Andrew Haywood

The Performance of Random Forests in an Operational Setting for Large Area Sclerophyll Forest Classification

Mapping and monitoring forest extent is a common requirement of regional forest inventories and public land natural resource management, including in Australia. The state of Victoria, Australia, has approximately 7.2 million hectares of mostly forested public land, comprising ecosystems that present a diverse range of forest structures, composition and condition. In this paper, we evaluate the performance of the Random Forest (RF) classifier, an ensemble learning algorithm that has recently shown promise using multi-spectral satellite sensor imagery for large area feature classification. The RF algorithm was applied using selected Landsat Thematic Mapper (TM) imagery metrics and auxiliary terrain and climatic variables, while the reference data was manually extracted from systematically distributed plots of sample aerial photography and used for training (75%) and accuracy (25%) assessment. The RF algorithm yielded an overall accuracy of 96% and a Kappa statistic of 0.91 (confidence interval (CI) 0.909–0.919) for the forest/non-forest classification model, given a Kappa maximised binary threshold value of 0.5. The area under the receiver operating characteristic plot produced a score of 0.91, also indicating high model performance. The framework described in this study contributes to the operational deployment of a robust, but affordable, program, able to collate and process large volumes of multi-sourced data using open-source software for the production of consistent and accurate forest cover maps across the full spectrum of Victorian sclerophyll forest types.

Understanding the Effects of ALS Pulse Density for Metric Retrieval across Diverse Forest Types

Pulse density, the number of laser pulses that intercept a surface per unit area, is a key consideration when acquiring an Airborne Laser Scanning (ALS) dataset. This study compares area-based vegetation structure metrics derived from multireturn ALS simulated at six pulse densities (0.05 to 4 pl m-2) across a range of forest types: from savannah woodlands to dense rainforests. Results suggest that accurate measurement of structure metrics (canopy height, canopy cover, and vertical canopy structure) can be achieved with a pulse density of 0.5 pl m-2 across all forest types when compared to a dataset of 10 pl m-2. For pulse densities <0.5 pl m-2, two main sources of error lead to inaccuracies in estimation: the poor identification of the ground surface and sparse vegetation cover leading to under sampling of the canopy profile. This analysis provides useful information for land managers determining capture specifications for large-area ALS acquisitions.

Semi-automating the Stand Delineation Process in Mapping Natural Eucalypt Forests

Summary For decades, aerial photo interpretation has been, and to a good extent is still, the method of choice for producing fine-scale native forest stand mapping. Recent computer techniques have eased the task of the interpreter, who is now able to delineate polygons through on-screen digitising in a geographical information system (GIS) environment. Even with these advances, a great deal of skill is required in the polygon delineation. In an effort to contribute to the automation of this process, we introduce an open-source object-based solution to the mapping of forest stand boundaries using attributes derived from digital aerial photography and laser scanning data acquired over a study area in the Victorian Central Highlands. This methodology transforms remotely sensed imagery (single or multichannel) and canopy raster layers derived from laser scanning (lidar) into polygon vector layers. It is intended that the resultant polygon layer should resemble the product derived by an aerial interpreter, without any prior knowledge of the scene. The derived product aims to produce a layer comprised of relatively homogeneous polygons all exceeding a minimum size. The derived product is meant to be a preliminary template aimed at reducing time and effort in manual digitisation. The relationship between spectral, texture and laser scanning derived features for forest stand boundary delineation and human interpreted boundaries is not straight forward. The interpreter however, can aggregate and sometimes correct the automated delineated regions by simple drag-and-click operations This approach is relatively cheap and flexible, being a workable compromise between fully automated image interpretation which requires further research for acceptable levels of accuracy and reliability, and manual segmentation and classification. Preliminary results are encouraging, both in regard to automating the process and the delivery of robust delineation of stand boundaries in native forest landscapes. Future research will focus on appropriate input resolution to reduce computation requirements and improved data fusion methods to obtain more accurate forest stand delineation.

Using discrete-return airborne laser scanning to quantify number of canopy strata across diverse forest types

The vertical arrangement of forest canopies is a key descriptor of canopy structure, a driver of ecosystem function and indicative of forest successional stage. Yet techniques to attribute for canopy vertical structure across large and potentially heterogeneously forested areas remain elusive. This study introduces a new technique to estimate the Number of Strata (NoS) that comprise a canopy profile, using discrete-return Airborne Laser Scanning (ALS) data. Vertically resolved gap probability (P-gap) aggregated over a plot is generalized with a nonparametric cubic spline regression (P-s). Subsequently a count of the positive zero-crossings of second derivative of 1 - P-s is used to estimate NoS. Comparison with inventory derived estimates at 24 plots across three diverse study areas shows a good agreement between the two techniques (RMSE=041 strata). Furthermore, this is achieved without altering model parameters, indicating the transferability of the technique across diverse forest types. NoS values ranged from 0 to 4 at a further 239 plots, emphasizing the need for a method to quantify canopy vertical structure across forested landscapes. Comparison of NoS with other commonly derived ALS descriptors of canopy structure (canopy height, canopy cover and return height coefficient of determination) returned only a moderate correlation (r(2)<04). It is proposed the presented method provides a primary descriptor of canopy structure to complement canopy height and cover, as well as a candidate Ecological Biodiversity Variable for characterizing habitat structure.

Using Airborne Laser Scanning Data to Estimate Structural Attributes of Natural Eucalypt Regrowth Forests

Summary Airborne laser scanning (lidar) data provide the means to measure the vertical and horizontal structure of forest vegetation. The aim of this study was to investigate how metrics derived from laser scanning data could be used in simple regression models to estimate eucalypt top height, basal area and stems per hectare on 20 m x 20 m field plots. The study area was located in the Central Forest Management Area in Victoria. The target population was younger regrowth forests aged 20–60 y dominated by mountain ash (Eucalyptus regnans) and other ash species. A linear regression function was able to provide precise estimates of eucalypt top height (r 2 = 0.87; root mean square error (RMSE) = 3.9 m) using a single height percentile variable. On the strength of this result it should be possible to predict top height with a precision that is close to traditional field measurement methods. Regression estimates of eucalypt stand basal area were less precise (r 2 = 0.56; RMSE = 14.7 m2) than those of the top height model. The model included both a height percentile and intensity variable. Regression modelling was able to provide an estimate of (eucalypt stems per hectare)−2 using height percentiles, laser intensity and canopy structure as predictor variables (r 2 = 0.41; RMSE = 5.6).

Using ensemble margin to explore issues of training data imbalance and mislabeling on large area land cover classification

This work introduces new ensemble margin criteria, to evaluate the performance of Random Forests (RF), in the context of large area land cover classification, using imbalanced and noisy training data. Experiments using binary and multiclass classification problems reveal insights into the behaviour of RF over big data, in which training data contains noise and may not be evenly distributed among classes. The margin-based RF performance evaluation is conducted using remote sensing and ancillary spatial data, across a 7.2 million hectare study area.

Using Landsat Spectral Indices in Time-Series to Assess Wildfire Disturbance and Recovery

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.

The impact of sensor characteristics for obtaining accurate ground-based measurements of LAI

Calibration and validation of LAI products require accurate ground-based measurements. Many indirect ground-based sensors such as digital hemispherical photography (DHP), ceptometers, and terrestrial laser scanners (TLS) are used interchangeably to estimate reference values. However these sensors have biases in regards to the true LAI value, which can never be known in the field. Results from three representative woody ecosystems in Eastern Australia are presented from real field measurements. Significant differences were found between methods at the individual measurement and plot scale. Furthermore, one of the sites in South East Australia was measured and modeled in a 3D deterministic model. In this digital environment where the truth is known, sensors can be simulated to determine their bias.