Farmland_Detection_and_Categorization

CSCE 873: Computer Vision Project

Objective 1: Preprocessing (Ashish)

• Two screenshots in the root folder are for one year of land cover image overlayed on the top of NDVI image.
• All images are resampled to smaller size (resampled 9x9 pixel as 1 pixel).
  • If the classified image had same values in 9x9 window, then that pixel is populated as same value, otherwise that pixel is set as NaN.
• input\ includes
  • Classified crops 2013 - 2022.
  • NDVI images for each year resampled at matching lower resolution.
  • Surface apparent temperature matching image size of classified images.
  • Utilities: Resample: Contain post processing tasks like resampling TIFF's to lower resolution based on all pixels having same value, else the output is zero.
  • Utilities: ScatterPlotVsClasses: Generate the Land Cover classes Vs NDVI pixel scatter plots (before and after filtering).
  • automatic_mask_generator_SAM PDF & IPYNB: Create image segmentation using the uploaded TIFF file (Note: Large file can't be uploaded here). It runs of Google Collab.
    • automatic_mask_generator_SAM PDF: Read the script as a PDF file.

Objective 2: Crop vs no crop (Shubham)

• src/DataPreprocessing.ipynb is Jupyter Notebook to preprocess input data.
• input/objective_2_crop_vs_no_crop contains a sample to input images used to train UNet.
• src/UNetTrain.ipynb is to build the UNet model.
• src/UNetTrain.ipynb is to build the UNet model.

Objective 3: Crop categorization (Max, Quan, MaxSally)

• Crop_classification.ipynb is the Jupyter notebook to build the model.
• input/objective_2_crop_classification is the dataset used for the model. Each image is a sample image of a crop area with the top 3 labels in its name. Because the image encodes BrightnessTemp with the normalized value of a region, the image, when viewed under a typical image viewer, may appear strange.
  • The dataset in this repository is only a small sample of the actual dataset. Please use the link here to get entire dataset and put them in this directory for training. Link to the dataset
• result/objective2 contains the results of objective 2 such as training history and metrics vs epochs (top1 + top3 accuracy, F1 score, and loss) of validation and loss
• src/crop_classification_utilities.py: contains utilities function to print graphs from training history.

Objective 4: Growth length (Salem)

• Use src/cropgrowth_predictor.ipynb to train snd save the model.
• Trained model can be found at result/objective4_crop_growth_predictor/growth_stage_predictor.h5.
• src/cropgrowth_predictor-checkpoint.ipynb can be used to observe how the current model was trained.
• src/corn_filter.py was used to filter non-corn areas out of the dataset (This dataset was used in the Kc_filter.py script).
• Results for the above can be found in result/objective4_crop_growth_predictor/
• src/Kc_filter.py was used to filter out areas that do not match the appropriate corn growth stage obtained using Kc values and the Day of Year (DOY) the image was captured on.
• Results for the above can be found in input/objective4_crop_growth_predictor/ (This dataset was used to train the Model).