Climate-driven disasters like landslides are increasing in frequency, but the AI models we use to map them are becoming environmentally and financially unsustainable. Modern Earth observation teams face a massive bottleneck: processing high-resolution topographical data requires immense, energy-hungry cloud compute and expensive GPUs.

In this talk, I will break down a practical "Green AI" engineering solution to this compute crisis. I will introduce a U-Net++ channel-selection framework, recently published in Frontiers in Remote Sensing, that acts as an intelligent data filter, stripping away redundant topographical channels before they ever reach the GPU.

By optimizing the data pipeline, this framework successfully reduces GPU VRAM requirements by 40% without sacrificing mapping accuracy. This approach does more than just lower the hardware barrier for underfunded disaster response teams; it actively reduces the carbon footprint and energy consumption of deploying complex spatial models and also makes it faster for a disaster-response scenario.

Attendees will walk away understanding:
The environmental and financial toll of unnecessarily bloated spatial segmentation models.
The architecture behind the VRAM-saving channel-selection filter.
How to apply these techniques to build leaner, greener geospatial models that can run on accessible, low-power hardware during time-sensitive climate emergencies.