Forest disturbance characterization in the era of earth observation big data: A mapping review

Background and objective: Forests play a crucial role throughout the world as highly productive ecosystems, serving approximately a quarter of the human population by providing valuable services. However, these ecosystems are subject to various natural and human-induced disturbances such as insect outbreaks, fires, windthrow, snow damage, selective logging, and harvest, which significantly influence the composition and structure of forests. Recent studies have observed a notable increase in both natural and anthropogenic disturbances on a global scale. The availability of free Earth observation (EO) archives alongside the maturation of algorithms for analysing time series data have opened new opportunities to detect and understand forest disturbances in a vast spatio-temporal context. Over the past few decades, numerous EO-based approaches have been proposed to monitor forest ecosystems, benefiting from the open -data policies of multiple satellite constellations. This study conducted a mapping review to shed light on the state-of-the-art in forest disturbance characterization using EO big data. Study design: We searched major online databases, including Scopus, the Web of Science, MDPI, Science Direct and IEEE Xplore, extensively in order to identify scientific literature published from 1995 to 2023. Of the more than 2000 records screened, 104 publications met the specific inclusion criteria and were included in the review. The selected studies were categorized based on the type of spatial and spectral -temporal patterns used to characterize forest disturbances, the predictors employed to classify the target disturbances, the data fusion methods applied when using multisource EO data, the classification algorithms employed and the accessibility of the reference data used in the studies. Results: Our findings reveal that temporal patterns derived from spectral reflectance have been used three times more frequently than spatial patterns. The most common predictors were the spectral change magnitude and timing of disturbances derived from the short-wave infrared (SWIR) spectral region, along with measures of disturbed patch dimensions and elevation. The random forest emerged as the most used algorithm for classification. Furthermore, it was observed that less than 10% of the studies provided or made available the datasets used for their analyses. Conclusion: Although significant progress has been made in characterizing forest disturbances, the validity and credibility of results were found to be highly variable and much dependent on the choice of change detection methods, predictors and classification algorithms. The scarcity of research studies that provide access to their data poses a significant obstacle to the advancement of current methodologies and the development of novel approaches.