I am working on an image classification competition for an internship. Dataset images are segmented buildings pasted on black background. The task is to guess if the top is made of concrete, plastic, and so on for a total of 6 classes. The challenge is images vastly in aspect ratios some very wide some very tall. I padded them on a square sized image keeping the aspect ratio. But I am not getting anywhere. train, Val, test accuracies are in synch this seems too good to be true. Confusion matrix is also not showing any major shift. I have 74% accuracy across three sets + - 2%
But when I submit prediction on unlabeled data to rank I get an f1 score of only 16%. I am using frozen effnetB2. I never worked with satellite images before what am I missing
Satellite Image Classification | Papers With Code
https://paperswithcode.com/task/satellite-image-classification/codeless
Have a look at this work : https://uwescience.github.io/sat-image-analysis/resources.html
and https://www.youtube.com/watch?v=QbgIwOq7ksQ
In my opinion, Radar data can help you to get better results. This paper may help you in this regard:
Land cover mapping of urban environments using multispectral LiDAR data under data imbalance
In the case of accuracy assessment, I strongly suggest you to apply per-pixel assessment. More information is available:
Per-pixel land cover accuracy prediction: A random forest-based method with limited reference sample data
There is no background in a satellite image and every element contains geographical information. To segment and classify such images with higher accuracy, the relevant ground information is required.
The objects of interest can be segmented and classified based on ground information and applying a classifier such as box classifier which uses reference values from the multispectral satellite images. The identified portion of the images can be considered as foreground portion and rest as background. The soft computing approach also be applied for better classification. Bitplane level image characteristics can also be considered as features for image classification.
As satellite images covers a vast area of ground, the structural objects covering larger area such as roads, rivers can be classified with better accuracy.
Try this link:
Satellite Image Classification with Deep Learning
Mark Pritt, Gary Chern
Satellite Image Classification | Papers With Code
https://paperswithcode.com/task/satellite-image-classification/codeless
Ram Ratan sir, The images provided in the dataset are derived from already segmented images. organizer cropped the object of interest from the segmentation mask and pasted it on a pitch-black background. The size of this new image is basically the width and height of the object of interest. hence, datasets have a vast variety of different aspect ratios.
Dear Pushpak Bhoge
This paper may help you to understand the problem
Article Novel Feature Selection and Voting Classifier Algorithms for...
Please read these articles:
Satellite Image Classification with Deep Learning by :Mark Pritt, Gary Chern
Image classification by : Mário Caetano
COMPARING DIFFERENT SATELLITE IMAGE CLASSIFICATION METHODS: AN APPLICATION IN AYVALIK DISTRICT,WESTERN TURKEY.
By: Aykut AKGÜN , A.Hüsnü ERONAT and Necdet TÜRK
Satellite Image Classification Methods and Techniques: A Review
DOI: 10.5120/21088-3779
- Badges
- Science topic
- Similar topics
- Computer Science
- Images