap, (size, size), interpolation=cv2.INTER_LINEAR) return cmap def _classify_bloom(ndci_map: np.ndarray) -> dict: valid = ndci_map[~np.isnan(ndci_map)] if np.isnan(ndci_map).any() else ndci_map.ravel() if len(valid) == 0: return {"severity": "CLEAR", "ndci_mean": 0.0, "ndci_max": 0.0, "bloom_pct": 0.0, "severe_pct": 0.0} bloom_pct = float((valid > 0.10).mean() * 100) severe_pct = float((valid > 0.25).mean() * 100) if severe_pct > 15: severity = "CRITICAL" elif bloom_pct > 25: severity = "HIGH" elif bloom_pct > 8: severity = "MEDIUM" elif bloom_pct > 1: severity = "LOW" else: severity = "CLEAR" return { "severity": severity, "ndci_mean": float(np.nanmean(ndci_map)), "ndci_max": float(np.nanmax(ndci_map)), "bloom_pct": bloom_pct, "severe_pct": severe_pct, } # ── VLM inference ────────────────────────────────────────────────────────────── def _do_inference(pil_img: Image.Image, location: str, date: str, ndci_mean: float, bloom_pct: float) -> str: import torch processor, model = _load_model() user_text = ( f"Location: {location}\nDate: {date}\n" f"NDCI Mean: {ndci_mean:.3f} | Bloom Coverage: {bloom_pct:.1f}%\n\n" "Classification thresholds:\n" " CLEAR <1% | LOW 1–8% | MEDIUM 8–25% | HIGH >25% | CRITICAL if severe_pct >15%\n\n" "Analyze this Sentinel-2 NDCI colormap and issue an AlgaeGuard bloom assessment." ) messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": [ {"type": "image"}, {"type": "text", "text": user_text}, ]}, ] text = processor.apply_chat_template(messages, add_generation_prompt=True) inputs = processor(text=text, images=[pil_img], return_tensors="pt").to( next(model.parameters()).device ) with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=350, do_sample=False, temperature=None, top_p=None) return processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True) if HAS_SPACES: run_vlm = spaces.GPU(_do_inference) else: run_vlm = _do_inference # ── SimSat fetch ─────────────────────────────────────────────────────────────── def _fetch_simsat(lat: float, lon: float, date: str) -> Optional[tuple[np.ndarray, dict]]: url = os.environ.get("SIMSAT_API_URL", "").rstrip("/") if not url: return None params = { "lat": lat, "lon": lon, "timestamp": f"{date}T12:00:00", "spectral_bands": ["red", "rededge", "nir", "green", "blue"], "size_km": 10.0, "return_type": "array", "window_seconds": 864000, } try: r = requests.get(f"{url}/data/image/sentinel", params=params, timeout=60) r.raise_for_status() data = r.json() except Exception as e: print(f"SimSat error: {e}") return None meta = data.get("sentinel_metadata", {}) if not meta.get("image_available"): return None img_block = data["image"] raw = base64.b64decode(img_block["image"]) shape = img_block["metadata"]["shape"] dtype = np.dtype(img_block["metadata"]["dtype"]) bands = np.frombuffer(raw, dtype=dtype).reshape(shape).astype(np.float32) bands = np.moveaxis(bands, 0, -1) # (H, W, 5) ndci_map = _ndci_from_bands(bands) return ndci_map, _classify_bloom(ndci_map) # ── HTML helpers ─────────────────────────────────────────────────────────────── def _alert_card(severity: str, bloom_pct: float, ndci_mean: float, ndci_max: float, location: str, date: str, vlm_text: str = "", source: str = "") -> str: c = SEVERITY_CFG[severity] vlm_block = ( f'
' f'{vlm_text}
' ) if vlm_text else "" src_badge = ( f'[{source}]' ) if source else "" return f"""
AlgaeGuard · Sentinel-2 Assessment
{location}{src_badge}
{date}
{c['emoji']} {severity}
Bloom Coverage
{bloom_pct:.1f}%
NDCI Mean
{ndci_mean:.3f}
NDCI Max
{ndci_max:.3f}
Recommended Action: {c['action']}
{vlm_block}
""" def _empty_alert() -> str: return """
Select a location and date, then click Load to see the spectral assessment or Run VLM for the full AI report.
