Spectral discrimination and index development of roofing materials and conditions using field spectroscopy and worldview-3 satellite image

Monitoring roofing materials and conditions are important to improve urban management and support the well-being of an urban environment. However, it is difficult to map the within-class roofing surfaces in the heterogeneous urban environment using multispectral remotely sensed data due to a wide-range of materials exist. Furthermore, restriction in multispectral bands has led to spectral confusions when mapping the within-class roofing types. Normal supervised pixel-based classification scheme is often applied to map urban land use and land cover which the accuracy is dependent on the training site information. Hence, the development of new approach is indispensable to improve the within-class roofing materials mapping. A technique combining the information of spectral and spatial characteristic could be considered as an effective way to improve urban analysis and overcome issues of limited spectral band and coarse spatial resolution of remote sensing data. Therefore, this study utilize the combination of field spectroscopy hyperspectral data and very high resolution multispectral data of WorldView-3 (WV-3) to map the roofing materials and conditions.Hyperspectral data has the ability of providing adequate spectral information for discriminating within-class of roofing materials and conditions. This study utilizes field spectroscopy data as fundamental on analysing roofing spectral signature instead of using airborne hyperspectral data due to the expensive cost of data acquisition. However, handling hyperspectral data required effective method to reduce redundancy of data, yet maintaining the useful information. This research investigates a feature selection technique to discriminate between four different types of roof materials (i.e.: asbestos, concrete, clay and metal) and their conditions by using field spectroscopy within the range of 350 nm – 2500. Three feature selection algorithms of Genetic Algorithm (GA), Support Vector Machine (SVM) and Random Forest (RF) were used to select the most significant wavelengths since the algorithms works well with large size of data and widely applied for feature selection of hyperspectral remote sensing data. Results from feature selection specify that visible, Shortwave Infrared-1 (SWIR-1) and SWIR-2 (SWIR-2) spectral regions are important for roof materials and conditions separation. Comparatively, overall accuracy obtained from GA, SVM and RF algorithms are fairly high in percentage with GA and SVM both produced 96.3%, while RF yield 97.53% accuracy. Generally, the highest accuracy is produced using RF feature selection (97.53%), hence, describe the efficiency of RF algorithm for feature selection task using field spectroscopy data. Accuracy of spectra without feature selection was also investigated and the result was lower compared with classification using significant wavelengths. Result using all wavelengths mostly recorded 44.44%, while result using significant wavelengths mostly produced 100% accuracy. Therefore, the findings described the importance of selecting significant wavelengths to improve the spectral classification accuracy.Additionally, new spectral indices of Normalized Difference Concrete Condition Index (NDCCI) and Normalized Difference Metal Condition Index (NDMCI) have been developed in this study for detecting concrete and metal roofing condition status. Significant wavelengths located at visible to near infrared spectral region were used as basis for developing spectral indices to be applied onto very high resolution satellite imagery of WV-3 satellite data. The classification accuracy using spectral indices were compared with the normal supervised pixel-based classification of SVM. Results show that the spectral indices produce higher accuracy compared to SVM classification with NDCCI produced 84.44% compared to SVM classification of concrete condition accuracy by 73.06%. NDMCI produced 94.17% accuracy which is higher compared to SVM classification of metal condition accuracy of 62.5%. Therefore, the results indicate that the application of the developed spectral indices is effective for mapping roofing conditions in the heterogeneous urban environment. Feasibility of the spectral indices developed were assessed by validating the indices on second study area, generating results of 72.06 for NDMCI and 70.3% accuracy for NDCCI. Lower accuracy obtained could be due to different study areas; the first study area is residential and commercial area, meanwhile the second study area is university area. Generally, both spectral indices could be considered for roofing materials and conditions detection application, however NDMCI perform better than NDCCI.Generally, the findings of this study describes the importance of applying feature selection as a preliminary process before developing a spectral index to eliminate uninformative data and enhance the separation between impervious surfaces. Spectral index of NDCCI and NDMCI found to be effective in providing roof degradation status map in effective time-manner and parameter-free algorithm compared to normal supervised classification scheme. However, the proposed technique has some limitations which the classification is depended on roof brightness, hence, lead to misclassification between new metal roofs and new concrete roofs. This approach also required removal of natural land cover such as vegetation and water bodies to provide better delineation of impervious surfaces. Therefore, further improvement is needed to provide better classification accuracy for delineation of roofing types. The output from this study could contribute into new insight of urban planning and monitoring and sensor development for urban mapping