Underwater spatial intelligence confronts a setting in which the two assumptions that anchor terrestrial methods both fail at once: light no longer propagates cleanly, since wavelength-dependent attenuation and scattering degrade vision with range, and satellite positioning does not penetrate water, so absolute position must be inferred from acoustics, inertia, or the structure being surveyed. Figure 14.4 depicts the resulting perceive-localize-map-inspect loop. Navigation and SLAM solve the GPS-denied localization problem, surveyed comprehensively for autonomous underwater vehicles by Paull et al.[2991], with foundational deep-sea visual SLAM demonstrated on the wreck of the Titanic [2992] and modern visual-inertial-pressure localization benchmarked by datasets such as AQUALOC [2993]. Perception and enhancement contend with the optical medium itself: physically grounded models remove the water from an image to recover true scene color and range [2994], learned enhancement restores contrast in real time for downstream perception [2995], and benchmarks such as UIEB standardize the evaluation [2996], while large annotated image databases like FathomNet [2997] enable ocean-scale detection. Mapping and inspection reconstruct the seafloor, habitats, and submerged structures, with structure-relative perception and coverage planning enabling autonomous inspection of complex ship hulls [2998]. The operational layer spans low-cost platforms such as the BlueROV2,1 survey vehicles from Kongsberg2 and research institutions including WHOI3 and MBARI.4
