28 research outputs found
RF_Modelling_PO
No full textValavi et al. (2021) Modelling species presence-only data with random forests. Ecography 44: 1–12, 202
SDM_spatial_validation
No full textThis repository stores the code, data, and Appendix S1 for "Valavi et al. (2023) Flexible species distribution modelling methods perform well on spatially separated testing data" published in the Global Ecology and Biogeography journal
Predictive_perfromance_of_PO_SDMs
No full textThis repository stores data for "Valavi, R., Guillera‐Arroita, G, Lahoz‐Monfort, J.J. & Elith, J. (2021) Predictive performance of presence-only species distribution models: a benchmark study with reproducible code. Ecological Monographs. DOI:10.1002/ecm.1486
‐fold cross‐validation of species distribution models
No full text[EN]When applied to structured data, conventional random cross-validation techniques can lead to underestimation of prediction error, and may result in inappropriate model selection. We present the r package blockCV, a new toolbox for cross-validation of species distribution modelling. Although it has been developed with species distribution modelling in mind, it can be used for any spatial modelling. The package can generate spatially or environmentally separated folds. It includes tools to measure spatial autocorrelation ranges in candidate covariates, providing the user with insights into the spatial structure in these data. It also offers interactive graphical capabilities for creating spatial blocks and exploring data folds. Package blockCV enables modellers to more easily implement a range of evaluation approaches. It will help the modelling community learn more about the impacts of evaluation approaches on our understanding of predictive performance of species distribution models.[AR]هنگامی که روش اعتبارسنجی متقاطع بر روی دادههای دارای ساختار (مکانی٬ محیطی …) اعمال میشود٬ میتواند منجر به تخمین نادرست خطای پیشبینی و در نتیجه اشتباه در انتخاب مدلها شود.در اینجا کتابخانه blockCV در نرمافزار برنامهنویسی R را بعنوان یک ابزار جدید برای اعتبارسنجی متقاطع مدلهای توزیع گونهها ارائه میدهیم. گرچه این کتابخانه با ایدهی مدلسازی توزیع گونهها توسعه داده شده است اما میتواند برای انواع مدلسازیهای مکانی نیز مورد استفاده قرار گیرد.این کتابخانه توانایی ساخت زیرمجموعههای مجزای مکانی و محیطی از دادهها را دارد و همچنین شامل ابزاری برای اندازهگیری دامنه تاثیر خودهمبستگی مکانی در متغیرهای پیشبینیکننده میباشد که به کاربر دیدگاه بهتری از ساختار مکانی این دادهها را میدهد. این کتابخانه دارای ابزارهای گرافیکی برای بررسی و ساخت بلاکهای مکانی نیز میباشد.کتابخانه blockCV مدلسازان را قادر میسازد که دامنه وسیعتری از روشهای اعتبارسنجی را بکار گیرند و همچنین درباره تاثیرات روشهای مختلف اعتبارسنجی بر درک بهتر ما از قدرت پیشبینی مدلهای توزیع گونهها کمک میکند.R.V. is supported by an Australian Government Research Training Program Scholarship and a Rowden White Scholarship; G.G.-A. by an Australian Research Council (ARC) Discovery Early Career Researcher Award (DE160100904), and J.J.L.-M. and J.E. by ARC Discovery Project 160101003. J.E. appreciates the support of the ARC's Centre of Excellence for Environmental Decisions (CE11001000104). We also thank Babak Mirbagheri, Nick Golding, and reviewers and editors: Robert Anderson, David Roberts, an anonymous reviewer and Associate Editor, for their helpful suggestions and advice.Peer reviewe
SDM_NCEAS
No full textThese are data described in this publication: Elith, J., Graham, C.H., Valavi, R., Abegg, M., Bruce, C., Ferrier, S., Ford, A., Guisan, A., Hijmans, R.J., Huettmann, F., Lohmann, L.G., Loiselle, B.A., Moritz, C., Overton, J.McC., Peterson, A.T., Phillips, S., Richardson, K., Williams, S., Wiser, S.K., Wohlgemuth, T., Zimmermann, N.E. (2020). Presence-only and presence-absence data for comparing species distribution modeling methods. Biodiversity Informatics 15:69-8
On the predictive performance of correlative species distribution models
No full text© 2021 Roozbeh ValaviSpecies Distribution Modelling (SDM) is a widely used tool in ecological studies and wildlife conservation. Despite the vast literature on this topic developed to date, species distribution modelling remains an active area of research because there are so many potential uses for the models but also many challenges to fitting them well. Therefore, it is important to have a clear understanding of how these methods work and how they can be best used in each setting. My thesis focuses on advancing methodological aspects of SDM in particular assessing and improving the predictive performance, including providing tools and guidelines for their use.
One key aim of my PhD is to test a broad range of common SDM algorithms across several datasets. As a step towards that goal, in my first research chapter I compared newly developed algorithms and novel implementations of established ones to update a landmark study by Elith and colleagues in 2006, using their data. In summary, I evaluated 13 modelling methods (and several implementations of some) on an independently collected testing dataset. The dataset used includes presence-only records for 226 species across the world with presence-absence data for model evaluation. This dataset is now published, and the manuscript provided as an appendix to my thesis as I contributed to its publication. The result of the first research chapter showed that some models perform generally better than others. An ensemble of five tuned models was the best model in averaged and ranked performance. In contrast, the model implemented by biomod modelling framework with the default parameters was an average performer indicating that ensemble models per se are not the best solution to all modelling problems. Overall non-parametric models with the capability of optimising the bias-variance trade-off performed strongly. This includes boosted regression trees (BRT), MaxEnt and a variant of Random Forest (RF). All the data and code with working examples are provided with the manuscript of this chapter to make this study fully reproducible.
One striking result of my methods comparison study was the remarkable improvement of some machine learning models (e.g., RF) when the imbalanced nature of presence-background data is considered during model fitting. This triggered the idea behind my second research chapter, where I explored why common implementations of RFs fitted to presence-background data appear highly sensitive to how the imbalance between number of presence records and number of background samples is treated. It turned out that class imbalance was not the only source of the problem, but class overlap played a central role (class overlap is when different classes happen in the same environmental range – here classes are presences and background samples). Although the commonly used default implementation of RF in R programming is problematic when modelling presence-background data, I identified several solutions that work well with such data without losing the environmental representativeness of background samples. A novel part of this chapter was its clear demonstration of reasons for differing performance, and evidence that several implementations overfit imbalanced and overlapped data.
To enable new approaches for evaluating SDMs, I developed an R package that facilitates flexible use of block cross-validation with species distribution data. This is important for testing predictive performance at sites spatially or environmentally separated from the training sample. My package covers several common blocking methods and supports different types of species data; it is available freely and I published a manuscript, forming the next chapter, to explain its use. This underpins the last research chapter where I ask whether methods that did well in my first research chapter – many of which fit quite complex models – still perform well when evaluated with spatially separated testing data. They did. In comparing models with spatial vs random partitioning, the order of models did not change. The best model in both random and spatial partitioning was an ensemble model followed by MaxEnt, a variant of RF and BRT. These are models that are capable of fitting complex fitted functions compared to simple and smoother models such as generalised linear and additive models that did not perform as strongly. However, complexity is not in itself the key: some complex models performed poorly, and some smoother models, well. Recently popular methods for tuning MaxEnt did not improve its performance.
Finally, I have provided all the data and code from my thesis to make the methods accessible to other researchers
rvalavi/blockCV: v2.0.1
No full textprint() and cat() are removed from the functions and a verbose argument is added instea
9 arcsecond gridded HCAS 2.4 (2001-2018) base model estimation of habitat condition, 18-year epochs including updated NCI v2.0, annual epochs 2001 to 2021, and trends for continental Australia (with caveats)
No full textCaveat: this dataset extends the annual epochs used in the HCAS v2.1 data collection but may not represent an improvement in the regional accuracy of condition assessments. Please read the accompanying technical report for method details. This dataset has a limited distribution because it is under revision and will be superseded by HCAS v3.0 with anticipated publication in mid-2024.