Competed in Track 1 of the 2nd Hyperspectral Remote Sensing Data Processing and Application Challenge on Kaggle: detecting 18 object classes (people, e-bikes, cars, fruit and their plastic look-alikes, sports balls) in 16-band snapshot hyperspectral images, often dark and noisy street scenes. Competition rules allowed only a single detector, with no ensembles or weighted box fusion.

Built an end-to-end pipeline that demosaics raw mosaic frames into 16-band cubes and adapts COCO-pretrained detectors to 16-channel input. Benchmarked YOLO26, RT-DETR, D-FINE, DINO and Co-DETR under the competition metric; a D-FINE-S model (10.2M parameters) trained with the DEIM recipe won with 0.710 mAP@[.5:.95] out-of-fold. Band-shared photometric augmentation improved detection of the dark street classes.

Found that local validation was anti-correlated with the public leaderboard because test images differ from training images, so final model selection used matched-seed leaderboard probes. Training ran on an RTX 3090 and 4× A100 servers; the best public leaderboard score was 0.658 mAP.