CNN-remotesensing

The study of shrub detection with Google Earth imagery was based on Inception model and ResNet-152 model using Keras(as front-end) and Tensorflow(as back-end).

Due to space limitation in Github, the dataset and codes are provided via Google Drive: https://drive.google.com/drive/u/0/folders/1IfyivDERnNj-NW6q6CCam47zca-n0p6e

Short description:

############ INCEPTION v3 ############

Directory structure

Inceptionv3 1_Dataset classes (database to train with class images in folders. e.g. “Z” for vegetation and “S”

for soil)

image (image to classification test. e.g. “ZoneTest.jpg”) 2_Train inception (Model pretrained weights inception v3 ) Train_models (Model retrained weights inception v3 + new classes chips “Z” and “S” ) 3_Label (3 options) individual (show probabilities in one test image like images in 1_Dataset/classes/Z/

“label_image_new.py” Tensorflow 1.0 or greater)

preprocessing (show probabilities in bounding box for extent test image like

“ZoneTest.jpg”. Requires OpenCV2)

heatmap (show heatmap of extent test image like “ZoneTest.jpg”)

Requisites and libs for inception v3 model

Pillow Tensorflow >= 1.0 Opencv + numpy python 2.7

How to retrain model inception v3 with augmentation data

python 2_Train/retrain_au-scale.py --bottleneck_dir=2_Train/Train_models/bottlenecks -- how_many_training_steps 1000 --model_dir=inception -- output_graph=2_Train/Train_models/retrained_graph.pb -- output_labels=Train/Train_models/retrained_labels.txt --image_dir=1_Dataset/classes

How to label test individual image

python 3_Label/individual/label_image_new.py 1_Da taset/test/Z/Z1.jpg

How to label with preprocessing an image

python 3_Label/preprocessing/preprocessing.py

How to label with Heatmap (2 steps)

1-create_csv_from_slidingwindow.py 2-heatmap_from_csv.py

########### RESNET 152 ###########

Directory structure

ResNet152 1_Dataset Train Z S Test Z S

Requisites and libs for ResNet 152 model

Keras >= 2 https://keras.io/ Tensorflow >= 1.0 https://www.tensorflow.org/ Opencv + numpy python 2.7 #ResNet model train and validate test images. python resnet_152.py --image_dir=1_Dataset/ --imgs_rows=224 --imgs_cols=224 -- batch_size=8 --epochs=10

Corresponding e.guirado@ual.es and siham@ugr.es

Acknowledgements

The study of shrub detection with Google Earth imagery was based in Google developers for the Inception model and adaptation of ResNet-152 model used by Keras and Tensorflow.