bisht-prachi/lacuna_solar_survey

my attempt at the lacuna solar survey challenge: to develop a model capable of detecting and counting solar panels and boilers in aerial imagery of Madagascar # Solar Panel Count Prediction Predict the number of solar panel "boil" and "pan" units from aerial images and metadata using deep learning (EfficientNetV2 with metadata fusion). --- ## Overview This repository was built for the **Lacuna Solar Survey Challenge** on Kaggle. It uses a hybrid deep learning architecture combining image features and structured metadata to predict: - `boil_nbr`: Number of boiling units - `pan_nbr`: Number of panel units --- ## Directory Structure - solar-panel-prediction.ipynb Kaggle notebook - Train.csv - Test.csv - README.md - submission_*.csv --- ## Setting up ### 1. Clone the repo git clone github.com cd solar-panel-prediction ### 2. Set up environment pip install -r requirements.txt Make sure you have access to a CUDA-compatible GPU and your data paths are correctly set in the script. ## Training To train the model across 3 folds: python main.py The best model per fold will be saved as: best_model_fold0.pth best_model_fold1.pth best_model_fold2.pth ## Inference & Submission Predictions on the test set will be saved as: submission_original.csv (with float predictions) submission_integer.csv (with rounded integers) ## Model Architecture Backbone: tf_efficientnetv2_b3 from timm Metadata Processor: Fully connected + LayerNorm + Dropout Fusion: Image and metadata features are concatenated after attention Regressor: 2-head count predictor with Softplus output ## Tools Used PyTorch + AMP Albumentations for augmentation K-Fold Cross Validation Huber Loss + CosineAnnealing Scheduler Multihead Attention for metadata embedding ## Evaluation Metric Mean Absolute Error (MAE) on the boil_nbr and pan_nbr predictions. - Final Public MAE: 0.947831978 - Rank: 45/712 - Qualitative results: - ## License This project is licensed under the MIT License. ### Acknowledgements Inspired by the Lacuna Solar Survey Challenge on Zindi. EfficientNetV2 and pretrained models provide …