DeepGlobe Land Cover Classification Challenge

English | 简体中文

DATASET

DATA

• The training data for Land Cover Challenge contains 803 satellite imagery in RGB, size 2448x2448.
• The imagery has 50cm pixel resolution, collected by DigitalGlobe's satellite.
• You can download the training data in the download page with filetype of “Starting Kit”. Testing satellite images will be will be uploaded later.

Label

• Each satellite image is paired with a mask image for land cover annotation. The mask is a RGB image with 7 classes of labels, using color-coding (R, G, B) as follows.

  • Urban land: 0,255,255 - Man-made, built up areas with human artifacts (can ignore roads for now which is hard to label)
  • Agriculture land: 255,255,0 - Farms, any planned (i.e. regular) plantation, cropland, orchards, vineyards, nurseries, and ornamental horticultural areas; confined feeding operations.
  • Rangeland: 255,0,255 - Any non-forest, non-farm, green land, grass
  • Forest land: 0,255,0 - Any land with x% tree crown density plus clearcuts.
  • Water: 0,0,255 - Rivers, oceans, lakes, wetland, ponds.
  • Barren land: 255,255,255 - Mountain, land, rock, dessert, beach, no vegetation
  • Unknown: 0,0,0 - Clouds and others

• File names for satellite images and the corresponding mask image are id_sat.jpg and id_mask.png. is a randomized integer.

• Please note:

  • The values of the mask image may not be the exact target color values due to compression. When converting to labels, please binarize each R/G/B channel at threshold 128.
  • Land cover segmentation from high-resolution satellite imagery is still an exploratory task, and the labels are far from perfect due to the cost for annotating multi-class segmentation mask. In addition, we intentionally didn't annotate roads because it's already covered in a separate road challenge.

Evaluation Metric

• We will use pixel-wise mean Intersection over Union (mIoU) score as our evaluation metric.
  • IoU is defined as: True Positive / (True Positive + False Positive + False Negative).
  • mean IoU is calculated by averaging over all classes.
• Please note the Unknown class (0,0,0) is not an active class used in evaluation. Pixels marked as Unknown will simply be ignored. So effectively mIoU is averaging over 6 classes in total.

Result

Acknowledgment

This repo borrows code heavily from

• rishizek's repo tensorflow-deeplab-v3-plus
• DeepGlobe_2018_A_CVPR_2018_paper
• deepglobe.

Cite us

@INPROCEEDINGS{9064236,  
author={Y. {Li} and L. {Chen}},  
booktitle={2019 IEEE 5th International Conference on Computer and Communications (ICCC)},   
title={Land Cover Classification for High Resolution Remote Sensing Images with Atrous Convolution and BFS},   year={2019},  
volume={},  
number={},  
pages={1808-1813},}

How to implement this repo?

File Structure

• deepglobe_land
  • dataset
    • land_train (Data we downloaded from the website )
    • onechannel_label (Generated by running rgb2label.py)
    • voc_train_all.record (Generated by running create_tf_record_all.py )
  • ini_checkpoints
    • resnet_v2_101 (resnet_v2_101.ckpt and train.graph)
  • utils
  • rgb2label.py
  • create_tf_record_all.py
  • deeplab_model.py

Code functions and excution order

1. rgb2label.py the satellite images id_mask.png are RGB images. This code change the RGB images to onechannel images so we can use them to generate tfrecord files.

2. create_tf_record_all.py generate tfrecord files. The input and output directory were set in the code follows.

   parser.add_argument('--data_dir', type=str, default='./dataset/',
                       help='Path to the directory containing the PASCAL VOC data.')
   
   parser.add_argument('--output_path', type=str, default='./dataset',
                       help='Path to the directory to create TFRecords outputs.')
   
   parser.add_argument('--image_data_dir', type=str, default='land_train',
                       help='The directory containing the image data.')
   
   parser.add_argument('--label_data_dir', type=str, default='onechannel_label',
                       help='The directory containing the augmented label data.')

3. train.py

4. inference.py To apply semantic segmentation to your images.

Other useful codes

1. utils a toolkit

2. deeplab_model.py the deeplabV3+ model

3. tensorboard

   tensorboard --logdir MODEL_DIR
   

   If you want to run Tensorboard on a remote server. This stackoverflow discussion may be help