Supervisory Configuration of Deep Learning Networks for Plant Stress Detection and Synthetic Dataset Generation

In computer vision, plant stress detection involves the identification and classification of crop stresses. There are several approaches for the identification of green areas. The most recent approaches rely on machine-learning techniques or deep-learning networks to develop this task. Unfortunately, when attempting to use these networks to detect stressed plants, their performance drastically decreases. In most cases, these networks cannot detect plant stress. In addition, there are extensive repositories of plants on the internet. However, in most cases, these repositories do not include stressed plants. An alternative is to use networks to generate realistic synthetic images; nevertheless, these mathematical models frequently fail to produce accurate synthetic images (increasing supervision and collection times). Motivated by the latter, we propose a supervisory configuration of deep-learning networks to detect stressed plants and generate synthetic databases. This methodology consists of three phases. First, we collected a small set of Internet images of the stressed crops. Second, the process involves final layer training of the image generation model by introducing a new node into the network. Finally, we supervised the generative model using a classification neural network and a feedback loop. This supervision increased the quality of the generated synthetic images. Therefore, the experimental results were promising. The proposed configuration showed a 23.85% increase in average precision and a 10.8% increase in average recall compared with traditional classification architectures using the same synthetic dataset. These results demonstrated the feasibility of this configuration for the classification of stressed crops using synthetic datasets.