""" # ── Tab 1: Live Demo ─────────────────────────────────────────────────────────── def _get_dates(location_key: str) -> list[str]: return [e["date"] for e in TIMESERIES.get(location_key, [])] def tab1_load(location_key: str, date: str): """Load preset — no model, instant.""" entries = TIMESERIES.get(location_key, []) entry = next((e for e in entries if e["date"] == date), None) if entry is None: return None, _empty_alert() img_path = EXAMPLES_DIR / entry["image"] pil_img = Image.open(img_path).convert("RGB") if img_path.exists() else None loc = LOC_DISPLAY.get(location_key, location_key) alert = _alert_card(entry["severity"], entry["bloom_pct"], entry["ndci_mean"], entry["ndci_max"], loc, date, source="Spectral only") return pil_img, alert def tab1_run_vlm(location_key: str, date: str): """Load preset + run VLM inference.""" entries = TIMESERIES.get(location_key, []) entry = next((e for e in entries if e["date"] == date), None) if entry is None: return None, "

No data for this date.

" img_path = EXAMPLES_DIR / entry["image"] if not img_path.exists(): return None, "

Example image not found.

" pil_img = Image.open(img_path).convert("RGB") loc = LOC_DISPLAY.get(location_key, location_key) vlm_text = run_vlm(pil_img, loc, date, entry["ndci_mean"], entry["bloom_pct"]) alert = _alert_card(entry["severity"], entry["bloom_pct"], entry["ndci_mean"], entry["ndci_max"], loc, date, vlm_text, source="AlgaeGuard VLM") return pil_img, alert def tab1_update_dates(location_key: str): dates = _get_dates(location_key) return gr.Dropdown(choices=dates, value=dates[-1] if dates else None) # ── Tab 2: Historical Timeline ───────────────────────────────────────────────── _SEV_COLORS = { "CLEAR": "#4ade80", "LOW": "#a3e635", "MEDIUM": "#facc15", "HIGH": "#f97316", "CRITICAL": "#ef4444", } def tab2_build_chart(location_key: str): entries = TIMESERIES.get(location_key, []) if not entries: return go.Figure() dates = [e["date"] for e in entries] bloom_pcts = [e["bloom_pct"] for e in entries] severities = [e["severity"] for e in entries] ndci_means = [e["ndci_mean"] for e in entries] fig = go.Figure() # Background area fill fig.add_trace(go.Scatter( x=dates, y=bloom_pcts, fill="tozeroy", fillcolor="rgba(96,165,250,0.06)", line=dict(color="#60a5fa", width=1.5), mode="lines", name="Bloom %", showlegend=False, )) # Per-severity markers for sev in SEVERITY_ORDER: idx = [i for i, s in enumerate(severities) if s == sev] if not idx: continue fig.add_trace(go.Scatter( x=[dates[i] for i in idx], y=[bloom_pcts[i] for i in idx], mode="markers", name=sev, marker=dict(color=_SEV_COLORS[sev], size=9, symbol="circle", line=dict(color="#000", width=0.5)), text=[f"{sev}
{dates[i]}
Bloom: {bloom_pcts[i]:.1f}%
NDCI: {ndci_means[i]:.3f}" for i in idx], hovertemplate="%{text}", )) # Severity threshold lines thresholds = [ (1, "#a3e635", "LOW"), (8, "#facc15", "MEDIUM"), (25, "#f97316", "HIGH"), ] for y, color, label in thresholds: fig.add_hline(y=y, line=dict(color=color, dash="dot", width=1), annotation=dict(text=label, font=dict(color=color, size=10), bgcolor="#0a0a0f", x=1.0)) # Real-world crisis markers events = [ev for ev in REAL_WORLD_EVENTS if ev["location"] == location_key] for ev in events: fig.add_vline( x=ev["date"], line=dict(color="#ef4444", dash="dash", width=1.5), annotation=dict(text=f"⚠ {ev['label']}", textangle=-90, font=dict(color="#ef4444", size=10), bgcolor="#0a0a0f"), ) loc = LOC_DISPLAY.get(location_key, location_key) fig.update_layout( title=dict(text=f"HAB Timeline — {loc}", font=dict(color="#e2e8f0", size=15)), xaxis=dict(title="Date", gridcolor="#1e1e2e", tickformat="%b %Y", tickfont=dict(color="#888")), yaxis=dict(title="Bloom Coverage (%)", gridcolor="#1e1e2e", tickfont=dict(color="#888")), paper_bgcolor="#0a0a0f", plot_bgcolor="#0d0d1a", font=dict(color="#ccc", family="Courier New"), legend=dict(bgcolor="#0d0d1a", bordercolor="#333", x=0, y=1), hovermode="closest", margin=dict(r=80), ) return fig def tab2_events_html(location_key: str) -> str: events = [ev for ev in REAL_WORLD_EVENTS if ev["location"] == location_key] if not events: return "

No documented crises for this location in our dataset.

