2026-03-11 18:37:40 +09:00
--- a/backend/src/aerial/aerialService.ts
+++ b/backend/src/aerial/aerialService.ts
@ -61,7 +61,7 @@ export async function getMediaBySn(sn: number): Promise<AerialMediaItem | null>
 }

 export async function fetchOriginalImage(camTy: string, fileId: string): Promise<Buffer> {
-  const res = await fetch(`${IMAGE_ANALYSIS_URL}/get-original-image/${camTy}/${fileId}`, {
+  const res = await fetch(`${IMAGE_API_URL}/get-original-image/${camTy}/${fileId}`, {
    signal: AbortSignal.timeout(30_000),
  });
  if (!res.ok) throw new Error(`이미지 서버 응답: ${res.status}`);
@ -364,8 +364,7 @@ export async function updateSatRequestStatus(sn: number, sttsCd: string): Promis
 // OIL INFERENCE (GPU 서버 프록시)
 // ============================================================

-const OIL_INFERENCE_URL = process.env.OIL_INFERENCE_URL || 'http://localhost:5001';
-const IMAGE_ANALYSIS_URL = process.env.IMAGE_ANALYSIS_URL || OIL_INFERENCE_URL;
+const IMAGE_API_URL = process.env.IMAGE_API_URL ?? 'http://localhost:5001';
 const INFERENCE_TIMEOUT_MS = 10_000;

 export interface OilInferenceRegion {
@ -393,7 +392,7 @@ export async function stitchImages(
  for (const f of files) {
    form.append('files', new Blob([f.buffer], { type: f.mimetype }), f.originalname);
  }
-  const response = await fetch(`${IMAGE_ANALYSIS_URL}/stitch`, {
+  const response = await fetch(`${IMAGE_API_URL}/stitch`, {
    method: 'POST',
    body: form,
    signal: AbortSignal.timeout(300_000),
@ -409,9 +408,8 @@ export async function stitchImages(
 export async function requestOilInference(imageBase64: string): Promise<OilInferenceResult> {
  const controller = new AbortController();
  const timeout = setTimeout(() => controller.abort(), INFERENCE_TIMEOUT_MS);
-
  try {
-    const response = await fetch(`${OIL_INFERENCE_URL}/inference`, {
+    const response = await fetch(`${IMAGE_API_URL}/inference`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({ image: imageBase64 }),
@ -432,7 +430,7 @@ export async function requestOilInference(imageBase64: string): Promise<OilInfer
 /** GPU 추론 서버 헬스체크 */
 export async function checkInferenceHealth(): Promise<{ status: string; device?: string }> {
  try {
-    const response = await fetch(`${OIL_INFERENCE_URL}/health`, {
+    const response = await fetch(`${IMAGE_API_URL}/health`, {
      signal: AbortSignal.timeout(3000),
    });
    if (!response.ok) throw new Error(`status ${response.status}`);
--- a/backend/src/prediction/imageAnalyzeService.ts
+++ b/backend/src/prediction/imageAnalyzeService.ts
@ -102,7 +102,7 @@ export async function analyzeImageFile(imageBuffer: Buffer, originalName: string
      if (res.status === 400 && text.includes('GPS')) {
        throw Object.assign(new Error('GPS_NOT_FOUND'), { code: 'GPS_NOT_FOUND' });
      }
-      throw new Error(`이미지 분석 서버 오류: ${res.status}`);
+      throw new Error(`이미지 분석 서버 오류: ${res.status} - ${text}`);
    }

    serverResponse = await res.json() as ImageServerResponse;
--- a/frontend/src/common/components/map/MapView.tsx
+++ b/frontend/src/common/components/map/MapView.tsx
@ -161,7 +161,7 @@ interface MapViewProps {
  incidentCoord?: { lon: number; lat: number }
  isSelectingLocation?: boolean
  onMapClick?: (lon: number, lat: number) => void
-  oilTrajectory?: Array<{ lat: number; lon: number; time: number; particle?: number; model?: PredictionModel }>
+  oilTrajectory?: Array<{ lat: number; lon: number; time: number; particle?: number; model?: PredictionModel; stranded?: 0 | 1 }>
  selectedModels?: Set<PredictionModel>
  dispersionResult?: DispersionResult | null
  dispersionHeatmap?: Array<{ lon: number; lat: number; concentration: number }>
--- a/frontend/src/common/hooks/useSubMenu.ts
+++ b/frontend/src/common/hooks/useSubMenu.ts
@ -175,4 +175,50 @@ export function consumeHnsReportPayload(): HnsReportPayload | null {
  return v;
 }

