Background

1.1Patterns of ecosystem vulnerability in Tasmania¶

The impact of human activity on the natural environment is far-reaching, rapid, and undeniable Crutzen, 2006Steffen et al., 2021. Consequently, there is an urgent global imperative for comprehensive ecosystem monitoring at scales necessary for effective conservation management Farrior et al., 2016Keppel et al., 2015. This challenge is particularly acute for endemic species, which often occupy restricted habitats and exhibit limited adaptive capacity to environmental change, yet conservation assessments for endemic species remain inadequate globally Gallagher et al., 2023. The persistence of ancient lineages in geographically restricted refugia represents a critical conservation priority, as these climatically stable areas function as both evolutionary museums and potential sources for future biotic radiation Cai et al., 2023Harrison & Noss, 2017Jetz et al., 2004. Understanding these systems is increasingly urgent given the theory of alternative stable states, where ecosystems may be unable to return to their original ecological condition once critical thresholds are exceeded Lamothe et al., 2019.

Tasmania exemplifies global patterns of paleoendemism, harbouring exceptional concentrations of ancient lineages with highly restricted contemporary distributions. The island’s temperate vegetation communities include many iconic paleoendemic species such as Lagarostrobos franklinii (Huon pine), Athrotaxis selaginoides (King Billy pine), Athrotaxis cupressoides (Pencil pine), and Nothofagus gunnii (tanglefoot beech). While Nothofagus cunninghamii (Myrtle beech) is not endemic to Tasmania, it shares many similar characteristics and habitat with the endemics. These species represent broader patterns of Tasmanian paleoendemism, persisting through the use of different climate refugia strategies: wet rainforest refugia support Huon pine and Myrtle beech Buckley et al., 1997Drew et al., 2022Lindenmayer et al., 2000, while cool alpine environments provide habitat for tanglefoot beech and Athrotaxis species. Both Athrotaxis and Lagarostrobos are exceptionally long lived but with extremely slow growth rates Allen et al., 2017Allen et al., 2011Gibson, 1988Gibson, 1986.

Despite these species persisting for millions of years, their slow growth rates and poor dispersal confine them to refugial habitats. Historical anthropogenic impacts have disrupted this delicate balance through intensive logging, habitat inundation from hydroelectric development, and altered fire regimes Horne & Hickey, 1991. Contemporary climate change compounds these pressures through warming and drying trends that increase fire risk and penetrate historically fire-free refugia Bowman et al., 2019Holz et al., 2020Worth et al., 2017. Evidence of local extinctions and population collapses demonstrates the vulnerability of these slow-growing species to changing fire regimes Bliss et al., 2021Bowman et al., 2021Fletcher et al., 2018. Additional stressors such as habitat fragmentation and introduced pathogens like Phytophthora further compromise population viability Khaliq et al., 2019. The combination of historical population reduction and accelerated environmental change has created conditions where slow-growing species lack sufficient time for natural recruitment and recovery Cole et al., 2014, highlighting the critical need for precise spatial monitoring of species distributions and habitat condition.

1.2Identifying and quantifying vulnerable areas¶

Current vegetation mapping approaches employed by Natural Resources and Environment Tasmania (NRE Tasmania) reflect methodological constraints that limit their utility for contemporary conservation challenges. The primary approach relies on manual interpretation of aerial photography, utilising limited spectral information and subjective visual cues to delineate vegetation boundaries Department of Natural Resources and Environment Tasmania, 2020. Critically, the resources required by these approaches limit the frequency at which the mapping can be updated. For example, the most recent TASVEG dataset, v4.0 released in 2020 represents the first major update since 2013 and an average of five years between major versions since 2004 Department of Natural Resources and Environment Tasmania, 2020. This temporal lag prevents detection of rapid changes in population extent or condition, particularly problematic when current distributions may represent ecological “echoes” of past conditions where recruitment of new individuals is no longer possible due to altered fire regimes Keppel et al., 2015. Spatial resolution limitations force reliance on community-level classifications rather than species-level discrimination, masking important ecological patterns within populations and compromising targeted management interventions.

