KenzoBou/Locust-Breeding-Ground---Satellite-Analysis

This repository was originally created during the GEO AI Hackathon 2025, co-organised by Instadeep and Datacraft. This repo uses a Prithvi backbone to make segmentation prediction. The aim of this project is to early detect locust breeding grounds in Africa with HLS data and Instageo. # Locust-Breeding-Ground---Satellite-Analysis This repository was originally created during the GEO AI Hackathon 2025, co-organised by Instadeep and Datacraft. It fine-tunes Prithvi to make segmentation prediction to early detect locust breeding grounds in Afriaca with HLS data. The code features a complete pipeline for processing HLS satellite imagery, computing spectral indices, fine-tuning InstaDeep’s PrithviSeg model via InstaGeo, and producing locust-presence predictions. ## Repository Structure ## Methodology 1. **Data Subsetting** Selected the most recent 25% of training imagery chips and their segmentation maps to manage storage and preserve temporal order. 2. **Band Replacement** I replaced the SWIR2 band (channel 5 of each time step) with NDVI to match the six-band input expected by Prithvi’s patch embedding. ```python # NDVI = (NIR – Red) / (NIR + Red + 1e-6) ndvi = (nir - red) / (nir + red + 1e-6) This in-place modification allowed training on NDVI without increasing file size or chip count. ## Train/Validation Split Generated train_ds_subset.csv listing chip and seg_map paths, then created a 70/30 split into train_split.csv and validation_split.csv. ## Model Training Fine-tuned the PrithviSeg backbone on the transformed six-band inputs (Blue, Green, Red, NIR, NDVI, SWIR1) using Hydra: "python scripts/train_instageo.py \ --config configs/locust.yaml \ --root_dir . \ --train_csv train_split.csv \ --val_csv validation_split.csv \ --epochs 10 \ --batch_size 8" Evaluation & Inference Ran validation and generated test-set predictions: bash Copier Modifier ## Validation python scripts/inference_instageo.py \ --config configs/locust.yaml \ --root_dir . \ --test_csv validation_split.csv \ --checkpoint outputs/first_run/instageo_best_checkpoint.ckpt \ --mode eval ## Test-set inference python scripts/inference_instageo.py \ --config configs/locust.yaml \ --root_dir . \ --test_csv test_ds.csv \ --checkpoint outputs/first_run/instageo_best_checkpoint.ck …