Spatial intelligence—the ability to interpret environmental structure and estimate position and motion from visual observations—is essential for navigation in natural environments. This project builds on NAISS 2025/22-1304 (even though it is not formally a continuation) to develop reliable methods for forest conservation mapping and video-based localization, combining learned representations with geometry, physical constraints, and uncertainty assessment. The proposed computational work supports two ongoing funded projects: AI methods for vision-based terrestrial localization (https://research.chalmers.se/en/project/12242), funded by Chalmers Area of Advance Transport for 2026–2027, and Spatial Intelligence and Performance in Orienteering (https://research.chalmers.se/en/project/12656), funded by the Swedish Research Council for Sport Science for 2026. The former develops localization methods using visual observations, inertial data, and public maps to support navigation resilient to satellite-positioning disruptions. The latter investigates how spatial skills in orienteering can be measured and trained using AI-supported video tests, eye-tracking, and geospatial analysis. The proposed environmental interpretation and localization methods provide a computational foundation for these related research directions. For forest conservation mapping, we will investigate how canopy-height structure supports prediction across geographical regions and forest types. Comparisons of 1 m and 2 m resolution will establish the trade-off between predictive performance and computational cost. Geographically separated evaluation areas and continuous tiled inference will assess transferability and consistency, subject to the availability of representative conservation labels. For video-based localization, we will extend a pipeline combining monocular depth, ground segmentation, camera-height-based scale correction, and visual odometry. Experiments will assess variable camera height and mapped landmarks as means of improving metric scale and reducing accumulated trajectory drift. Evaluation on synthetic sequences and real forest videos will measure trajectory error, scale stability, failure frequency, and processing time, accounting for uncertainty in real-world reference trajectories. A smaller complementary activity will continue kidney MRI segmentation experiments on the existing public dataset from 60 subjects, examining task-informed augmentation and confidence-weighted ensembles under mixed distortions. This provides an additional setting for testing whether uncertainty estimates identify spatial prediction errors. NAISS resources will support raster preprocessing, repeated neural-network training, depth inference, and trajectory reconstruction. Held-out evaluation, targeted ablations, and repeated runs will assess robustness and variability. Expected outcomes include validated mapping and localization methods, reproducible workflows, and evidence on the balance between spatial accuracy, generalization, memory use, and processing cost.