Deep learning-based automated crop field delineation for african agriculture

Overview

This solution uses advanced deep learning techniques and high-resolution satellite imagery to automatically detect and delineate individual smallholder farmer fields across large spatial scales in Africa. It addresses the critical gap in agricultural data in regions where traditional agricultural statistics are difficult and expensive to collect. By accurately mapping field boundaries at scale, the technology supports evidence-based investments in smallholder agriculture, food security planning, and rural development.

The approach is designed for data-scarce contexts common across the African continent, where conventional machine learning methods have struggled to achieve reliable results. It enables organizations working in agricultural development, land management, and policy planning to obtain precise, scalable field-level data.

Technical specifications

Key features:

• Convolutional neural network architecture (U-Net) trained to classify satellite imagery pixels into three categories: field boundary, field interior, and non-field
• Binary label training approach that delivers high-accuracy segmentation
• Online labeling platform used to generate training data efficiently
• Image segmentation process combined with vectorized output to produce polygonized representations of individual fields
• Achieves greater than 90% accuracy with precision above 0.6 and recall above 0.9 in test comparisons
• Significantly higher prediction certainty along field boundaries compared to traditional Random Forests approaches
• In a Ghana case study, returned an average field size of 0.8 hectares, closely matching expected field sizes, versus 6.8 hectares for Random Forests

Technology readiness level

The technology has progressed beyond initial proof of concept. Early validation using a Random Forests model with active learning achieved 83% accuracy but with low precision and recall. The newer U-Net deep learning model has demonstrated substantially improved performance in test comparisons, with much higher certainty in field boundary predictions and significantly improved segmentations.

Next-stage validation is planned for Nigeria, with a one-year implementation timeline using an online labeling platform and convolutional neural network training. The ultimate goal is to extend the approach across all of Africa to inform investments in smallholder agriculture. The research team is seeking partnership funding for key personnel and cloud computing resources to support this scaling effort.