Dear all,
My goal is to predict night light intensity based on day time satellite imagery. I am using data from VIIRS and Landsat 8 sensors. My study area is shown in the attached image and I am following the methodology of the attached paper. According the authors their day time images are of 400*400 pixel (dimensions). Also, they classified the nighttime images into 3 categories (low, medium, high intensity lights). In my study area very few day time pixels corresponds to high intensity lights compared to the other 2 categories (e.g. I can find only 40 day time pixels for category 3 but I can sample way more for the other two categories). How should I sample (select) my day time images for each category, taking into account that each sample will have different dimensions (e.g. I can find only 40 day time pixels for category 3 but I can sample way more for the other two categories)? These samples are going to be my input to the VGG-16 architecture.
Hi Nikolaos Tziokas!
You can randomly choose the size 224x224, conduct an experiment, evaluate the accuracy of the model. And also choose a smaller image size and also conduct an experiment and evaluate the quality of the model. And compare at what sizes of images your model works better. Neural networks are creativity and a lot of experiments. Also, I recommend using VGG-19, Resnet, Alexnet, etc.
I agree with the comment from Shahboz Sharifovich Qodirov
Neural architectures require lots of experimentations and tunings. To answer your question, you should follow the strict dimension protocol, if there is some for some of the models. Otherwise, if you are flexible in selecting the dimension, then you can try experimenting with it, in a range of 200-500. Don't make it too small. That might give you faster training but may lead to distorted results.
Hello Nikolaos Tziokas ,
As Shahboz Sharifovich Qodirov and Hritam Basak suggested , you could make the implementation of several DNN and ANN architectures such as ResNet50, ResNet101 or ResNet150 as these higher order architectures make the use of skipping the layers and train faster for evaluating results and thus there are less possibilities of overfitting the model. Also using Softmax and ReLu while downsampling the images helps. For higher orders and large datasets, I would suggest using Voxel for segmentation if necessary and Neurosim for effective memeory optimization.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size