Monitoring changes in inland waters is crucial due to their links with ecosystems, human populations, and the economy. Satellite missions have long supported this monitoring, and the launch of the Surface Water and Ocean Topography (SWOT) satellite in 2022 has provided unprecedented observations of inland water bodies. However, early SWOT data, specifically the Pixel Cloud (PIXC) product, show limitations in classifying surface types. This study applies Machine Learning (ML) methods, namely K-nearest Neighbor (KNN), Random Forest (RF), and eXtreme Gradient Boosting (XG Boost), to improve surface classification in PIXC data. We trained classification models using features from SWOT and one feature from Landsat imagery: the water occurrence value. Our method was tested on eight reservoirs in Iran: 15-Khordad, Dez, Karun4, Zayandehrud, Agh Chai, Doosti, Shahid Rajaei, and Garan-across available SWOT overpasses from 2023 to 2024. Reservoir surface areas derived from our ML-based classifications were validated against in situ data and compared with the original PIXC and the LakeSP product. Results show consistent improvements across all ML methods. Compared to the original PIXC classification, correlation increased by 79.2\% (KNN), 78.7\% (RF, XG Boost), while NRMSE improved by 24.6\%-27.8\%. KGE scores improved by over 130\% for all methods. Compared to LakeSP, our method reduced NRMSE by 37.8\% (XG Boost), increased correlation up to 123\%, and improved KGE scores by over 175\%. No single ML method consistently outperformed the others, but all showed substantial improvements, demonstrating the potential of ML to enhance SWOT-based inland water monitoring.

In the 30 years of its availability, satellite altimetry has established itself as an important tool for understanding the Earth system. Originally developed for oceanography and geodesy, it has also proven valuable for monitoring water level variation of lakes and rivers. However, when using altimetry for inland waters, there is always a critical issue: retracking i.e. the procedure in which the range from the satellite to the water surface is (re)estimated. The current retracking methods heavily rely on single waveforms, which results in a high sensitivity to every individual peak in the waveform and in a strong dependency on the waveform's shape. Here, we propose the Bin-Space-Time (BiST) retracking method that moves beyond finding a single point in a 1D waveform and instead seeks a retracking line within a 2D radargram, for which the temporal information over different cycles is also considered. The retracking line divides the radargram into two segments: the left (Front) and right-hand side (Back) of the retracking line. Such a segmentation approach can be interpreted as a binary image segmentation problem, for which spatiotemporal information can be incorporated. We follow a Bayesian approach, exploiting a probabilistic graphical model known as a Markov Random Field (MRF). There, the problem is arranged as a Maximum A Posteriori estimation of an MRF (MAP-MRF), which means finding a retracking line that maximizes a posterior probability density or minimizes a posterior energy function. Our posterior energy function is obtained by a prior energy function and a likelihood energy function, both of them depending on signal intensity and bin: 1) The prior: the bin-space energy function defined between firstorder neighbouring pixels of a radargram modeling the spatial dependency between their labels for given intensities and bins and 2) The likelihood: the temporal energy function of a pixel for labeling Front or Back given its overall temporal evolution. The realization of the field with the minimum sum of the bin-space and the temporal energy functions is then found through the maxflow algorithm. Consequently, the retracking line, the boundary between the Back and Front region is obtained. We apply our method to both pulse-limited and SAR altimetry data over nine lakes and reservoirs in the USA with different sizes and different altimetry characteristics. The resulting water level time series are validated against in situ data. Across the selected case studies, on average, the BiST retracker improves the RMSE by approximately 0.5 m compared to the best existing retracker. The main benefit of the proposed retracker, which operates in bin, space, and time domains, is its robustness against unexpected waveform variations, making it suitable for diverse inland water surfaces.