Comparative Analysis of Supervised and Self-Supervised Transformer Models for Water Stress and Phytophthora Detection in Cupressus

Cupressus is a genus of conifers whose importance lies primarily in the ecological, environmental, and forestry fields. Therefore, it is important to maintain continuous monitoring of these species to promptly detect conditions such as water stress and infections caused by Phytophthora. In this work, a comparative analysis was carried out between supervised Transformer models (ViT-B/16 and Swin-Small) and self-supervised Transformer models (ViT-B/16+MAE, Swin-Base+SimMIM, ViT-B/14+DINOv2, and ViT-B/16+MaskFeat). Two experimental scenarios were considered. The first corresponds to clean data—that is, preprocessed and normalized images without additional degradations. The second corresponds to images incorporating noise in the form of visual perturbations via Gaussian blur, random adjust sharpness and color jitter, with the aim of evaluating and comparing the robustness of the models against image quality degradations. A model comparison was performed by evaluating performance using quantitative metrics such as accuracy, specificity, recall, precision, and F1-score. Additionally, a complementary analysis was conducted using confusion matrices, Receiver Operating Characteristic (ROC) curves, Precision–Recall (PR) curves, and activation maps obtained via Grad-CAM. The results show that the supervised Swin-Small model emerged as the best-performing model overall in both scenarios considered, obtaining an accuracy of 0.9969 in the clean data scenario and 0.7979 in the degraded data scenario. Meanwhile, ViT-B/16+MAE performed best among the self-supervised models, reaching an accuracy of 0.9962 in the clean data scenario and 0.7091 in the degraded data scenario. These results support the conclusion that Swin-Small, trained via supervised learning, offers the best balance between performance and robustness against visual degradations, whereas the evaluated self-supervised models failed to outperform the supervised approach under the analyzed conditions of this set of experiments.

Authors

Institutions

Publication Details

Journal
Forests
Published
2026-11-20
DOI
https://doi.org/10.3390/f17111393
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Comparative Analysis of Supervised and Self-Supervised Transformer Models for Water Stress and Phytophthora Detection in Cupressus

José Cruz Martínez-Perales, Miriam Pescador‐Rojas, B. Luna-Benoso, Úrsula Samantha Morales-Rodríguez et al.
4 links
Forests
Smart Agriculture and AI
article

Comparative Analysis of Supervised and Self-Supervised Transformer Models for Water Stress and Phytophthora Detection in Cupressus

José Cruz Martínez-Perales, Miriam Pescador‐Rojas, B. Luna-Benoso, Úrsula Samantha Morales-Rodríguez, Josué Aldebarán González López
article en

Abstract

Cupressus is a genus of conifers whose importance lies primarily in the ecological, environmental, and forestry fields. Therefore, it is important to maintain continuous monitoring of these species to promptly detect conditions such as water stress and infections caused by Phytophthora. In this work, a comparative analysis was carried out between supervised Transformer models (ViT-B/16 and Swin-Small) and self-supervised Transformer models (ViT-B/16+MAE, Swin-Base+SimMIM, ViT-B/14+DINOv2, and ViT-B/16+MaskFeat). Two experimental scenarios were considered. The first corresponds to clean data—that is, preprocessed and normalized images without additional degradations. The second corresponds to images incorporating noise in the form of visual perturbations via Gaussian blur, random adjust sharpness and color jitter, with the aim of evaluating and comparing the robustness of the models against image quality degradations. A model comparison was performed by evaluating performance using quantitative metrics such as accuracy, specificity, recall, precision, and F1-score. Additionally, a complementary analysis was conducted using confusion matrices, Receiver Operating Characteristic (ROC) curves, Precision–Recall (PR) curves, and activation maps obtained via Grad-CAM. The results show that the supervised Swin-Small model emerged as the best-performing model overall in both scenarios considered, obtaining an accuracy of 0.9969 in the clean data scenario and 0.7979 in the degraded data scenario. Meanwhile, ViT-B/16+MAE performed best among the self-supervised models, reaching an accuracy of 0.9962 in the clean data scenario and 0.7091 in the degraded data scenario. These results support the conclusion that Swin-Small, trained via supervised learning, offers the best balance between performance and robustness against visual degradations, whereas the evaluated self-supervised models failed to outperform the supervised approach under the analyzed conditions of this set of experiments.

ForestsVol. 17(11)
Instituto Politécnico Nacional (MX)
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon