EGU22-10060
https://doi.org/10.5194/egusphere-egu22-10060
EGU General Assembly 2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Automated mapping of Soil Surface Components (SSCs) in highly heterogeneous environments with Unoccupied Aerial Systems (UAS) and Deep Learning: working towards an optimised workflow

Eva Arnau-Rosalén1, Ramón Pons-Crespo2, Ángel Marqués-Mateu2, Jorge López-Carratalá3, Antonis Korkofigkas4, Konstantinos Karantzalos5, Adolfo Calvo-Cases6, and Elias Symeonakis1
Eva Arnau-Rosalén et al.
  • 1Department of Natural Sciences, Manchester Metropolitan University, Manchester, United Kingdom
  • 2Department of Cartographic Engineering, Geodesy and Photogrammetry, Universitat Politècnica de València, València, Spain
  • 3Catholic University of Murcia-UCAM, Guadalupe, Murcia, Spain
  • 4School of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece
  • 5Remote Sensing Lab., National Technical University of Athens, Athens, Greece
  • 6Inter-University Institute for Local Development (IIDL), Department of Geography, University of Valencia, Valencia, Spain

Pattern recognition remains a complex endeavour for ‘structure/function’ approaches to ecosystem functioning. It is particularly challenging in dryland environments where spatial heterogeneity is the inherent functional trait related with overland flow redistribution processes. Within this context, the concept of Soil Surface Components (SSCs) emerged, representing Very-High-Resolution (VHR) hydrogeomorphic response units. SSCs are abstraction entities where spatial patterns of the soil surface and erosional functional processes are linked, according to a large pool of experimental evidence.  

Τhis abstraction complexity, particularly in the abiotic domain, has  so far mandated the use of on-screen visual photointerpretation for the mapping of SSCs, thus limiting the extent of the study cases and their potential for providing answers to the ongoing research discourse. Although significant advances have been achieved with regards to the VHR mapping of vegetation traits with either shallow or deep machine learning algorithms, mapping the full range of SSCs requires bridging the existing gap related with the abiotic domain.

The current confluence of technical advances in: (i) Unoccupied Aerial Systems (UAS), for VHR image acquisition and high geometric accuracy; (2) photogrammetric image processing (e.g. Structure from Motion, SfM), for accurately adding the third dimension, and (3) Deep Learning (DL) architectures that consider the spatial context (i.e. Convolutional Neural Networks, CNN), offers an unprecedented opportunity for achieving the pattern recognition quality required for the automated mapping of SSCs.

We decompose this complex issue with a stepwise approach in an attempt to optimise protocols across all stages of the entire process. For the initial step of image acquisition, we focus on the design of optimal UAS flight parameters, particularly with regards to flight height and image resolution, as this relates to the scale of the analysis: a critical issue for hillslope and catchment scale surveys. At the core of the methodological framework, we then approach the challenge of mapping the patchy mosaic of SSCs as a hierarchical image segmentation problem, decomposed into classification (i.e. discrete) and regression (i.e. continuous fields) tasks, required for dealing with the biotic (e.g. vegetation) and abiotic (e.g. fractional cover of rock fragments) domains, respectively.

Our pilot study area is a hillslope transect near Benidorm, a representative case in semi-arid environment of SE Spain. In this area, the mapping of SSCs was previously undertaken via visual image interpretation. We obtain satisfactory results that allow for the differentiation of plant physiognomies (i.e. annual herbaceous, shrubs, perennial tussock grass and trees). Regarding the abiotic SSCs, in addition to the identification of rock outcrops, we are also able to quantify the fractional cover of rock fragments (RF): an improvement to the visual photointerpretation of only three intervals of RF coverage. A number of challenges remain, such as the position of RF and the transferability of our methodological framework to sites with different lithological and climatological properties.

How to cite: Arnau-Rosalén, E., Pons-Crespo, R., Marqués-Mateu, Á., López-Carratalá, J., Korkofigkas, A., Karantzalos, K., Calvo-Cases, A., and Symeonakis, E.: Automated mapping of Soil Surface Components (SSCs) in highly heterogeneous environments with Unoccupied Aerial Systems (UAS) and Deep Learning: working towards an optimised workflow, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10060, https://doi.org/10.5194/egusphere-egu22-10060, 2022.