+// ─── 유출유 예측 보고서 실 데이터 전달 ──────────────────────────
+export interface OilReportPayload {
+  incident: {
+    name: string;
+    occurTime: string;
+    location: string;
+    lat: number | null;
+    lon: number | null;
+    pollutant: string;
+    spillAmount: string;
+    shipName: string;
+  };
+  pollution: {
+    spillAmount: string;
+    weathered: string;
+    seaRemain: string;
+    pollutionArea: string;
+    coastAttach: string;
+    coastLength: string;
+    oilType: string;
+  };
+  weather: {
+    windDir: string;
+    windSpeed: string;
+    waveHeight: string;
+    temp: string;
+  } | null;
+  spread: {
+    kosps: string;
+    openDrift: string;
+    poseidon: string;
+  };
+  coastal: {
+    firstTime: string | null;
+  };
+  hasSimulation: boolean;
+}
+
+let _oilReportPayload: OilReportPayload | null = null;
+export function setOilReportPayload(d: OilReportPayload | null) { _oilReportPayload = d; }
+export function consumeOilReportPayload(): OilReportPayload | null {
+  const v = _oilReportPayload;
+  _oilReportPayload = null;
+  return v;
+}
+
 export { subMenuState }
--- a/frontend/src/tabs/prediction/components/OilSpillView.tsx
+++ b/frontend/src/tabs/prediction/components/OilSpillView.tsx
@ -8,12 +8,12 @@ import { BoomDeploymentTheoryView } from './BoomDeploymentTheoryView'
 import { BacktrackModal } from './BacktrackModal'
 import { RecalcModal } from './RecalcModal'
 import { BacktrackReplayBar } from '@common/components/map/BacktrackReplayBar'
-import { useSubMenu, navigateToTab, setReportGenCategory } from '@common/hooks/useSubMenu'
+import { useSubMenu, navigateToTab, setReportGenCategory, setOilReportPayload, type OilReportPayload } from '@common/hooks/useSubMenu'
 import type { BoomLine, AlgorithmSettings, ContainmentResult, BoomLineCoord } from '@common/types/boomLine'
 import type { BacktrackPhase, BacktrackVessel, BacktrackConditions, ReplayShip, CollisionEvent } from '@common/types/backtrack'
 import { TOTAL_REPLAY_FRAMES } from '@common/types/backtrack'
 import { fetchBacktrackByAcdnt, createBacktrack, fetchPredictionDetail, fetchAnalysisTrajectory } from '../services/predictionApi'
-import type { CenterPoint, HydrDataStep, ImageAnalyzeResult, PredictionDetail, SimulationRunResponse, SimulationSummary, WindPoint } from '../services/predictionApi'
+import type { CenterPoint, HydrDataStep, ImageAnalyzeResult, OilParticle, PredictionDetail, SimulationRunResponse, SimulationSummary, WindPoint } from '../services/predictionApi'
 import { useSimulationStatus } from '../hooks/useSimulationStatus'
 import SimulationLoadingOverlay from './SimulationLoadingOverlay'
 import { api } from '@common/services/api'
@ -109,7 +109,7 @@ export function OilSpillView() {
  const flyToTarget = null
  const fitBoundsTarget = null
  const [isSelectingLocation, setIsSelectingLocation] = useState(false)
-  const [oilTrajectory, setOilTrajectory] = useState<Array<{ lat: number; lon: number; time: number; particle?: number; model?: PredictionModel }>>([])
+  const [oilTrajectory, setOilTrajectory] = useState<OilParticle[]>([])
  const [centerPoints, setCenterPoints] = useState<CenterPoint[]>([])
  const [windData, setWindData] = useState<WindPoint[][]>([])
  const [hydrData, setHydrData] = useState<(HydrDataStep | null)[]>([])
@ -604,6 +604,62 @@ export function OilSpillView() {
    }
  }

+  const handleOpenReport = () => {
+    const OIL_TYPE_CODE: Record<string, string> = {
+      '벙커C유': 'BUNKER_C', '경유': 'DIESEL', '원유': 'CRUDE_OIL', '윤활유': 'LUBE_OIL',
+    };
+    const accidentName =
+      selectedAnalysis?.acdntNm ||
+      analysisDetail?.acdnt?.acdntNm ||
+      incidentName ||
+      '(미입력)';
+    const occurTime =
+      selectedAnalysis?.occurredAt ||
+      analysisDetail?.acdnt?.occurredAt ||
+      accidentTime ||
+      '';
+    const wx = analysisDetail?.weather?.[0] ?? null;
+
+    const payload: OilReportPayload = {
+      incident: {
+        name: accidentName,
+        occurTime,
+        location: selectedAnalysis?.location || analysisDetail?.acdnt?.location || '',
+        lat: incidentCoord?.lat ?? selectedAnalysis?.lat ?? null,
+        lon: incidentCoord?.lon ?? selectedAnalysis?.lon ?? null,
+        pollutant: OIL_TYPE_CODE[oilType] || oilType,
+        spillAmount: `${spillAmount} ${spillUnit}`,
+        shipName: analysisDetail?.vessels?.[0]?.vesselNm || '',
+      },
+      pollution: {
+        spillAmount: `${spillAmount.toFixed(2)} ${spillUnit}`,
+        weathered: simulationSummary ? `${simulationSummary.weatheredVolume.toFixed(2)} m³` : '—',
+        seaRemain: simulationSummary ? `${simulationSummary.remainingVolume.toFixed(2)} m³` : '—',
+        pollutionArea: simulationSummary ? `${simulationSummary.pollutionArea.toFixed(2)} km²` : '—',
+        coastAttach: simulationSummary ? `${simulationSummary.beachedVolume.toFixed(2)} m³` : '—',
+        coastLength: simulationSummary ? `${simulationSummary.pollutionCoastLength.toFixed(2)} km` : '—',
+        oilType: OIL_TYPE_CODE[oilType] || oilType,
+      },
+      weather: wx
+        ? { windDir: wx.wind, windSpeed: wx.wind, waveHeight: wx.wave, temp: wx.temp }
+        : null,
+      spread: { kosps: '—', openDrift: '—', poseidon: '—' },
+      coastal: {
+        firstTime: (() => {
+          const beachedTimes = oilTrajectory.filter(p => p.stranded === 1).map(p => p.time);
+          if (beachedTimes.length === 0) return null;
+          const d = new Date(Math.min(...beachedTimes) * 1000);
+          return `${String(d.getHours()).padStart(2, '0')}:${String(d.getMinutes()).padStart(2, '0')}`;
+        })(),
+      },
+      hasSimulation: simulationSummary !== null,
+    };
+
+    setOilReportPayload(payload);
+    setReportGenCategory(0);
+    navigateToTab('reports', 'generate');
+  };
+
  return (
    <div className="relative flex flex-1 overflow-hidden">
      {/* Left Sidebar */}
@ -876,7 +932,7 @@ export function OilSpillView() {
      </div>

