Fig 1.
UAV-driven remote sensing of above-ground biomass in rice crops based on NIR imagery.
Fig 2.
Crop setup: (a) Rice crops. (b) Each plot was designed with an area of 4.95m2. (c) Destructive biomass sampling. (d) An example of a Ground-Truth biomass (BM) dataset. The crop field was designed with three spatial repetitions (Rep) containing 2 contrasting rice genotypes.
Fig 3.
Detailed presentation of the UAV system utilized in this work.
Fig 4.
The two clusters in the RGN space, the foreground (TF) and background (TB), created following Algorithm 1.
The segmented RGN image was captured using the Parrot-Sequoia, stacking the respective multi-spectral camera bands, after aligning the images according to the camera intrinsics and drone altitude.
Fig 5.
Uniform random distribution of grouped and classified pixels on RGN image.
White pixels are associated with vegetation, while black pixels are associated with soil.
Fig 6.
(a-topleft) The original image in RGN color space, (b-topright) the hard segmentation output of the GFkuts algorithm after 5 iterations, (c-bottomleft) the soft segmentation result of the GF refinement, (d-bottomright) and the adaptive thresholding output to create a binary mask of the canopy.
Fig 7.
F1-score and accuracy for all tested algorithms reported in Table 1.
Table 1.
Image segmentation performance.
Mean results over 400 NIR images (image sub-regions of 10 × 10 pixels).
Fig 8.
Rice canopy detail after the segmentation process: (a) segmentation results for each tested algorithm. (b) reconstructed image using four channel data space depicted in (c) RGN+(red-edge) space.
Table 2.
Near-infrared vegetation indices for non-destructive above-ground biomass estimations.
The term ρf refers to the reflectance value at the frequency f).
Fig 9.
Correlation matrix for the extracted features for (a) up-land and (b) low-land production systems.
The terms D-BM and F-BM correspond to the dry and fresh biomass, respectively. WC is the water content, while the rest of the features correspond to the Vegetation Indices (VIs) defined in Table 2.
Fig 10.
Vegetation Index computation: (a) VI variance through an entire phenological cycle. (b) An example of the VI-feature dynamics during a single growing stage. The inset shows the rice-leaf healthy status based on different wavelength readings.
Fig 11.
UAV crop coverage: (a) 3D flight trajectory. The UAV was set to fly at 20m over the crop at a maximum speed of 1.5ms−1. The black dots at ground-level correspond to the GPS-tracks of aerial imagery samples. (b) Crop fields—CIAT base station. (c) Parrot-Sequoia multispectral camera bands.
Fig 12.
AGBD estimation results: (a,d) plot segmentation comparative results between the K-means and the proposed GFKuts approach. (b,e) ANN-driven estimations in biomass VS Ground-truth measurements. (c,f) ElasticNet-driven identification of the biomass readings according to the planted rice varieties under lowland and upland production systems.