\n\nThis data collection comprises, for continental Australia, the 9-arcsecond gridded Habitat Condition Assessment System (HCAS) version 2.4 (2001-2018) base model (HCAS v2.4) estimation of habitat condition for terrestrial biodiversity, a series of 18-year epochs (including updated NCI v2.0), annual epochs between 2001 and 2021 derived from the base model, and a series of trend analyses derived from the annual epochs. Several other datasets support use and interpretation of the base model, epochs and trends. \n\nThe core collection in the HCAS v2.4 product suite comprises 32 datasets (four 18-year epochs, 21 annual epochs, seven trends), and a larger number of input and supporting datasets to inform use and interpretation, including four corresponding 18-year epochs of the NCI v2.0. The majority are raster datasets in GeoTIFF format (*.tif) at 0.0025° resolution (approx. 250 m grids), Geographic Datum of Australia (GDA) 1994 (EPSG:4283). All GeoTIFFs are also provided as Cloud Optimised GeoTIFFs (COGs) in relevant subfolders labelled ‘COG’. COGs are regular GeoTIFF files with an internal organisation to enable HTTP GET range requests to ask for just the needed parts of a file.\n\nThe HCAS v2.4 base model and epochs datasets are habitat condition indices that vary continuously from a theoretical minimum of 0.0 (ecosystem integrity removed) to a maximum of 1.0 (ecosystem integrity in reference condition). The index represents the contribution that a given site (grid cell) makes to the effective area of ecosystem integrity remaining within any given spatial reporting unit, expressed as a proportion of the contribution made by a site in reference condition.\n\nThis HCAS v2.4 product suite was developed to extend the annual epochs through to 2021, using remote sensing generated from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite. The MODIS satellite is reaching its end of life and future HCAS updates will be developed using Landsat satellite imagery, as well as other satellite inputs. The MODIS satellite will continue to operate, in a declining orbit, through to December 2025. \n\nNote: The remote sensing data originally derived for these analyses included the calendar year 2022. However, investigations into reasons for a downward trend in the continental average of the 2022 condition epoch compared with the 2021 epoch, led to the discovery of a data download error in the remote sensing imagery. Missing tiles across some regions of Australia led to erroneous habitat condition estimates using the 2022 data. This affected the 2022 annual epoch, and the 2005–22 and 2001–2022 long-term epochs). The effect is muted in the continental average of the long-term epochs but obvious in analyses using the annual epoch. Therefore, results using the 2022 data were removed from this data collection. Because the analyses using these data for the EN01 measure in the DCCEEW 2022-23 Annual Report were presented as continental and land zonal averages, the data bias error in 2022 does not alter that overall result.\n\nTechnical report: Williams KJ, Valavi R, Lehmann EA, Van Niel T, Harwood T, Newnham G, Paget M, Donohue R, Giljohann KM, Liu N, Lyon P, Pinner L, Malley C, Deb D and Ferrier S (2023) Habitat Condition Assessment System (HCAS): developing HCAS v2.4 with annual epochs updated to 2021. Publication number EP2023-4902. CSIRO, Canberra, Australia. https://publications.csiro.au/publications/publication/PIcsiro:EP2023-4902. \n\nDetails are summarised in the product brief provided with this data collection. \nLineage: The HCAS v2.4 product suite was developed using a combination of:\na) time series remotely sensed ecosystem characteristic variables, derived from the CSIRO MODIS fractional cover dataset (Guerschman and Hill, 2018) and MODIS persistent and annual vegetation cover (Donohue et al., 2014; Donohue et al., 2009), derived from MODIS Collection 6.0 ;\nb) gridded environmental covariates, including climate from the 9s climatology for continental Australia 1976-2005 (Harwood et al., 2016a), soil and landscape data from the Soil and Landscape Grid of Australia aggregated to 9s (Gallant et al., 2018), used to predict reference ecosystem condition; and \nc) Spatially inferred reference sites sampled to represent Australia’s varied environments, used as training data to build the model of reference ecosystem condition and as benchmarks for calculating proximity to reference condition for all test sites.\n\nSpecifically, the HCAS v2.4 reference ecosystem model was developed using a generalised additive model (GAM) with c.239,000 reference sites, 27 environmental covariates and seven remotely sensed ecosystem characteristic variables summarised over 18 years (2001 to 2018). The base model was developed using the 2001–18 long term epoch for comparability with the original HCAS v2.1 (Harwood et al., 2021). The same seven ecosystem characteristic fractional cover summary variables derived from the MODIS satellite were used. This is the last year in which the MODIS satellite can provide reliable Earth observations, as it reaches its end of life. Future HCAS versions will transition to alternative ongoing Earth observation platforms (e.g. Landsat). \n\nThe HCAS v2.4 differs from the original HCAS v2.1 in the following ways:\n• The number, location and ecological representativeness of the reference sites (including expert nominated inclusions and exclusions)\n• The algorithm used to partition remotely sensed persistent and recurrent green cover fractions (the algorithm used to derive litter