Surface reflectance retrieval of remote sensing images

Abstract

Atmospheric correction (AC) is a critical preprocessing step in the analysis of satellite imagery, aimed at correcting the effects of the atmosphere on reflectance values recorded by satellite sensors. The atmosphere scatters and absorbs radia tion, which can distort the true reflectance values of Earth’s surface. This correc tion process involves modifying the recorded top of atmosphere (TOA) radiances to accurately reflect surface reflectance (SR), also known as bottom of atmosphere (BOA) reflectance, thereby providing more precise and reliable remote sensing data for various applications such as land cover classification, climate monitoring, and environmental management. Traditional AC methods rely on physics-based models, such as the 6S model, which are computationally intensive, complex, in f lexible to generalize over various land covers and atmospheric conditions and require extensive prior knowledge of atmospheric conditions. In contrast, deep learning (DL) models offer a data-driven alternative that is comparatively less compute intensive, less complex, easily generalizable and can potentially capture complex, nonlinear relationships between TOA radiances and SR without explic itly relying on atmospheric parameters. This thesis presents a novel approach to ACofsatellite images using DL techniques, specifically focusing on SR retrieval. This research begins with developing an encoder-decoder DL architecture to emulate the 6S model. This model can accurately generate SR images, effec tively capturing the intricate transformations performed by traditional physics based methods. The performance of this DL approach has been rigorously vali dated, demonstrating its potential for further extension and application. Building on this foundation, the thesis explores incorporating seasonal infor mation into AC. AC is influenced by the aerosol optical depth and column water vapour, which change with the seasons. Hence, it becomes imperative for the DL model to capture seasonal variations occurring in the atmosphere. A season aware atmospheric correction model (SAAC-Net) is introduced, utilizing convo lutional neural networks (CNNs) with residual connections. SAAC-Net is an end to-end deep neural network trained on seasonally and spatially diverse Landsat 8 satellite images. This model leverages synthetic seasonal bands to account for seasonal variations, enhancing the generalisation capability across different land cover types. The efficiency of this model is demonstrated through extensive ex periments conducted over the Indian subcontinent, and a model sensitivity study has also been performed. The model’s effectiveness is validated against Landsat 8’s Land Surface Reflectance Code (LaSRC), achieving a root mean square error (RMSE) of approximately 0.042, a structural similarity index measure (SSIM) of around 0.97, and a correlation coefficient (r) of approximately 0.99. For ground measurements from RadCalNet, the proposed model shows an RMSE of about 0.053, an SSIM of around 0.90, and r of approximately 0.88. Further advancing the field, the research addresses the challenges of tempo ral data shifts due to atmospheric changes. This study presents empirical evi dence of dataset shift presence by analyzing the joint distribution of TOA and BOAreflectances using Landsat 8 images from 2013 and 2020. We employ three approaches: the Kolmogorov-Smirnov (KS) test, Wasserstein distance calcula tion, and evaluation using a trained deep learning model. These methods quantify dataset shifts from both distributional and numerical perspectives. Aspatio-temporal atmospheric correction model (STAC) is developed to inte grate spatial, spectral, and temporal aspects, providing a comprehensive solution for dynamic atmospheric conditions. The results are compared with the SAAC Net and Landsat 8 LaSRC AC model. Experimental results demonstrate that the STAC model outperforms SAAC-Net, with a 30% improvement in the RMSE across various spectral bands. This study also validates the model’s generalization ability by evaluating its performance on diverse land covers (LCs) not included in the training data. The average RMSE of the STAC model across all LCs is 0.036, significantly lower than the 0.044 RMSE of the state-of-the-art DL model, SAAC Net. Additional experiments using the RadCalNet ground-measured dataset show that the mean relative difference (MRD) for the STAC model across all bands is 0.027, substantially lower than the MRD of SAAC-Net (0.083) and Landsat LaSRC (0.073). The culmination of these efforts provides a robust framework for SR retrieval using DL, underscoring the potential of data-driven methods to enhance remote sensing applications. This thesis makes significant advancements in AC by offer ing a DL based alternative to traditional, computationally intensive physics-based models. Overall, this work lays the groundwork for future research in extending deep learning to remote sensing, aiming to find computationally efficient alternatives to physics-based solutions. It emphasizes the transformative potential of artificial intelligence in improving the accuracy and efficiency of remote sensing applica tions