Deficiency Identification of Greenhouse Lettuce using Explainable AI

Abstract

The Convolutional Neural Network (CNN) based solutions are used to identify the nutrient deficiencies of the crops based on the color variances of the leaves. However, one of the major problems in the CNN based solutions are the lack of ability to explain the results obtained. This research is focused on overcoming this challenge by combining the results obtained from CNN with TensorFlow Inference Engine to provide humanely understandable results for deficiency identification of crops. Therefore, greenhouse lettuce is selected as the crop for the study. Greenhouse farming became popular with the technological evolution in the last few decades. This aims to provide an optimal nutrient composition to the corps, to protect the crops from pests without applying pesticides, and to provide the optimal environmental conditions to the corps such astemperature, humidity, etc. Lettuce is one of the mostly consumed vegetable crops among the green house crops but facing a yield loss due the nutrient deficiencies. Therefore, the early identification of deficiencies becomes crucial. A custom test bed is created to gather data/images and using those data/images YOLOv3 object detection model was trained to detect Calcium, Nitrogen, and Magnesium nutrient deficiencies of greenhouse lettuce. The results demonstrate a mean average precision of 94.38% on training data and 75.53% on custom data. The trained weights were combined with the TensorFlow Inference Engine to provide explainable results using a local knowledge base of deficiencies.