Abstract

There are major factors which complicate the monitoring of air pollution in Delhi. Optical sensors are down in monsoons, and static warning levels cannot tell apart the daily winter thermal inversion from the real pollution episode. We propose a solution to these problems within the framework of prediction and anomaly detection. In the prediction phase, we first computed three spectral indices from Sentinel-2 surface reflectance bands and applied Daubechies db4 wavelet filters to these indices via a parallel deep branch alongside a ResNet-50 backbone. This branch separates each index into four sub-bands, producing a 64-dimensional wavelet feature vector that is concatenated with CNN spatial features and evaluated across three model engines: Random Forest, XGBoost, and CNN-LSTM. Wavelet decomposition was essential for improving model accuracy by separating the satellite image into distinct high frequency and low frequency sub-bands to isolate atmospheric haze from sharp urban structural details, with the CNN-LSTM achieving the largest gain (Test R2 rising from 0.6821 to 0.7265, RMSE from 49.25 to 45.68 µg/m3). The residuals from predicted PM2.5 were obtained and PCA components were analysed. The resultant p-dimensional vector is used in multivariate EWMA recursion for finding the anomalies.