Soil Image Classification with Transfer Learning

In this video, I walk you through my full solution to the Soil Image Classification Kaggle challenge by Annam AI — where we achieved a perfect leaderboard score of 1.00!

💡 We used EfficientNet-B3 and ResNet-50, fine-tuned on a custom dataset of 4 soil types: Alluvial, Red, Black, and Clay. Our final model was an ensemble of both, boosted with transfer learning, focal loss, and data augmentations.

📊 Key Highlights:

Custom PyTorch Dataset and DataLoader classes

Image augmentations and preprocessing using ImageNet stats

Stratified train/val split to handle class imbalance

Fine-tuning EfficientNet and ResNet with ReduceLROnPlateau and Early Stopping

Ensemble averaging and macro F1 evaluation

Confusion matrix, error analysis, and final predictions

🚀 Final Model Scores:

ResNet-50: F1 = 1.0000, Acc = 98.78%

EfficientNet-B3: F1 = 0.9783, Acc = 97.14%

Ensemble: F1 = 0.9957, Accuracy = 99.59%

👉 Watch till the end for the test set submission and analysis!

🔗 Dataset: Kaggle

Don’t forget to like, subscribe, and drop your questions or feedback in the comments!

#soilclassificationkaggle
#soilimageclassification
#kaggleimageclassificationproject
#cnnsoilclassification
#efficientnetsoilclassification
#resnettransferlearningtutorial
#kaggle #transferlearning #ensemble
#imageclassificationensemble #pytorch
#pytorch transfer learning tutorial
#soildatasetkaggleannamai
#cnn vs transfer learning
#focallosspytorch
#ensemble learning kaggle
#bestkagglenotebooksoil