Moderate evidence
More
A simpler model beat deep learning at predicting signal loss in simulated Iraqi weather
Study title: Machine Learning Outperforms Deep Learning for Atmospheric Attenuation Prediction in Free-Space Optical Communications under Iraqi Weather Conditions
- Evidence type
- Benchmark or lab evaluation
- Sample
- n = 1,500
- Sources
- 1 sourcePeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-1258cite or check it
Study title: Machine Learning Outperforms Deep Learning for Atmospheric Attenuation Prediction in Free-Space Optical Communications under Iraqi Weather Conditions
A study evaluated 15 machine learning, deep learning and hybrid models for predicting atmospheric attenuation in free-space optical communications under five Iraqi weather regimes. Using a 1500-sample synthetic dataset derived from physical propagation models, Random Forest achieved R2 0.9654 and RMSE 1.324 dB/km, outperforming the best deep learning model at R2 0.7766 and slightly edging the best hybrid at 0.9571.
Read the study abstract
Free-space optical communications systems offer high bandwidth, increased security and license-free operation but are highly affected by the performance degradation due to the atmospheric attenuation caused by scattering and absorption. The prediction of attenuation accuracy is even more important in Iraq where the environment is hot, dusty, foggy and rainy in a random fashion. The aim of this study is to assess the performance of machine learning, deep learning and hybrid modeling techniques for the prediction…

The gain
Random Forest predicted atmospheric attenuation for free-space optical links more accurately than deep learning models on a 1500-sample synthetic Iraqi weather dataset, with dust concentration and visibility as top drivers and faster training and inference.
In brief
The result matters because FSO links offer high bandwidth and license-free operation but degrade with scattering and absorption, especially in hot, dusty, foggy and rainy conditions. SHAP analysis identified dust concentration and visibility as dominant predictors, and RF also trained and inferred faster, though the authors note synthetic-data limits mean real atmospheric measurements are needed before operational deployment.
Main points
- Synthetic dataset of 1500 samples built from physical propagation models for five regimes: clear sky, fog, rain, dust storms and snow.
- Fifteen models tested: six machine learning, six deep learning including LSTM and CNN-LSTM, and three hybrid approaches.
- Best deep learning R2 was 0.7766 versus RF 0.9654 and best hybrid 0.9571, with no significant difference between RF and best hybrid at p = 0.083.
The full story
Researchers generated 1500 synthetic samples using established physical propagation models to represent clear sky, fog, rain, dust storms and snow conditions relevant to Iraq.
They compared Random Forest, Extreme Gradient Boosting, LightGBM, SVR, Linear Regression, KNN against MLP, DNN, LSTM, 1D CNN, CNN-LSTM, Attention-based Network and three hybrids, using SHAP to rank features and statistical testing to compare top models.
Sources
1 on file- Journal of Visualized Experimentsdoi:10.3791/73069
ace
What readers say