Projects
Siamese-BiMamba
Attention-free linear-time O(N) State Space Model for bitemporal satellite building damage assessment. Combines a shared Bidirectional Mamba encoder with TemporalMix gated local fusion. Achieved 82.86% Damage F1 / 91.23% AUPRC on xBD at just 1.86M parameters.
PCAE (Partial Convolution Autoencoder)
1D Convolutional Autoencoder for hyperspectral anomaly detection in food safety inspection. Integrates a two-phase self-supervised + one-class learning framework with memory-bank nearest-neighbor scoring. Achieved 87.75% mean AUPR using 13.6x fewer parameters than SOTA baselines.
SteelCross-Net
Dual-backbone hybrid model fusing Swin-Large (global context Query) and ConvNeXt-Large (local texture Key/Value) via Multi-Head Cross-Attention. Fused with 18 deterministic statistical features (GLCM, FFT, spatial gradients) acting as latent anchors against cluster fragmentation.
Multi-Task Cassava Disease Identification
1D hyperspectral MTL framework coupling spectral disease classification with a self-supervised Masked Autoencoder (MAE) reconstruction auxiliary task. Introduces Modified Focal Loss (MFL) with soft labels to tackle severe class imbalance on 3,600+ band spectrometry data.
MarineDex
Fine-grained marine species classification pipeline and mobile application built on MobileNetV2 feature extraction and Hyperband tuning. Reached 92.34% validation accuracy across 16 marine taxa classes.