Clouds are a persistent source of uncertainty in satellite analysis of the Nilgiris. Frequent monsoon cover, thin haze, bright exposed soil, tea plantations, and steep terrain can all confuse a cloud detector. A reliable cloud mask is therefore a prerequisite for crop monitoring, forest assessment, landslide studies, water management, and change detection across Ooty, Coonoor, Kotagiri, Gudalur, and surrounding blocks.
U-Net remains a strong starting point because it combines broad image context with precise pixel-level boundaries. The practical objective is not simply to train a neural network, but to create a repeatable pipeline that produces trustworthy masks on imagery from dates and locations the model has not seen.
Define the segmentation task first
Decide what the model should label before collecting data. A binary mask can classify each pixel as cloud or clear, but operational workflows often benefit from additional classes:
- Opaque cloud: dense cloud that blocks surface information.
- Cloud shadow: dark shadow cast on the ground by cloud.
- Cirrus or thin cloud: semi-transparent cloud that may still affect analysis.
- Clear land and water: usable surface pixels.
- Uncertain or no-data pixels: areas excluded from training and evaluation.
For a first implementation, binary cloud segmentation is appropriate. If the output will support agricultural or ecological analysis, add cloud shadow as a separate class rather than treating it as clear land. Also define a minimum mapping unit: isolated one-pixel predictions may be sensor noise, while aggressive filtering can remove small but meaningful cloud fragments.
Choose and prepare Nilgiris imagery
Sentinel-2 is a practical source because it offers multispectral bands and a five-day revisit under suitable acquisition conditions. Landsat can complement it where a longer historical record matters. Download scenes covering the full area of interest and retain images across summer, southwest monsoon, northeast monsoon, and clearer winter periods. Seasonal diversity is more valuable than simply adding many nearly identical scenes.
Before training, standardise the data:
- Convert imagery to surface reflectance where possible.
- Reproject scenes to a consistent coordinate reference system.
- Resample selected bands to one ground sampling distance, commonly 10 metres for Sentinel-2 workflows.
- Stack useful bands, such as visible, near-infrared, and short-wave infrared channels.
- Clip to the Nilgiris boundary, while retaining a small surrounding buffer to capture edge conditions.
- Record acquisition date, processing level, cloud metadata, and tile identifier.
Avoid random pixel splits. Adjacent pixels are highly correlated, so a random split can produce an unrealistically high score. Instead, split by scene, date, or geographic tile. A model trained on western Nilgiris tiles should be tested on eastern tiles and different seasons to measure genuine generalisation.
Build labels that reflect real ambiguity
Cloud masks can come from existing quality layers, automated cloud-probability products, or manual annotation. Existing masks accelerate bootstrapping, but inspect them carefully: they may miss thin clouds, misclassify bright rooftops or exposed rock, and handle cloud shadows inconsistently.
For manual labels, annotate representative examples rather than only obvious storm clouds. Include:
- Dense and broken cloud fields.
- Thin cirrus over tea estates and forest.
- Cloud edges and mixed pixels.
- Bright soil, concrete, rock, and snow-like highlights.
- Fog and low cloud over ridgelines.
- Dark forest, water, and cloud shadow.
Create a small adjudication set reviewed by a second annotator. Store masks as integer class rasters, not colour screenshots, and keep the original image and label aligned pixel for pixel. If you are new to model design, review customizable neural network architectures for beginners before changing U-Net’s depth or feature widths.
Design a U-Net suitable for multispectral data
The encoder extracts increasingly abstract features through convolution and downsampling. The decoder restores resolution, while skip connections bring back fine boundaries from earlier layers. For Sentinel-2, change the input channels from three to the number of bands you actually use. A compact model is often preferable to a very deep one when labelled Nilgiris data is limited.
A robust baseline includes:
- 256×256 or 512×512 pixel patches, selected according to GPU memory.
- Convolution blocks with batch normalisation or group normalisation.
- Dropout in deeper layers to reduce overfitting.
- Bilinear upsampling followed by convolution, or learned transposed convolution.
- A one-channel sigmoid output for binary masks, or softmax for multiple classes.
