European Journal of Sustainable Agroecosystems

Background: Nutrient stress, especially nitrogen (N), phosphorus (P), potassium (K), and sulphur (S) shortages, continues to be one of the biggest yield-limiting barriers and has significant economic implications in rice cultivation both in irrigated and rain-fed environments. Conventional diagnostic procedures like visual evaluation and destructive tissue tests can be lengthy and only produce results after the crop has already experienced irreversible damage due to stress. The advent of the remote sensing (RS) approaches such as satellite systems, UAVs and proximal methods combined with machine learning (ML) and deep learning (DL) analytics is transforming nutrient stress diagnosis by making it possible to diagnose the stress condition early, and without destroying samples.
Aim: The goal of the current review is to summarize the available knowledge of peer-reviewed studies carried out from 2015 to 2025 regarding the integration of RS and ML/DL methods in the identification of nutrient stress in rice with respect to the identification of the contemporary methods of research, comparing the efficiency of the algorithms in combination with various sensors, as well as identifying unanswered questions in research.
To find relevant literature, the authors searched major databases including Scopus, Web of Science, PubMed, IEEE Xplore, ScienceDirect and Google Scholar using relevant keywords such as remote sensing, machine learning, nutrient stress, precision agriculture, and Oryza sativa. A total of seventy-one studies met the requirements and were accepted for the review after screening.
Key findings: Vegetation indices derived from drones and satellites (i.e., hyperspectral and multispectral) can effectively assess nitrogen status while the deficiencies of phosphorus, potassium and sulphur remain relatively less studied and have more complex spectral characteristics. Machine learning algorithms and ensemble methods such as RF and SVM as well as DL and transformer techniques perform better than traditional vegetation indices when combining data sources.
Remaining challenges involve the heavy nitrogen bias in research, lack of harmonized multi-nutrient datasets for field research, constraints in making models transferable across cultivars and regions, inability to integrate spectral data with soil data and climate variables, and the gap in performance between trials in laboratory conditions and actual agricultural applications. To summarize, RS-ML has a strong technological edge as a method for optimizing nutrient use in rice production, but it is still under-utilized. Future studies should focus on building a standardized dataset of multi-nutrient benchmarks, simple but interpretable models applicable in small-scale agriculture, and the connection of RS-ML to precision agriculture through the implementation of variable-rate fertilization equipment.