Source-linked AI summary
Plant Growth Estimation with a Camera-Based Vegetation Index Mapping System for Agricultural Ground Vehicles
Lukas Pindl, Michael Maier, Timo Oksanen
TL;DR
The paper addresses high-resolution, real-time plant-growth mapping from agricultural ground vehicles. It projects multispectral camera images onto an estimated ground plane and compares the resulting maps with a commercial sensor. The systems show similar trends, while the camera approach supports ground-pixel filtering and real-time processing even on a laptop.
Problem
Existing plant-growth mapping options include expensive, low-resolution ground sensors and aerial imagery with stitching, lower resolution, or cloud-coverage limitations.
Method
A tractor-mounted multispectral camera uses perspective projection and estimated vehicle positions to merge NDVI images into a field map.
Results
The camera-based and commercial sensors show similar trends, with a parcel-level Pearson correlation of 0.841, while camera processing is fast enough for two cameras at 1 Hz on the laptop setup.
Takeaways & Limitations
The approach provides real-time, high-resolution plant-growth maps and can filter ground pixels to separate visible plant coverage from plant NDVI differences.
Takeaways & Limitations
The projection assumes a flat ground plane, does not yet account for radial distortion, and has been tested with only one camera despite support for merging multiple cameras.
Abstract
from arXiv · showhide
This work presents a camera-based sensor for mapping plant growth over an agricultural field. The sensor can be used on ground-based vehicles like a tractor and relies on RTK-GNSS to correctly merge many multispectral images onto one map. NDVI is used as an index to estimate plant growth, but the approach can be used with other indices as well depending on the camera. The images from the camera are projected onto the estimated ground plane using perspective projection. This computationally simple approach allows for real-time processing of all images on the field even with low-end hardware. The results are compared to a commercially established vehicle-mounted sensor. Absolute values are hard to compare, but both sensors show similar trends. The camera-based approach also allows for filtering of ground and non-ground areas, potentially reducing the impact of crop density on the average measured NDVI.
1. INTRODUCTION
The paper motivates a tractor-mounted camera system for high-resolution plant-growth mapping to support nitrogen-use efficiency, addressing limits of expensive, low-resolution vehicle sensors and weather-sensitive aerial imaging.
- Vegetation indices such as NDVI can identify field areas with different yield potential and support decisions about where to apply more or less fertilizer.
- Existing ground-vehicle sensors can accurately map plant growth independently of weather but are typically expensive, low-resolution, and based on a single optical sensor.
- The proposed contribution is a camera mounted to a tractor that generates high-resolution maps of the whole field in real time.
- Multispectral imaging is also used from drones and satellites, but drone images require later stitching while satellite imagery has lower resolution and can be blocked by clouds.
2. METHODS & MATERIALS
The system computes vegetation indices from a tractor-mounted multispectral camera, projects pixels onto a ground plane, and georeferences measurements with vehicle-position estimates for field mapping.
- 2.1 Applied Methods: NDVI compares plant reflectance in the red and near-infrared spectra, with values near 0 indicating soil or dead plants and values near 1 indicating highly healthy plants.
- 2.1 Applied Methods: Each image pixel is mapped to the ground using a pinhole camera model, camera rays, coordinate transformation, and intersection with the assumed ground plane.
- 2.1 Applied Methods: The ground-plane projection assumes locally flat terrain, becomes less reliable on strongly curved fields or at long distances, and therefore uses points within 20 m of the camera.
- 2.2 Used Material: A Sentera multispectral camera captured 1024x768 images while mounted on an agricultural spreader, with a PAR sensor tracking sunlight intensity.
- 2.2 Used Material: The study used 40 winter-wheat parcels with five nitrogen amounts, comparing the camera-based setup against downward-facing commercial N-Sensors.
3. ALGORITHM
The system captures multispectral images, computes NDVI, projects pixels onto the ground, and merges observations into a grid map using GNSS-based positioning. It supports optional ground filtering and weighted map updates, while multiple-camera merging remains untested.
