Guneet Mutreja

A blog-style exposition describing AlphaEarth embeddings and how embedding-space vector operations can enable continuous measurement of processes like deforestation and predictive hotspot detection, arguing for a move beyond pixel-centric remote sensing approaches.

Introduces HiRes-FusedMIM, a dual-encoder masked image modeling pre-trained model that jointly leverages high-resolution RGB and DSM data to improve building-level tasks such as classification and segmentation; includes evaluations showing benefits of DSM inclusion and dual-encoder architectures.

Presents a self-supervised learning framework (using EfficientNet with CBAM and SimCLR pretraining on domain-specific aerial datasets) for classifying building roof types, demonstrating strong accuracy and generalization while using fewer parameters than transformer-based models.