A comprehensive evaluation of change detection methods on PlanetScope constellation data

Abstract

Recent advances in satellite imaging have seen the thriving of PlanetScope data. A leading feature of this data is the high temporal frequency at which imagery is captured, although at a low to medium spatial resolution. Additionally, these images are often accompanied by cloud masks and other auxiliary information, further enhancing their usability. The simplistic access provided by PlanetScope’s online portal emphasizes its popularity. Such advancements naturally lead to inquiries about the efficiency of PlanetScope data in change detection analyses. The principal aim of this investigation is to evaluate the suitability and efcacy of PlanetScope imagery when applied to established change detection models, with the ultimate goal of determining the utility of PlanetScope data for the purpose of change detection. For evaluation purposes, the chosen models were divided into two categories: those based on traditional change detection methodologies, and those leveraging advanced deep learning algorithms. This research utilized three previously developed models, applying them to the PlanetScope dataset to ascertain its adaptability and effectiveness. These models were subsequently juxtaposed for comparative analysis. The initial model was infuenced by [Kondmann et al., 2021], which developed a half-sibling regression technique founded on the principles proposed by [Schölkopf et al., 2016]. This model was subsequently termed the Sibling Regression for Optical Change Detection (SiROC). In contrast, the latter two models are grounded in deep learning models. Both models incorporate the Siamese network architecture; one utilizes it in conjunction with a fully convolutional network as outlined by [Daudt et al., 2018a], while the other pairs it with the Transformer architecture, drawing inspiration from [Bandara and Patel, 2022]. Upon analyzing the results derived from the simulation PlanetScope dataset, it was perceived that certain models experienced minor improvements, whereas others demonstrated signifcant advancements, based on the evaluation metrics employed. Specifcally, the Transformer-based model excelled, achieving performance metrics exceeding 90%. Conversely, the SiROC model ranked lowest on three out of the four predetermined metrics. Notably, the fully convolutional approach shows noticeable variations when juxtaposed with the results from its original dataset.