Estimation of Vegetation Indices With Random Kernel Forests

Vegetation indexes help perform precision farming because they provide useful information regarding moisture, nutrient content, and crop health. Primary sources of those are satellites unmanned aerial vehicles equipped with expensive multispectral sensors. Reducing the price obtaining such would increase availability farming. Several studies have proposed deep neural network methods to estimate from RGB color images. However, these report relatively large errors for mature plants, when highly non-linear relationships between image bands vegetation arise. One could apply multilayer random forest-based models (Deep Forests) solve this problem, but limited discriminative power ability on catching features. The cornerstone Deep Forests is that at each layer enrich original features embeddings containing empiric class probabilities previous layers, although deliver information. In paper, we propose methods, which combine ideas Forests, Random multivariate trees, global pruning tackle problems. We applied oblique (linear) kernel (non-linear) trees as basic classifiers Forest improve power. also utilized a method refine loss optimization. This helps generate more expressive Forest, significantly improves results data analysis. experiments, compared AlexNet ResNet-based networks several classification datasets well NDVI prediction task. experiments show Forest-based competitive small medium size feature-set. task indicate robust senescence plants outperform network-based solutions.