-
Notifications
You must be signed in to change notification settings - Fork 1
/
modelling_info.Rmd
82 lines (48 loc) · 7.21 KB
/
modelling_info.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
---
title: "Modelling sugar cane pest species distributions"
author: "Alex Slavenko"
date: "27-05-2024"
output:
html_document:
theme: "flatly"
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = FALSE, warning = FALSE, message = FALSE)
library(tidyverse)
library(sf)
sp_all <- read_csv("database/sugarcane_pests_database_2024-05-30.csv") %>% count(species)
```
### Species data
#### Occurrence data
Species data were extracted from the Global Biodiversity Information Facility (**GBIF**; [gbif.org](https://www.gbif.org)) database. The records were checked against the known distribution of each pest in the Centre for Agriculture and Bioscience International (**CABI**; [cabi.org](https://www.cabi.org)) database.
Seven priority sugar cane pest species had occurrence data available in GBIF: *Chilo auricilia*, *Chilo infuscatellus*, *Eumetopina flavipes*, *Perkinsiella saccharicida*, *Scirpophaga excerptalis*, *Sesamia grisescens*, and *Yamatotettix flavovittatus*. However, the actual number of unique records varied greatly between species:
```{r echo = FALSE, message = FALSE, warning = FALSE}
library(knitr)
kable(sp_all)
```
Due to data limitations, some model outputs my be unreliable and should be interpreted with caution, particularly for species with low sample sizes.
#### Background sample
To model species presence-only data with statistical models a random sample of background data needs to be extracted from the landscape of interest. However, if sampling bias occurs in the presence-only data, it can lead to biased model results (Phillips *et al.* 2009). We accounted for this by generating a bias layer for each species and sampling background points weighted by the bias layer, so that background points are sampled with the same spatial density as presence data.
For each species we created a spatial kernel density estimate around its occurrence records, projected onto a raster masked to only include countries in which the species is known to occur. We then generated 1000 background samples for each species by randomly sampling from the masked raster, with the sampling probability for each cell weighted by the spatial kernel density estimate.
### Species distribution modelling
<!-- We used an ensemble approach for species distribution modelling (Araujo & New 2007) - we fit five different models (Generalised Additive Model [GAM], Gradient Boosting Machine [GBM], Generalised Linear Modle [GLM], Maximum Entropy [MaxEnt] and Random Forests [RF]) and calculated a weighted average of the model predictions based on model accuracy. -->
Since there are only few records per species, we fitted a logistic binary regression using a Hierarchical Generalized Additive Model (**HGAM**; Pedersen *et al.* 2019) with Point Process weighting (Fithian & Hastie 2013) to increase the predictive power of the model.
We downloaded global layers of 19 bioclimatic variables representing the mean, extremes, and variation in temperature and precipitation from WorldClim v2.1 (Fick *et al.*. 2017), and a layer of the Landsat Enhanced Vegetation Index (EVI; Masek *et al.*. 2006, Vermote *et al.*. 2016). We then used Principal Coordinate Analysis (PCA) to convert the layers into orthogonal Principal Components (PCs) to reduce collinearity in the predictor variables. We used the 4 first PCs, which collectively explain 88% of the variation in the bioclimatic variables, as predictor variables.
<!-- We then used block cross-validation (Muscarella *et al.*. 2014) to evaluate model performance by splitting the data into four geographical blocks (bottom-right, bottom-left, top-right, and top-left). We calculated the True Skill Statistic (TSS; Allouche *et al.*. 2006) for each model and each block, and generated ensemble models for each species by averaging all models and blocks weighted by their TSS scores. -->
We then used 10-fold cross-validation to train the model on random samples of 90% of the data, and evaluate model performance by making predictions on the remaining 10%. We calculated the area under the receiver operating characteristic curve (AUC) for each model trained on each fold, and selected as the best model the one which achieved the highest AUC. The best model achieved excellent predictive performance, with an AUC score of 0.91.
<!-- We then used the ensemble model for each species to make predictions for all cells in Australia and projected the resultant habitat suitability scores onto a map of Australia. -->
We then used the best model to make predictions for all cells in Australia and projected the resultant habitat suitability scores onto a map of Australia.
### Host plants
To see how climatic suitability coincides with host plant suitability in Australia, we downloaded a dataset showing the location and extent of select agricultural, mining and forest product commodities from the Catchment Scale Land Use of Australia – Commodities – Update December 2020 (ABARES 2021). We filtered this to include sugar cane, as well as other crops identified as host plants for some of the six pests: barley, maize, oats, rice, sorghum, and wheat. The extent of each of these commodities' growing regions can be plotted on top of the climatic suitability maps.
### References
ABARES (2021) *Catchment Scale Land Use of Australia – Update December 2020*. Australian Bureau of Agricultural and Resource Economics and Sciences, Canberra.
Allouche, O., Tsoar, A. & Kadmon, R. (2006) Assessing the accuracy of species distribution models: prevalence, Kappa and the True Skill Statistic (TSS). *Journal of Applied Ecology* **43**, 1223-1232.
Araújo, M.B. & New, M. (2007) Ensemble forecasting of species distributions. *Trends in Ecology & Evolution* **22**, 42-47.
Fick, S.E. & Hijmans, R.J. (2017) WorldClim 2: new 1km spatial resolution climate surfaces for global land areas. *International Journal of Climatology* **37**, 4302–4315.
Fithian, W. & Hastie, T. (2013) Finite-sample equivalence in statistical models for presence-only data. *The Annals of Applied Statistics* **7**, 1917–1939.
Masek, J.G., Vermote, E.F., Saleous, N., Wolfe, R., Hall, F.G., Huemmrich, F., Gao, F., Kutler, J. & Lim, T.K. (2006) A Landsat surface reflectance data set for North America, 1990-100. *IEEE Geoscience and Remote Sensing Letters* **3**, 68-72.
Muscarella, R., Galante, P.J., Soley-Guardia, M., Boria, R.A., Kass, J.M., Uriarte, M. & Anderson, R.P. (2014) ENMeval: An R package for conducting spatially independent evaluations and estimating optimal model complexity for Maxent ecological niche models. *Methods in Ecology and Evolution* **5**, 1198-1205.
Pedersen, E.J., Miller, D.L., Simpson, G.L. & Ross, N. (2019) Hierarchical generalized additive models in ecology: an introduction with mgcv. *PeerJ* **7**, e6876.
Phillips, S.J., Dudik, M., Elith, J., Graham, C.H., Lehmann, A., Leathwick, J. & Ferrier, S. (2009) Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data. *Ecological Applications* **19**, 181–197.
Vermote, E., Justice, C., Claverie, M., & Franch, B. (2016) Preliminary analysis of the performance of the Landsat 8/OLI land surface reflectance product. *Remote Sensing of Environment* **185**, 46-56.