Wele Gedara Chaminda Bandara, Jeya Maria Jose Valanarasu, Vishal M. Patel

[Paper Link] (Preprint submitted to IEEE Transactions on Geoscience and Remote Sensing)

Hyperspectral pansharpening aims to synthesize a low-resolution hyperspectral image (LR-HSI) with a registered panchromatic image (PAN) to generate an enhanced HSI with high spectral and spatial resolution. Recently proposed HS pansharpening methods have obtained remarkable results using deep convolutional networks (ConvNets), which typically consist of three steps: (1) up-sampling the LR-HSI, (2) predicting the residual image via a ConvNet, and (3) obtaining the final fused HSI by adding the outputs from first and second steps. Recent methods have leveraged Deep Image Prior (DIP) to up-sample the LR-HSI due to its excellent ability to preserve both spatial and spectral information, without learning from large data sets. However, we observed that the quality of up-sampled HSIs can be further improved by introducing an additional spatial-domain constraint to the conventional spectral-domain energy function. We define our spatial-domain constraint as the $L_1$ distance between the predicted PAN image and the actual PAN image. To estimate the PAN image of the up-sampled HSI, we also propose a learnable spectral response function (SRF). Moreover, we noticed that the residual image between the up-sampled HSI and the reference HSI mainly consists of edge information and very fine structures. In order to accurately estimate fine information, we propose a novel over-complete network, called HyperKite, which focuses on learning high-level features by constraining the receptive from increasing in the deep layers.

@inproceedings{bandara2021hyperspectral,
title={Hyperspectral Pansharpening Based on Improved Deep Image Prior and Residual Reconstruction},
author={Bandara, Wele Gedara Chaminda and Valanarasu, Jeya Maria Jose and Patel, Vishal M},
journal={arXiv preprint arXiv:2107.02630},
year={2021}
} 

In this paper, we used three publically available datasets and the link to each dataset is given below:

1. Pavia Center Dataset: Click Here
2. Botswana Dataset: Click Here
3. Chikusei Dataset: Click Here

Once you downloaded the dataset in .mat format save them in respective folders: ./datasets/pavia_centre/, ./datasets/botswana/, and ./datasets/chikusei/.

The requirements.txt file lists all Python libraries required for DIP-HyperKite. To install requirements first create a new conda environment and then install all required libraries using following commands.

conda create --name hyperkite

conda activate hyperkite

conda install --file requirements.txt

Next, we generate LR-HSIs, PAN images, and Ref-HSIs required to train the pansharpening model using the famous Wald's protocol. For this, you simply needs to run the process_pavia.m, process_botswana.m, and process_chikusei.m files in the ./datasets/pavia_centre/, ./datasets/botswana/, and ./datasets/chikusei/, respectively.

To generate the up-sampled version of LR-HSIs, you need to run the following code. Please make sure to change the first few lines of the ./configs/config_dhp.json as you want. Basically you want to change the experiment_name and dataset you want to run. CUDA_VISIBLE_DEVICES=0 python train_dhp.py --config ./configs/config_dhp.json

You need to repreat this for all the three datasets.

Once you generated the DIP upsampled HSIs, now you are ready to train the HyperKite network. You can train the HyperKite network by executing the following command. CUDA_VISIBLE_DEVICES=0 python train.py --config ./configs/config.json

All the results will be saved in ./Experiments/ folder. You can visualize all the performance metrics by executing the following command. tensorboard --logdir ./Experiments/Vxx/pavia_center/