      {/* Right Panel */}
-      {activeSubTab === 'analysis' && <RightPanel onOpenBacktrack={handleOpenBacktrack} onOpenRecalc={() => setRecalcModalOpen(true)} onOpenReport={() => { setReportGenCategory(0); navigateToTab('reports', 'generate') }} detail={analysisDetail} summary={simulationSummary} />}
+      {activeSubTab === 'analysis' && <RightPanel onOpenBacktrack={handleOpenBacktrack} onOpenRecalc={() => setRecalcModalOpen(true)} onOpenReport={handleOpenReport} detail={analysisDetail} summary={simulationSummary} />}

      {/* 확산 예측 실행 중 로딩 오버레이 */}
      {isRunningSimulation && (
--- a/frontend/src/tabs/reports/components/ReportGenerator.tsx
+++ b/frontend/src/tabs/reports/components/ReportGenerator.tsx
@ -2,7 +2,7 @@ import { useState, useEffect } from 'react';
 import {
  createEmptyReport,
 } from './OilSpillReportTemplate';
-import { consumeReportGenCategory, consumeHnsReportPayload, type HnsReportPayload } from '@common/hooks/useSubMenu';
+import { consumeReportGenCategory, consumeHnsReportPayload, type HnsReportPayload, consumeOilReportPayload, type OilReportPayload } from '@common/hooks/useSubMenu';
 import { saveReport } from '../services/reportsApi';
 import {
  CATEGORIES,
@ -32,6 +32,8 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {

  // HNS 실 데이터 (없으면 sampleHnsData fallback)
  const [hnsPayload, setHnsPayload] = useState<HnsReportPayload | null>(null)
+  // OIL 실 데이터 (없으면 sampleOilData fallback)
+  const [oilPayload, setOilPayload] = useState<OilReportPayload | null>(null)

  // 외부에서 카테고리 힌트가 변경되면 반영
  useEffect(() => {
@ -44,6 +46,9 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
    // HNS 데이터 소비
    const payload = consumeHnsReportPayload()
    if (payload) setHnsPayload(payload)
+    // OIL 예측 데이터 소비
+    const oilData = consumeOilReportPayload()
+    if (oilData) setOilPayload(oilData)
  }, [])

  const cat = CATEGORIES[activeCat]
@ -65,8 +70,19 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
    report.status = '완료'
    report.author = '시스템 자동생성'
    if (activeCat === 0) {
-      report.incident.pollutant = sampleOilData.pollution.oilType
-      report.incident.spillAmount = sampleOilData.pollution.spillAmount
+      if (oilPayload) {
+        report.incident.name = oilPayload.incident.name;
+        report.incident.occurTime = oilPayload.incident.occurTime;
+        report.incident.location = oilPayload.incident.location;
+        report.incident.lat = String(oilPayload.incident.lat ?? '');
+        report.incident.lon = String(oilPayload.incident.lon ?? '');
+        report.incident.shipName = oilPayload.incident.shipName;
+        report.incident.pollutant = oilPayload.pollution.oilType;
+        report.incident.spillAmount = oilPayload.pollution.spillAmount;
+      } else {
+        report.incident.pollutant = sampleOilData.pollution.oilType;
+        report.incident.spillAmount = sampleOilData.pollution.spillAmount;
+      }
    }
    try {
      await saveReport(report)
@ -82,6 +98,24 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
    const sectionHTML = activeSections.map(sec => {
      let content = `<p style="font-size:12px;color:#666;">${sec.desc}</p>`;

