Precision Agriculture with ML

In agtech, deep neural networks are commonly used for field image analysis. U-Net–class architectures are widely considered an industry standard for segmentation due to their strength in pixel-level tasks.

However, we faced two key constraints. The client had no GPU infrastructure in place, and U-Net–based models require significant training and retraining time. Project timelines were tight, driven by the seasonality of field operations.

A more resource-efficient alternative was required. We deliberately chose a Support Vector Machine (SVM) approach as a faster and lighter method that could still meet quality requirements given the client’s infrastructure and timeline constraints.

Morphology-related features—such as RGB-derived indices, texture, and shape—tend to be consistent across healthy plants of the same crop type.

Using reference images for each crop and growth stage, the SVM classifies pixels into groups, separating healthy vegetation from damaged plants and weeds. Images flagged with anomalies are routed to additional processing to enable more precise identification and characterization of field issues.

In collaboration with the client’s agronomists, we assembled reference photo sets for each crop at multiple growth stages, from early growth stages to full maturity. We annotated the images and incorporated them into the model training process. The cycle continued until the training dataset covered all twelve crop types cultivated by the company.

To evaluate model stability and validate the selected configuration, we applied cross-validation across multiple data splits to reduce overfitting and detect systematic errors early. Once performance proved stable, we integrated the model into the client’s field monitoring system.

• Capture. Agricultural drones equipped with RGB cameras capture field imagery using predefined ground sampling distance (GSD) and frame overlap parameters. The images are uploaded to the server and assembled into orthophotos at a fixed resolution.

• Preprocessing. The system normalizes lighting and scale to reduce false positives caused by shadows and brightness variation.
  Feature Extraction. For each pixel, the model extracts color and texture features.

• Classification. The SVM assigns pixels to classes. Adjacent labels are merged into polygons representing affected zones, and boundaries are refined using morphological operations.

We implemented a separate forecasting module based on XGBoost to estimate vegetation level, crop maturity, and expected yield. The model incorporates historical weather data and phenological stage information.

The module generates field- and season-level analytics, including forecast-versus-actual yield comparisons. These insights enable management to estimate expected yield per plot and determine optimal harvest timing to prioritize fully matured crops.