A Near Real-Time Method for Forest Change Detection Based on a Structural Time Series Model and the Kalman Filter

Abstract

The increasing availability of dense time series of earth observation data has incited a growing interest in time series analysis for vegetation monitoring and change detection. Vegetation monitoring algorithms need to deal with several time series characteristics such as seasonality, irregular sampling intervals, and signal artefacts. While common algorithms based on deterministic harmonic regression models account for intra-annual seasonality, inter-annual variations of the seasonal pattern related to shifts in vegetation phenology due to different temperature and rainfall are usually not accounted for. We propose a transition to stochastic modelling and present a near real-time change detection method that combines a structural time series model with the Kalman filter. The model continuously adapts to new observations and allows to better separate phenology-related deviations from vegetation anomalies or land cover changes. The method is tested in a forest change detection application aiming at the assessment of damages caused by storm events and insect calamities. Forest changes are detected based on the cumulative sum control chart (CUSUM) which is used to decide if new observations deviate from model-based forecasts. The performance is evaluated in two test sites, one in Malawi (dry tropical forest) and one in Austria (temperate deciduous, coniferous and mixed forests) based on Sentinel-2 time series. Both forest areas are characterized by a distinct, but temporally varying leaf-off season. The presented change detection method shows overall accuracies above 99%, users’ accuracies of 76.8% to 88.6%, and producers’ accuracies of 68.2% to 80.4% for the forest change stratum (minimum mapping unit: 0.1 ha). Results are based on visually interpreted points derived by stratified random sampling. A further analysis revealed that increasing the time series density by merging data from two Sentinel-2 orbits yields better forest change detection accuracies in comparison to using data from one orbit only. The resulting increase in users’ accuracy amounts to 7.6%. The presented method is capable of near real-time processing and could be used for a variety of automated forest monitoring applications.

Estimating VAIA Windstorm Damaged Forest Area in Italy Using Time Series Sentinel-2 Imagery and Continuous Change Detection Algorithms

“…In this study, we showed that forest area damaged by the VAIA windstorm can be mapped with both the algorithms we tested BEAST, and CCDC. The analyses we performed showed that S2 NBR TS and CDC algorithm (i.e., BEAST and CCDC) are able to map forest areas damaged by windthrows confirming the results, in terms of PA, UA, OA, obtained in Malawi and Austria by Puhm et al [37], using S2 data a near real-time CDC that combines structural time series model and Kalman filter.…”

“…Within the vast array of algorithms developed to map land use changes using TS optical satellite imagery, those that are able to detect interannual changes using trajectory analysis appears to be adequate to detect interannual changes of forest area [36,37]. Some of the most promising CD approaches are: the continuous change detection and classification (CCDC) [38]; the exponentially weighted moving average change detection (EWMACD) [39]; or the Bayesian estimator of abrupt change, seasonal change, and trend (BEAST) [40] methods.…”

“…In addition, these algorithms are able to split the time series into three adaptative components (i.e., trend, seasonal, and remainder) [42,44,45] for each investigated year. Many of these studies use dense Landsat or MODIS TS as input data [40,46], while, the use of S2 and CCDC algorithm to map forest areas damaged by windstorms is limited [37]. Many recent studies on VAIA windstorm underlined that S2 is adequate to detect windstorm since S2 program offers innovative features for forest remote sensing by combining high spatial resolution (i.e., 13 bands, from 0.443 to 2.190 µm with the visible (i.e., R, G, B,) and the near infrared bands at 10-m spatial resolution and four red-edge bands at 20-m spatial resolution), wide coverage and a quick revisit time (i.e., every 5 days after the launch of Sentinel-2B satellite in 2017) [15,18,22,23].…”

Phenology Modelling and Forest Disturbance Mapping with Sentinel-2 Time Series in Austria

“…Recent studies stressed the capabilities of an updating S2 time series that predicts forest phenology using a recursive Kalman filter [31]. Unlike these studies, we use a Savitzky-Golay filtering (SGF) approach, as former studies showed the advantages of SGF to smooth out signal noise but retain temporal details.…”

“…The ground resolution is up to 10 m and the revisit time is less than five days. The use of S2 data that is free of charge is quite established in agricultural monitoring [28,29] and non-forest phenology modelling [30], whereas advanced S2-based TSAs focusing on forest ecosystems are so far rare [31].…”