+      // OIL 섹션에 실 데이터 삽입
+      if (activeCat === 0 && oilPayload) {
+        if (sec.id === 'oil-pollution') {
+          const rows = [
+            ['유출량', oilPayload.pollution.spillAmount, '풍화량', oilPayload.pollution.weathered],
+            ['해상잔유량', oilPayload.pollution.seaRemain, '오염해역면적', oilPayload.pollution.pollutionArea],
+            ['연안부착량', oilPayload.pollution.coastAttach, '오염해안길이', oilPayload.pollution.coastLength],
+          ];
+          const simBanner = !oilPayload.hasSimulation
+            ? '<p style="font-size:10px;color:#f97316;margin-bottom:8px;">시뮬레이션이 실행되지 않아 오염량은 입력값 기준으로 표시됩니다.</p>'
+            : '';
+          const trs = rows.map(r =>
+            `<tr><td style="padding:6px 8px;border:1px solid #ddd;color:#888;">${r[0]}</td><td style="padding:6px 8px;border:1px solid #ddd;font-weight:bold;text-align:right;">${r[1]}</td><td style="padding:6px 8px;border:1px solid #ddd;color:#888;">${r[2]}</td><td style="padding:6px 8px;border:1px solid #ddd;font-weight:bold;text-align:right;">${r[3]}</td></tr>`
+          ).join('');
+          content = `${simBanner}<table style="width:100%;border-collapse:collapse;font-size:12px;">${trs}</table>`;
+        }
+      }
+
      // HNS 섹션에 실 데이터 삽입
      if (activeCat === 1 && hnsPayload) {
        if (sec.id === 'hns-atm') {
@ -261,9 +295,9 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
                      </div>
                      <div className="grid grid-cols-3 gap-3">
                        {[
-                          { label: 'KOSPS', value: sampleOilData.spread.kosps, color: '#06b6d4' },
-                          { label: 'OpenDrift', value: sampleOilData.spread.openDrift, color: '#ef4444' },
-                          { label: 'POSEIDON', value: sampleOilData.spread.poseidon, color: '#f97316' },
+                          { label: 'KOSPS', value: oilPayload?.spread.kosps || sampleOilData.spread.kosps, color: '#06b6d4' },
+                          { label: 'OpenDrift', value: oilPayload?.spread.openDrift || sampleOilData.spread.openDrift, color: '#ef4444' },
+                          { label: 'POSEIDON', value: oilPayload?.spread.poseidon || sampleOilData.spread.poseidon, color: '#f97316' },
                        ].map((m, i) => (
                          <div key={i} className="bg-bg-1 border border-border rounded-lg p-4 text-center">
                            <p className="text-[10px] text-text-3 font-korean mb-1">{m.label}</p>
@ -274,23 +308,30 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
                    </>
                  )}
                  {sec.id === 'oil-pollution' && (
-                    <table className="w-full table-fixed border-collapse">
-                      <colgroup><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /></colgroup>
-                      <tbody>
-                        {[
-                          ['유출량', sampleOilData.pollution.spillAmount, '풍화량', sampleOilData.pollution.weathered],
-                          ['해상잔유량', sampleOilData.pollution.seaRemain, '오염해역면적', sampleOilData.pollution.pollutionArea],
-                          ['연안부착량', sampleOilData.pollution.coastAttach, '오염해안길이', sampleOilData.pollution.coastLength],
-                        ].map((row, i) => (
-                          <tr key={i} className="border-b border-border">
-                            <td className="px-4 py-3 text-[11px] text-text-3 font-korean bg-[rgba(255,255,255,0.02)]">{row[0]}</td>
-                            <td className="px-4 py-3 text-[12px] text-text-1 font-mono font-semibold text-right">{row[1]}</td>
-                            <td className="px-4 py-3 text-[11px] text-text-3 font-korean bg-[rgba(255,255,255,0.02)]">{row[2]}</td>
-                            <td className="px-4 py-3 text-[12px] text-text-1 font-mono font-semibold text-right">{row[3]}</td>
-                          </tr>
-                        ))}
-                      </tbody>
-                    </table>
+                    <>
+                      {oilPayload && !oilPayload.hasSimulation && (
+                        <div className="mb-3 px-3 py-2 rounded text-[10px] font-korean" style={{ background: 'rgba(249,115,22,0.08)', border: '1px solid rgba(249,115,22,0.3)', color: '#f97316' }}>
+                          시뮬레이션이 실행되지 않아 오염량은 입력값 기준으로 표시됩니다.
+                        </div>
+                      )}
+                      <table className="w-full table-fixed border-collapse">
+                        <colgroup><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /><col style={{ width: '25%' }} /></colgroup>
+                        <tbody>
+                          {[
+                            ['유출량', oilPayload?.pollution.spillAmount || sampleOilData.pollution.spillAmount, '풍화량', oilPayload?.pollution.weathered || sampleOilData.pollution.weathered],
+                            ['해상잔유량', oilPayload?.pollution.seaRemain || sampleOilData.pollution.seaRemain, '오염해역면적', oilPayload?.pollution.pollutionArea || sampleOilData.pollution.pollutionArea],
+                            ['연안부착량', oilPayload?.pollution.coastAttach || sampleOilData.pollution.coastAttach, '오염해안길이', oilPayload?.pollution.coastLength || sampleOilData.pollution.coastLength],
+                          ].map((row, i) => (
+                            <tr key={i} className="border-b border-border">
+                              <td className="px-4 py-3 text-[11px] text-text-3 font-korean bg-[rgba(255,255,255,0.02)]">{row[0]}</td>
+                              <td className="px-4 py-3 text-[12px] text-text-1 font-mono font-semibold text-right">{row[1]}</td>
+                              <td className="px-4 py-3 text-[11px] text-text-3 font-korean bg-[rgba(255,255,255,0.02)]">{row[2]}</td>
+                              <td className="px-4 py-3 text-[12px] text-text-1 font-mono font-semibold text-right">{row[3]}</td>
+                            </tr>
+                          ))}
+                        </tbody>
+                      </table>
+                    </>
                  )}
                  {sec.id === 'oil-sensitive' && (
                    <>
@ -304,9 +345,9 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
                  )}
                  {sec.id === 'oil-coastal' && (
                    <p className="text-[12px] text-text-2 font-korean">
-                      최초 부착시간: <span className="font-semibold text-text-1">{sampleOilData.coastal.firstTime}</span>
+                      최초 부착시간: <span className="font-semibold text-text-1">{oilPayload?.coastal?.firstTime ?? sampleOilData.coastal.firstTime}</span>
                      {' / '}
-                      부착 해안길이: <span className="font-semibold text-text-1">{sampleOilData.coastal.coastLength}</span>
+                      부착 해안길이: <span className="font-semibold text-text-1">{oilPayload?.pollution.coastLength || sampleOilData.coastal.coastLength}</span>
                    </p>
                  )}
                  {sec.id === 'oil-defense' && (
@ -318,11 +359,20 @@ function ReportGenerator({ onSave }: ReportGeneratorProps) {
                    </div>
                  )}
                  {sec.id === 'oil-tide' && (
-                    <p className="text-[12px] text-text-2 font-korean">
-                      고조: <span className="font-semibold text-text-1">{sampleOilData.tide.highTide1}</span>
-                      {' / '}저조: <span className="font-semibold text-text-1">{sampleOilData.tide.lowTide}</span>
-                      {' / '}고조: <span className="font-semibold text-text-1">{sampleOilData.tide.highTide2}</span>
-                    </p>
+                    <>
+                      <p className="text-[12px] text-text-2 font-korean">
+                        고조: <span className="font-semibold text-text-1">{sampleOilData.tide.highTide1}</span>
+                        {' / '}저조: <span className="font-semibold text-text-1">{sampleOilData.tide.lowTide}</span>
+                        {' / '}고조: <span className="font-semibold text-text-1">{sampleOilData.tide.highTide2}</span>
+                      </p>
+                      {oilPayload?.weather && (
+                        <p className="text-[11px] text-text-3 font-korean mt-2">
+                          기상: 풍향/풍속 <span className="text-text-2 font-semibold">{oilPayload.weather.windDir}</span>
+                          {' / '}파고 <span className="text-text-2 font-semibold">{oilPayload.weather.waveHeight}</span>
+                          {' / '}기온 <span className="text-text-2 font-semibold">{oilPayload.weather.temp}</span>
+                        </p>
+                      )}
+                    </>
                  )}

