University of Twente Student Theses


Mapping probabilities of arable fields using modis, sentinel-1 and sentinel-2 based image features in Ghana

Trivedi, Manushi Bhargav (2020) Mapping probabilities of arable fields using modis, sentinel-1 and sentinel-2 based image features in Ghana.

[img] PDF
Abstract:To cope up with the current demand and supply chains of agricultural products, monitoring the food production capacity via the national or regional level agricultural database on forecasted crop productions is necessary. The location and extent of cultivated cropped area is the critical input for mapping precise crop yields. Current remote sensing methods addressing the combined use of the open-source satellite data are very few and are efficient for northern countries but cannot be adopted where Arable Fields Area (AFA) is typically characterized as irregularly shaped and difficult to distinguish. Recent remote sensing-based available global land cover products are inconsistent across the globe. They do not address the complexity of agriculture landscapes, and it majorly focuses on the use of high or very high spatial resolution images (<5m). Hence, this research has focused on identifying the potential of freely available earth observation data based satellite image features for estimating probabilities of AFA in the African agricultural landscape (Eastern region in Ghana as a case study). Higher temporal MODIS images are used for capturing long-term vegetation climatology pattern to stratify the landscape. Based on which dry and wet seasons and strata with cropping intensity via Crop Productive Zones (CPZs) on the regional level has been derived. It has addressed the landscape heterogeneity at the pixel level via homogenous stratification. The use of the median composite of Sentinel-1 (S1) and Sentinel-2 (S2) images across the dry and wet seasons and, over the years (2017-2019), is exploited due to its spatial, temporal, spectral, and polarimetric capability to differentiate AFA with other vegetated and non-vegetated areas. In lines with it, topographical and textural image features are also examined for explaining additional local and regional level arable field distribution. In this study, one temporal (CPZs), two topographic (Slope and Elevation), 14 spectral (optical and red-edge vegetation indices), ten polarimetric (Dry and Wet VV, VH, VV/VH, VV+VH, VV-VH) and 110 texture (Dry and Wet S1 & S2 variance, homogeneity, dissimilarities, entropy, contrast) image features have been studied extensively. The relevant image features have been identified for mapping AFA and its probabilities have been mapped using the RandomForest (RF) algorithm. A total of 36 important features have been selected, out of which 33 features are texture features (majorly Variance texture), one is topographic (Elevation), one is temporal (CPZs), and one is polarimetric (Dry VV) image feature. Where topographic and texture features, in general, improve prediction by reducing 0.14 to 0.10%, temporal CPZs feature reduces 0.05%, and polarimetric feature reduces 0.04% error in Brier Score (BS). In general, the topographic elevation and, optical and radar-based texture feature outperformed the spectral features. It also outperformed a temporal and polarimetric image features in a way as well. It is also important to note that the challenge of extreme cloud cover in optical images have been addressed via the median composite of the images over long-term (three years) (2017-2019) and seasonal changes (different dry and wet period) for the different region within the study area have been identified and implemented well. However, the performance of RF for predicting the extremes probabilities is skeptical or leveraged at extreme points, and it needs further improvement via the change in sampling design and image processing (especially image integration).
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