Search

Abstract Understanding forest dynamics at the landscape scale is critical given the demands of sustainable forest management and climate change mitigation. This study proposes an approach for holistically characterising and analysing forest dynamics across large areas and long-time periods using information derived from Landsat time-series. To achieve this, we first developed a two-phase classification process to predictively map (1) disturbance and recovery levels and (2) disturbance causal agents for multiple detected disturbance events. The model explanatory data included a range of trajectory-based metrics derived from Landsat time-series, while model training and validation data were derived from a human-interpreted reference dataset. While previous studies have often described forest dynamics using some specific spectral change metrics, we demonstrated an ensemble approach to map disturbance and recovery trends (by treating them as a single metric) and to characterise not only abruptly occurring change events (e.g., clear-fell logging and wildfire) but also events of low severity (e.g., prescribed burning and selective logging). In addition, we adopted a space-time data cube concept to simultaneously report both newly detected disturbance events (detected disturbances) as well as events that have previously occurred but are ongoing (ongoing disturbances). This ongoing element of forest dynamics is often under-reported. The method was tested over 3.7 million ha of public land sclerophyll forests, where multiple disturbance events have occurred over the last 30 years (1987–2016). Our models of disturbance and recovery levels obtained overall accuracies of 78.6% and 72.3% for primary and secondary disturbance events, respectively. The overall accuracies for the models of disturbance causal agents were 80.7% and 73.0%, respectively. The data cube reported an average annual disturbance rate of 4.2% per year. This was dominated by newly detected disturbance (2.7% per year) as distinct from ongoing disturbance that was, however, considerable (1.5% per year). Our approach presented herein can improve the understanding of forest dynamics over long time periods and large areas and has potential for supporting land managers. Graphical abstract Unlabelled Image Highlights • Forest change analysis over large areas and long-time series using Landsat data • Mapping forest disturbance and recovery levels and associated casual agents • Improving reporting of forest dynamics by adopting a data cube concept [ABSTRACT FROM AUTHOR]

Satellite earth observation is being increasingly used to monitor forests across the world. Freely available Landsat data stretching back four decades, coupled with advances in computer processing capabilities, has enabled new time-series techniques for analyzing forest change. Typically, these methods track individual pixel values over time, through the use of various spectral indices. This study examines the utility of eight spectral indices for characterizing fire disturbance and recovery in sclerophyll forests, in order to determine their relative merits in the context of Landsat time-series. Although existing research into Landsat indices is comprehensive, this study presents a new approach, by comparing the distributions of pre and post-fire pixels using Glass's delta, for evaluating indices without the need of detailed field information. Our results show that in the sclerophyll forests of southeast Australia, common indices, such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Burn Ratio (NBR), both accurately capture wildfire disturbance in a pixel-based time-series approach, especially if images from soon after the disturbance are available. However, for tracking forest regrowth and recovery, indices, such as NDVI, which typically capture chlorophyll concentration or canopy 'greenness', are not as reliable, with values returning to pre-fire levels in 3-5 years. In comparison, indices that are more sensitive to forest moisture and structure, such as NBR, indicate much longer (8-10 years) recovery timeframes. This finding is consistent with studies that were conducted in other forest types. We also demonstrate that additional information regarding forest condition, particularly in relation to recovery, can be extracted from less well known indices, such as NBR2, as well as textural indices incorporating spatial variance. With Landsat time-series gaining in popularity in recent years, it is critical to understand the advantages and limitations of the various indices that these methods rely on. [ABSTRACT FROM AUTHOR]

• Fire disturbance and forest recovery trends were explored over 2 billion ha of forest. • A data driven approach is proposed, using both MODIS and Landsat satellites. • Widespread changes in the last 18 years were not found, however regional trends were observed. • Results can support national and international monitoring and reporting requirements. The regular and consistent measurements provided by Earth observation satellites can support the monitoring and reporting of forest indicators. Although substantial scientific literature espouses the capabilities of satellites in this area, the techniques are under-utilised in national reporting, where there is a preference for aggregating ad hoc data. In this paper, we posit that satellite information, while perhaps of low accuracy at single time steps or across small areas, can produce trends and patterns which are, in fact, more meaningful at regional and national scales. This is primarily due to data consistency over time and space. To investigate this, we use MODIS and Landsat data to explore trends associated with fire disturbance and recovery across boreal and temperate forests worldwide. Our results found that 181 million ha (9 %) of the study area (2 billion ha of forests) was burned between 2001 and 2018, as detected by MODIS satellites. World Wildlife Fund biomes were used for a detailed analysis across several countries. A significant increasing trend in area burned was observed in Mediterranean forests in Chile (8.9 % yr−1), while a significant decreasing trend was found in temperate mixed forests in China (-2.2 % yr−1). To explore trends and patterns in fire severity and forest recovery, we used Google Earth Engine to efficiently sample thousands of Landsat images from 1991 onwards. Fire severity, as measured by the change in the normalized burn ratio (NBR), was found to be generally stable over time; however, a slight increasing trend was observed in the Russian taiga. Our analysis of spectral recovery following wildfire indicated that it was largely dependent on location, with some biomes (particularly in the USA) showing signs that spectral recovery rates have shortened over time. This study demonstrates how satellite data and cloud-computing can be harnessed to establish baselines and reveal trends and patterns, and improve monitoring and reporting of forest indicators at national and global scales. [ABSTRACT FROM AUTHOR]