                  {/* ── HNS 대기확산 섹션들 ── */}
--- a/prediction/image/DOCKER_USAGE.md
+++ b/prediction/image/DOCKER_USAGE.md
@ -16,6 +16,7 @@
 8. [로그 확인 및 디버깅](#8-로그-확인-및-디버깅)
 9. [컨테이너 관리](#9-컨테이너-관리)
 10. [주의사항](#10-주의사항)
+11. [CPU 전용 환경 실행](#11-cpu-전용-환경-실행)

 ---

@ -278,10 +279,10 @@ docker system prune -f

 ## 10. 주의사항

-### GPU 필수
- AI 모델(`epoch_165.pth`)은 `cuda:0` 디바이스로 로드된다.
- GPU 없이 실행하면 서버 기동 시 오류가 발생한다.
- CPU 전용 환경에서 테스트하려면 `Inference.py`의 `device='cuda:0'`을 `device='cpu'`로 수정해야 한다.
+### GPU 자동 감지
+- 서버 기동 시 `torch.cuda.is_available()`로 GPU 유무를 자동 감지한다.
+- GPU가 있으면 `cuda:0`, 없으면 `cpu`로 자동 폴백된다.
+- 환경변수 `DEVICE`로 device를 명시 지정할 수 있다 (예: `DEVICE=cpu`, `DEVICE=cuda:1`).