Ground validation protocols introduce systematic sampling bias through preferential access to readily accessible areas, typically excluding remote terrain where many endemic populations persist Burg et al., 2015Clarke et al., 2012Morrison, 2016. The labour-intensive nature of manual interpretation creates scalability constraints, limiting monitoring frequency and extent. Additionally, these approaches provide little to no information about vegetation health, demographic structure, population trends, and neglect the value significant individual trees. These attributes are critical to adaptive conservation planning given that old and large individual trees are essential to ecosystem health Lindenmayer et al., 2012Lindenmayer et al., 2014, population size data are necessary for designing effective conservation areas Burgman et al., 2001, and the quality of presence data directly affects species distribution modelling accuracy Fei & Yu, 2016.

1.3The promise of remote sensing and machine learning for ecosystem monitoring¶

Remote sensing revolutionised automated ecosystem monitoring and at broader spatial scales than possible with traditional field methods Roughgarden et al., 1991. Early vegetation mapping applications utilised broad-band multispectral sensors such as Landsat and SPOT imagery Rouse et al., 1973Rouse Jr et al., 1975, achieving reasonable success for broad vegetation class discrimination. However, the spatial resolution of the datasets limited the types of objects that could be discriminated in scenes with high spectral variance such as forests Woodcock & Strahler, 1987. Both Fisher Fisher, 1997 and Cracknell Cracknell, 1998 stressed the importance the pixel-object relationship, and the problem of mixed signals of sub-IFOV objects. Treitz and Howarth Treitz & Howarth, 2000 found that the spatial variation of the classes must be resolvable in the imagery and that multiple spatial scales should be considered when aiming to discriminate forest ecosystems in multispectral imagery. Similarly, Cochrane Cochrane, 2000 demonstrated that spectral responses in vegetation are “non-unique” in hyperspectral data with a 10 m spatial resolution, and suggests that multiple high spatial resolution samples from a single individual may mitigate this problem.

The pixel-based approach to classification is not suitable for high spatial resolution imagery, where within-class spectral variance can exceeded between-class differences due to shadowing, illumination effects, and sub-pixel heterogeneity Wulder et al., 2004. Object-based image analysis (OBIA) addressed some of these limitations by accounting for the spatial characteristics of a target Blaschke et al., 2014Blaschke, 2010. Treating the target as an object captures the within-class spectral variability mitigates the so called “salt and pepper” effect when using pixel-based classification methods in high resolution imagery Yu et al., 2006. OBIA outperforms pixel-based approaches for vegetation discrimination tasks in very high-resolution imagery in many cases Immitzer et al., 2012Sibaruddin et al., 2018. However, Dingle Robertson and King Dingle Robertson & King, 2011 found no significant difference between the two approaches, emphasising that the performance of OBIA greatly depends on an effective segmentation and feature engineering.

Machine learning approaches, such as Random Forest (RF) and Support Vector Machine (SVM) algorithms, improved classification accuracy over the physically-based Spectral Angle Mapper (SAM) in early demonstrations Clark et al., 2005 in both pixel and object-based approaches. While a data driven approach significantly improved the accuracy of species discrimination, these techniques struggled with the multi-scale spatial-spectral characteristics of organic targets such as vegetation Dingle Robertson & King, 2011.

1.4The rise of deep neural networks and representation learning in remote sensing¶

The advent of deep neural networks fundamentally transformed computer vision through the introduction of representation learning, where the algorithm automatically discovers hierarchical features directly from data without manual feature engineering LeCun et al., 2015. Convolutional neural networks (CNNs) revolutionised dense segmentation tasks through architectures such as fully convolutional networks Long et al., 2015 and U-Net Ronneberger et al., 2015. More recently, Vision Transformers (ViTs) have demonstrated superior performance compared to CNNs in image classification tasks but require much larger datasets to avoid overfitting Kolesnikov et al., 2021. These learned representations consistently outperform engineered features by capturing complex spatial-spectral patterns that are difficult to encode manually, particularly for fine-grained species discrimination Zhong et al., 2024. However, the substantial data requirements of deep learning models initially posed significant challenges for remote sensing applications, where labelled training data is often scarce and expensive to acquire.