" rows = "" for ev in events: c = SEVERITY_CFG[ev["severity"]] rows += f""" {ev['date']} {ev['label']} {ev['desc']} """ return f""" {rows}
Date Event Impact
""" def tab2_update(location_key: str): return tab2_build_chart(location_key), tab2_events_html(location_key) # ── Tab 3: Custom Inference ──────────────────────────────────────────────────── def _geocode(name: str) -> tuple[Optional[float], Optional[float], str]: try: from geopy.geocoders import Nominatim geo = Nominatim(user_agent="algaeguard-hab") result = geo.geocode(name, timeout=10) if result: return result.latitude, result.longitude, result.address except Exception: pass return None, None, "Could not geocode — try entering coordinates directly" def _folium_map(lat: float, lon: float, zoom: int = 4) -> str: import folium m = folium.Map(location=[lat, lon], zoom_start=zoom, tiles="CartoDB dark_matter", width="100%", height=300) folium.Marker( [lat, lon], popup=f"Target
{lat:.4f}°, {lon:.4f}°", tooltip="Target location", icon=folium.Icon(color="red", icon="exclamation-sign"), ).add_to(m) for loc in HAB_LOCATIONS: folium.CircleMarker( [loc["lat"], loc["lon"]], radius=6, color="#60a5fa", fill=True, fill_opacity=0.35, tooltip=loc["display"], ).add_to(m) return m._repr_html_() def tab3_geocode(location_name: str): if not location_name.strip(): return "", "", "Enter a location name to geocode" lat, lon, addr = _geocode(location_name.strip()) if lat is None: return "", "", addr return str(round(lat, 4)), str(round(lon, 4)), addr def tab3_infer(image_input, location_name: str, lat_str: str, lon_str: str, date: str): pil_img = None ndci_map = None stats = {} lat, lon = None, None # Resolve coordinates if lat_str and lon_str: try: lat, lon = float(lat_str), float(lon_str) except ValueError: pass # Determine location display name loc_display = location_name.strip() or (f"{lat:.4f}°N, {lon:.4f}°E" if lat else "Unknown") date_str = date or datetime.now().strftime("%Y-%m-%d") # Path A: SimSat live fetch (requires SIMSAT_API_URL env var) simsat_available = bool(os.environ.get("SIMSAT_API_URL")) source = "" map_html = "" if lat and lon and not image_input and simsat_available: result = _fetch_simsat(lat, lon, date_str) if result: ndci_map, stats = result cmap_bgr = _ndci_to_colormap(ndci_map) pil_img = Image.fromarray(cv2.cvtColor(cmap_bgr, cv2.COLOR_BGR2RGB)) source = "SimSat · Sentinel-2" else: source = "SimSat — no image available for this date/location" # Path B: uploaded image if image_input is not None: if isinstance(image_input, np.ndarray): pil_img = Image.fromarray(image_input).convert("RGB") else: pil_img = image_input.convert("RGB") source = "Uploaded NDCI colormap" # Build map if lat and lon: try: map_html = _folium_map(lat, lon) except Exception: map_html = "" if pil_img is None: msg = ( "

No image available.
" + ("Upload an NDCI colormap image." if not simsat_available else "Upload an NDCI colormap or provide coordinates + date.") + "

" ) return None, map_html, msg # VLM inference ndci_mean = stats.get("ndci_mean", 0.0) bloom_pct = stats.get("bloom_pct", 0.0) severity = stats.get("severity", "MEDIUM") ndci_max = stats.get("ndci_max", 0.0) vlm_text = run_vlm(pil_img, loc_display, date_str, ndci_mean, bloom_pct) # Re-parse severity from VLM output if we don't have spectral stats if not stats: for sev in ["CRITICAL", "HIGH", "MEDIUM", "LOW", "CLEAR"]: if sev in vlm_text.upper(): severity = sev break alert_html = _alert_card(severity, bloom_pct, ndci_mean, ndci_max, loc_display, date_str, vlm_text, source) return pil_img, map_html, alert_html # ── CSS ──────────────────────────────────────────────────────────────────────── CSS = """ body, .gradio-container { background: #080810 !important; } .tab-nav button { font-family: 'Courier New', monospace !important; font-size: 0.95em !important; } .tab-nav button.selected { color: #60a5fa !important; border-color: #60a5fa !important; } #header { text-align: center; padding: 24px 0 8px; } #header h1 { font-family: 'Courier New', monospace; font-size: 2.2em; background: linear-gradient(135deg, #60a5fa, #34d399); -webkit-background-clip: text; -webkit-text-fill-color: transparent; margin: 0; } #header p { color: #666; font-family: 'Courier New', monospace; font-size: 0.85em; margin: 6px 0 0; } .gr-button-primary { background: #1d4ed8 !important; border: none !important; } .gr-button-secondary { background: #1e1e2e !important; border: 1px solid #333 !important; } """ # ── Build UI ─────────────────────────────────────────────────────────────────── SIMSAT_LIVE = bool(os.environ.get("SIMSAT_API_URL")) LOC_CHOICES = [(l["display"], l["name"]) for l in HAB_LOCATIONS] with gr.Blocks(theme=gr.themes.Base( primary_hue="blue", neutral_hue="slate", font=[gr.themes.GoogleFont("Inter"), "monospace"], ), css=CSS, title="AlgaeGuard — HAB Detection from Orbit") as demo: gr.HTML(""" """) with gr.Tabs(): # ── Tab 1: Live Demo ─────────────────────────────────────────────────── with gr.Tab("🛰️ Live Demo"): gr.Markdown( "Select a monitored water body and date. **Load** shows spectral stats instantly. " "**Run VLM** loads the fine-tuned model and generates the full operator report (~30–45s on CPU, ~4s on GPU)." ) with gr.Row(): with gr.Column(scale=1): t1_loc = gr.Dropdown(choices=LOC_CHOICES, value="lake_erie", label="Water Body", interactive=True) t1_date = gr.Dropdown(choices=_get_dates("lake_erie"), value=_get_dates("lake_erie")[-1], label="Date", interactive=True) with gr.Row(): t1_load_btn = gr.Button("Load", variant="secondary") t1_vlm_btn = gr.Button("⚡ Run VLM", variant="primary") t1_img = gr.Image(label="NDCI Colormap (JET)", height=360) with gr.Column(scale=2): t1_alert = gr.HTML(value=_empty_alert()) t1_loc.change(tab1_update_dates, t1_loc, t1_date) t1_load_btn.click(tab1_load, [t1_loc, t1_date], [t1_img, t1_alert]) t1_vlm_btn.click(tab1_run_vlm, [t1_loc, t1_date], [t1_img, t1_alert]) # Pre-load first example on startup demo.load(tab1_load, inputs=[gr.State("lake_erie"), gr.State(_get_dates("lake_erie")[-1])], outputs=[t1_img, t1_alert]) # ── Tab 2: Historical HAB Timeline ───────────────────────────────────── with gr.Tab("📈 HAB Timeline"): gr.Markdown( "218 real Sentinel-2 observations across 6 water bodies (2022–2024). " "Red dashed lines mark documented crisis events." ) t2_loc = gr.Dropdown(choices=LOC_CHOICES, value="lake_erie", label="Water Body", interactive=True) t2_chart = gr.Plot(label="Bloom Coverage Over Time") t2_events = gr.HTML() t2_loc.change(tab2_update, t2_loc, [t2_chart, t2_events]) demo.load(tab2_update, inputs=gr.State("lake_erie"), outputs=[t2_chart, t2_events]) # ── Tab 3: Custom Inference ───────────────────────────────────────────── with gr.Tab("🔍 Custom Inference"): gr.Markdown( "**Path A — Upload** an NDCI colormap PNG and run inference directly.