Development of a hybrid software system for monitoring erosion processes and evaluating the effectiveness of land reclamation measures using recommender algorithms

Abstract

The study addresses the problem of operational monitoring of erosion processes and justification of reclamation measures, bridging the gap between isolated models and practical management needs. The objective is to develop a hybrid software system that integrates physical-geographical modeling and machine learning with recommender algorithms for monitoring, forecasting, and selecting measures across large territories. The scientific novelty lies in the integration of RUSLE and gradient boosting into a unified architecture with a geoinformation database and a multicriteria optimization module, ensuring data compatibility and automation of decision-making. Methods are based on a modular architecture: data acquisition and preprocessing, analytical core, visualization, and decision support. The system uses SRTM 30 m DEM, multispectral images from Landsat 8/9 and Sentinel 2, soil maps, precipitation data from meteorological observations, and ERA5. Implementation is carried out in Python with GDAL/OGR, GeoPandas, Rasterio, Scikitlearn, and TensorFlow, with PostgreSQL/PostGIS as the storage solution. The hybrid model computes R, K, LS, C, and P factors using RUSLE and adjusts forecasts with a boosting model trained on field measurements. The recommender module ranks measures (contour plowing, terracing, shelterbelts, grassing) according to an integral efficiency and cost index. Validation in a pilot area demonstrated improved accuracy compared to baseline RUSLE: for arable land increased from 0.62 to 0.85, for pastures from 0.71 to 0.89, and for forests from 0.78 to 0.92; RMSE and MAE decreased by several tens of percent. The system significantly accelerated processing: the full analysis cycle for 1000 decreased from 391,5 to 36,6 minutes (a 10,69-fold speedup). The recommender module provides a rational compromise between ecological effectiveness and budget, suggesting technically and economically feasible measures. Discussion and sensitivity analysis confirm the dominant influence of factors C and LS: variations produce forecast changes over 30%, while K and R have a lesser effect. The practical significance consists in transitioning from degradation assessment to proactive planning, including scenario modeling and prioritization of measures. Limitations are related to the quality of input data and the need for model updates; future prospects include expanding observation sources, MLOps integration, and enhanced economic assessments. The software system is ready for industrial integration.