# what is the best approach in dealing with large dimension custom data for training and predicting deep learning models

i am trying to implement semantic segmentation for satellite images.My custom dataset has dimensions(height,width)in range (3000, 3000)what is the best approach for feeding(for training) and predicting on this dataset. Should i take random cropping or use resizing to a target size or something else? Will i loose resolution in case of resizing? I am asking because pretrained networks(like unet) have height and width in the range 300-400 only.

• That's not a lot of information about the problem and its inherent all-important signal:noise ratio. If you are in the typical signal:noise situation, the only hope to having any kind of stability is to do a great deal of data reduction (unsupervised learning) before doing any supervised learning (analysis making use of $Y$). – Frank Harrell May 9 '19 at 11:21
• @FrankHarrell --what more information should i provide? Also could you explain what you mean by data reduction through unsupervised learning? – Ankit Sharma May 9 '19 at 11:48
• Give a sense of whether you are predicting something that is mechanistic or has a strong random component. Unsupervised learning example methods: principal components analysis, sparse principal components, variable clustering followed by principal component scoring etc. I have a chapter of this in RMS. – Frank Harrell May 9 '19 at 17:17
• @FrankHarrell I am predicting on high resolution satellite images. Trying to segment them into ground classes like vegetation trees, buildings, roads,etc – Ankit Sharma May 11 '19 at 9:45
• Although there is a sample size issue here, if the underlying signal:noise ratio (e.g., $R^2$) is very large, you may be able to use an aggressive machine learning algorithm on all the features if you do rigorous validation (e.g., 100 repeats of 10-fold cross-validation). – Frank Harrell May 11 '19 at 12:07