A COMPREHENSIVE ASSESSMENT AND ANALYSIS OF CITY DYNAMIC EVOLUTION AND DRIVING FACTORS FOR DETROIT

Author

• Shu Chen, Southern Illinois University CarbondaleFollow

Date of Award

12-1-2024

Degree Name

Doctor of Philosophy

Department

Environmental Resources & Policy

First Advisor

Wang, Guangxing

Abstract

City dynamics are characterized by urban growth (urban expansion or sprawling) and city decay or shrinking. The former is due to the increase of population and social-economic development, while the latter is because of population decrease and social-economic development slowing down. Substantial research has been conducted in the field of urban sprawling but there have been fewer reports that deal with detection and monitoring of urban decline. Especially, there is a lack of systematical studies to characterize city decay and an approach that can be used to quantify the city decline. Thus, there is a need to overcome the gaps that currently exist in the field of characterizing and monitoring city shrinking.In this study, a methodological framework was proposed based on multi-source remote sensed images and social-economic data by developing a novel mixed training sample-based spectral unmixing analysis method and two new comprehensive assessment indices. The former was used to detect city land use and land cover (LULC) changes, while the latter is utilized to monitor the dynamics of city social-economic aspects. A mixed training sample-based Convolutional Neural Networks (CNN) model was proposed for spectral unmixing analysis. Two comprehensive assessment indices consisted of an urban decline or decay index (UDI) and an image-derived comprehensive assessment index. The mixed training sample-based spectral unmixing analysis method was compared with traditional pure training sample-based spectral unmixing method through four models including multiple linear regression (MLR), random forest (RF), artificial neural network (ANN) and CNN. Moreover, the mixed training sample-based CNN spectral unmixing was integrated with the classification rule based on maximum fractional abundance, which led to a CNN-MMA classification method. The CNN-MMA classification was then compared with traditional maximum likelihood (ML) classification, RF classification, ANN classification, and an integration of pure training sample-based CNN spectral unmixing and the classification rule based on maximum fractional abundance (CNN-PMA) to detect the changes of LULC types. The methodologic framework was examined to detect and monitor Detroit decline from 1985 to 2020 using social-economic data including population, unemployment rate, poverty rate and per capital income (PCI), and multi-source remotely sensed images (Landsat, Sentinel, aerial photographs, etc.), including 400 subjectively selected pure training samples, 400 randomly drawn mixed training samples and 400 validation samples. Several conclusions were drawn. First of all, the methodology successfully detected the changes of Detroit city LULC (water, tree, urban and grassland) types from 1985 to 1990, 2000, 2008, 2010 and 2020 and revealed the city decay. Overall, the urbanized area dominated Detroit but remained similar area estimates during the period of the study time. Tree (including shrub) cover area increased and grassland decreased. The population and PCI decreased but the unemployment and poverty rate increased. Thus, both the UDI and the image-derived comprehensive assessment index showed a trend of the city decline in Detroit. Secondly, the proposed mixed training sample-based spectral unmixing analysis method led to more accurate estimates of water, tree, urban and grassland fractional abundances than the traditional pure training sample-based spectral unmixing method. The improvements were statistically significant at p< 0.5 by reducing the root mean square error (RMSE) of 13.4%, 14.4%, 25.6% and 29.5%, respectively, corresponding to the MLR, RF, ANN and CNN models. In cases of using both the mixed training samples and the pure training samples, the CNN models led to significantly more accurate estimates of the LULC fractional abundances than the MLR, RF and ANN. The mixed training sample-based CNN spectral unmixing model offered the most accurate estimates of the fractional abundances, implying that the mixed training sample-based CNN spectral unmixing method offered the potential of overcoming the limitations of traditional classification methods due to mixed pixels for accurately monitoring the dynamics of the LULC types in the complex city landscape. Thirdly, integrating the mixed training sample-based CNN spectral unmixing model and the rule of maximum abundance resulted in a novel classification method CNN-MMA. The CNN-MMA outperformed the traditional classification methods ML, RF, CNN and CNN-PMA by increasing the overall accuracy of 14% to 29%, 6% to 17%, 6% to 10%, and 2% to 10%, respectively. However, the CNN-MMA classification still led to overestimations of the built-up area and underestimation of tree (including shrub) cover area mainly because this method used the maximum estimates of fractional abundances for hard classification and ignored the areal portions of the minor LULC fractional abundances within the mixed pixels. Moreover, the proposed UDI integrated four social-economic factors including reciprocal of population, unemployment rate, poverty rate and reciprocal of PCI using principal component analysis (PCA) plus entropy with the weights of 0.743, 0.125, 0.124 and 0.08, respectively. This implied that population was most important, then unemployment rate and poverty rate, and PCI. The study also resulted in an image-derived comprehensive assessment index by integrating normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI) and modified normalized difference water Index (MNDWI) that contained the information of vegetated area, built-up area and water area. A statistically significant correlation between UDI and the image-derived comprehensive assessment index was found. Both the UDI and the image-derived comprehensive assessment index captured the spatial patterns and temporal trends of the decline in Detroit from 1990 to 2020. The city decay was especially obvious in the central and south parts and during the period of time from 2010 to 2020 due to the COVID-19 pandemic that greatly increased the unemployment and poverty rate. Overall, this study highlighted the benefits of integrating remote sensing images with socioeconomic data to comprehensively assess the city decline of Detroit. The proposed methodological framework overcame the gaps that currently exist in the field of detecting and monitoring urban shrinking and provided the potential of developing a near real-time system for comprehensively monitoring the dynamics of the city in terms of both physical and social-economic changes.