Hyperspectral cameras capture hundreds of colour bands invisible to the human eye, revealing crops, minerals and pollution. Training AI to read these images depends on many choices. This study tests one that is often overlooked: the optimiser.
The problem
Hyperspectral images carry rich spectral and spatial information that supports agriculture, mining and environmental monitoring. Deep learning can classify them, but how the model is trained affects its accuracy and how well it generalises.
What the research did
We used HybridSN, a well-known model that combines 2D and 3D convolutions, as a testbed. We compared numerous optimisers, the algorithms that adjust a network during training, on three standard datasets, measuring overall accuracy, average accuracy and Kappa.
What we found
The optimiser made a clear difference to results. DiffMoment performed best on the Indian Pines dataset, and the best choice varied across datasets, underlining the importance of selecting the optimiser carefully.
Who benefits
- Agricultural and environmental scientists using hyperspectral imaging
- Remote-sensing engineers
- Machine learning practitioners tuning models
Limits and what comes next
The study uses one model architecture and three benchmark datasets; results may differ for other models and real-world imagery.
The team
This research was led by Anish Sarkar, Utpal Nandi and Bachchu Paul (Vidyasagar University) with Sudipta Kr Ghosal, Moirangthem Marjit Singh, Jyotsna Kumar Mandal and me.
Frequently asked questions
What is a hyperspectral image?
An image that records many narrow bands of light, far beyond the red, green and blue the eye sees.
What is an optimiser in AI?
The algorithm that updates a neural network’s settings during training to reduce errors.
Does the optimiser really matter?
Yes. This study found it noticeably changes classification accuracy.