Cannot open the resultant JSON file after NIR analysis due to huge file size

RavinduDharmasena commented 5 years ago

Description

I have tested a pixelwise NIR analysis and it completed without any errors. Output of that is a JSON file contains NIR analysis data. But the case is the file size is 153.9 MB (I don't know the reason the huge file size) and because of that I can't check the file as the every editinf application is becoming unresponsive when it is opening. In the zip file given below contains the python code and that huge generated json and the image I used in this script. When I compressed the JSON file the file size is 262.2 KB.

test.zip

What is the reason for this ? How can I solve this ?

nfahlgren commented 5 years ago

pcv.analyze_nir_intensity stores a frequency table (the same data that is plotted in the output histogram) in the JSON file. The output values are the number of pixels of a given intensity or intensity range. The intensity values or range of values are created by binning intensities and counting pixels within each range. The number of equally-spaced bins is controlled by the bins input. The default number of bins is 256, the total number possible intensity values in an 8-bit image (2^8). In your code you are setting bins to the height*width of the image, so there are >12 million bins in the frequency table, that's why your file is so large.

RavinduDharmasena commented 5 years ago

So am I doing the pixel wise nir analyzing right in the script ?. Is the photo a 8-bit one ?. Even though I understand what you said, still I can't figure out those questions.

nfahlgren commented 5 years ago

Yes, in your script it is running the pcv.analyze_nir_intensity function on all pixels of the red channel of the input image (none are excluded by the mask). Your image is 8-bit, so you would want bins=256 as the largest number of bins.

Note that bins can be a smaller number if one wants to summarize the frequency table more. For example, when working with a 16-bit image that has 65,536 possible intensity values, I would personally still use bins=256 or something similar because having 65,536 output values for every image is pretty large and the frequency table can be summarized without a lot of loss of information.

RavinduDharmasena commented 5 years ago

@nfahlgren , Thank you so much for pointing out the huge mistake I made in the script. Is there a way to get the NIR value all pixels out from those possible intensity values ? or any other way to get the nir intensities in pixel wise using plantCV ?.Still I'm stuck on that.

RavinduDharmasena commented 5 years ago

This is the histogram I created using the data generated from above script

histogram

But I need to get the individual NIR intensity of each pixel. So is there a way to do that ?. Column B is the frequency (I forgot to rename it)

HaleySchuhl commented 5 years ago

Hello @RavinduDharmasena ,

I just wanted to add that the pcv.analyze_nir_intensity() function is equipped to handle both 8-bit and 16-bit images since it will check the type of image and sets the maximum pixel intensity based on this information.

if gray_img.dtype == 'uint16':
    maxval = 65536
else:
    maxval = 256

The user input of bins, like Noah mentioned, is the resolution of the histogram information. The shape of the histogram would be the same even with less than 256 bins.

What are you trying to do with the individual pixel values? The image you've been working with seems to have NIR data stored in the red channel but looking at the red channel as an array is the only way to keep every single pixel intensity. Are there specific regions of the image you're interested in the pixel NIR intensities?

RavinduDharmasena commented 5 years ago

@HaleySchuhl , I am trying to create a NDVI (Normalized Difference Vegetation Index) map and calculate the healthy and unhealthy plant percentage in a tea state using aerial NIR (Near Infra Red) and RGB maps created using aerial drone images. For that I have to following task.

Getting the reflectance of red color and NIR from NIR and RGB maps created using aerial drone footages pixel wisely.
Calculation of NDVI index for each pixel using those reflectance data.
Creation of NDVI map using the NDVI data generated.
Calculation of approximate percentage of plants that are healthy and unhealthy in tea state

In first task I tried to get the red channel intensity by using following equation Red channel reflectance = ((1/255) * red channel value)/100

To get the NIR intensity of each pixel I'm trying to use the plantCV. That is the point that I'm currently stucked.

For NDVI calculation there is a formula. It is given below

NDVI = (NIR - RED)/(NIR)+ RED)

In above equation NIR means near infra red reflectance while RED means red color reflectance. So if I can accomplish first step, 2nd one is nothing

Then using those NDVI values, I have to create a NDVI map and should resize it.

So, for calculation red color intensity, image creation resizing and purposes I will use pillow library. For 2nd step I need only numpy library while 1st step needs pillow, plantCV and numpy.

For 4th step also can be accomplished by solving following formulas.That is the final step.

Percentage of plants having =( No. of pixels with healthy plants NDVI indexes/Total pixel amount) * 100 good health

Percentage of plants having =( No. of pixels with unhealthy plants NDVI indexes/Total pixel amount) * 100 bad health

Thats the thing I am going to do. Any ideas,weakness,threats,suggesstion are highly appreciated and I really need to do pixel wise NIR analysis because of that.

And other really important thing is does anyone knows a NDVI spectrum library or how to create a NDVI spectrum ?

nfahlgren commented 5 years ago

Hi @RavinduDharmasena,

Thanks for sending this, we have a good sense for what you want to accomplish now. The pcv.analyze_nir_intensity only summarizes a grayscale image (using a masked region if provided) as a histogram.

In principle if you had a NIR NumPy ndarray and a RED ndarray you could calculate the per-pixel NDVI index directly with NumPy. This assumes that the NIR and RED ndarrays are the same dimensions and that the images are aligned already.

Do you have two images or is the image you posted above Green, Blue, Red+NIR?

RavinduDharmasena commented 5 years ago

Thank you @nfahlgren , you are absoutely right about the assumptions and process. Actually it is very hard to find a dataset that contains the NIR and RGB image of same location. But luckily I found a such dataset recently. But actually still I didn't tested it out and it is like 1GB dataset. Given below is the link.

https://ivrl.epfl.ch/research-2/research-downloads/supplementary_material-cvpr11-index-html/

Image I uploaded is an NIR image. I don't think there is a separate metadata for NIR intensity in that. Actually I tried to calculate the NDVI index using only the NIR images as well. Thats why I uploaded NIR image only.

nfahlgren commented 5 years ago

I could be wrong but it looks to me like your test image was acquired with an RGB camera with the IR filter removed (these tend to have a reddish tint because the red channel gets overexposed with IR light). That would mean the red channel is RED+NIR instead of just RED (or just NIR). I don't know of a method to isolate or estimate the NIR component of that channel without a RED-only reference.

RavinduDharmasena commented 5 years ago

@nfahlgren, Actually I got that image from OpenDroneMap dataset github repository.According to them this image is NIR. Link is given below.

https://github.com/OpenDroneMap/ODMdata

Here are 2 RGB and NIR tiff files of same location

country.zip

RavinduDharmasena commented 5 years ago

@nfahlgren, This are the images (related with that NIR image) metadata from OpenDroneMap command line toolkit. Coordinate system is WGS84. Looks like you are correct. If that image cannot be used try the country.zip file uploaded in previous comment.

nfahlgren commented 5 years ago

I'm not sure what that dataset is, there doesn't seem to be any documentation for it. It's labeled rising_valley_nir but the images are RGB. The metadata you have says they were taken with a Canon Powershot S110, which is a standard color camera. My guess is that they modified it to remove the IR filter. So it's seeing NIR but the data is entangled with visible red light also.

The other example dataset you posted with the separate RGB and NIR pictures could be used for NDVI calculations that you want to do.

You could read in the RGB image and isolate the RED channel. Separately read in the NIR image, then you have a NumPy ndarray for RED and NIR. I tested the following in Jupyter

from plantcv import plantcv as pcv
import cv2
import numpy as np

# Set debug mode to plot
pcv.params.debug = "plot"

# Read the RGB image
rgb, _, _ = pcv.readimage("0000_rgb.tiff", mode="RGB")
# Read the NIR image
nir, _, _ = pcv.readimage("0000_nir.tiff", mode="GRAY")

# Isolate the RED channel
blue, green, red = cv2.split(rgb)
pcv.plot_image(red)

# Convert the RED and NIR data types to FLOAT64 instead of UINT8 so we can do math
red = red.astype(np.float64)
nir = nir.astype(np.float64)

# Calculate NDVI
ndvi = (nir - red) / (nir + red)

# Plot the grayscale NDVI image
pcv.plot_image(ndvi)

# Make a pseudocolored NDVI image
ndvi_pseudo = pcv.visualize.pseudocolor(gray_img=ndvi, min_value=-1, max_value=1, cmap="viridis")

RavinduDharmasena commented 5 years ago

@nfahlgren, Thank you so much for the effort. There is a small issue when it is running in my conda environment as given below

Traceback (most recent call last): File "./research_dss/test/abc/test1/ndvi_newmodulw.py", line 11, in rgb, , _ = pcv.readimage("0000_rgb.tiff", mode="RGB") File "/home/ravindu/anaconda3/envs/research_dss/lib/python3.7/site-packages/plantcv/plantcv/readimage.py", line 44, in readimage fatal_error("Failed to open " + filename) File "/home/ravindu/anaconda3/envs/research_dss/lib/python3.7/site-packages/plantcv/plantcv/fatal_error.py", line 14, in fatal_error raise RuntimeError(error) RuntimeError: Failed to open 0000_rgb.tiff

I added the #!/usr/bin/env python at the beginning of the .py file Images are also in same directory

nfahlgren commented 5 years ago

The error says it can't open the image, so it either can't find it (it's only looking in your terminal's current working directory) or it doesn't have permission to open it.

RavinduDharmasena commented 5 years ago

@nfahlgren, Thank you so much. It works perfectly. But can you explain the meaning of _,_ in rgb, _, _ ?.

HaleySchuhl commented 5 years ago

@RavinduDharmasena The function pcv.readimage returns the image object, path, and filename. Since there isn't a need to store the path or filename you can store them to _ as placeholders.

RavinduDharmasena commented 5 years ago

@HaleySchuhl Thank you so much for short and sweet explanation. And again @nfahlgren I should give a big thank you for you. I think it's the time to close this issue. As final statement before closing this issue I just need to give a suggestion.Can you place tutorials for vegetation index maps generation like NDVI map using plantCV in plantCV doc..

cesarsm98 commented 4 years ago

@HaleySchuhl , I am trying to create a NDVI (Normalized Difference Vegetation Index) map and calculate the healthy and unhealthy plant percentage in a tea state using aerial NIR (Near Infra Red) and RGB maps created using aerial drone images. For that I have to following task.

Getting the reflectance of red color and NIR from NIR and RGB maps created using aerial drone footages pixel wisely.

Calculation of NDVI index for each pixel using those reflectance data.

Creation of NDVI map using the NDVI data generated.

Calculation of approximate percentage of plants that are healthy and unhealthy in tea state

In first task I tried to get the red channel intensity by using following equation Red channel reflectance = ((1/255) * red channel value)/100

To get the NIR intensity of each pixel I'm trying to use the plantCV. That is the point that I'm currently stucked.

For NDVI calculation there is a formula. It is given below

NDVI = (NIR - RED)/(NIR)+ RED)

In above equation NIR means near infra red reflectance while RED means red color reflectance. So if I can accomplish first step, 2nd one is nothing

Then using those NDVI values, I have to create a NDVI map and should resize it.

So, for calculation red color intensity, image creation resizing and purposes I will use pillow library. For 2nd step I need only numpy library while 1st step needs pillow, plantCV and numpy.

For 4th step also can be accomplished by solving following formulas.That is the final step.

Percentage of plants having =( No. of pixels with healthy plants NDVI indexes/Total pixel amount) * 100 good health

Percentage of plants having =( No. of pixels with unhealthy plants NDVI indexes/Total pixel amount) * 100 bad health

Thats the thing I am going to do. Any ideas,weakness,threats,suggesstion are highly appreciated and I really need to do pixel wise NIR analysis because of that.

