2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/SAR.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-sar	SAR (Synthetic Aperture Radar)	10_Wiki/Topics	verified	self	Synthetic Aperture Radar SAR imagery	none	A	0.9	applied	sar remote-sensing radar geospatial deep-learning		2026-05-10	pending	language framework Python SNAP / PyTorch

SAR (Synthetic Aperture Radar)

매 한 줄

"매 active microwave imaging — 매 cloud/night 무관 의 24/7 Earth observation". 매 1950s aircraft radar 로 시작 → 매 Sentinel-1 (ESA, free), Capella, ICEYE, Umbra 등 매 commercial smallsat constellation 으로 daily revisit. 매 deep learning (despeckling, segmentation, change detection) 이 매 traditional SAR processing pipeline 을 재편.

매 핵심

매 물리

• Active sensor: own microwave (C-band 5.4GHz, X-band 9.6GHz, L-band 1.3GHz) 송신, backscatter 측정.
• Synthetic aperture: 매 platform motion 으로 매 large virtual antenna 합성 → sub-meter resolution.
• Polarimetry: HH/HV/VH/VV → material/structure 정보.
• Interferometry (InSAR): phase difference 로 매 mm-level surface deformation.

매 Data products

• SLC (Single Look Complex): phase 보존, InSAR 용.
• GRD (Ground Range Detected): amplitude only, 일반 분석.
• Speckle noise: multiplicative, log-normal — 매 deep despeckling 의 핵심 challenge.

매 응용

1. Maritime surveillance (ship detection, dark-vessel via AIS-cross).
2. Disaster response (flood mapping, earthquake deformation).
3. Agriculture (crop type, soil moisture).
4. Defense (change detection, target classification).
5. Subsidence monitoring (city, mining, dam).

💻 패턴

Sentinel-1 download (sentinelsat)

from sentinelsat import SentinelAPI
from datetime import date

api = SentinelAPI("user", "pass", "https://apihub.copernicus.eu/apihub")
products = api.query(
    footprint="POLYGON((127 37, 128 37, 128 38, 127 38, 127 37))",
    date=(date(2026, 1, 1), date(2026, 5, 1)),
    platformname="Sentinel-1",
    producttype="GRD",
)
api.download_all(products)

Speckle filtering (Refined Lee)

import numpy as np
from scipy.ndimage import generic_filter

def refined_lee(img, size=7):
    def filter_fn(window):
        mean = window.mean()
        var = window.var()
        cu = 0.523  # SAR ENL-derived
        ci = np.sqrt(var) / mean if mean else 0
        w = max(0, (ci**2 - cu**2) / (ci**2 * (1 + cu**2)))
        return mean + w * (window[len(window)//2] - mean)
    return generic_filter(img, filter_fn, size=size)

Deep despeckling (SAR-CNN, 2026 SOTA)

import torch.nn as nn

class SARCNN(nn.Module):
    def __init__(self, depth=17):
        super().__init__()
        layers = [nn.Conv2d(1, 64, 3, padding=1), nn.ReLU()]
        for _ in range(depth - 2):
            layers += [nn.Conv2d(64, 64, 3, padding=1),
                       nn.BatchNorm2d(64), nn.ReLU()]
        layers.append(nn.Conv2d(64, 1, 3, padding=1))
        self.net = nn.Sequential(*layers)

    def forward(self, x):
        # log domain — speckle becomes additive
        return torch.log(x + 1e-6) - self.net(torch.log(x + 1e-6))

Ship detection (CFAR + YOLO-SAR)

from ultralytics import YOLO

# Fine-tuned on SSDD/HRSID dataset
model = YOLO("yolov8-sar-ship.pt")
results = model("sentinel1_grd_tile.tif", conf=0.4, imgsz=1024)
for box in results[0].boxes:
    lon, lat = pixel_to_geo(box.xywh[0][:2])
    print(f"Ship @ {lat:.4f},{lon:.4f} conf={box.conf.item():.2f}")

InSAR coherence + interferogram (snappy)

# ESA SNAP via snappy: master/slave coregistration → ifg
from snappy import GPF, ProductIO

master = ProductIO.readProduct("S1A_master.zip")
slave = ProductIO.readProduct("S1A_slave.zip")
coreg = GPF.createProduct("Back-Geocoding", params, [master, slave])
ifg = GPF.createProduct("Interferogram", {"includeCoherence": True}, coreg)
ProductIO.writeProduct(ifg, "ifg.dim", "BEAM-DIMAP")

Flood change detection

import rasterio
import numpy as np

with rasterio.open("pre.tif") as a, rasterio.open("post.tif") as b:
    pre, post = a.read(1), b.read(1)
# Log-ratio
lr = np.log10(post / (pre + 1e-6))
flood_mask = lr < -0.5  # darker = water in VV

매 결정 기준

상황	Approach
Free, weekly revisit	Sentinel-1 GRD
Sub-daily, sub-meter	Capella / ICEYE / Umbra commercial
Deformation (mm)	InSAR time series (Sentinel-1, ALOS-2 L-band)
Foliage penetration	L-band (ALOS, NISAR 2026)
Maritime wide-area	Sentinel-1 EW + AIS fusion

기본값: Sentinel-1 GRD + deep despeckling + YOLO-SAR for object tasks.

🔗 Graph

🤖 LLM 활용

언제: report generation from detection outputs, multi-modal SAR+optical fusion via VLM (Prithvi-SAR, 2026), tasking-orchestration agents. 언제 X: pixel-level despeckling/segmentation — use specialized CNN/transformer, not LLM.

❌ 안티패턴

• Speckle ignored: training optical CNN directly on SAR amplitude — speckle dominates loss.
• No log/dB conversion: SAR has 60+ dB dynamic range; visualize/train in dB scale.
• Geocoding skipped: pixel coords ≠ geographic — terrain correction required.
• Single polarization: dual-pol (VV+VH) gives material discrimination essentially free.

🧪 검증 / 중복

• Verified (ESA Copernicus, NASA NISAR docs, Capella tech papers).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — SAR physics, deep despeckling, ship/flood detection