OLS, RF and SVM Regressor Models for Temporal Change Detection: A SAR Data-Based Exploratory Study

Abstract

Time series-based change detection is still a challenging topic due to the lack of availability of dense time series. Especially in publicly available satellite remote sensing data, dense time series are often not available. In this study, we use machine learning techniques to assess the feasibility of using these models to expand a time series in the future to measure the change in backscatter due to deforestation in the Khandwa region. The Ordinary Least Squares, Random forest and Support vector machine based regressor were used to predict and extend time series Sentinel-1 SAR backscatter values spanning from 10th January 2019 to 30th May 2022. The predicted values were compared with the actual backscatter value to assess the deviation between the expected and actual values, which was substantial due to the reduced backscatter caused by deforestation at the end of 2022 and early 2023 in the region. The results show a severe overfitting in the case of the random forest regressor, The Ordinary least square R2=-1.09 was able to capture the seasonal variations well but failed on baseline accuracy. The support vector machine R2=0.24, meanwhile, had the better overall fit, but still, the errors were not small. These initial results indicate that using machine learning models to generate simulated time series data is still a challenge. Especially, the complexity of SAR data proves to be challenging during modelling, and more work on such change detection methods is needed. © 2026 IEEE.