- Overlap during inference so objects near patch borders are not cut off.
For thin cloud and class imbalance, binary cross-entropy alone is often insufficient. Start with a combined loss such as weighted binary cross-entropy plus Dice loss. Focal loss can help when cloud pixels are rare, but tune it against a validation set rather than assuming it will improve results.
Train with geography-aware validation
Use augmentations that reflect satellite observations: horizontal and vertical flips, ninety-degree rotations, modest brightness changes, and noise. Avoid transformations that create physically implausible spectral relationships. Do not apply ordinary RGB colour augmentation blindly to multispectral bands.
Use early stopping, a learning-rate scheduler, and checkpoints based on validation IoU or Dice rather than training accuracy. Track per-class performance and inspect false positives. A model can achieve high pixel accuracy simply by predicting clear land everywhere when clouds occupy a small fraction of the image.
A useful experiment log records:
- Bands and normalisation method.
- Patch size and overlap.
- Label source and version.
- Train, validation, and test geography.
- Loss, optimiser, learning rate, and batch size.
- Threshold used to convert probabilities into masks.
- Metrics by season, terrain type, and cloud category.
Evaluate what the mask will be used for
Report intersection over union, Dice or F1 score, precision, recall, and boundary quality. Precision matters when discarding valid imagery is costly; recall matters when downstream analysis must not use contaminated pixels. Evaluate cloud and shadow separately if they have separate labels.
Review predictions as map panels, not only summary numbers. Inspect ridge lines, forest canopies, tea gardens, urban areas, and cloud edges. Calculate performance separately for monsoon and clearer periods. Also test the model on a completely held-out acquisition date from a different satellite tile if possible.
Set an uncertainty rule for production. Pixels close to the decision threshold can be marked uncertain rather than forced into cloud or clear classes. This is especially useful for thin haze and mixed pixels. Maintain a human review queue for low-confidence tiles and feed corrected examples back into later training rounds.
Deploy efficiently and control cloud costs
For research, a GPU notebook is enough for the first baseline. For repeat processing, export the model, tile large scenes, run batch inference, stitch predictions, and apply morphology only after probability maps are saved. Preserve the original probabilities so thresholds can be changed without rerunning the model.
A lightweight inference service can run on a modest VM or scheduled batch job. Read how to deploy AI applications with minimal cloud costs when comparing serverless, managed GPU, and local deployment options. If imagery is sensitive or workflows involve government datasets, also consider the controls described in best AI tools for private cloud data intelligence.
For reproducibility, version the model, preprocessing code, label files, and scene catalogue together. Store outputs with acquisition date, model version, threshold, and quality flags. Never overwrite an older mask without retaining its provenance.
Common failure modes
- Bright land labelled as cloud: add hard-negative examples from roads, roofs, bare soil, and rock.
- Cloud shadows missed: include a shadow class and use near-infrared and short-wave infrared bands.
- Monsoon performance collapse: add wet-season scenes and geographically separate validation data.
- Patch-edge artefacts: use overlap, padding, and weighted blending during reconstruction.
- Excellent validation, poor field results: check for scene leakage and split by geography or date.
- Overconfident predictions: calibrate thresholds and preserve an uncertainty class.
A practical 2026 implementation plan
Start with 30–50 carefully selected scenes, create a reviewed mask for a representative subset, and train a compact multispectral U-Net. Establish a geography-aware baseline before experimenting with attention gates, pretrained encoders, or transformer hybrids. Improve the dataset using error analysis, not architecture changes alone.
Once the mask is stable, connect it to the actual decision workflow: crop time-series cleaning, forest-change analysis, or hazard mapping. Document the acceptable error rate for that use case, publish sample predictions, and make the pipeline rerunnable. For grant-backed work, this evidence is more valuable than a large model with an opaque training process. Teams building broader verification pipelines may also find Validator Cloud AI for model verification useful as a reference for auditability.
AI Grants India supports Indian builders working on applied AI and geospatial problems. Explore AI Grants India for funding and programme information as you move from a research prototype to a field-ready system.