Land Suitability Assessments for Agriculture using the Data By-products of Mining Exploration

Abstract

This thesis was proposed in the context of the social licence to operate for a new gold mine in West Africa that necessitated the relocation of subsistence farmers and asked the question: Can land suitability assessments for agriculture be carried out effectively using the data by-products of mining exploration, without the necessity for and cost of additional data acquisition? The project objectives were to develop a land suitability assessment process that, firstly, was locally relevant with outputs usable by farmers in the vicinity of a new gold mine in the tropical south-west of Burkina Faso and, secondly, is transferable to other sites with different terrain, climate and styles of agriculture. The compensation maps prepared as part of the application for the mining licence provided known occurrences for locally grown crops, so facilitating the use of data-driven species distribution methods to produce crop suitability maps. However, the clustered occurrences presented the likelihood of spatial sampling bias affecting models and the evaluation of results was complicated by the lack of test data outside of these areas. The maximum entropy algorithm (Maxent) was used to produce crop suitability models using a methodology that took advantage of the geographical separation of the presence data sites to develop cross-validation models trained on different sets of presence points. The accuracy of model predictions (using the area under the curve (AUC) of the receiver operator curve for test data) and similarity of resulting suitability maps (from correlations) were compared to assess model accuracy, robustness and sensitivity to sampling bias; however, region wide validation of results was not possible with the available data. The methodology was then applied to two local sites in South Australia for which region-wide verification data were available in order to validate the methods used and to demonstrate their transferability to other sites with different terrain, climate and styles of agriculture. Neither the categorical soil map supplied by the exploration company nor the publicly available global maps of commonly used soil properties were useful as model predictors. Algorithmic methods were devised to process both sources of soil data into a new set of hybrid soil layers (combining the fine spatial detail of the supplied map with the multi-dimensionality of the soil property maps) that proved effective in modelling. The thesis contributes to the application of species distribution modelling by presenting this new method for converting categorically valued maps into continuously valued raster layers for effective use by modelling algorithms. The thesis also demonstrates effective cost-free and language-independent mapping solutions that overcome the local challenges of illiteracy and poor access to technology. Paper maps were designed for map users without access to other technology, and interactive maps were produced for map users with access to electronic devices, with and without internet access. The method of predicting local agricultural land suitability presented in the thesis has been shown to be transferrable to other sites. It is particularly well suited to mining applications in developing countries where detailed data on local agriculture are collected as part of the environmental and social impact assessments. As such, it could become a model for future mining projects and contribute to more successful collaborations between the mining sector and local communities in developing countries.