Vegetation indexes help perform precision farming because they provide useful information regarding moisture, nutrient content, and crop health. Primary sources of those indexes are satellites and unmanned aerial vehicles equipped with expensive multispectral sensors. Reducing the price of obtaining such information would increase the availability of precision farming. Several studies have proposed deep neural network methods to estimate the indexes from RGB color images. However, these studies report relatively large errors for mature plants, when highly non-linear relationships between image RGB bands and vegetation indexes arise. One could apply multilayer random forest-based models (Deep Forests) to solve this problem, but they have limited discriminative power and ability on catching non-linear relationships between image features. The cornerstone of the Deep Forests is that at each layer they enrich original features with embeddings containing empiric class probabilities from previous layers, although these probabilities deliver limited information. In this paper, we propose methods, which combine ideas of Deep Forests, Random Forests of multivariate trees, and global pruning of Random Forests to tackle these problems. We applied oblique (linear) and kernel (non-linear) trees as basic classifiers of the Deep Forest to improve discriminative power. We also utilized a method to refine Random Forests with a global loss optimization. This method helps to generate more expressive embeddings at each layer of the Deep Forest, which significantly improves results of the data analysis. In the experiments, we compared those methods with AlexNet and ResNet-based neural networks on several image classification datasets as well as on the NDVI prediction task. The experiments on image classification show that the proposed Deep Forest-based methods provide competitive results on datasets with small and medium size of feature-set. The results of the NDVI prediction task indicate that these methods are robust to senescence of plants and outperform neural network-based solutions.
DOI: 10.1109/ACCESS.2023.3261129
Download PDF or read onlin at the IEEE Xplore library: https://ieeexplore.ieee.org/document/10078876
ResearchGate: https://www.researchgate.net/publication/369490986_Estimation_of_vegetation_indices_with_Random_Kernel_Forests
D. A. Devyatkin. Estimation of Vegetation Indices With Random Kernel Forests // in IEEE Access, vol. 11, pp. 29500-29509, 2023.