Using Earth Observation Data and Machine Learning for Discriminating and Mapping Tree Species in the Welsh Forestry Sector

Abstract

Knowledge of the distribution of tree species from global to local levels is essential for forest management, particularly in terms of ensuring sustainable production and use and assessing the impacts of change. Such information is also critical for understanding and quantifying contributions to biodiversity, carbon, and water/ soil quality. Earth observation acquired by sensors operating in multiple modes (optical, radar and lidar) is becoming increasingly important for this purpose, particularly given improvements in spatial resolution and frequencies of coverage. Included in these are the new suite of PlanetScope sensors that are providing 3-5 m observations of the Earth’s surface on a daily basis (cloud permitting) in the visible to near-infrared wavelength regions.

Focusing on temperate forests in Wales (United Kingdom), the research aims to investigate the use of monthly time-series of PlanetScope data acquired over single years for discriminating between and mapping five tree species that are especially important for biodiversity/carbon storage (oak and beech) and timber production (Sitka spruce, Norway spruce and Douglas fir). Such information is particularly important to support forest and land management (e.g., in evaluating risk in the case of plant disease) and address local to international policy requirements and obligations (e.g., for enhancing the condition of forests and connectivity of habitats, increasing resilience, including in response to climate change).

The methods focused on undertaking field campaigns in North and South Wales (Coed-y-Brenin Forest and Wentwood and Wye Valley forests, respectively) to identify and/or confirm the dominance of the five species selected for the study.
From these locations, in collaboration with the Welsh Government and Forest Research, more than 600 samples of five tree species were collected and thanks to delivered time-series data by Planet Labs of two sites, profiles of visible, near-infrared and derived indices data were extracted from acquisitions collated into monthly composites for five years (2018, 2019, 2020, 2021, 2022) with the latter (2021, 2022) also including some red edge bands. A comparison of the reflectance spectra for the different months indicated the best discrimination of oak and beech during leaf flush, particularly within indices using at least two visible and near-infrared bands (e.g., MCARI1). Discrimination between Douglas fir, Sitka spruce and Norway spruce was most evident from spring to autumn, but was often compromised because of high spectral overlap between these. Comparison of temporal trajectories indicated that these were inconsistent between sites and years, largely because of differences in environmental settings, climate, age and growth stages.

Given the large number and complexity of temporal observations and spectral indices derived, Machine Learning approaches to image classification were explored with the help of the PyCaret software. Of these, the CatBoost algorithm was
considered to provide the best capacity for discriminating the five tree species, with overall accuracies generally exceeding 90% for the model but also in reserved validation datasets. When applied to the time series of PlanetScope data for each site, and using three years of data and site-specific training, the accuracies were greatest and a close correspondence with existing mapping (i.e., Natural Resources Wales Subcompartment Database) was observed. However, when applied across different sites and years, classification accuracy declines significantly, resulting in limited mapping effectiveness and transferability.

For Wales, the research concluded that Machine Learning algorithms allowed the mapping of the five species with accuracies that were deemed acceptable. However, the model is not sufficiently transferable to other locations and certainly not nationally for Wales. Continued use of the PlanetScope data, based on the outcomes from this work, would require investment in the on-site collection or confirmation of forest species type for the area of interest and using more than one year (ideally) of PlanetScope monthly composites. Further work would also be needed to expand the approach to other species common to Wales, noting that many occur in stands of mixed age and species composition, with several being rare or confined to understorey layers.

Date of Award	2025
Original language	English
Awarding Institution	Aberystwyth University
Sponsors	Llywodraeth Cymru KESS-2
Supervisor	Richard Lucas (Supervisor) & Pete Bunting (Supervisor)