Soil_Classification_annam

Soil Image Classification Challenge – Solution README

Overview

This repository contains my submission for the Soil Image Classification Challenge organized by Annam.ai at IIT Ropar. The task involves building a machine learning model to classify soil images into one of four categories:

• Alluvial Soil
• Black Soil
• Clay Soil
• Red Soil

The competition aims to promote balanced classification performance by maximizing the minimum F1-score across all classes.

---

Model Used

I used the ConvNeXt-Tiny model, a modern convolutional architecture that has shown state-of-the-art performance on image classification tasks. It combines hierarchical convolutional structures with design insights from transformers, offering both accuracy and efficiency.

---

Workflow Summary

1. Data Preprocessing

• Resizing: All images were resized to 224x224 pixels to match the input requirements of ConvNeXt-Tiny.
• Normalization: Pixel values were normalized using ImageNet statistics.
• Augmentation (during training):
  • Random Horizontal Flip
  • Random Rotation
  • Color Jitter
  • Random Crop with Padding
• These augmentations help improve generalization, especially for soil types with subtle differences.

2. Model Architecture

• Base Model: ConvNeXt-Tiny from timm library, pretrained on ImageNet.
• Modifications:
  • Replaced the classifier head with a fully connected layer for 4 output classes.
  • Applied dropout before the final classifier for regularization.

3. Training Setup

• Loss Function: CrossEntropyLoss
• Optimizer: AdamW with weight decay
• Scheduler: Cosine Annealing with warm restarts
• Batch Size: 32
• Epochs: 30 (with early stopping based on validation F1-min score)

4. Evaluation Strategy

• The model's performance was evaluated using the minimum F1-score across all classes on the validation set.
• Validation set was stratified to maintain class distribution.

---

Results

On the validation set:

• Minimum F1-score: 1.000
• Confusion matrix and per-class F1-scores are plotted in the notebook to visualize performance across soil types.

---

File Structure

• convinexttiny.ipynb: Main notebook containing all code for:

  • Data loading and preprocessing
  • Model architecture
  • Training and evaluation
  • Prediction generation

• submission.csv: Final prediction file in the required format.

---

Reproducibility

To reproduce the results:

1. Install dependencies:

   pip install -r requirements.txt

2. Run the notebook:

   jupyter notebook convinexttiny.ipynb

3. Ensure the dataset is placed in the correct directory as referenced in the notebook.

---

Conclusion

This solution leverages modern convolutional architectures to effectively classify soil images. Emphasis was placed on:

• Balanced performance across all soil types
• Proper data augmentation and preprocessing
• Clear code structure and modular implementation

Please refer to the code comments for additional insights on decisions and tuning strategies.

THE SECOND CHALLENGE:

Soil Image Classification Challenge

Overview

The Soil Image Classification Challenge is a machine learning competition organized by Annam.ai at IIT Ropar. The objective is to build a classifier that determines whether an input image contains soil or non-soil content using visual features.

Deadline: May 25, 2025, 11:59 PM IST
Metric: F1 Score (binary classification)

---

Dataset Description

Objective

Predict whether an image is a soil image (label = 1) or not (label = 0).

Structure

• Train Set

  • Images located in train/
  • CSV with columns: image_id, soil_type
  • All soil images → labeled as 1

• Test Set

  • Only image_ids are provided
  • Participants must submit predicted labels

• Soil Types in Training:

  • Alluvial Soil
  • Black Soil
  • Clay Soil
  • Red Soil

Note: For this binary classification, all above types are treated equally as “soil”.

---

Model Architecture

This solution uses MobileNetV2, a lightweight convolutional neural network pretrained on ImageNet.

Architecture

• Base: MobileNetV2 (frozen layers)
• Custom Head:
  • GlobalAveragePooling2D
  • Dense(64, relu)
  • Dense(1, sigmoid) — for binary classification

Training Strategy

1. Data Preprocessing:

   • Images resized to 224×224
   • Normalized to [0, 1]
   • Augmented using ImageDataGenerator

2. Synthetic Negative Samples:

   • Augmented training images used to create visually diverse fake “non-soil” examples
   • Labelled as 0

3. Training Phases:

   • Phase 1: Train head layers with base model frozen
   • Phase 2: Fine-tune top 30 layers of MobileNetV2 with a reduced learning rate

4. Loss Function: Binary Crossentropy

5. Optimizer: Adam

6. Callbacks: EarlyStopping (monitored val_loss)

---

Evaluation Metric

• F1 Score: Balances Precision and Recall
• Encourages strong, balanced classification performance
• Final evaluation includes:
  • F1-score on private test set
  • Review of submitted code and documentation

---

Submission Format

Submit a submission.csv with columns:

image_id,label
IMG_001.jpg,1
IMG_002.jpg,0
...

Also include your:

• Final notebook/script with:
  • Preprocessing steps
  • Model training
  • Evaluation
  • Prediction generation
• Code must be clean, readable, and well-commented

---

Reproducibility

To reproduce the results:

pip install -r requirements.txt
python preprocess.py    # optional, if separate
python train.py
python predict.py

Model and prediction logic is implemented in mobilenetupdated.ipynb

---

Results

• Achieved F1 Score = 1.000
• Efficient training due to lightweight MobileNet backbone
• Robust handling of imbalance via synthetic negatives

---

Tools & Libraries

• Python, NumPy, Pandas, Matplotlib
• TensorFlow / Keras
• Scikit-learn (F1-score)
• ImageDataGenerator for augmentation

---

Notes

• Avoid training on test data
• Ensure predictions are reproducible
• Submit early to avoid last-minute issues

---

Acknowledgements

Organized by Annam.ai in collaboration with IIT Ropar. This challenge fosters practical application of AI in agriculture and environmental sciences.