Machine learning algorithms such as Random Forest (RF) are being increasingly applied on traditionally geographical topics such as population estimation. Even though RF is a well performing and generalizable algorithm, the vast majority of its implementations is still 'aspatial' and may not address spatial heterogenous processes. At the same time, remote sensing (RS) data which are commonly used to model population can be highly spatially heterogeneous. From this scope, we present a novel geographical implementation of RF, named Geographical Random Forest (GRF) as both a predictive and exploratory tool to model population as a function of RS covariates. GRF is a disaggregation of RF into geographical space in the form of local sub-models. From the first empirical results, we conclude that GRF can be more predictive when an appropriate spatial scale is selected to model the data, with reduced residual autocorrelation and lower Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) values. Finally, and of equal importance, GRF can be used as an effective exploratory tool to visualize the relationship between dependent and independent variables, highlighting interesting local variations and allowing for a better understanding of the processes that may be causing the observed spatial heterogeneity.

No abstract

In this letter the recently developed Extreme Gradient Boosting (Xgboost) classifier is implemented in a veryhigh-resolution (VHR) object-based urban Land Use-Land Cover application. In detail, we investigated the sensitivity of Xgboost to various sample sizes, as well as to feature selection (FS) by applying a standard technique, Correlation Based Feature Selection. We compared Xgboost with benchmark classifiers such as Random Forest (RF) and Support Vector Machines (SVM). The methods are applied to VHR imagery of two Sub-Saharan cities of Dakar and Ouagadougou and the village of Vaihingen, Germany. The results demonstrate that, Xgboost parametrized with a Bayesian procedure, systematically outperformed RF and SVM, mainly in larger sample sizes.

This study presents the development of a semi-automated processing chain for urban object-based land-cover and land-use classification. The processing chain is implemented in Python and relies on existing open-source software GRASS GIS and R. The complete tool chain is available in open access and is adaptable to specific user needs. For automation purposes, we developed two GRASS GIS add-ons enabling users (1) to optimize segmentation parameters in an unsupervised manner and (2) to classify remote sensing data using several individual machine learning classifiers or their prediction combinations through voting-schemes. We tested the performance of the processing chain using sub-metric multispectral and height data on two very different urban environments: Ouagadougou, Burkina Faso in sub-Saharan Africa and Liège, Belgium in Western Europe. Using a hierarchical classification scheme, the overall accuracy reached 93% at the first level (5 classes) and about 80% at the second level (11 and 9 classes, respectively).

Up-to-date and reliable land-use information is essential for a variety of applications such as planning or monitoring of the urban environment. This research presents a workflow for mapping urban land use at the street block level, with a focus on residential use, using very-high resolution satellite imagery and derived land-cover maps as input. We develop a processing chain for the automated creation of street block polygons from OpenStreetMap and ancillary data. Spatial metrics and other street block features are computed, followed by feature selection that reduces the initial datasets by more than 80%, providing a parsimonious, discriminative, and redundancy-free set of features. A random forest (RF) classifier is used for the classification of street blocks, which results in accuracies of 84% and 79% for five and six land-use classes, respectively. We exploit the probabilistic output of RF to identify and relabel blocks that have a high degree of uncertainty. Finally, the thematic precision of the residential blocks is refined according to the proportion of the built-up area. The output data and processing chains are made freely available. The proposed framework is able to process large datasets, given that the cities in the case studies, Dakar and Ouagadougou, cover more than 1000 km 2 in total, with a spatial resolution of 0.5 m.

To classify Very-High-Resolution (VHR) imagery, Geographic Object Based Image Analysis (GEOBIA) is the most popular method used to produce high quality Land-Use/Land-Cover maps. A crucial step in GEOBIA is the appropriate parametrization of the segmentation algorithm prior to the classification. However, little effort has been made to automatically optimize GEOBIA algorithms in an unsupervised and spatially meaningful manner. So far, most Unsupervised Segmentation Parameter Optimization (USPO) techniques, assume spatial stationarity for the whole study area extent. This can be questionable, particularly for applications in geographically large and heterogeneous urban areas. In this study, we employed a novel framework named Spatially Partitioned Unsupervised Segmentation Parameter Optimization (SPUSPO), which optimizes segmentation parameters locally rather than globally, for the Sub-Saharan African city of Ouagadougou, Burkina Faso, using WorldView-3 imagery (607 km2). The results showed that there exists significant spatial variation in the optimal segmentation parameters suggested by USPO across the whole scene, which follows landscape patterns—mainly of the various built-up and vegetation types. The most appropriate automatic spatial partitioning method from the investigated techniques, was an edge-detection cutline algorithm, which achieved higher classification accuracy than a global optimization, better predicted built-up regions, and did not suffer from edge effects. The overall classification accuracy using SPUSPO was 90.5%, whilst the accuracy from undertaking a traditional USPO approach was 89.5%. The differences between them were statistically significant (p < 0.05) based on a McNemar’s test of similarity. Our methods were validated further by employing a segmentation goodness metric, Area Fit Index (AFI)on building objects across Ouagadougou, which suggested that a global USPO was more over-segmented than our local approach. The mean AFI values for SPUSPO and USPO were 0.28 and 0.36, respectively. Finally, the processing was carried out using the open-source software GRASS GIS, due to its efficiency in raster-based applications.

No abstract

Land cover Classified maps obtained from deep learning methods such as Convolutional neural networks (CNNs) and fully convolutional networks (FCNs) usually have high classification accuracy but with the detailed structures of objects lost or smoothed. In this work, we develop a methodology based on fully convolutional networks (FCN) that is trained in an end-to-end fashion using aerial RGB images only as input. Skip connections are introduced into the FCN architecture to recover high spatial details from the lower convolutional layers. The experiments are conducted on the city of Goma in the Democratic Republic of Congo. We compare the results to a state-of-the art approach based on a semi-automatic Geographic object image-based analysis (GEOBIA) processing chain. State-of-the art classification accuracies are obtained by both methods whereby FCN and the best baseline method have an overall accuracy of 91.3% and 89.5% respectively. The maps have good visual quality and the use of an FCN skip architecture minimizes the rounded edges that is characteristic of FCN maps. Additional experiments are done to refine FCN classified maps using segments obtained from GEOBIA generated at different scale and minimum segment size. High OA of up to 91.5% is achieved accompanied with an improved edge delineation in the FCN maps, and future work will involve explicitly incorporating boundary information from the GEOBIA segmentation into the FCN pipeline in an end-to-end fashion. Finally, we observe that FCN has a lower computational cost than the standard patch-based CNN approach especially at inference.