<p>Using pre-trained convolutional neural networks (CNNs) architectures have proven effective in high-resolution remote sensing, even when homogeneous and few data samples are used. However, it is still uncertain how well models trained with limited spatial information can transfer the learning process from one domain to a new domain (transductive transfer learning). This paper evaluates transductive transfer learning in CNN regression models using RGB-based data captured by unoccupied aerial vehicles on five sites, using the prediction of <i>Pinus radiata</i> canopy coverage as a case study. We trained five models, one per site, analyzing their internal performance using fine-tuning and feature extraction training approaches. Then, we evaluated their transfer learning ability to new unseeing sites. We found that the trained models perform accurately within their domain, as previous research demonstrates (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="521_2025_11267_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> > 0.90 using fine-tuning). However, we depicted varying performances during transfer learning, with <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="521_2025_11267_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> ranging from − 0.68 to 0.63 for feature extraction and from − 4.42 to 0.77 for fine-tuning. Our results show that these poor performances are independent of the training approach (fine-tuning or feature extraction), the number of observations, or the complexity of the model. In contrast, the success during transfer learning is closely linked to the similarity between the source and target domains, which is often unknown when predicting new data. These results depict the importance of carefully planning the future use of such models for their sustainability and generalization over time.</p>

Convolutional neural network models with low spatial variability hamper the transfer learning process

期刊论文
发表日期：

2025年5月5日

查看全文

Alejandra Bravo-Diaz,
Sebastián Moreno,
Javier Lopatin

摘要

Using pre-trained convolutional neural networks (CNNs) architectures have proven effective in high-resolution remote sensing, even when homogeneous and few data samples are used. However, it is still uncertain how well models trained with limited spatial information can transfer the learning process from one domain to a new domain (transductive transfer learning). This paper evaluates transductive transfer learning in CNN regression models using RGB-based data captured by unoccupied aerial vehicles on five sites, using the prediction of Pinus radiata canopy coverage as a case study. We trained five models, one per site, analyzing their internal performance using fine-tuning and feature extraction training approaches. Then, we evaluated their transfer learning ability to new unseeing sites. We found that the trained models perform accurately within their domain, as previous research demonstrates ( $R^2$ $R^{2}$ > 0.90 using fine-tuning). However, we depicted varying performances during transfer learning, with $R^2$ $R^{2}$ ranging from − 0.68 to 0.63 for feature extraction and from − 4.42 to 0.77 for fine-tuning. Our results show that these poor performances are independent of the training approach (fine-tuning or feature extraction), the number of observations, or the complexity of the model. In contrast, the success during transfer learning is closely linked to the similarity between the source and target domains, which is often unknown when predicting new data. These results depict the importance of carefully planning the future use of such models for their sustainability and generalization over time.