Giulia Antonia Resente

• Quantitative Wood Anatomy (QWA) is defined as the analysis of the xylem anatomical features in trees, shrubs and herbaceous plants to investigate plants functioning, growth and environment. By combining the recognition of wood anatomical structures together with measurement techniques to quantify anatomical features like cell size, cell wall thickness, or vessel density, QWA provides comparable data among growth ring time series. The combination between QWA and dendrochronology allows for the establishment of wood anatomical trait time series which are particularly valuable in the frame of past climatic reconstructions and, in parallel, to predict plant functioning under future climate projections. Due to the intensification of the climate crisis, QWA is becoming an increasingly important tool for understanding the impacts on forest and shrub ecosystems and establish counteractive strategies. Current methodologies for quantitative wood anatomical analyses provide manual or semi-automated methods, therefore requiring significant user input in terms of settings adjustments and manual editing. These characteristics hinders these tools from being the ideal solution to tackle the current rising demand for wood anatomical data. The time spent for such analyses and the effort employed to gain meaningful results raise the interest in the implementation of AI in QWA. Quantitative wood anatomical analyses may considerably improve in precision and efficiency following AI incorporation to the overall workflow, because of AI ability to identify complex patterns and relationships within wood structure. Furthermore, the automatization introduced by AI methods is supposed to improve the time-consuming task of manually editing traditional image analyses output. For these reasons this dissertation addresses two research topics: i) applying AI detection skills to improve quantitative wood anatomical analyses on thin-sections from wooden cores (Chapters I and II) and ii) introduce AI for the detection of concentric rings in shrub thin-sections and facilitate their measurements (Chapters III and IV). The successful development of two distinct tools responding respectively to aim i) and ii) demonstrated that it is possible to improve the current state of the art by joining the two fields of AI and QWA for a variety of purposes. We introduced the development of CARROT (Cell And Ring RecOgnition Tool) in order to streamline quantitative wood anatomical analyses for a faster and automated workflow (Chapter I), and of INBD (Iterative Next Boundary Detection) to address the methodological gap of concentric ring automated detection and the relative computation (Chapter III). These tools showed not only the ability to provide meaningful results in the execution of the main tasks, but also to generally outperform manual or classic image analysis (Chapters I and IV). In view of the results obtained by both approaches, we promote the use of CARROT and INBD underscoring on one side the advantage of employing automatized methods to save time during analyses, and on the other side, the relevance of their broad applicability. Both tools operate several essential tasks with fairly high accuracy, handling two growth structures (trees and shrubs), and in the case of CARROT four wood anatomical types (conifer, ring-porous, semi-ring-porous, and diffuse-porous). A great potential of application resides in the implementation of a user interface for both tools (Chapter II and IV), promoting wood anatomical analysis improvement through user-friendly interfaces in an open-source environment. Practical applications of both tools were also performed. In Chapter II, CARROT was employed with the purpose of studying wood anatomical changes in surviving pedunculate oaks, after the flooding and the permanent rewetting of a formerly drained peatland. In this context, CARROT was found to be once again meeting the expectations in terms of high cell recognition performance, coping with the segmentation of both very wide earlywood vessels and very small latewood vessels. In Chapter IV, INBD cross-dating potential was tested to frame the realistic application of the tool. Results showed that cross-dating statistics were higher for INBD ring width measurements compared to those obtained manually, and that in most cases INBD was outperforming manual measurements even prior to any cross-dating attempt. In general, we could observe that both methods would greatly benefit from the implementation of larger training datasets, which would enhance their accuracy across diverse wood anatomical dataset. For INBD specifically, future developments should focus on including functions that allow users to correct wrongly detected outputs and consequently recalculate the data. Currently, recent attempts in merging AI and wood anatomy mainly take advantage of AI strengths to focus on species recognition or ring identification from cores, without effectively addressing quantitative wood anatomical research questions. For these reasons, CARROT and INBD can be regarded as cutting-edge techniques in quantitative wood anatomical analyses, for their innovative method and for their effective feasibility. Their employment would not only improve results in terms of accuracy and time, but also allow researchers to shift the focus towards the interpretation of the results and their discussion, rather than the current constraints of obtaining such results. Overall, the integration of CARROT and INBD and similar tools into quantitative wood anatomical research framework yield the possibility of expanding such studies, constituting a meaningful resource to broaden ecological investigations.