Editing Ai Climate (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''Satellite deforestation detection with segmentation:'''
<syntaxhighlight lang="python">
import torch
import segmentation_models_pytorch as smp
from torch.utils.data import DataLoader
from torchvision import transforms

# U-Net with ResNet-50 backbone for satellite imagery segmentation
model = smp.Unet(
    encoder_name="resnet50",
    encoder_weights="imagenet",
    in_channels=4,      # RGB + Near-Infrared (NIR) bands from Sentinel-2
    classes=2,          # Forest vs. deforested
)

# Training setup
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
loss_fn = smp.losses.DiceLoss(mode='binary')  # Better for imbalanced segmentation

# Inference on new satellite tile
def detect_deforestation(tile_tensor, threshold=0.5):
    """
    tile_tensor: (1, 4, H, W) satellite image with R, G, B, NIR bands
    Returns: binary mask of deforested areas
    """
    model.eval()
    with torch.no_grad():
        logits = model(tile_tensor)
        probs = torch.sigmoid(logits)
        return (probs > threshold).squeeze(0)  # (1, H, W) binary mask

# Computing deforestation area
def area_hectares(mask, pixel_size_m=10):  # Sentinel-2 = 10m/pixel
    """Convert binary mask to area in hectares."""
    n_pixels = mask.sum().item()
    return n_pixels * (pixel_size_m ** 2) / 10_000  # Convert m² to hectares
</syntaxhighlight>

; AI for climate: application landscape
: '''Weather forecasting''' → GraphCast, Pangu-Weather, FourCastNet, NeuralGCM
: '''Deforestation monitoring''' → Global Forest Watch (Hansen et al.) + daily alerts via Planet AI
: '''Energy demand forecasting''' → LightGBM, LSTM, TFT on smart meter data
: '''Solar/wind forecasting''' → XGBoost, LSTM, NWP post-processing
: '''Species identification''' → iNaturalist (CV), BirdNET (audio ML)
: '''Climate downscaling''' → Super-resolution CNNs on GCM output (ClimATE-ViT)
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">