 ### 첫 기동 시간
 - AI 모델 로드: 약 **10~30초** 소요 (GPU 메모리에 로딩)
@ -297,3 +298,79 @@ docker system prune -f
  ports:
    - "5002:5001"  # 호스트 5002 → 컨테이너 5001
  ```
+
+---
+
+## 11. CPU 전용 환경 실행
+
+GPU(NVIDIA)가 없는 환경에서는 CPU 전용 설정을 사용한다.
+
+### 사전 요구사항 (CPU 모드)
+
+| 항목 | 최소 버전 | 확인 명령어 |
+|------|----------|-------------|
+| Docker Engine | 24.0 이상 | `docker --version` |
+| Docker Compose | 2.20 이상 | `docker compose version` |
+| NVIDIA 드라이버 | **불필요** | — |
+
+### 빠른 시작 (CPU)
+
+```bash
+# prediction/image/ 디렉토리로 이동
+cd prediction/image
+
+# 환경변수 파일 준비 (필요 시)
+cp .env.example .env
+
+# CPU 이미지 빌드 + 실행
+docker compose -f docker-compose.cpu.yml up -d --build
+
+# 서버 상태 확인
+curl http://localhost:5001/docs
+```
+
+### 빌드 명령어 (CPU)
+
+```bash
+# CPU 이미지만 빌드
+docker compose -f docker-compose.cpu.yml build
+
+# 캐시 없이 빌드
+docker compose -f docker-compose.cpu.yml build --no-cache
+```
+
+> **참고**: CPU 기반 PyTorch 이미지는 GPU 이미지(~8GB) 대비 약 70% 용량이 절감된다.
+> 단, CPU 추론은 GPU 대비 처리 속도가 느리므로 대용량 이미지 분석 시 시간이 더 소요된다.
+
+### 실행 명령어 (CPU)
+
+```bash
+# 백그라운드 실행
+docker compose -f docker-compose.cpu.yml up -d
+
+# 포그라운드 실행 (로그 바로 출력)
+docker compose -f docker-compose.cpu.yml up
+
+# 중지
+docker compose -f docker-compose.cpu.yml down
+```
+
+### 로컬 직접 실행 (Docker 없이)
+
+```bash
+# GPU 있으면 자동으로 cuda:0 사용, 없으면 cpu로 폴백
+python api.py
+
+# device 강제 지정
+DEVICE=cpu python api.py
+DEVICE=cuda:1 python api.py
+```
+
+### GPU/CPU 모드 확인
+
+서버 기동 로그에서 사용 device를 확인할 수 있다:
+
+```
+[Inference] 사용 device: cpu      ← CPU 모드
+[Inference] 사용 device: cuda:0   ← GPU 모드
+```
--- a/prediction/image/Dockerfile
+++ b/prediction/image/Dockerfile
@ -1,11 +1,12 @@
 # ==============================================================================
 # wing-image-analysis — 드론 영상 유류 분석 FastAPI 서버
 #
-# Base: PyTorch 2.1 + CUDA 12.1 + cuDNN 8 (devel 빌드 — GDAL 컴파일 필요)
+# Base: PyTorch 1.9.1 + CUDA 11.1 + cuDNN 8
+#       (mmsegmentation 0.25.0 / mmcv-full 1.4.3 호환 환경)
 # GPU:  NVIDIA GPU 필수 (MMSegmentation 추론)
 # Port: 5001
 # ==============================================================================
-FROM pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel
+FROM pytorch/pytorch:1.9.1-cuda11.1-cudnn8-devel

 ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONDONTWRITEBYTECODE=1 \
@ -32,6 +33,14 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # rasterio는 GDAL 헤더 버전을 맞춰 빌드해야 한다
 ENV GDAL_VERSION=3.4.1

+# ------------------------------------------------------------------------------
+# mmcv-full 1.4.3 — CUDA 11.1 + PyTorch 1.9.0 pre-built 휠
+# (소스 컴파일 없이 수 초 내 설치)
+# ------------------------------------------------------------------------------
+RUN pip install --no-cache-dir \
+    mmcv-full==1.4.3 \
+    -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html
+
 # ------------------------------------------------------------------------------
 # Python 의존성 설치
 # ------------------------------------------------------------------------------
--- a/prediction/image/Dockerfile.cpu
+++ b/prediction/image/Dockerfile.cpu
@ -0,0 +1,112 @@
+# ==============================================================================
+# wing-image-analysis — 드론 영상 유류 분석 FastAPI 서버 (CPU 전용)
+#
+# Base: python:3.9-slim + PyTorch 1.9.0 CPU 빌드
+#       (mmsegmentation 0.25.0 / mmcv-full 1.4.3 호환 환경)
+#       python:3.9 필수 — numpy 1.26.4, geopandas 0.14.4가 Python >=3.9 요구
+# GPU:  불필요 (CPU 추론)
+# Port: 5001
+# ==============================================================================
+FROM python:3.9-slim
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    DEVICE=cpu
+
+WORKDIR /app
+
+# ------------------------------------------------------------------------------
+# 시스템 패키지: GDAL / PROJ / GEOS (rasterio, geopandas 빌드 의존성)
+# libspatialindex-dev: geopandas 공간 인덱스
+# opencv-contrib-python-headless 런타임 SO 의존성 (python:3.9-slim에 미포함):
+#   libgl1        — libGL.so.1
+#   libglib2.0-0  — libgthread-2.0.so.0, libgobject-2.0.so.0, libglib-2.0.so.0