Deep learning has demonstrated remarkable success in tree species segmentation by automatically learning hierarchical spatial-spectral features that capture complex patterns invisible to traditional methods. Notable achievements include the mapping of over 1.8 billion individual trees across 1.3 million km² using U-Net architectures with satellite imagery Brandt et al., 2020, and fine-grained vegetation mapping achieving 84% accuracy from high-resolution RGB imagery alone Kattenborn et al., 2019. Three-dimensional convolutional neural networks (3D-CNNs) have proven particularly effective for hyperspectral data, simultaneously exploiting both spatial context and spectral relationships to achieve classification accuracies exceeding 95% Zhang et al., 2020. However, these advances come with significant challenges: deep learning models require vast amounts of training data across diverse acquisition conditions Brandt et al., 2020Brandt et al., 2025, suffer from limited interpretability as “black box” systems, and often struggle with generalisation beyond their training domains.

The challenge of model transferability across different sensors, geographic locations, and temporal conditions represents a critical bottleneck in operational remote sensing applications. Domain shift, arising from sensor characteristics, atmospheric conditions, seasonal variations, and geographic differences, can severely compromise model performance when deployed beyond training conditions Brandt et al., 2025. Spatial autocorrelation in training data further exacerbates this issue, with studies demonstrating that violations of spatial independence can inflate apparent model performance by up to 30%, leading to overly optimistic assessments of generalisation capability Kattenborn et al., 2022. Knowledge distillation (KD) has emerged as a promising solution for cross-sensor learning, enabling the transfer of knowledge from high-fidelity (spatial and spectral) teacher models to lower-fidelity student networks Himeur et al., 2025. Shin et al. (2023) showed that KD from a MS-trained teacher model improves the accuracy of a RGB-trained student model by approximately 2% (OA = 94.04%). However, it is necessary to highlight that the teacher and student models in Shin et al. (2023) share the same RGB dataset which is a spectral subset of the Sentinel-2 based EuroSAT dataset Helber et al., 2019. The teacher uses multiple modalities with an encoder branch for: RGB, NIR, RE, SWIR, whereas the student uses a single encoder branch: RGB.

Multimodal data fusion is a critical advancement beyond single-sensor approaches, combining complementary information from spectral, structural, and temporal domains to enhance species discrimination capabilities. True multimodal fusion extends beyond feature-level fusion (data stacking) Zhong et al., 2024, towards architectures that process each modality through dedicated pathways before integration. Late-fusion strategies, where modality-specific streams are processed independently at their native resolutions before final combination, have demonstrated superior performance in species classification tasks compared to unimodal-multi-source models Hong et al., 2021Tiel et al., 2025. While these approaches have demonstrated the value of integrating multi-source remote sensing data, the true potential of multimodality is in the ability to integrate symbolic or contextual information Baltrušaitis et al., 2019. This can be by integrating OpenStreetMap information Hong et al., 2021, or semantic vegetation attributes and hierarchical taxonomy Sumbul et al., 2018. Harmon et al. (2023) demonstrates that incorporating domain knowledge through rule-based constraints improves the classification of rare tree species in RGB imagery.

The integration of ecological understanding and environmental context is a paradigm shift from pure computer vision approaches towards ecologically informed deep learning models. By incorporating environmental variables such as climate, soil properties, and topographic characteristics, these models can leverage the fundamental ecological principle that species distributions are constrained by environmental niches and habitat suitability Brodrick et al., 2019. This approach enables DNNs to move beyond simple biophysical component identification towards the direct recognition of ecological patterns and processes at scales previously impossible to consider Brodrick et al., 2019. The combination of remote sensing imagery and environmental datasets in multimodal DNNs has used to improve and finding patterns in SDMs at a regional scale Ryo et al., 2021Tiel et al., 2025, but there is limited research on how this information can be exploited for species segmentation tasks.