This repo is an illustration of the model training details for the Kaggle Competition:DSTL Satellite Imagery Feature Detection. I won a Bronze Medal in this competition.
The goal of this competition is to detect and classify 10 types of objects in the given regions(1km*1km) of satellite imagery provided by the Defence Science and Technology Laboratory (Dstl). The dataset consists of 450 images, 25 of them have training labels. The satellite images are provided in both 3-band and 16-band formats. The 16-band format includes Panchromatic (450-800 nm), 8 Multispectral (red, red edge, coastal, blue, green, yellow, near-IR1 and near-IR2)(400 nm - 1040 nm) and 8 SWIR (1195 nm - 2365 nm). In this competition, information of all 20 bands are utilized in training the model.
The idea of U-net is applied in this competition for training the CNN for feature detection of satellite imagery. U-net was first developped by a German research team from University of Freiburg for Biomedical Image Segmentation. U-net then shows its perfomance for image segmention problem in the Kaggle competition Ultrasound Nerve Segmentation. In this competition, I trained different CNNs based on the idea of U-net and got good results.
The input data can be downloaded here.