Using a self-organizing map to predict invasive species: sensitivity to data errors and a comparison with expert opinion

Abstract

Summary1. Predicting which species are more likely to invade a region presents significant difficulties to researchers and government agencies. Methods for estimating the risk of establishment are often qualitative and rely on consultation with experts and stakeholders. The inherent subjectivity of this process can lead to ambiguities in any estimate of a species' risk of establishment. 2. Using global presence ⁄ absence data of insect crop pests employed a self-organizing map (SOM) to categorize regions based on similarities in species assemblages. This technique enabled them to generate a list of species and rank them based on an index of the risk of establishment. However, the sensitivity of this risk list to errors in the presence ⁄ absence data has never been tested. 3. We evaluated the sensitivity of the SOM method by altering the original presence ⁄ absence data by increasing amounts and compared estimates of risk with those generated by a national coordinating body (Plant Health Australia) utilizing expert stakeholder opinion. 4. The risk list was unaffected by alterations of up to 20% of data over all regions. The error rate we detected in the data was within these limits. 5. Comparison with the expert stakeholder methodology revealed significant differences in the estimates of establishment risk. Further analysis of the Australian data revealed that a number of regions with strong trade links to Australia supported species assemblages similar to those in Australia, suggesting they are possible sources of pest species with high probability of establishment. 6. Synthesis and applications. This analysis confirms that the SOM methodology is a robust tool in the quantification of risk of establishment. In addition, SOMs can deliver a level of objectivity, which can complement current consultative processes employed by many biosecurity agencies around the world, providing a better overall assessment of invasion risk. This assessment can inform research and development funding decisions and incursion management plans for both government and host industries. While SOMs are utilized in this work for the prioritization of pest insects they can potentially be applied to any taxa (pest or native) or at any scale in which the data are available.

Global threat to agriculture from invasive species

“…Each of the 1,297 elements of the neuron weight vector of a BMU corresponds to each of the 1,297 invasive species in the analysis and will have a value between 0 and 1, which is a measure of the strength of association of the invasive species with the assemblage of invasive species of any country assigned to that BMU. The strength of association for a species can be interpreted as an index of establishment likelihood for that species in a region (10,19,20). On completion of the analysis, an establishment index can be determined for every species in every country included in the analysis.…”

“…The SOM is able to assess the similarity between locations (in this case, countries) based on species assemblages for all countries simultaneously, generating establishment indices for all species in all locations in which they are not currently present. This method has been shown to be resilient to significant errors in species distributional data (19) and highly effective at ranking those species that can establish in a region above those that cannot (20).…”

Identification of Outlier Loci Responding to Anthropogenic and Natural Selection Pressure in Stream Insects Based on a Self-Organizing Map

“…In Paini et al [34] binary data (presence or absence of pest species in different regions) were similarly used for training pest occurrence in different sampling sites. A different number of species (variables) was selected for data alteration (flipping over) according to different proportion levels (0.05, 0.10, 0.20, and 0.30).…”

“…Being partly inspired by Paini et al [34], we adopted a recognition process instead of retraining. We intended to investigate the local effect of altered data on trained clusters specifically for each individual instead of checking overall output variability in addressing stability of SOM performance.…”

“…In revealing the input-output relationships, Paini et al [34] conducted a SOM sensitivity test by altering presence/absence data of species occurrence (variables) in different sampling sites (sample units) in pest management. They revealed that small changes in a limited range of data supported the SOM's robust predictions of pest invasion risks.…”