Urban land use and land cover classification using ENVINet-5: a comparative analysis

In recent years, Deep Learning (DL) methods have found extensive application in remote sensing and urban analysis. Software packages like ENVI, now offer the off-the-shelf DL Model called ENVINet-5, simplifying DL model training and spatial data extraction without requiring programming expertise. To evaluate and compare ENVINet-5's information extraction capabilities and classification performance with more conventional methods, including Spectral Angle Mapper (SAM), Random Forest (RF), and Support Vector Machine (SVM), this study conducted a comparative assessment in the rapidly developing port city of Mombasa, Kenya. The assessment used high spatial resolution WorldView-3 and medium spatial resolution Sentinel-2 images, along with diverse land cover and land use class levels. Results suggested ENVINet-5's classification performance in urban analysis using Sentinel-2 images was below expectations, with less than 60% accuracy compared to 78.0%, 89.2% and 79.9% accuracy achieved by SAM, RF, and SVM algorithms, respectively. While ENVINet-5 outperformed RF and SVM when applied to high spatial resolution 8-bit pixel depth WorldView-3 images and refined class design, high numbers of misclassified and unclassified pixels were observed. Downscaling radiometric resolution during correction in WorldView-3, coupled with the exceptionally detailed information it provides for the large-scale site, could potentially limit the capabilities of ENVINet-5.