Nutrient Deficiency Detection and Yield Loss Prediction in Black Pepper Using U2Net and Ensemble of Shallow CNN and MobileNetV2

Abstract

Background: Digital image analysis combined with deep learning offers powerful tools for detecting plant nutrient deficiency (ND), a critical challenge in precision agriculture. Aims: This study aims to develop an ensemble transfer learning approach for ND detection in black pepper (BP) and evaluate its effectiveness for classification and yield-loss (YL) prediction. Methods: An ensemble of MobileNetV2 and a custom-made shallow convolutional neural network was implemented, with U2Net-based background removal to improve feature extraction. The model was validated on the BP Dataset (DS)—BPNutriDef03 (4469 images)—and tested on a rice DS (4399 images). The framework included (1) ND classification using leaf imagery analysis and (2) YL forecasting based on nutrient deficiency severity (NDS). Results: The ensemble achieved classification accuracies of 99.22% for BP and 95.14% for rice. The yield prediction based on the NDS model estimated the YL of 27.83% for BP and 33.42% for rice. Conclusions: The proposed approach demonstrates robust performance and generalizability, offering a scalable, automated decision-support system for ND monitoring and yield prediction in precision crop management.