Answers are grounded in this paper and cite their sources.
Despite the importance of woody vegetation landscape features (individual trees, rows of trees, hedgerows, riparian vegetation etc.) in halting the loss of biodiversity in agricultural areas, data on their extent remains lacking. We classified woody vegetation (i.e. the woody vegetation landscape features and forest) from the Slovenian national orthophoto. We used convolutional neural networks with the goal of providing the most accurate classification in the agricultural land. We tested whether adding patches, covered with forest, to the training dataset would improve the classification of woody vegetation in the agricultural land. The results show significant differences in the classification performance of the models when comparing patch sizes, model architectures, band combinations and maximum forest cover of the training patches. The top-performing architecture is HRNet (High-Resolution Network); the patches that provide the best results include red, green and near-infrared spectral bands, have sides of length 256 pixels and are covered with forest to a maximum extent of 60%. With these parameters, a model achieved Jaccard index of 80.13% in agricultural areas. We used this model to classify woody vegetation in Slovenian outstanding landscapes, which proved the model is applicable in the majority of the country. Most misclassifications are in specific environments (e.g. surfaces covered with reed and saltworks), which were not represented in the training data. The comprehensive model assessment in diverse Slovenian landscapes also informs similar efforts worldwide.
No takes yet. Share an insight, caveat, or question.
Gabrič et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: