Urban remote sensing: from local to global and back Paolo Gamba University of Pavia, Italy
A few words about Pavia Historical University (1361) in a nice town
slide 3 Geoscience and Remote Sensing Society
Geoscience and Remote Sensing Society
Geoscience and Remote Sensing Society
Outline A brief introduction to urban mapping at the country level Challenges for urban mapping at the national scale National urban mapping using Landsat data National urban mapping applications Conclusions and further work
Why global urban area maps? According to the United Nations, cities occupy around 2% of the total land, however 7
Spatial resolution and urban areas Using free satellite imagery ground spatial resolution on the order of 500 to 30 m are achievable, and continental/regional analyses may be performed. Using multispectral satellites at high resolution (from 10 to 2.5 m posting) we may work on the single town scale and the urban environments. Finally, with Very High Resolution satellites (1 m or less), the single urban element (a building, a street,...) may be individuated and studied.
National-level remote sensing A global view of urban areas may be useful to understand the processes behind urbanization. Trends for land use transformation must be monitored, forecasted and controlled to prevent the degradation of the environment. Economic variable collected nation-wide can be spatially disaggregated using urban areas as proxies (population, industrial/commercial activities, pollution sources, )
URS at town scale At town scale, urban remote sensing may improve the understanding of the town structure, characterizing blocks and changing patterns of land cover/land use. Environmental models at urban scale are increasingly requiring geographically distributed inputs, available using current satellite systems.
VHR Urban Remote Sensing At the most detailed scale, many analysis are possible at the feature level, e.g. buildings, roads, other artificial structures: detecting and modeling these structure for generic location-based service use; Characterizing the three dimensional structure of the urban landscape and providing more sophisticated added value products.
A multiscale fusion problem For urban areas, multiple spatial scale of analysis must be used, and information at each of these scales reconciled so that features extracted at one scale match with their generalization at coarse scales and, at the same time, help infer more refined features at finer spatial resolution(s) P. Gamba, Human settlements: a global challenge for EO data processing and interpretation, Proceedings of IEEE, doi: 10.1109/JPROC.2012.2189089
First challenge: scale
Second challenge: diversity
Third challenge: sparsity
Fourth challenge: data availability
Fifth challenge: processing
Let s focus on urban areas
Landsat archives & Sentinel data Landsat archives cover a long time span, and the whole globe. The Sentinel constellation by ESA is going to complement it Landsat spatial resolution (30 m) may coarse for detailed urban applications, but their granularity and continuity in time is ideal for urban development mapping. The usual approach to capture global maps from multispectral data is the use of indexes (e.g., NDVI, NDSI, ), but each index captures only one specific material. Thus, we need something different, able to cope with other challenges, such as: collected radiances change with seasons and weather conditions; spectral properties of urban materials are not consistent on wide geographic areas; urban material are extremely various even within the same urban area.
The whole procedure Selection of a Landsat scene Selection (and combination) of multiple scenes NDSV computation NDSV ij = (B i -B j )/(B i +B j ) Classification using single or multiple classifiers Wide area classification by combing multiple training sets Spatial regularization Multitemporal map comparison
NDSV Selection of a Landsat scene Selection (and combination) of multiple scenes forest agricultural fields NDSV computation Classification using single or multiple classifiers Wide area classification by combing multiple training sets Spatial regularization Multitemporal map comparison urban water
Multiple classifiers Selection of a Landsat scene Selection (and combination) of multiple scenes NDSV computation Classification using single or multiple classifiers Wide area classification by combing multiple training sets Spatial regularization Multitemporal map comparison Classif. Classif. Classif. Classif. Classif. Map
Sao Paulo: (Multiple) classifier(s) Manual GT INPE
Spatial regularization Selection of a Landsat scene Classification using single or multiple classifiers Spatial regularization NDSV computation Selection (and combination) of multiple scenes Wide area classification by combing multiple training sets Multitemporal map comparison Per-pixel classifiers are prone to salt and pepper noise Standard approach is using spatial regularization (e.g., MRF) GEE: morphological filtering adapted to the mapping scale Morphologi cal opening (erosion + dilation) Structuring element
Regularization via morphology
Multi-temporal data sets Selection of a Landsat scene Selection (and combination) of multiple scenes Classif. Map NDSV computation Classification using single or multiple classifiers Wide area classification by combing multiple training sets Classif. Map Spatial regularization Multitemporal map comparison
At least one year for cloudy skies One year collection versus a combination of quarterly collections
Median/Composite/Greenest Median: collections obtained by considering all scenes on a give area and assigning to the final image the median (more probable) value: good idea to reduce clouds, but it does not work always. A decision level fusion is required. Composite: collections where a cloud score is computed to assess the quality of each scene and aggregate cloud-free pixels from multiple images: slightly better than median filter at reducing cloud cover; slightly worse at the radiometry of the scene; more or less same classification accuracy Greenest: pre-computed yearly collections where the pixel value is selected from the scene where that pixels has the highest NDVI value.
Results: JiangSu province Multiple classifiers and no post-classification Landsat GT VHR GT:
Results: Istanbul 1995 2005 2015
What is this for? Wide area Global (risk, population, exposure, economic, ) analysis Single date Multitemporal analysis (past trends, future developments, ) Finer spatial resolution more details than existing maps
Better time granularity
More details
New information @ 30 grid
Application: population mapping Aggregate Census Blocks to One Administrative Unit Up Process Census Data and Create 10 km Buffered Covariate Boundary, Conformal Projection Extract Vector-Based Covariate Data by Buffered Census Boundary From VMAP0/1 Project and Resample/Aggregate Other Raster Data to Match Census Buffer Rasterize Vector Data, Matching Gridded Census Buffer Supplement or Replace Base Data With Improved Sources Create Distance-To Rasters For Rasterized Point, Line and Binary Class-Based Raster Covariates Aggregate Covariates by Census Blocks and Estimate Random Forest Model Predict Per-Pixel Population Density With Random Forest Algorithm Stevens, F.R., Gaughan, A.E., Linard, C., and A.J. Tatem. Disaggregating census data for population mapping using Random Forests with remotely-sensed and other ancillary data. In Review: Plosone Aggregate Per-Pixel Population to Finer Level Administrative Units and Compare To Afri/AsiaPop GRUMP/GPW Project Population Density Prediction Grid and Redistribute Census Block Totals Weighted by Predictions
Results: the island of Java Landcover Types Number of Trees (RF) #Variables Mean of Squared Residuals % of Variance Explained MDA (2007, Imagery from 2005) 500 8 0.29 81 MDA+MODIS 500 m (2010) 500 6 0.26 83 MDA+GEE (2009) 500 9 0.24 84
Population change maps
Conclusions Remote sensing at the national scale implies a number of challenges: the diversity issue, connected to the spectral and spatial features of urban areas, that may be reduced using a normalized spectral index and exploiting anisotropic textural features; the sparsity issue, and the one due to limited data availability, that may be solved, for the past, using multitemporal data fusion, i.e. exploiting data in different days and, for more recent dates, exploiting different sensors; the computational factor, that is becoming less problematic, because cloud and parallel computing are already the state-of-.the-art of RS data analysis. It is worth noting, however, that remote sensing at the regional/national level opens many new possibilities and applications!