Abstract :
[en] By 2050, global demand for wheat is expected to increase by 60%. As a major producer of winter wheat, China faces severe structural challenges such as water scarcity and low efficiency in nitrogen fertilizer utilization. Achieving a sustainable increase in winter wheat production is critical important for China. Traditional agricultural monitoring methods are inefficient, a situation that has been significantly improved by the emergence of unmanned aerial vehicle (UAV) remote sensing technology. However, this technology still faces bottlenecks such as isolated models and a weak association to physiological mechanisms in practical applications. This research aims to systematically address a series of scientific problems from macro yield prediction to internal growth diagnosis through a gradually deepening technical path, providing a reliable and efficient technical solution for precision agriculture.
This research began by strengthening the predictive foundation. After that it utilized semi-supervised machine learning algorithms to overcome the limitation of data scarcity, and then analyzed key parameters to achieve collaborative inversion of coexisting parameters. Four steps were carried out as a systematic work.
The first step investigated the effectiveness of UAV remote sensing data for predicting winter wheat yield. A LSTM-random forest hybrid model based on the spectral characteristics of a single annual time series was constructed, it made winter wheat yield prediction successfully. This verified the effectiveness of the time-series remote sensing method, but also revealed its dependence on the amount of labeled data.
To overcome the data constraints, multiple years of data with various experimental treatments were introduced in the second step, and a self-training semi-supervised learning strategy was adopted to automatically expand training dataset for yield prediction. The original spectral reflectance was directly used as input features. These approches significantly enhanced the robustness and generalization ability of the yield prediction model.
By predicting the nitrogen content of the plants, the research in the third step was extended to the key internal physiological processes that drive yield formation. Based on the integration of high-dimensional spectral features, the plant nitrogen content was precise inverted, which opened up the "black box" of yield formation.
In the last stage, multi-task learning frameworks were introduced to solve the joint prediction problem of the biomass represented by dry weight and fresh weight which are two distinct but closely related components. This method effectively enhanced the training and prediction efficiency without reducing the overall prediction accuracy of biomass, thereby uncovered the intrinsic relationship between dry and fresh weights.
This study systematically constructed and validated a technical framework applicable to winter wheat growth monitoring. The monitoring system, progressing from "yield" to "nitrogen content" and then to "biomass components", achieved a shift from result prediction to process diagnosis. It provides a quantifiable technical tool for understanding the interactions among water, nitrogen, and biomass accumulation, holding significant theoretical value and application prospects.
The methodological system developed in this study provides core methodological support for promoting precision agriculture from single-point prediction to comprehensive diagnosis, and is significant for achieving optimal resource allocation and the strategic goal of food security.
