Crops Classification in Small Areas Using Unmanned Aerial Vehicles (UAV) and Deep Learning Pre-trained Models from Detectron2

Abstract

Crop identification has been a subject of study for a long time. With the help of satellites and manned aircraft, surveys are carried out in order to determine the areas planted with different crops. This is done for different purposes, and the methods used are expensive and time-consuming, which has restricted their use to very special cases. For small farmers, determining the planting areas for different species could lead to better coordination and distribution of the supply of agricultural products. The main objective of this work is to explore the use of drones to capture images, and Deep Learning pre-trained models techniques to classify crops in small areas, in order to provide a solution to problems related to the estimation of planting areas, for different horticultural products grown on a small scale. This is an exploratory work, in which the Detectron2 Zoo was evaluated. Detectron2 has been developed by Facebook IA and offers a great number of pre-trained models. These models can serve in many object detection tasks, and evaluating their performance might be the first step in the search for a practical solution to many problems, before building a model from scratch. In this work, we did not make changes to the architectures of the original models, but we explored some variations of the corresponding hyperparameters. We selected three classes of areas to detect in our experiments: uncultivated areas (no crop), green onion crops (onion), and foliage flowers crops (foliage). The results show that it is possible to detect small-scale crops and classify them with the help of Detectron2 models. However, we would still have to perform further evaluations in order to compare these models with other state-of-the-art methods, and we would still have to modify the architectures of these models in order to improve their performance. In future work, it is recommended to delve into aspects such as the count of plants in order to determine their quantity with greater precision and thus being able to carry out planning activities in a much more efficient way.