+#   libsm6        — libSM.so.6
+#   libxext6      — libXext.so.6
+#   libxrender1   — libXrender.so.1
+#   libgomp1      — libgomp.so.1 (OpenMP, numpy/opencv 병렬 처리)
+# ------------------------------------------------------------------------------
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    gdal-bin \
+    libgdal-dev \
+    libproj-dev \
+    libgeos-dev \
+    libspatialindex-dev \
+    libgl1 \
+    libglib2.0-0 \
+    libsm6 \
+    libxext6 \
+    libxrender1 \
+    libgomp1 \
+    gcc \
+    g++ \
+    git \
+  && rm -rf /var/lib/apt/lists/*
+
+# rasterio는 GDAL 헤더 버전을 맞춰 빌드해야 한다
+ENV GDAL_VERSION=3.4.1
+
+# ------------------------------------------------------------------------------
+# GDAL Python 바인딩 (osgeo 모듈) — 시스템 GDAL 버전과 일치해야 한다
+# python:3.9-slim은 conda 없이 pip 환경이므로 명시적 설치 필요
+# ------------------------------------------------------------------------------
+RUN pip install --no-cache-dir GDAL=="$(gdal-config --version)"
+
+# ------------------------------------------------------------------------------
+# PyTorch 1.9.0 CPU 버전 설치
+# (mmsegmentation 0.25.0 / mmcv-full 1.4.3 호환)
+# ------------------------------------------------------------------------------
+RUN pip install --no-cache-dir \
+    torch==1.9.0+cpu \
+    torchvision==0.10.0+cpu \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+# ------------------------------------------------------------------------------
+# mmcv-full 1.4.3 CPU 휠 (CUDA ops 없는 경량 빌드, 추론에 충분)
+# ------------------------------------------------------------------------------
+RUN pip install --no-cache-dir \
+    mmcv-full==1.4.3 \
+    -f https://download.openmmlab.com/mmcv/dist/cpu/torch1.9.0/index.html
+
+# ------------------------------------------------------------------------------
+# Python 의존성 설치
+# ------------------------------------------------------------------------------
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# ------------------------------------------------------------------------------
+# 로컬 mmsegmentation 설치 (mx15hdi/Detect/mmsegmentation/)
+# 번들 소스를 먼저 복사한 뒤 editable 설치한다
+# ------------------------------------------------------------------------------
+COPY mx15hdi/Detect/mmsegmentation/ /tmp/mmsegmentation/
+RUN pip install --no-cache-dir -e /tmp/mmsegmentation/
+
+# ------------------------------------------------------------------------------
+# 소스 코드 전체 복사
+# 대용량 데이터 디렉토리(Original_Images, result 등)는
+# docker-compose.cpu.yml의 볼륨 마운트로 외부에서 주입된다
+# ------------------------------------------------------------------------------
+COPY . .
+
+# ------------------------------------------------------------------------------
+# .dockerignore로 제외된 런타임 출력 디렉토리를 빈 폴더로 생성
+# (볼륨 마운트 전에도 경로가 존재해야 한다)
+# ------------------------------------------------------------------------------
+RUN mkdir -p \
+    /app/stitch \
+    /app/mx15hdi/Detect/Mask_result \
+    /app/mx15hdi/Detect/result \
+    /app/mx15hdi/Georeference/Mask_Tif \
+    /app/mx15hdi/Georeference/Tif \
+    /app/mx15hdi/Metadata/CSV \
+    /app/mx15hdi/Metadata/Image/Original_Images \
+    /app/mx15hdi/Polygon/Shp
+
+# ------------------------------------------------------------------------------
+# 런타임 설정
+# ------------------------------------------------------------------------------
+EXPOSE 5001
+
+# workers=1: 모델을 프로세스 하나에서만 로드 (메모리 공유 불가)
+CMD ["uvicorn", "api:app", "--host", "0.0.0.0", "--port", "5001", "--workers", "1"]
--- a/prediction/image/api.py
+++ b/prediction/image/api.py
@ -180,6 +180,8 @@ async def run_script(
    except subprocess.TimeoutExpired:
        raise HTTPException(status_code=500, detail="Script execution timed out")
    except Exception as e:
+        import traceback
+        traceback.print_exc()
        raise HTTPException(status_code=500, detail=str(e))


--- a/prediction/image/docker-compose.cpu.yml
+++ b/prediction/image/docker-compose.cpu.yml
@ -0,0 +1,46 @@
+version: "3.9"
+
+# CPU 전용 docker-compose 설정