And other really important thing is does anyone knows a NDVI spectrum library or how to create a NDVI spectrum ?

Hello RavinduDhasmasena,

I am working on a project related with NDVI maps. The idea of the project looks very similar to what you are saying in this response.

I would like to know some things about what you have doing.

Have you been able to get the red and infrared band with python. In that case, how you have taken both?
Do you align the RGB and NIR maps to get the index? I am getting kind of stuck in this step.
How do you create a NDVI map once you have NDVI computed?
How do you know the number of pixels of an image?

Btw, is there any way I can get into contact with you to discuss more in detail about the project?

Thank you in advance, César Sebastián

RavinduDharmasena commented 4 years ago

@cesarsm98, I am really happy to heard that you are also trying to do the same stuff. So answers for your questions given below.

Q1 : Have you been able to get the red and infrared band with python. In that case, how you have taken both?

Yes. You can do it using an image library like OpenCV or PIL. But from RGB images you can get only red color intensity because in an RGB camera the NIR is gets filtered by the IR filter in the camera. So to get the NIR intensity, you have to get an NIR image of same location. The both images should have taken in same position and in same time. That is the most hard part if you are not using a multispectral camera like me because of the expensiveness. If you are using multispectral camera, you can get the RGB band and NIR band from one camera.

Q2: Do you align the RGB and NIR maps to get the index? I am getting kind of stuck in this step.

Yes. You have to do that if you are not using a multispectral camera. You have to use a RGB camera and a NIR camera to get the data if you are not using multispectral one. Then the pixels in each photo doesn't cover the same area. So if you are calculating NDVI using original RGB and NIR image data, you will get wrong NDVI values. So you have to align the both images using a technique such as feature mapping and then a cropping. OpenCV has those functionalities as well.

Q3: How do you create a NDVI map once you have NDVI computed Step 1:Resize the both images to a same resolution. If both images are in same resolution, you don't have to do this. Step 2: Take the RGB pixel data into a numpy array by reading the RGB image (numpy is a package use to create multi dimensional arrays). In this case you will get a 2D array. In that array also in each index, there is tuple ( Unmutable 1D array) ehich contains the RGB color values of each pixel. Step 3: Take the NIR pixel data in to a numpy array by reading the NIR image. In this also case you will get an 2D array. But in each index, the NIR intensity is directly available in px. Step 4: Create a numpy 2D array which has the same size of the RGB and NIR arrays to store the NDVI values. Step 5: Split the numpy array that contains the RGB data to 3 arrays by separating the red, green and blue color data into 3 separate arrays. Step 6: Iterate through the NIR and Red color arrays in same loop and calculate the NDVI for each pixel using the NDVI calculation equation. Step 7: Create a numpy 2D array which has the same size of the RGB and NIR arrays to store the pixel color values for each NDVI values. Step 8: Save the array created in step 7 to a image.

AT the end of these steps, you will get a grayscale image and that is the original NDVI map. But if you want a color version of NDVI map, you have to skip Step 8 and follow the Step 9 and 10 Step 9: At the end of calculating the NDVI for each pixel in the loop you have to create a tuple that contains the RGB pixel values and you have to push it to the array created in the Step 8. Step 10: Save the array created in the Step 8 as an image.

Q4: How do you know the number of pixels of an image? Step 1: Get the image into a numpy array. Step 2:Multiply the values in size attribute in the numpy array.

Now you have the number of pixels in the image. If you are using a multispectral camera an photo for each band in same location will be saved in a photo capture. In that case you can skip first 5 steps.

If you have any question please ask it in this issue or you can contact me using my mail. it is ravinduta@gmail.com. If you are commenting on this issue I will get a email as well. But the answers might be little bit late because of the workload in my job. But anyway I always try to give the best answer for regarding this.

danforthcenter / plantcv

Cannot open the resultant JSON file after NIR analysis due to huge file size #410

Description