Web-based tool for measuring soil, plant, and residue cover

Overview

This project delivers a web-based tool that uses convolutional deep neural networks to automatically quantify the fraction of soil, plant canopy, and crop residue from downward-facing images. The solution addresses the need for accurate, scalable land cover measurement in agricultural and environmental research. By applying semantic segmentation techniques, the tool can distinguish between bare soil, plant material, and crop stubble while accounting for morphology and image context, outperforming traditional hard thresholding approaches. The web interface, built with Python Streamlit, makes the technology accessible to researchers, agronomists, and land managers without requiring deep learning expertise.

Technical specifications

• Deep learning architecture: Uses a SegNet encoder-decoder convolutional neural network for pixel-level classification of land cover components
• Classification targets: Bare soil, plant canopy, and crop residue (stubble)
• Image requirements: Downward-facing photographs of agricultural fields
• Validation accuracy: Achieved greater than 90 percent accuracy on labeled validation sets, with the ability to match or improve upon human labeling
• Web interface: Python Streamlit application for accessible, browser-based use
• Training framework: Tensorflow-based model with tunable constants and network layers
• Dataset approach: Initial training on nearly 3,000 labeled images, with planned expansion to approximately 10,000 images to capture diverse crops, lighting conditions, and field scenarios

Technology readiness level

The technology has been partially validated through testing of a SegNet encoder-decoder model on a dataset of nearly 3,000 laboratory-collected and labeled images. The model successfully classifies images similar to its training set and has demonstrated the ability to improve on human labeling accuracy. Current development focuses on expanding the dataset to account for diverse crop stubble types, color schemes, contrast variations, and shadow conditions across different times of day. Future work includes field-based data collection, quality control, manual pixel-level labeling, model retraining with improved parameters, and deployment as a web-accessible tool. The project is seeking support for data collection, labeling efforts, and computing infrastructure including a new deep learning workstation.