- Image Acquisition and Processing: Images are captured through Ethernet-triggered requests while multithreaded processing retrieves each image and records trigger times for position calculations.The trigger-to-availability delay can exceed one second, so background processing allows other work to continue while images become available.
- Image Acquisition and Processing: NDVI is calculated per pixel from camera channels, with dark pixels and potentially oversaturated readings excluded to reduce unreliable values.The camera’s 8-bit resolution makes dark-area ratios noisy, while channel oversaturation can produce strongly erroneous NDVI values.
- Map Projection and Merging: Each NDVI image is projected onto the ground plane using camera pose and vehicle position, with ray-ground intersections computed in a vectorized implementation.The method generally works without adjusting the ground plane for plant height, although such adjustment may help for very high or dense plants.
- Map Projection and Merging: Projected pixels are averaged into 20 cm grid cells and merged with existing map data, filling empty cells and combining overlaps by weighted averaging.The merge weight k determines how much the new value contributes relative to the old value.
- Map Analysis: Parcel-level analysis averages mapped pixels within buffered parcel areas to compare mean NDVI across nitrogen application strategies.A usual 0.8 m buffer excludes expected transitional areas near parcel edges.
4. DATA COLLECTION
Data were collected on a sunny day about one month after first fertilization while the tractor traversed the tramlines in both directions. The camera faced left, allowing coverage of all parcels during the route.
- Data Collection: The experiment was conducted on a sunny day in late April, approximately one month after the parcels’ first fertilization.
- Data Collection: The tractor followed the rightmost and left tramlines and then returned along the same path to cover all parcels with the left-facing camera.
5. RESULTS
The generated maps reflected parcel layout and nitrogen-related NDVI patterns, while filtered maps had higher average NDVI but similar distributions. The camera-based system correlated strongly with the commercial sensor, and GPU processing was much faster than laptop processing.
- Parcel Mapping: The final NDVI map visibly matched the parcel layout, with more highly fertilized areas appearing brighter.The displayed map was cropped for emphasis and did not apply a ground-pixel filter.
- Nitrogen Response: NDVI clearly increased from 0 to 40 kg/ha of nitrogen before the curve gradually flattened at higher application amounts.Blue points represent parcel means, orange points average each nitrogen step, and bars show standard deviation.
- Ground Filtering: Activating ground-pixel filtering increased average NDVI while preserving a general shape and distribution similar to the unfiltered case.
- Sensor Comparison: Pearson correlation between the camera-based and commercial sensors was 0.841 using parcel means, with similar standard deviations despite difficult absolute-value comparison.Different wavelengths and calibration make direct comparison of absolute NDVI values difficult.
- Processing Performance: GPU-supported processing projected 459 images in 6.14 s, averaging 13.4 ms per image, whereas the laptop averaged 174 ms per image.The timings include image reading, ground projection, and merging with the existing map.
6. DISCUSSION AND CONCLUSION
The camera-based mapping captured the expected nitrogen-response trend and showed similar parcel-level trends to a commercial sensor, while ground-pixel filtering raised average NDVI without changing the overall pattern. The approach also supports separating visible plant coverage from plant NDVI, but projection accuracy and response variance remain limitations.
- Ground-pixel filtering: Ignoring ground pixels increased average NDVI while preserving the general shape and distribution of the unfiltered nitrogen-response comparison.In this instance, filtering did not have a major overall impact, but it may help applications focused on plant-only NDVI.
- Commercial-sensor comparison: Pearson correlation between the camera-based and commercial sensor measurements was 0.841 across parcel means, with similar standard deviations despite difficult absolute-value comparisons.Differences in wavelengths and calibration limit direct comparison of absolute NDVI values.
- Ground-pixel filtering: Ground-pixel filtering can separate visible plant coverage effects from differences in the NDVI of the plants themselves.This provides an additional analysis option when lower biomass would otherwise lower the area-averaged NDVI through more ground pixels.
- Limitations and future work: Further work should investigate the high variance in nitrogen response and improve projection by accounting for radial distortion and non-flat ground planes.The current projection assumes a perfectly flat ground plane, with accuracy becoming worse for strongly curved fields and long distances.