Land Cover Classification in Complex and Fragmented Agricultural Landscapes of the Ethiopian Highlands

Abstract

Ethiopia is a largely agrarian country with nearly 85% of its employment coming from agriculture. Nevertheless, it is not known how much land is under cultivation. Mapping land cover at finer resolution and global scales has been particularly difficult in Ethiopia. The study area falls in a region of high mapping complexity with environmental challenges which require higher quality maps. Here, remote sensing is used to classify a large area of the central and northwestern highlands into eight broad land cover classes that comprise agriculture, grassland, woodland/shrub, forest, bare ground, urban/impervious surfaces, water, and seasonal water/marsh areas. We use data from Landsat spectral bands from 2000 to 2011, the Normalized Difference Vegetation Index (NDVI) and its temporal mean and variance, together with a digital elevation model, all at 30-m spatial resolution, as inputs to a supervised classifier. A Support Vector Machines algorithm (SVM) was chosen to deal with the size, variability and non-parametric nature of these data stacks. In post-processing, an image segmentation algorithm with a minimum mapping unit of about 0.5 hectares was used to convert per pixel classification results into an object based final map. Although the reliability of the map is modest, its overall accuracy is 55%-encouraging results for the accuracy of agricultural uses at 85% suggest that these methods do offer great utility. Confusion among grassland, woodland and barren categories reflects the difficulty of classifying savannah landscapes, especially in east central Africa with monsoonal-driven rainfall patterns where the ground is obstructed by clouds for significant periods of time. Our analysis also points out the need for high quality reference data. Further, topographic analysis of the agriculture class suggests there is a significant amount of sloping land under cultivation. These results are important for future research and environmental monitoring in agricultural land use, soil erosion, and crop modeling of the Abay basin.

Optimisation of Savannah Land Cover Characterisation with Optical and SAR Data

“…This was anticipated and explained by both the low coverage of the area by such land cover types (8%), as well as the spectral confusion with the fallow land [32,64]. The latter was also noted by Eggen et al [67] in their Landsat-based Ethiopian study, where they report very low accuracies for the barren class (54% user's and 48% producer's accuracy).…”

“…It is commonly understood that savannah landscapes are difficult to map, mostly due to the low inter-class but high intra-class spectral variability, confounded by seasonal and within-pixel variation [32,67]. Our best performing model, incorporating 130 parameters and estimated from dry and wet season optical and radar data, was able to achieve a high overall accuracy (91.1 ± 1.7%), as well as user's and producer's accuracies ranging from 80% to 98% for woody cover, grasses and crops.…”

Post-Disaster Recovery Assessment with Machine Learning-Derived Land Cover and Land Use Information

“…The variability of urban-rural setting in Tacloban's landscape required a classification algorithm equipped to deal with complex relationships between spectral information and land surface conditions, while being robust with a small training data set. SVM is also known to be specifically well-suited for applications of multispectral imagery [36]. It is less data intensive compared to other machine learning algorithms such as Neural Networks, which require a large amount of training data [35].…”

Dynamic Land Cover Mapping of Urbanized Cities with Landsat 8 Multi-temporal Images: Comparative Evaluation of Classification Algorithms and Dimension Reduction Methods

“…The first method used in this study was based on pixel-based supervised implementation of the SVM algorithm. The SVM uses an iterative learning process to define linear hyperplanes for separation between classes [53]. The initial implementation of the SVM was working as a binary classifier, and a pairwise classification approach was developed to cope with multiclass processing requirements, such as satellite image classification [54].…”