Comparative classification analysis of post-harvest growth detection from terrestrial LiDAR point clouds in precision agriculture

In precision agriculture, detailed geoinformation on plant and soil properties plays an important role, e.g., in crop protection or the application of fertilizers. This paper presents a comparative classification analysis for post-harvest growth detection using geometric and radiometric point cloud features of terrestrial laser scanning (TLS) data, considering the local neighborhood of each point. Radiometric correction of the TLS data was performed via an empirical range-correction function derived from a field experiment. Thereafter, the corrected amplitude and local elevation features were explored regarding their importance for classification. For the comparison, tree induction, Naïve Bayes, and k-Means-derived classifiers were tested for different point densities to distinguish between ground and post-harvest growth. The classification performance was validated against highly detailed RGB reference images and the red edge normalized difference vegetation index (NDVI705), derived from a hyperspectral sensor. Using both geometric and radiometric features, we achieved a precision of 99% with the tree induction. Compared to the reference image classification, the calculated post-harvest growth coverage map reached an accuracy of 80%. RGB and LiDAR-derived coverage showed a polynomial correlation to NDVI705 of degree two with R2 of 0.8 and 0.7, respectively. Larger post-harvest growth patches (>10 10 cm) could already be detected by a point density of 2 pts./0.01 m2. The results indicate a high potential of radiometric and geometric LiDAR point cloud features for the identification of post-harvest growth using tree induction classification. The proposed technique can potentially be applied over larger areas using vehicle-mounted scanners.