+# GPU(nvidia-container-toolkit) 없이도 실행 가능
+# 실행: docker compose -f docker-compose.cpu.yml up -d --build
+
+services:
+  image-analysis:
+    build:
+      context: .
+      dockerfile: Dockerfile.cpu
+    image: wing-image-analysis:cpu
+    container_name: wing-image-analysis
+    ports:
+      - "5001:5001"
+    environment:
+      - DEVICE=cpu
+
+    volumes:
+      # ── mx15hdi (EO 드론 카메라) ────────────────────────────────────────
+      # 입력: 업로드된 원본 이미지
+      - ./mx15hdi/Metadata/Image/Original_Images:/app/mx15hdi/Metadata/Image/Original_Images
+      # 출력: 메타데이터 CSV
+      - ./mx15hdi/Metadata/CSV:/app/mx15hdi/Metadata/CSV
+      # 출력: 지리참조 GeoTIFF (컬러 / 마스크)
+      - ./mx15hdi/Georeference/Tif:/app/mx15hdi/Georeference/Tif
+      - ./mx15hdi/Georeference/Mask_Tif:/app/mx15hdi/Georeference/Mask_Tif
+      # 출력: 유류 폴리곤 Shapefile
+      - ./mx15hdi/Polygon/Shp:/app/mx15hdi/Polygon/Shp
+      # 출력: 블렌딩 추론 결과 / 마스크 이미지
+      - ./mx15hdi/Detect/result:/app/mx15hdi/Detect/result
+      - ./mx15hdi/Detect/Mask_result:/app/mx15hdi/Detect/Mask_result
+      # ── starsafire (열화상 카메라) ──────────────────────────────────────
+      - ./starsafire/Metadata/Image/Original_Images:/app/starsafire/Metadata/Image/Original_Images
+      - ./starsafire/Metadata/CSV:/app/starsafire/Metadata/CSV
+      - ./starsafire/Georeference/Tif:/app/starsafire/Georeference/Tif
+      - ./starsafire/Georeference/Mask_Tif:/app/starsafire/Georeference/Mask_Tif
+      - ./starsafire/Polygon/Shp:/app/starsafire/Polygon/Shp
+      - ./starsafire/Detect/result:/app/starsafire/Detect/result
+      - ./starsafire/Detect/Mask_result:/app/starsafire/Detect/Mask_result
+      # ── 스티칭 결과 ─────────────────────────────────────────────────────
+      - ./stitch:/app/stitch
+
+    # GPU deploy 섹션 없음 — CPU 전용 실행
+
+    restart: unless-stopped
--- a/prediction/image/docker-compose.yml
+++ b/prediction/image/docker-compose.yml
@ -9,7 +9,6 @@ services:
    container_name: wing-image-analysis
    ports:
      - "5001:5001"
-    env_file: .env

    volumes:
      # ── mx15hdi (EO 드론 카메라) ────────────────────────────────────────
--- a/prediction/image/mx15hdi/Detect/Inference.py
+++ b/prediction/image/mx15hdi/Detect/Inference.py
@ -1,5 +1,6 @@
 import os, mmcv, cv2, json
 import numpy as np
+import torch
 from pathlib import Path
 from PIL import Image
 from tqdm import tqdm
@ -13,9 +14,19 @@ _MX15HDI_DIR = _DETECT_DIR.parent       # mx15hdi/

 def load_model():
    """서버 시작 시 1회 호출. 로드된 모델 객체를 반환한다."""
+    # 우선순위: 환경변수 DEVICE > GPU 자동감지 > CPU 폴백
+    env_device = os.environ.get('DEVICE', '').strip()
+    if env_device:
+        device = env_device
+    elif torch.cuda.is_available():
+        device = 'cuda:0'
+    else:
+        device = 'cpu'
+    print(f'[Inference] 사용 device: {device}')
+
    config = str(_DETECT_DIR / 'V7_SPECIAL.py')
    checkpoint = str(_DETECT_DIR / 'epoch_165.pth')
-    model = init_segmentor(config, checkpoint, device='cuda:0')
+    model = init_segmentor(config, checkpoint, device=device)
    model.PALETTE = [
        [0, 0, 0],       # background
        [0, 0, 204],     # black
--- a/prediction/image/requirements.txt
+++ b/prediction/image/requirements.txt
@ -4,18 +4,26 @@ uvicorn[standard]==0.29.0

 # 이미지 처리
 numpy==1.26.4
-opencv-python-headless==4.9.0.80
+# opencv-contrib-python-headless: headless(GUI 불필요) + contrib(Stitcher 등) 통합
+opencv-contrib-python-headless==4.9.0.80
 Pillow==10.3.0
 piexif==1.1.3
+scikit-image==0.19.3
+matplotlib==3.5.1

 # 지리 데이터 처리
 rasterio==1.3.10
 geopandas==0.14.4
 shapely==2.0.4
 pyproj==3.6.1
+# osgeo(GDAL Python 바인딩)는 시스템 GDAL 버전과 맞춰야 하므로 Dockerfile에서 설치

-# AI/ML — PyTorch는 base 이미지에 포함, mmsegmentation은 로컬 소스에서 설치
-mmcv==2.1.0
+# AI/ML — PyTorch는 base 이미지에 포함, mmcv/mmsegmentation은 Dockerfile에서 설치
+# mmcv-full==1.4.3 은 torch/CUDA 버전에 맞는 pre-built 휠이 필요하여 Dockerfile에서 직접 설치
+
+# OCR (메타데이터 추출: Export_Metadata_mx15hdi.py)
+paddlepaddle==2.6.2
+paddleocr==2.7.0.2

 # 유틸리티
 pandas==2.2.2