{Tofael Ahamed M Nazim Uddin Ehtashamul Haque Nabil Anan Orka}

Abstract

Early detection of plant nutritional deficiencies, followed by prompt corrective measures, is crucial for maintaining crop yield and produce quality. However, detecting these early signs in plant leaves often proves challenging, even with computer-aided diagnostic tools, because of their subtlety. As a solution, this study introduces Nutrispace, a new color space that enhances deep learning-based nutritional stress recognition by accentuating early signs of nutritional deficiency in leaf images. In this study, we evaluated Nutrispace’s efficacy by comparing its performance to RGB, HSV, and CIELAB using three lightweight classifiers: EfficientNetB0, MobileNetV2, and DenseNet121. To better understand Nutrispace’s effective range, we tested these classifiers with four image input sizes: 32 × 32, 64 × 64, 128 × 128, and 256 × 256. Our test dataset comprised images of ash gourd (Benincasa hispida), bitter gourd (Momordica charantia), and snake gourd (Trichosanthes cucumerina) leaves with early nitrogen and potassium deficiencies, as well as healthy controls. Our findings show that Nutrispace consistently improved accuracy across all 12 test cases, with improvements ranging from 1% to more than 8% compared to RGB. The performance improvement was more significant for higher-resolution cases, with Nutrispace achieving the maximum test accuracy of 90.62% on 256 × 256 images. Overall, Nutrispace performed effectively, irrespective of classifier structure and input size