For supervised method , i need training data....for that I have ground truth data which is applicable for 2 Landsat 8 images taken 16 days apart.
As per all the literature, i have studied CNNs need large dataset for training the NN. But in my case it's available for only 2 images...even though i have 10 images.
I can use same training labels for land covers like water body, road, settlement for images of 1 year ...since they don't change but my focus region is crop fields which change every 2 months. Please guide me on this?
Dear Paminder,
The issue of having a large sample data for different period of times is really a critical issue in creating different maps such as Land cover and crop maps. There are many ways to overcome this issue especially with respect to crops. because as you said urban settlements such as buildings, roads...etc they do not change in short period of time like crops. In this case I can mask these features. I will be left with water bodies, rocks and fellow lands this will be identified by your two surveys and can be masked two. What is left is the forest and fruits trees they can be also identified by the samples and can be masked. With respect to non perennial crops some can also be identified with the samples you have of course you will have problem with the others. You can overcome this by using the 10 images after doing atmospheric correction to create phenological graphs and consult with an experienced agriculture and crop experts to identify them. This way you can create a data set of the location of different features that can exist in the images and train CNN to recognize them. Good luck.
Thank you for detailed explanation...I think it's workable as you suggest. Since i am training a CNN using these values....this needs to be verified by an expert.
Any suggestion regarding some existing dataset which can be used to train CNN, i came across UC mercedes dataset from USGS.
My dataset is Landsat 8 and LISS-IV.
Dear Parminder,
As for the training dataset, largely speaking, the larger the better. But it is not always easy to obtain such big training dataset based on field observation and visual interpretation of satellite imagery. If it is > 5-10% of the total images, it should be considered "good".
To catch the phenological change of crops, you can obtain several images during the the crop growing-maturity-harvesting period as Awad said (fortunately, Landsat images are free of charge), or you can just get the spring and summer ones which provide the strongest contrast in phenological feature of crops. This will help you map out what you want.
To ensure a very high accuracy of land use/cover mapping, you can also employ DEM and land surface temperature (LST) using different classifer but not only CNN. My experience shows that NN is not as good as SVM (Support Vector Machine) and RF (Random Forest) classifiers for small areas. For regional-scale and national-scale land cover mapping, Maximum Likelihood is still a good choice as it runs very fast and can ensure high accuracy if you have not very powerful computer for processing. One of my papers discussing how to enhance regional-scale land cover mapping may provide you ideas.
Good try and all the best,
Weicheng
Hello ma'am, I am trying to do an image segmentation of satellite image using CNNs. But I don't know it's ground truth data necessary for training. Please can you tell me how can I get the ground truth data for a satellite image. The satellite image is for the panna forest area.
You could use DataTurks [ https://dataturks.com/ ], it gives you web collaborative interface where you can upload image, invite your colleagues and get data annotated in no time. You can see some sample projects here :
https://dataturks.com/projects/trending
@ Gajendra ...Thanks for the response. I created a customised dataset. I have completed this task.
Hi
how to generate a dataset for training CNN for Wireless data communication Mobile celler Phone for massive devices connectivity?
- Badges
- Science topic
- Similar topics
- Agricultural Science
- Agronomy
- Crop Science
- Crop