This is an invited talk organised by the Business and Industrial Section of the Royal Statistical Society. The talk will follow immediately after the Annual General Meeting of the section. The talk will be of particular interest to both academics and practitioners working in the areas of precise agriculture and geospatial intelligence, sensors and satellite imagery applications.
Speaker: Professor Clement Atzberger, Mantle Labs, University of Natural Resources and Life Sciences Vienna
The demand for reliable agricultural information such as crop acreage and crop health status is continuously increasing. The estimates need to be updated frequently, provided early in the growing season and allow for a global coverage at low costs. These requirements can only be met by satellite-based Earth Observation (EO) data. The task of EO-based crop identification is well understood but nevertheless challenging compared to a ground-based (visual) inspection and identification, as the satellites do not capture highly discriminative visual clues such as leaf size, growth form, fruits, flowers etc which are used by humans to identify a given species. Instead, EO techniques have to rely solely on the spectral signature of the observed pixel, which is recorded as the integral within a pixel much larger than an individual plant (e.g. 10 x 10 m for the most widely used sensors). This spectral signature, and its temporal evolution, is itself determined by bio-physical determinants such as leaf area index (e.g. total leaf surface area within the pixel), leaf pigmentation, chlorophyll and water content, and only slightly modified by the mentioned visual clues. Hence, the pheno-typic expression of these bio-physical variables - including their co-variation over time - determine what one is able to record from space. This implies that crop species can only be distinguished from each other if they have a crop-specific setting and/or evolution of these bio-physical variables. Practitioners have moreover to deal with complex observation conditions, such as changing sun and view angles and the fact that images are often (partly) cloudy. Cloud related problems increase with the size of the study area, as the probability of cloud-free conditions for any particular region on Earth is inversely related to the size of the area that needs to be monitored. Leveraging ideas taken from the one-pixel camera concept, and taking advantage of the well known sparsity of the multi-temporal and multi-spectral imagery, we demonstrate how cloud-agnostic features can be extracted from cloudy time series and used for crop type identification and the reconstruction of gap-free time series.
Professor Clement Atzberger
Clement has more than 25 years experience in Remote Sensing and is responsible for the core remote sensing and solution design at Mantle Labs. He is a Full Professor at BOKU university (University of Natural Resources and Life Sciences Vienna), and Head of the Institute of Geomatics with long-lasting experience in Earth Observation and crop monitoring. He has published a large number (>100) of SCI publications and is Editor-in-Chief of the Section “Agriculture and Vegetation” of the renowned Remote Sensing journal (MDPI) – his current h-index is 40 with >1000 new citations every year.