Reliability of Machine Learning Maps

| |

Machine learning techniques are becoming more ubiquitous not only in our normal lives but also in research. In ecology, researchers increasingly strive to use the power of data to derive forecasts and estimates of how our environment is going to change as climate change and a host of other anthropogenic factors affect our planet.

While this is generally positive, the danger is such machine driven maps may not always tell an accurate story. Creating effective machine learning maps will require careful due diligence and close attention to the scientific process. 

Challenges of Using Machine Learning in Mapping Ecological Data

Limits to spatially-derived maps are evident, particularly as more research attempts to create global-scale maps that attempt to aid our decision making and understanding of our planet. For instance, recent works focusing on global tree restoration potential, global soil nematode abundances, or global soil maps are some examples.

Problems have been encountered particularly for datasets which have high clustering of data in some regions but are sparse in other regions. Researchers attempt to use spatial cross-validation techniques to validate results, but this proves difficult or even futile when clustering of input training and validation data are mostly constrained to limited regions.

While spatially-derived machine learning maps, such as those that use popular random forest techniques, may show results that seem plausible, using maps that show distance between an estimated machine learning output and input dataset can reveal a significant problem in how given global-scale maps are derived.

Regional Geographic Bias in Machine Learning

Researchers, particularly in Western states, have often focused their data inputs and field research in North America and Europe, which has created difficulties in accepting the accuracy of globally-derived maps that depend on other input regions, even if such maps are desirable from a decision and information point of view.

Two columns with maps on the left and graphs on the right showing the difference in predicted ecological distributions.
Meyer & Pebesma, 2022: “Spatial distribution and distribution of nearest neighbor distances (right; sample-to-sample distance in pink, prediction-location-to-sample distance in blue) for three different publicly available datasets: (B) specific leaf area from the Try database24 as used for the global mapping in Moreno-Martinez et al. (2018)25, and (C) the nematodes dataset compiled by Van den Hoogen et al. (2019).” Cropped figure. Full figure:, CC BY 4.0.

As argued in a recent article, one way to address this problem of over clustering data could be to create maps that literally remove or indicate areas where predictor values are significantly different from field collected data or simply remove results where information is too sparse to create reliable maps that take input data for a machine derived map. The problem is maps that do not remove given data sparse regions from their forecasted outputs have the potential to misconstrue an accurate ecological story, which could have negative impacts for us in making important global-scale decisions.[1]

Limits of Machine Learning for Developing Ecological Maps

Other articles have also hinted at the dangers of using some machine learning techniques for deriving maps. For instance, in an article assessing global lake ice using different machine learning techniques, including multinomial logistic regression, support vector machine, random forest, and gradient boosting trees, researchers demonstrate that these methods are all useful for producing maps that have accuracy over 94%.

However, all of these techniques, except random forest, were prone to being over sensitive in their predictability using cross-validation k-fold techniques on hyperparameters.[2] 

Machine Learning and Data Gaps in Mapping

Researchers have also demonstrated that there could be ways to improve data gaps by using Generative Adversarial Networks (GANs) to improve data quality and predictability. While GANs are known for their applicability in creating images, sometimes even fake images, GANs can also be used to derive data which can then be used in training of machine learning maps using two competitive networks that compare a derived dataset to an empirical dataset.

When the derived data are as good as the empirical data, the GAN model can then be used to create derived data at quality similar to empirical data. Researchers have shown that GAN-derived data can improve landslide susceptibility maps, as one example. While GANs do show potential in their scientific applicability in filling in data gaps, sufficient input data are still needed to create accurate derived data.[3]  

Machine Learning in Ecological Mapping Has Limits

While maps derived from machine learning are an increasing result of researchers using these maps to give us more knowledge on what could happen at global and regional scales for a variety of environmental phenomena, the danger is our input data used to create such maps are not always sufficient, particularly for regions away from North American and Europe.

There are some potential workarounds, such as using GANs to aid in data deficiencies, but over clustering of information in any given region is likely to bias output used to estimate more data sparse regions. Providing outputs that give viewers and other researchers clear warning where data outputs are potentially less trustworthy should be part of outputs given in any research using spatial maps derived from machine learning. Simply giving a map with results could lead to deceptive outputs unless we begin to demonstrate uncertainty as part of our outputs.  


[1]    For more on an article discussing the problems of machine learning derived maps and ways in which data shortcoming can be addressed, see: Meyer, Hanna, and Edzer Pebesma. 2022. ‘Machine Learning-Based Global Maps of Ecological Variables and the Challenge of Assessing Them’. Nature Communications 13 (1): 2208.

[2]    For more on different machine learning techniques and how results could be affected in mapping lake ice, see:  Wu, Yuhao, Claude R. Duguay, and Linlin Xu. 2021. ‘Assessment of Machine Learning Classifiers for Global Lake Ice Cover Mapping from MODIS TOA Reflectance Data’. Remote Sensing of Environment 253 (February): 112206.

[3]    For an example of using GANs to derive data used to train machine learning maps, see:  Al-Najjar, Husam A.H., and Biswajeet Pradhan. 2021. ‘Spatial Landslide Susceptibility Assessment Using Machine Learning Techniques Assisted by Additional Data Created with Generative Adversarial Networks’. Geoscience Frontiers 12 (2): 625–37.


Share this article

Enter your email to receive the weekly GIS Lounge newsletter: