Page 683 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Mix of data types, including: 4.8: Smart home/
Visual Data: High-resolution Red-Green-Blue (RGB) ortho- cities
mosaics and classified index maps (e.g., NDVI, NHFD).
Remote Sensing Data: High-resolution multispectral images (e.g., World-
view-3), synthetic aperture radar (SAR) imagery from ALOS-2/PALSAR-2.
Geospatial Data: GIS layers (elevation, slope, hydrology, land cover, excavation
Metadata zones) and Global Positioning System (GPS)-based ground control points and
(Type of topographic grids
Data)
Geophysical data: Ground Penetrating Radar (GPR) and magnetic survey results
used for subsurface anomaly detection and validation of remote sensing predictions.
Textual Metadata: Archaeological excavation reports, documen-
tation of processing techniques and methodological details.
Articles have been published in open-access journals. Further data sharing may
be possible upon request.

The use case employs several technologies and approaches for model training
and fine-tuning, including:
Unsupervised Learning: Applied ISODATA and K-means++ clustering to group
multispectral and SAR data based on spectral and spatial similarity without
predefined labels. Used PCA and Kernel PCA for dimensionality reduction and to
identify dominant geospatial patterns. Kernel PCA was combined with a Gaussian
Radial Basis Function (RBF) kernel to enhance predictive modeling of archaeo-
logical zones.
Deep Learning: implemented U-Net (CNN) for semantic segmentation of dune
patterns from labeled satellite imagery. The model was trained on 120 RGB
images using an 80/20 train-test split, with binary cross-entropy loss and binary
Model Train- accuracy as the evaluation metric. Training was stabilized over 30 epochs, achiev-
ing and ing a validation accuracy of 95%.
Fine-Tuning
Processing Frameworks: Model development and training were conducted using
the Keras framework on an NVIDIA A40 Graphics Processing Unit(GPU) within a
high-performance computing environment.
Workflow integrated into ArcGIS Pro for spatial analysis, deployment, and result
visualization.
Model refinement was supported by feedback from field validation, including
results from ground-penetrating radar (GPR) and magnetic surveys, which helped
adjust the model’s predictive accuracy in real-world conditions.
These combined approaches allowed for the effective training and fine-tuning
of models to accurately detect and predict archaeological features using remote
sensing data.

647