Convolutional Neural Network Based Nutrient Deficiency Classification in Leaves of Elaeis guineensis Jacq

Abstract

Nutrient deficiency is one of the main causes of the decline in oil palm production. In fact, oil palm farmers cannot diagnose the symptoms of the nutrient deficiency by themselves. Generally, the collected leaf samples need to be analyzed using laboratory equipment which consumes time and budget. In this research, the leaf samples including fronds 17 and 25 were collected from 37 oil palm trees. Frond 17 was used to analyze the amounts of Nitrogen (N), Phosphorus (P), Potassium (K), Magnesium (Mg), and Boron (B). Based on biochemical tests on palm leaves collected, the relationship between oil palm leaf characteristics and its amounts of nutrients was studied. Deep learning models were developed using Convolutional Neural Networks (CNN) to diagnose nutrient deficiency in oil palm from the leaves’ images. The nutrient results were classified into 3 groups: deficiency, normal, and excess. Totally, 682 images from frond 25 from trees across a farm were used for image data collection. Various CNN benchmark architectures were used to analyze the performance of nutrient deficiency classification but the separable convolutional CSBio2020 and BettaNet architectures were found to be ideal enough for the set of data collected. Separate models were trained to predict the levels of each nutrient. The models’ average accuracy is 77.2% for CSBio2020 and 80.4% for BettaNet. The average precision, recall, and F1 score are 0.75, 0.75, and 0.747, respectively, for CSBio2020 and 0.76, 0.813, and 0.775, respectively, for BettaNet. Many studies in the area dealt with few nutrients but this paper has all the nutrients analyzed and a deep learning architecture with parameters optimized to fit all in one with more optimal performance when compared to other existing methodologies.