My PROJECTS

IEEE IGARSS Data Fusion Contest 2021

Project 1 : A MULTI-MODAL FUSION APPROACH FOR SEMANTIC CHANGE DETECTION

Collaborators: Khishma Modoosoodun Nicolas and Stephanie Tumampos

As part of my study at the University of South Brittany I was privileged to participate in the data fusion contest for the year 2021 organized by IEEE Geoscience and Remote Sensing Society (GRSS) in collaboration with Microsoft.

Aim: This study will focus on the automatic land cover change detection from multitemporal, multiresolution and multispectral satellite imagery. The project will identify the gain and loss of 4 classes to identify the changes between two images, 2013 and 2017. These changes are gain and loss of water, tree canopy, low vegetation and impervious surfaces.

Data: The chosen area of study and interest is Maryland in the United States of America. The dataset is composed of data layers with 2250 tiles of 3800 by 3800 pixels.

Methods: UNET architecture is used to perform multi-modal fusion of NAIP and Landsat-8 datasets for semantic change detection. We trained UNET model experiments with Fusion and without Fusion. Training was performed for 10 epochs using all the tiles and validation was performed using 50 tiles.

Architecture Design

Figure 1. Architecture design

The UNET architecture was modified to perform fusion of

the two images that is NAIP and Landsat-8. The network has two input layers, since Landsat-8 has very low resolution the the NAIP image chip size was 240 by 240 and for LS-8 was 8 by 8 to for the kernels to extract features in the same location. The rest of the architecture consist of a series of convolutions and max-pooling in the encoder part, followed by fusion layer then afterwards we have the up sampling in the decoder part to make predictions of different land cover classes. This predictions are made for both years 2013 and 2017 and to get change maps we substract 2017 predictions minus 2013 predictions.

Results:

From our experiments we carried out. The highest iou score achieved was 0.39 and figure 2 shows the results for the predictions from our experiments. Ut both refers to training model for both years at the same time and Unet separate refers to training the the model with the two years separately

Figure 2. Predictions

Conclusion

To improve the accuracy, further studies on how to perform data augmentation techniques on the data before training the model can be explored to improve the regularization and reduce overfitting.

Project 2: Classification of Hyperspectral Images using Deep CNN