European Journal of Sustainable Agroecosystems

Background: The problem of nutrient deficiency in crops is one of the main causes which prevent the agricultural sector from achieving optimal results. Meanwhile, conventional methods of diagnosing crops nutrients are characterized by labor-intensiveness, long-time consumption and the absence of timely decision support in the assertion of precision nutrient management.
Objective: The objective of this study is to develop the Edge AI-based computer vision system assisting in real time detecting and classifying nutrition deficiencies in crops under the agricultural conditions.
Methods: The experimental field included wheat (Triticum aestivum L.), maize (Zea mays L.), and rice (Oryza sativa L.) according to 5 fertilizer treatment methods: control, nitrogen deficiency, phosphorus deficiency, potassium deficiency, and balanced fertilizer. The Edge AI-based decision support system used real-time data acquisition, on-device deep learning inference, and cloud analytics.
Results: The Edge AI system possessed a mean detection accuracy of 96.3% for the classification of nutrient deficiencies together with 127 milliseconds of latency for each inference, which allowed performing the processing in real-time at 7.9 frames/second. The system showed better performance metrics with precision values stretching from 94.1% to 97.8%, recall values ranging from 93.5% to 96.9%, and F1-scores accounting to 0.943-0.972 when describing the performance of the AI with various nutrient deficiencies. Multispectral imaging facilitated a rise in detection accuracy of deficiencies by 8.2% in comparison to the performance achieved with RGB-alone analysis. Field validation made with three different crops proved the viability of the deep learning model that performed uniformly in terms of getting reliable results under different environmental conditions.
Conclusions: This Edge AI framework gives an opportunity to speed up the process of diagnosing of nutrient deficiencies, reduce the need for people involved in this process by 78%, and also provide the capability to perform better precision nutrient management decisions in just 15 minutes after the observation in the field.