\n\n" + ("**Path B — Live satellite fetch:** Enter a location + date. " "AlgaeGuard will pull Sentinel-2 bands from SimSat, compute NDCI, " "and run the VLM automatically." if SIMSAT_LIVE else "**Path B (SimSat):** Set the `SIMSAT_API_URL` Space secret to enable " "live satellite fetch for any coordinates on Earth.") ) with gr.Row(): with gr.Column(scale=1): t3_img_upload = gr.Image(label="Upload NDCI Colormap (optional)", type="pil", height=260) gr.Markdown("**— or enter location —**") t3_loc_name = gr.Textbox(label="Location name", placeholder="e.g. Lake Balaton, Hungary") with gr.Row(): t3_lat = gr.Textbox(label="Latitude", placeholder="41.66") t3_lon = gr.Textbox(label="Longitude", placeholder="-83.55") t3_geo_btn = gr.Button("Geocode →", variant="secondary", size="sm") t3_geo_msg = gr.Textbox(label="Resolved address", interactive=False, lines=1) t3_date = gr.Textbox(label="Date (YYYY-MM-DD)", value=datetime.now().strftime("%Y-%m-%d")) t3_run_btn = gr.Button("⚡ Run AlgaeGuard", variant="primary") with gr.Column(scale=2): t3_map = gr.HTML(label="Location map") t3_result_img = gr.Image(label="NDCI Colormap", height=220, visible=True) t3_alert = gr.HTML() t3_geo_btn.click(tab3_geocode, [t3_loc_name], [t3_lat, t3_lon, t3_geo_msg]) t3_run_btn.click(tab3_infer, [t3_img_upload, t3_loc_name, t3_lat, t3_lon, t3_date], [t3_result_img, t3_map, t3_alert]) # ── Tab 4: About ──────────────────────────────────────────────────────── with gr.Tab("ℹ️ About"): gr.Markdown(f""" ## AlgaeGuard — Autonomous HAB Detection from Orbit **AI in Space Hackathon** · Liquid AI Challenge · DPhi Space Track ### The Problem The 2014 Toledo water crisis left 400,000 people without safe drinking water. Sentinel-2 NDCI data shows the bloom forming and intensifying on Lake Erie for **two full weeks before** that crisis. The data existed. The detection didn't happen in time because every existing pipeline routes raw imagery to the ground, queues it for analyst review, and produces a report 24–72 hours later. AlgaeGuard solves this by running inference **on-board**: satellite overpass → NDCI computation → VLM classification → 200-byte JSON alert downlinked to water utility ops. One orbit pass, one alert, latency under 90 minutes. ### Architecture ``` Satellite Overpass → Band Extraction → NDCI Colormap → LFM2.5-VL-450M → Alert SimSat polling B04/B05/B08/B03 JET 512×512 LoRA SFT JSON (T+0 min) spectral indices (T+10 min) on-board (T+90 min) ``` ### Model Performance (v2) | Metric | v1 | v2 | |--------|----|----| | Holdout Accuracy | 66.7% | **76.9%** (20/26) | | Eval Loss | 0.467 | **0.066** | | MEDIUM F1 | 0.00 | **0.87** | | CLEAR↔CRITICAL errors | present | **0** | All 6 errors are adjacent-class (e.g. HIGH predicted as MEDIUM). Zero CLEAR↔CRITICAL confusions across the holdout set. ### Published Artifacts | Artifact | Link | |----------|------| | Fine-tuned model | [debrajsingha/algaeguard-lfm2-5-vl-450m](https://huggingface.co/debrajsingha/algaeguard-lfm2-5-vl-450m) · 856MB · CC BY 4.0 | | Training dataset | [debrajsingha/algaeguard-hab-ndci](https://huggingface.co/datasets/debrajsingha/algaeguard-hab-ndci) · 398 samples | | Source code | [github.com/debpks/algaeguard-llm](https://github.com/debpks/algaeguard-llm) | | SimSat | [github.com/debpks/SimSat](https://github.com/debpks/SimSat) | ### Stack | Component | Technology | |-----------|-----------| | Satellite imagery | DPhi Space SimSat (Sentinel-2 simulation) | | Spectral indices | NumPy — NDCI, FAI, NDWI | | Base VLM | Liquid AI LFM2.5-VL-450M | | Fine-tuning | Liquid AI LEAP SDK (LoRA r=16, 15 epochs) | | Training compute | Modal A10G GPU | | Monitoring | WandB — `algaeguard_hab_detection` | """) demo.launch()