Introduced a closed-form least squares initialization method of deep neural networks, reducing training epochs by 30% and improving early accuracy by 12%.

Proposed a backpropagation-free training framework with closed-form, layer-wise optimization, achieving up to 76% lower MSE and improved accuracy over Adam/AdamW on benchmark datasets.

Developed a hybrid CNN–Vision Transformer model with novel feature integration mechanisms, yielding a 10% accuracy improvement in satellite image classification for wetland mapping.

Designed a multiscale deformable CNN achieving a 15% gain in wetland classification accuracy using satellite imagery.

Engineered an unsupervised framework using endmember-driven clustering to overcome the "curse of dimensionality" in hyperspectral datasets with limited labels. The system outperformed supervised SOTA benchmarks by 9%, achieving high precision with as few as five training samples per class.

Engineered an evolutionary AI framework using Particle Swarm Optimization to dynamically segment and integrate high-dimensional hyperspectral bands. This automated approach consistently outperformed seven industry-standard benchmarks.