Using Hybrid Models of AI for Identification of Trees by UAV Images of Forests: I. Machine-learning Component of the Models

Abstract

Artificial intellect models (machine learning, logical reasoning, etc.) are currently the focus of many remote sensing approaches for forest inventory management. Although they return satisfactory results in many tasks, some challenges remain, especially in the case of the highly dense distribution of trees in forests. In this paper, we propose a novel hybrid approach using together deep learning models and symbolic logic methods for identifying single-tree species in highly dense areas. The use of deep learning methods in solving high dimensional problems in face recognition has some issues due to low accuracy and interpretability of results. The paper proposes a hybrid approach for solving complex image classification problems. This approach involves the use of both machine learning methods and symbolic knowledge. The paper presents the structure and formal model of the hybrid system, which includes a new component, an operations manager. The first part of the paper proposes a new architecture of deep neural networks with attentional mechanisms built on blocking meta-functions. The corresponding module has been developed in Python language. The results of the module's work are provided to the knowledge base. As a result of symbolic conclusions, the teaching module is reorganized. The experiments conducted showed the effectiveness of the presented approach.