Soil Texture Classification: A Modern Approach Comparison

• 

  Training Acc/Loss

Inspiration

- Cost-free alternative to soil test tailored for third world farmers
- Designed for web app and uploading on mobile devices for ease of access
- Low-cost model for potential philanthropic community hosting

What it does

- Let user upload a photo and return result of the predicted soil type base on pre-trained dataset.

How we built it

- Pytorch
- OpenCV

Challenges we ran into

- Connection of frontend and backend

Accomplishments that we're proud of

- Maintain high accuracy with relatively fast training speed
- Build a moderate dataset from limited data through manipulation and augmentation

What we learned

- Comparison of accuracy, speed, size, and training speed among different models.
- Comparison to results from previous researches in this field, including:
    - 1D CNN, from [Soil Texture Classification with 1D Convolutional Neural Networks based on Hyperspectral Data](https://www.researchgate.net/publication/333513235_Soil_Texture_Classification_with_1D_Convolutional_Neural_Networks_based_on_Hyperspectral_Data)
        - Their implementation has a `Random Forest` approach, stating common practices in remote sensing
    - DCNN, from [Convolutional Neural Networks based classifications of soil images](https://dl.acm.org/doi/abs/10.1007/s11042-022-12200-y)
        - The paper has a much higher accuracy, yet their base on `TF` and  `Keras` left their performance to be a bit worse
    - CNN in [Soil-MobiNet: A Convolutional Neural Network Model Base Soil Classification to Determine Soil Morphology and Its Geospatial Location](https://www.researchgate.net/publication/372684705_Soil-MobiNet_A_Convolutional_Neural_Network_Model_Base_Soil_Classification_to_Determine_Soil_Morphology_and_Its_Geospatial_Location)
        - Their model has a training accuracy of `97%` and testing accuracy of `93%`, much higher than our model, with a much larger data set (`4864` original images)

What's next for Soil Classification: A Modern Approach

- As the model is fast and relatively small, running locally on a mobile device pre-trained will be a great addition.
- In that sense, no public server hosting is required, with also saving the bandwidth cost of uploading images.

Built With

• opencv
• python
• pytorch