The prediction of forest biomass at the landscape scale can be achieved by integrating data from field plots with satellite imagery, in particular data from the Landsat archive, using k-nearest neighbour (kNN) imputation models. While studies have demonstrated different kNN imputation approaches for estimating forest biomass from remote sensing data and forest inventory plots, there is no general agreement on which approach is most appropriate for biomass estimation across large areas. In this study, we compared several imputation approaches for estimating forest biomass using Landsat time-series and inventory plot data. We evaluated 18 kNN models to impute three aboveground biomass (AGB) variables (total AGB, AGB of live trees and AGB of dead trees). These models were developed using different distance techniques (Random Forest or RF, Gradient Nearest Neighbour or GNN, and Most Similar Neighbour or MSN) and different combinations of response variables (model scenarios). Direct biomass imputation models were trained according to the biomass variables while indirect biomass imputation models were trained according to combinations of forest structure variables (e.g., basal area, stem density and stem volume of live and dead-standing trees). We also assessed the ability of our imputation method to spatially predict biomass variables across large areas in relation to a forest disturbance history over a 30-year period (1987–2016). Our results show that RF consistently outperformed MSN and GNN distance techniques across different model scenarios and biomass variables. The lowest error rates were achieved by RF-based models with generalized root mean squared difference (gRMSD, RMSE divided by the standard deviation of the observed values) ranging from 0.74 to 1.24. Whereas gRMSD associated with MSN-based and GNN-based models ranged from 0.92 to 1.36 and from 1.04 to 1.42, respectively. The indirect imputation method generally achieved better biomass predictions than the direct imputation method. In particular, the kNN model trained with the combination of basal area and stem density variables was the most robust for estimating forest biomass. This model reported a gRMSD of 0.89, 0.95 and 1.08 for total AGB, AGB of live trees and AGB of dead trees, respectively. In addition, spatial predictions of biomass showed relatively consistent trends with disturbance severity and time since disturbance across the time-series. As the kNN imputation method is increasingly being used by land managers and researchers to map forest biomass, this work helps those using these methods ensure their modelling and mapping practices are optimized. [ABSTRACT FROM AUTHOR]

Abstract Time series analysis of Landsat data is widely used for assessing forest change at the large-area scale. Various change detection algorithms have been proposed, each employing different techniques to characterise abrupt disturbance events and longer term trends. However, results can vary significantly, depending on the algorithm, parameters and the spectral index (or indices) chosen. This mismatch in results has led to researchers hypothesizing that an ensemble based approach may increase accuracy. In this study we assess two change detection algorithms (LandTrendr and the R package strucchange), each with three indices (the Normalized Difference Vegetation Index or NDVI, the Normalized Burn Ratio or NBR, and Tasseled Cap Wetness or TCW). We test their ability to detect abrupt disturbances in sclerophyll forests over a 29 year time period, and subsequently evaluate a number of ensembles, using simple fusion rules and Random Forests models. A total of 4087 manually interpreted reference pixels, sampled from 9 million ha of forest, were used for training and validation. In addition, we assess the effects of priming the Random Forests classifier with confusing cases (commission errors from the time series algorithms). Our results clearly show that ensembles combining multiple change detection techniques out-perform any one method. Our most accurate Random Forests model, using an ensemble of all 6 algorithm outputs, along with 3 bi-temporal change rasters (change in NBR, NDVI and TCW), had an overall error rate of 7%, compared with the most accurate single algorithm/index approach (LandTrendr with NBR), which had an overall error of 21%. Our findings also indicate that acceptable results (14% error) can be achieved without the use of traditional change detection algorithms, by using robust reference data and Random Forests classification. However, by priming the classifier with confusing cases informed by the change detection algorithms, commission errors decreased substantially, at the expense of slight increases in omission errors. In fact, a Random Forests ensemble, using the primed training data and only 3 bi-temporal change rasters, was more accurate than any one individual algorithm, with an overall error of 11%. By including some additional metrics derived from Landsat time series (e.g. 2 year changes and overall means) this error was further reduced to 8%. Given that most change detection algorithms have large processing requirements, this suggests that algorithms can be applied to a sample of pixels only, for the sole purpose of training a machine learning classifier. We demonstrate the feasibility of this previously unexplored approach, by creating annual disturbance maps over a large area of forest (9 million ha) and long time period (29 years). Highlights • Ensembles out-perform individual algorithms in detecting forest disturbances • Change detection algorithms can be run on a subset of pixels to feed an ensemble. • High quality reference data is key to producing accurate disturbance maps. • Priming a Random Forests model with confusing cases improves performance. [ABSTRACT FROM AUTHOR]