Atmospheric correction of Sentinal-2 satellite Data using Deep learning

Remote sensing relies heavily on Atmospheric Correction (AC) to ensure accurate estimation of land Surface Reflectance (SR) for various applications.Conventional AC methods, while effective, are computationally expensive and require extensive atmospheric parameters that can be challenging to estimate accurately.This research proposes a novel deep learning model for AC that eliminates the need for explicit atmospheric parameter estimation.Our approach utilizes a Pix2Pix architecture trained on a diverse dataset of Sentinel-2 images covering all states in India, collected via Google Earth Engine.The model includes four bands (red, green, blue, and visible near-infrared) and directly predicts SR values from Top-of-Atmosphere (TOA) reflectance.The model demonstrated promising results, accurately estimating SR values across various scenarios.Evaluation metrics showed significant improvements, with mean Structural Similarity Index (SSIM) increasing from -0.0025 to 0.961 and mean Peak Signal-to-Noise Ratio (PSNR) rising from 11.0188 dB to 42.14 dB post-training.This approach not only simplifies the AC process but also achieves comparable or superior performance to traditional physicsbased methods.The experimental findings underscore the potential of deep learning as a robust and efficient alternative for atmospheric correction in remote sensing applications, offering possibilities for faster processing of large satellite image datasets.This study contributes to the application of artificial intelligence in remote sensing, paving the way for more accessible and efficient atmospheric correction methods.Future work could explore the model's adaptability to other sensors, incorporation of temporal data, and integration with traditional physics-based models.