University of Twente Student Theses


Mapping rrable field fractions with multisensor remote sensing data-driven gradient boosted and classical GAMs

Gragn, Yebelay Gonfa (2021) Mapping rrable field fractions with multisensor remote sensing data-driven gradient boosted and classical GAMs.

Link to full-text:
(only accessible for UT students and staff)
Abstract:Information on crop production estimates is the basis for supporting the current and future food security initiatives, especially for developing countries. However, for most developing countries obtaining crop production estimates is a challenge due to several reasons. One of the reasons is a challenge in identifying and extracting information about the extent and location of agricultural areas. These agriculture areas have different characteristics that make them challenging to quantify their extent. For instance, in Ethiopia's Oromia region, the arable fields are characterized by small size, irregular shape, often trees inside fields, irregular cropping patterns, and heterogeneity in weather conditions. These challenges increase the uncertainty in the delineation of the field extent. In this research, a combined method is developed to map arable field fractions with opensource earth observation data that minimize the uncertainty in estimating the field extent. Gradient boosted and Classical GAM models used with Sentinel-1 backscatter matrices, Sentinel-2 optical, topographic features, and hyper temporal images (i.e., Proba-V). The hyper-temporal imagery is used primarily to extrapolate a 1km NDVI (i.e., a 1km arable field fraction map is extrapolated and used as an input variable for the model). The hyper-temporal images are also used for identifying the wet and dry seasons for downloading Sentinel-1&2 image features. Eight Sentinel-1 image features (i.e., dry and wet season VV (Vertical transmit, Vertical receive), VH (Vertical transmit, Horizontal receive), VV/VH ratio, and NRMP) and eleven Sentinel-2 optical image features (i.e., three red-edge bands, 2 SWIR, two dry and wet season NDVI, two dry and wet season Normalized Difference Tillage Index (NDTI), and two dry and wet season Land Surface Wetness Index (LSWI)) are used in the model. In addition to Sentinel-1&2 image features, topographic variables (i.e., elevation, slope, relative DEM, and topographic wetness index) are included in the model. Gradient boosted regression is used to select the most important predictor variables, and the Classical GAM is used to predict arable field fractions from these important predictor variables. Based on the boosted GAM model and stability selection, six informative variables (i.e., dry season VH, elevation, red edge (Band-5), dry season VV/VH ratio, Slope, and a 1km arable field estimate) out of twenty-four explanatory variables are selected. The overall deviance of the model was 87%. The partial deviance explained by Sentinel-1 dry season VH was 33.3% which is the most explanatory variable in discriminating arable field fractions. The partial deviance of elevation, Band-5, 1km arable field fractions, slope and dry season VV/VH ratio was 17.2%, 14.4%,13.3%,6.2% and 3.34% respectively. Classical GAM is fitted with the most informative variables selected using the Gradient boost and stability selection method. Finally, a 20m arable field fraction map was extrapolated for the Oromia region. The developed method can be applied to extrapolate 20m arable field fractions for the rest regional states of Ethiopia and country-level wall-to-wall mapping by considering agroecological variations.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page