García Sanabria Plant Classifier: versioned image datasets and transfer-learning models for six plant species of Parque García Sanabria (Santa Cruz de Tenerife)

This deposit contains three versioned image datasets and the corresponding transfer-learning models for the classification of six plant species found in Parque García Sanabria (Santa Cruz de Tenerife, Spain), plus an "other" rejection class. The three dataset versions (196, 206 and 174 images) are not successive enlargements of a single collection: each was rebuilt from the error analysis of the previous model, tested in the field across 220 photographs. Version 2 corrects a spurious background correlation identified in version 1, in which a uniform sky background caused unrelated objects to be classified as Dracaena draco. Version 3 documents a rejected experiment: removing the weakest class solved the generic-trunk failure mode but caused a severe regression on Phoenix canariensis. It is included precisely because it failed. All three versions are retained deliberately, so that the background hypothesis — introduced jointly with other corrections in v2, and therefore plausible but unproven — can be tested in isolation. Models were fine-tuned from ResNet34 weights pre-trained on ImageNet, using fastai. Labels are encoded as directory names. Version 3 is not stored as a separate directory: it is version 2 with one class excluded, as described in datasets/v3-NOTE.md. Dataset images and model weights: CC BY 4.0. Code and documentation: MIT. Full methodology, field-test results and known limitations are documented in README.md and in docs/. This is a learning project with a small dataset. Reported accuracies are noisy and not comparable across versions, for the reasons stated in the documentation.