and bare fractions remained the same)\n• The number of environmental covariates used to model reference ecosystem patterns\n• Corrected variance standardisation of remotely sensed ecosystem characteristic variables for principal components \n• The use of univariate generalised additive modelling (GAM) to model the remotely sensed ecosystem characteristic principal components and predict the ecosystem reference state \n• The same reference sites used as training data were also used as benchmarks\n• The scaling method applied to the uncalibrated output to obtain the 0-1 score\n• The method of summarising remotely sensed ecosystem characteristic variables for use in annual condition epochs of habitat condition (summary variables for the long term epochs remained the same)\n• Annual epochs were extended by three years to 2021 (21 in total, calendar years) \n• Additional 18-year epochs were generated: 2002–19, 2003–20, 2004–21. \n\nTechnical report: Williams KJ, Valavi R, Lehmann EA, Van Niel T, Harwood T, Newnham G, Paget M, Donohue R, Giljohann KM, Liu N, Lyon P, Pinner L, Malley C, Deb D and Ferrier S (2023) Habitat Condition Assessment System (HCAS): developing HCAS v2.4 with annual epochs updated to 2021. Publication number EP2023-4902. CSIRO, Canberra, Australia. https://publications.csiro.au/publications/publication/PIcsiro:EP2023-4902. \n\nThis technical report provides details about the inputs, processing methods and outputs, with a focus on work leading to development of HCAS v2.4 and comparison with the original HCAS v2.1.\n\nThis data collection also includes an update to the National Connectivity Index version 2.0 (NCI v2) using the HCAS v2.4 as an input.\n\nDetails are summarised in the product brief provided with this data collection
iflint1/roadkill_manuscript: Maximising the informativeness of new records in spatial sampling design
No full text<p>This repository reproduces our results from the Supplementary Materials of our manuscript on Maximising the informativeness of new records in spatial sampling design (2023) published in the Methods in Ecology and Evolution journal.</p>
HCAS 3.1 (1988-2022) base model estimate of habitat condition (90m grid) and National Connectivity Index (NCI) 2.0, annual epochs of HCAS from 1990 to 2022 for continental Australia with uncertainty
No full textCitation: Valavi R, Lehmann EA, Liu N, Levick S, Giljohann KM, Williams KJ, Johnson S, Botha EJ, Munroe SEM, Collings S, Searle R, Van Niel TG, Newnham G, Paget M, Joehnk K, Hosack GR, Harwood TD, Malley C, Gunawardana D, Sivanandam P, Richards AE, Tetreault Campbell S and Ferrier S (2025) HCAS 3.1 (1988-2022) base model estimate of habitat condition (90m grid) and National Connectivity Index (NCI) 2.0, annual epochs of HCAS from 1990 to 2022 for continental Australia with uncertainty. Updated data collection 63571. CSIRO, Canberra, Australia. DOI: https://data.csiro.au/collection/csiro:63571. \n\nThis updated HCAS v3.1 data collection derives in part from Geoscience Australia’s archive of Landsat Earth observation Collection 3 Analysis Ready Data (Commonwealth of Australia, 2021) Derivative Products version 3.1.0 (Geoscience Australia, 2022) and version 3.0.0 (Lymburner, 2022) between 1988 and 2022, and comprises for continental Australia:\n• the 90 m gridded (Australian Albers projection, EPSG 3577, Geographic Datum of Australia 1994) Habitat Condition Assessment System version 3.1 (1988-2022) base model (HCAS v3.1) estimation of habitat condition for terrestrial biodiversity\n• 33 annual short-term epochs of HCAS from 1990 to 2022, formed using 3-year antecedent rolling averages of remotely sensed variables. \n• model-based uncertainty quantification as 95% confidence interval limits for the estimates of habitat condition for the long-term epoch (1988-2022) and selected short-term epochs (Lehmann et al., 2025)\n• a classification of long- and short-term epochs of habitat condition into six modification levels based on experts’ best estimate of condition for each category of the Vegetation Assets, States and Transitions (VAST) narrative framework (Thackway and Lesslie, 2006; 2008)\n• updated National Connectivity Index (NCI) v2.0 (Giljohann et al., 2022) for the long-term epoch using the HCAS v3.1 base model \n• long-term epoch of connectivity-adjusted condition (NCIC) derived from the geometric mean of the corresponding NCI and HCAS outputs (1988-2022)\n• several other datasets to support use and interpretation of the long- and short-term epochs of HCAS habitat condition products and derivatives, including reference sites and remote sensing inputs. \n\nThe HCAS v3.1 short-term epochs derive from the base model. The NCI and NCIC are developed by DCCEEW and published in this HCAS data collection, for the convenience of end users. \n\nThe majority are raster datasets in GeoTIFF format (*.tif) at 90 m grid resolution, Geographic Datum of Australia (GDA) 1994 (Australian Albers, EPSG:3577). All GeoTIFFs are provided as Cloud Optimised GeoTIFFs (COGs). In most cases, a corresponding *.png map image is also provided for quick views. Example summary tables by IBRA bioregions are also provided. \n\nThe HCAS v3.1 base model and epoch habitat condition data varies continuously from a theoretical minimum of 0.0 (ecosystem integrity removed) to a maximum of 1.0 (ecosystem integrity in reference condition). The index represents the contribution that a given site (grid cell) makes to the effective area of ecosystem integrity remaining within any given spatial reporting unit, expressed as a proportion of the contribution made by a site in reference condition. \n\nData extent is defined by any data pixel that intersected a coastline polygon as defined by the land fraction dataset (Liu 2024). The data and no-data extent of each grid layer encompasses the area within the Australian Continental Exclusive Economic Zone (EEZ) (Alcock et al., 2020), excluding territories of Cocos, Christmas, Norfolk, Macquarie, Heard, and McDonald Islands, as well as Antarctica (Liu and Newnham, 2024). \n\nData are described in "Habitat Condition Assessment System (HCAS) version 3.1: A guide to the 90-metre data collection with uncertainty. Technical report EP2025-1549" (published in 2025, available from CSIRO's publication repository, see related links).\nLineage: The HCAS v3.1 product suite was developed at 90 m grid resolution in Australian Albers projection (GDA94) using a combination of:\n• 14 annualised time series of remotely sensed ecosystem characteristic variables, 1988 to 2022 (Levick et al., 2025), sourced from the Digital Earth Australia Surface Reflectance NBART Landsat Analysis Ready Data Collection 3 (Commonwealth of Australia, 2021) derivative products v3.1.0 (Geoscience Australia, 2022) and v3.0.0 (Lymburner, 2022)\n• 58 environmental covariates selected from more than 120 'non-anthropogenic' candidates largely sourced from the Terrestrial Ecosystem Research Network (TERN) compilation (Malone et al., 2025; Searle, 2023) with additional custom variables developed by Liu et al. (2025); \n• spatially inferred reference sites as training data (Giljohann et al., 2025), sampled to represent the most intact remaining examples of Australia’s varied ecosystems and their environments (Valavi et al., 2025a) \n• spatially inferred reference sites as benchmark data (Giljohann et al., 2025), sampled to represent both remotely sensed ecosystem characteristic variables and their environments from among the most intact remaining examples of Australia’s ecosystems (Valavi et al., 2025a).\n\nThe HCAS v3.1 reference ecosystem model (REM) was developed using a generalised additive model (GAM) with 488,619 reference sites as training data, 46 environmental covariates (selected from 58), and 14 remotely sensed ecosystem characteristic variables summarised over 35 years (1988 to 2022). The REM was used in the HCAS v3.1 base model with the 1988 to 2022 long-term epoch and 544,668 benchmark reference sites. Short-term epochs for each of the 14 remotely sensed ecosystem characteristic variables were derived as 3-year antecedent rolling averages (the target year and the two prior years), 1990 to 2022 (33 epochs). These processing steps are detailed in (Valavi et al., 2025a). \n\nModel-based uncertainty estimates were derived from 100 bootstrap runs applied to the sample of training and benchmark reference sites to derive 95% confidence interval limits (upper 97.5th percentile of condition estimate, lower 2.5th percentile of condition estimate, and the resulting 95% confidence interval width) for each 90x90m pixel of the long-term epoch (1988-2022) and nine selected short-term epochs ending: 1990, 1995, \n\nAll data products were derived using a 90 m spatial grid in Australian Albers (GDA94 / Australian Albers). \n\nThe core collection in the HCAS v3.1 product suite comprises 34 datasets of habitat condition (the base model long-term epoch and 33 short-term epochs) with uncertainty quantification, and other input, derived and supporting datasets to inform use and interpretation, including the long-term epoch (1988-2022) of the NCI (v2.0 method) and the derived connectivity-adjusted condition, NCIC. \n\nThe "Habitat Condition Assessment System (HCAS) version 3.1: A guide to the 90-metre data collection with uncertainty" (updated and republished in 2025 as EP2025-1549) documentation, accompanying this data collection (see related links), outlines how the HCAS v3.1 differs from the HCAS v3.0 (Valavi et al. 2024; Williams et al. 2024) and HCAS v2.3 (Harwood et al., 2023; Williams et al., 2023a) data products. The updated product guide describes the uncertainty quantification products and provides further detail about limitations of particular note. \n\nThis data collection also includes an update to the National Connectivity Index version 2.0 (NCI v2) base model using the corresponding HCAS v3.1 inputs. The NCI v2.0 method (Giljohann et al. 2022) has not changed, only the input condition data. \n\nTechnical reports (in preparation) provide details about the inputs, processing methods, outputs and uncertainty quantification applied in developing HCAS version series 3. For the latest publications see: https://research.csiro.au/biodiversity-knowledge/projects/hcas/. Contact us at [email protected] or [email protected]
