Radio Environment Map Construction with Joint Space-Frequency Interpolation at ICAIIC 2020 February 19th, 2020 Koya Sato1, Kei Inage2, and Takeo Fujii3 Tokyo University of Science, Japan1 Tokyo Metropolitan College of Industrial Technology, Japan2 The University of Electro-Communications, Japan3 ※This research is supported by the Ministry of Internal Affairs and Communications in Japan

Background: Radio Map • Radio Map: Map storing average received signal power • Stored at cloud Examples of Applications • Television white space (TVWS) [K. Sato et al., IEEE TCCN, 2017] • Coverage analysis in cellular systems [S. Bi et al., IEEE Wireless Commun. 2019] • Fingerprint-based localization [B. Huang et al.,, IEEE JIoT, 2019] Accurate RMs boost performances of above applications. Ø How can we construct accurate RMs? 2

Crowdsensing-based Radio Map Construction • Mobile terminals sense radio environments • RMs are used for communications design Cloud Transmitter Location: (xk , yk ) Power: Pk [mW] Sensing nodes Upload 5G 5G 5G Average Received Power Location: (xj , yj ) Power: Pj [mW] Location: (xi , yi ) Power: Pi [mW] Radio Map References • S. Bi et al., IEEE Wireless Commun., 2019 • K. Sato et al., IEEE Access, July 2019 3

Conventional Discussions on RM construction We need to interpolate tooth-missed values from measurements 4 Measured values True values Related approaches • Path loss model • Kriging (or Gaussian Process Regression (GPR)): Our focus • Machine learning Interpolate These methods interpolate over space domain in one frequency

Cost Issues in Crowdsensing and Motivation Measurable bands are strictly limited!! SHF band (3-30GHz) UHF band (300-3000MHz) Bands that smartphones can measure Problem • Existing RM interpolation is performed over only spatial domain Ø Clouds can construct RMs over (very limited) bands that mobile terminals can measure Main Focus Expanding the applications of RMs in terms of frequency bands Ø We extend the spatial interpolation to joint space-frequency interpolation 5

Conventional Spatial Interpolation (1-D Case) 6 1. Path loss modeling via ordinary least squares (OLS) 2. Shadowing extraction from datasets 3. Kriging for shadowing interpolation Logarithmic of Distance Received Signal Power [dBm] Dataset over f0 Path Loss Shadowing Note: Kriging in this context is almost equal to Gaussian Process Regression (GPR)

Conventional Spatial Interpolation (1-D Case) 7 1. Path loss modeling via ordinary least squares (OLS) 2. Shadowing extraction from datasets 3. Kriging for shadowing interpolation Note: Kriging in this context is almost equal to Gaussian Process Regression (GPR) Received Signal Power [dBm] Estimated path loss Effect of distance Logarithmic of Distance

Conventional Spatial Interpolation (1-D Case) 8 1. Path loss modeling via ordinary least squares (OLS) 2. Shadowing extraction from datasets 3. Kriging for shadowing interpolation Note: Kriging in this context is almost equal to Gaussian Process Regression (GPR) Shadowing W [dB] Received Signal Power [dBm] Estimated path loss Effect of distance Logarithmic of Distance

Conventional Spatial Interpolation (1-D Case) 9 1. Path loss modeling via ordinary least squares (OLS) 2. Shadowing extraction from datasets 3. Kriging for shadowing interpolation Note: Kriging in this context is almost equal to Gaussian Process Regression (GPR) Shadowing W [dB] Received Signal Power [dBm] Weight factor →Optimized by Kriging Logarithmic of Distance

Conventional Spatial Interpolation (1-D Case) 10 1. Path loss modeling via ordinary least squares (OLS) 2. Shadowing extraction from datasets 3. Kriging for shadowing interpolation Note: Kriging in this context is almost equal to Gaussian Process Regression (GPR) Shadowing W [dB] Received Signal Power [dBm] Weight factor →Optimized by Kriging Question: Is it possible to extend this method to joint frequency-space interpolation? Logarithmic of Distance

Related Discussion: Frequency Correlation of Shadowing 11 Some authors evaluated from experiments... B. V. Van Laethem et al. (PERLett., 2012) • Measurement campaign for GSM systems • They reported... Ø 0.84 between 900MHz and 1800 MHz Ø 0.79 between 900MHz and 2100MHz Perahia et al. (IEEE VTC2001-Spring, Jan. 2001) • Comparison between uplink and downlink in FDMA systems • They reported... Ø 0.7-0.9 between uplink and downlink Shadowing has high frequency correlation even if the frequency difference is over 1GHz! Main Idea Shadowing values over the other bands are treated as the ones over the interpolated frequency

Proposed Method (1-D Case) 12 1. Path loss modeling via two-dimensional OLS 2. Shadowing extraction from datasets 3. Shadowing interpolation over frequency domain 4. Kriging for shadowing interpolation Logarithmic of Distance Received Signal Power [dBm] Dataset over f1 Dataset over f2

Proposed Method (1-D Case) 13 1. Path loss modeling via two-dimensional OLS 2. Shadowing extraction from datasets 3. Shadowing interpolation over frequency domain 4. Kriging for shadowing interpolation Received Signal Power [dBm] Path loss over f0 Effect of frequency Effect of distance Logarithmic of Distance

Proposed Method (1-D Case) 14 1. Path loss modeling via two-dimensional OLS 2. Shadowing extraction from datasets 3. Shadowing interpolation over frequency domain 4. Kriging for shadowing interpolation Received Signal Power [dBm] Shadowing over the other bands are treated over the ones at f0 Logarithmic of Distance

Proposed Method (1-D Case) 15 1. Path loss modeling via two-dimensional OLS 2. Shadowing extraction from datasets 3. Shadowing interpolation over frequency domain 4. Kriging for shadowing interpolation Received Signal Power [dBm] Weight factor →Optimized by Kriging Logarithmic of Distance

Simulation Setup 16 • Measured bands: 400MHz and 1500MHz • Interpolated bands: 800MHz • Measurement locations and frequencies are randomly selected 200m xi Tx Measured point Path loss model Okumura- Hata Stdev. of shadowing 8 [dB] Shadowing correlation 20 [m] Frequency correlation 0.80 Antenna height at Tx 30[m] Transmission power 30[dBm] 100m 100m

1500MHz (Measured) Example of Joint Space-Frequency Interpolation 17 Number of datasets: 100 800MHz (Estimated) 400MHz (Measured) 800MHz (True)

Accuracy Performance 18 More accurate • Root mean squared error (RMSE) at the center of simulation area Ø RMSE was calculated via Monte-Carlo simulation

Conclusion Main Contributions • We extended the RM construction from spatial interpolation to joint space-frequency interpolation • Numerical results indicated that we can construct the RM from datasets over other frequencies Future works • Validation using actual datasets 19