Abstract
The species Pinus radiata is highly invasive in native forests in Chile, drastically affecting the functioning and structure of ecosystems. Hence, it is imperative to develop robust approaches to detect P. radiata invasions at different scales. Models based on convolutional neural networks (CNN) have proven to be a promising alternative to detect plant invasions in high-resolution remote sensing data, such as those obtained by drones. However, studies have been limited in their spatial variability and their assessments of transferability or transfer learning to new sectors, hindering the ability to use these models in a real-world setting. We train models based on CNN architectures using unpiloted aerial vehicle data and evaluate their ability to transfer learning outside the training domain using regression approaches. We compared models trained with low spatial variability (mono-site) with those with high spatial variability (multi-site). We further sought to maximize the transference of learning outside the training domain by searching among different architectures and models, maximizing the evaluation in an independent data set. The results showed that transfer learning is better when multi-site models with higher spatial variability are used for training, obtaining a coefficient of determination R2 between 60% and 87%. On the contrary, mono-site models present a wide variability of performance attributed to the dissimilarity of information between sites, limiting the possibilities of using these models for extrapolations or model generalizations. We also obtained a significant difference between within-domain generalization using test data versus transfer learning outside the training domain, showing that testing data alone cannot depict such discrepancy without further data. Finally, the best models for transfer learning on new data domains often do not agree with those selected by the standard training/validation/testing scheme. Our findings pave the way for deeper discussions and further investigations into the limitations of CNN models when applied to high-resolution imagery.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.