TY - JOUR

T1 - Прогнозирование урожая пшеницы с использованием гиперспектральных данных дистанционного зондирования

AU - Митрофанова, Ольга Александровна

AU - Митрофанов, Евгений Павлович

AU - Буре, Владимир Мансурович

AU - Гармашов, Никита Сергеевич

PY - 2026/8/4

Y1 - 2026/8/4

N2 - The relevance of the research is due to the need to develop effective methods for operationalforecasting of grain crop yields. Traditional ground-based methods are time-consuming, whileremote sensing data (RSD), especially hyperspectral imaging, provide a unique opportunityfor non-invasive monitoring of crop conditions. However, the potential of RSD is often limiteddue to the high dimensionality of the data, multicollinearity, and insufficient knowledge ofculture-specific spectral features. The aim of the work was to analyze the informative valueof predictors and the possibilities of using machine learning methods to predict wheat yieldsusing hyperspectral data. The following tasks have been solved to achieve it: collection andpreparation of two heterogeneous datasets (from controlled test sites and from random points on an inhomogeneous field); development and implementation of an algorithm for automatedsearch for new vegetation indices (VI) by iterating through combinations of spectral chan-nels; conducting statistical data analysis to assess the significance of predictors; comparativeanalysis of the effectiveness of 11 feature selection methods and a variety of classificationand regression algorithms. As a result of the automated search, more than 120 million com-binations were processed and a number of new types were identified. Statistical analysisconfirmed stronger direct correlations of yield with agrochemical indicators and multispec-tral indexes, while information from hyperspectral channels, although less explicit, turnedout to be concentrated in certain areas of the spectrum. Forecasting experiments on thesetest sites have shown the critical importance of the feature selection stage. The best accuracyin the classification problem was achieved by the Gradient Boosting model with ANOVA se-lection of 10 features (Accuracy 0.77, F1-score 0.70), and in the regression problem by thesupport vector machine with RFE selection based on Random Forest (MAE 2.8, MAPE16.7 %). The most significant features included the newly proposed indexes (VI1–VI9)

AB - The relevance of the research is due to the need to develop effective methods for operationalforecasting of grain crop yields. Traditional ground-based methods are time-consuming, whileremote sensing data (RSD), especially hyperspectral imaging, provide a unique opportunityfor non-invasive monitoring of crop conditions. However, the potential of RSD is often limiteddue to the high dimensionality of the data, multicollinearity, and insufficient knowledge ofculture-specific spectral features. The aim of the work was to analyze the informative valueof predictors and the possibilities of using machine learning methods to predict wheat yieldsusing hyperspectral data. The following tasks have been solved to achieve it: collection andpreparation of two heterogeneous datasets (from controlled test sites and from random points on an inhomogeneous field); development and implementation of an algorithm for automatedsearch for new vegetation indices (VI) by iterating through combinations of spectral chan-nels; conducting statistical data analysis to assess the significance of predictors; comparativeanalysis of the effectiveness of 11 feature selection methods and a variety of classificationand regression algorithms. As a result of the automated search, more than 120 million com-binations were processed and a number of new types were identified. Statistical analysisconfirmed stronger direct correlations of yield with agrochemical indicators and multispec-tral indexes, while information from hyperspectral channels, although less explicit, turnedout to be concentrated in certain areas of the spectrum. Forecasting experiments on thesetest sites have shown the critical importance of the feature selection stage. The best accuracyin the classification problem was achieved by the Gradient Boosting model with ANOVA se-lection of 10 features (Accuracy 0.77, F1-score 0.70), and in the regression problem by thesupport vector machine with RFE selection based on Random Forest (MAE 2.8, MAPE16.7 %). The most significant features included the newly proposed indexes (VI1–VI9)

KW - canonical correlation analysis

KW - hyperspectral data

KW - machine learning

KW - precision agriculture

KW - principal component analysis

KW - remote sensing

KW - vegetation indexes

KW - wheat

KW - yield forecasting

UR - https://www.mendeley.com/catalogue/b9931def-6b94-3171-b146-b624477bf64c/

U2 - 10.21638/spbu10.2026.210

DO - 10.21638/spbu10.2026.210

M3 - статья

VL - 22

SP - 316

EP - 334

JO - ВЕСТНИК САНКТ-ПЕТЕРБУРГСКОГО УНИВЕРСИТЕТА. ПРИКЛАДНАЯ МАТЕМАТИКА. ИНФОРМАТИКА. ПРОЦЕССЫ УПРАВЛЕНИЯ

JF - ВЕСТНИК САНКТ-ПЕТЕРБУРГСКОГО УНИВЕРСИТЕТА. ПРИКЛАДНАЯ МАТЕМАТИКА. ИНФОРМАТИКА. ПРОЦЕССЫ УПРАВЛЕНИЯ

SN - 1811-9905

IS - 2

ER -