IncepV3Dense: Deep Ensemble Based Average Learning Strategy for Identification of Micro-Nutrient Deficiency in Banana Crop

Abstract

The Nutrition of a crop is very essential for the health conditions during its growth stages and yield. A plant development is dependent on various nutrients absorbed from the natural environment or fertilizer supplements. The shortage or lack of essential nutrients is one of the crucial factors which impacts the overall crop yield. Computer vision based phenomics have become an emerging area in agricultural research. In this part of research work, we propose a significant image classifier model which contributes as an Ensemble based Convolutional Neural Network(E-CNN) that can diagnose banana crop’s micro-nutrient deficiency with improved accuracy using the leaf images. We selected Six popular deep learning pre-trained models namely VGG-19, InceptionResNetV2, InceptionV3, Xception, DenseNet169 and DenseNet201 and performed the modification of parameters in the top dense layers to experiment with the public available mendeley dataset containing banana crop leaf images with nutrient deficiencies. The diagnostic accuracy along with precision, recall, F1 score and support score was observed. On comparing the classifying accuracy parameters of the six mutated pretrained models, the modified DenseNet169 model attains the highest testing accuracy. The performance analysis was also done using confusion matrices. Finally, we created three binary ensembled models such as Xception+InceptionV3, Dense169+Xception and InceptionV3+Dense169 based on their top performance accuracy scores for the detection of micro nutrient deficiency in banana crop using the concept of averaging strategy. The proposed mutated ensemble based model InceptionV3+Dense169 (IncepV3Dense) attains a validation accuracy of 98.62% and f1 score of 93% for detecting banana crop micro nutrient deficiency.

Keywords

• Image classification
• transfer learning
• deep ensemble learning
• micro-nutrient deficiency