Harnessing hyperspectral imaging and deep learning for terrestrial habitat mapping in arid landscapes: A case study in Saudi Arabia
Ali Elgendy, Hesham Morgan, Brandon Tran, Rejoice Thomas, Tamer Ismail, Yehya Kh. Shehadeh, Ahmed ElGharib, Ahmed Abdullah Al-Dughairi, Ali El Mubarak, Khaled Allam Harhash, Hesham El‐Askary
Abstract
Arid ecosystems remain under-mapped at actionable scales despite their ecological importance. Decision makers lack reliable, high-resolution habitat maps in drylands to prioritize protection and target restoration. This research integrates spaceborne hyperspectral imaging from the Environmental Mapping and Analysis Program (EnMAP) with deep learning semantic segmentation models to produce an updated level of habitat classification based on the International Union for Conservation of Nature (IUCN) for part of the Imam Turki bin Abdullah Royal Reserve, Saudi Arabia. Using ground control points and the full EnMAP spectral cube without band selection, U-Net and DeepLabV3+ architectures were each implemented with VGG19 and ResNet-101 encoder backbones, resulting in four model configurations for comparative evaluation. Among the tested models, U-Net with a VGG19 backbone achieved the highest performance, attaining an F1 score of 0.90, demonstrating superior capability for habitat mapping in arid environments. The resulting map segmented the dendritic wadi network, Rawdat depressions, and sand-plateau contacts with sharp boundaries. Aligned with the IUCN habitat scheme, Rawdat, seasonal vegetation-bearing desert carbonate sinkholes, are introduced as a new IUCN habitat class. The map produces verifiable indicators relevant to Sustainable Development Goals 13 (Climate Action) and 15 (Life on Land), supporting protection, restoration targeting, and monitoring. Therefore, the proposed deep learning hyperspectral framework can be applied to other arid and semi-arid regions worldwide to upgrade terrestrial ecosystem mapping and conservation planning. • EnMAP with U-Net yield precise arid habitat map using all 224 hyperspectral bands. • U-Net-VGG19 surpassed both DeepLabV3+ variants and U-Net-ResNet101 (F1 = 0.90). • Rawdat is introduced as a distinct, IUCN-aligned Level-3 habitat class. • Map captures Wadis, Rawdat depressions, Sand-Plateau contacts with sharp edges.