The massive loss of nitrogen (N) and phosphorus (P) from farmland ditches contributes to non-point source pollution, posing a significant global environmental challenge. Effectively removing these nutrients remains difficult in intensive agricultural systems. To address this, a novel composite ecological ditch system (CEDS) was developed by modifying traditional drainage ditches to integrate a grit chamber, zeolite, and ecological floating beds. Dynamic monitoring of N and P levels in water, plants, and zeolite was conducted to evaluate the system’s nutrient interception performance and mechanisms. The results showed the following: (1) Water quality improved markedly after passing through the CEDS, with nutrient concentrations decreasing progressively along the flow path. The system intercepted 41.0% of N and 31.9% of P, with inorganic N and particulate P as the primary forms of nutrient loss. (2) Zeolite removes N primarily through ion exchange, and P likely through chemical reactions, with maximum capacities of 3.47 g/kg for N and 1.83 g/kg for P. (3) Ecological floating beds with hydroponic cultivation enhanced nutrient uptake by the roots of Canna indica and Iris pseudacorus, with N uptake surpassing P. (4) Nutrient interception efficiency was positively correlated with temperature, ditch inlet concentrations, and rice runoff concentrations, but negatively with precipitation. This study demonstrates the CEDS’s potential for improving farmland water quality and suggests further enhancements in design and management to increase its economic and aesthetic value.

Agricultural nonpoint source pollution is emerging as one of the increasingly serious environmental concerns all over the world. This study conducted field experiments in Zengcheng District, Guangzhou City, from 2019 to 2023 to explore the mechanisms by which different crop types, fertilization modes, and meteorological conditions affect the loss of nitrogen and phosphorus in agricultural nonpoint source pollution. In rice and corn, the CK and PK treatment groups showed significant fitting advantages, such as the R² of rice-CK reaching 0.309. MAE was 0.395, and the R² of corn-PK was as high as 0.415. For compound fertilization groups such as NPK and OF, the model fitting ability decreased, such as the R² of rice-NPK dropping to 0.193 and the R² of corn-OF being only 0.168. In addition, the overall performance of the model was limited in the modeling of total phosphorus. A relatively good fit was achieved in corn (such as NPK group R² = 0.272) and in vegetables and citrus. R² was mostly below 0.25. The results indicated that fertilization management, crop types, and meteorological conditions affected nitrogen and phosphorus losses in agricultural runoff. Cornfields under conventional nitrogen, phosphorus, and potassium fertilizer (NPK) and conventional nitrogen and potassium fertilizer treatment without phosphorus fertilizer (NK) treatments exhibited the highest nitrogen losses, while citrus fields showed elevated phosphorus concentrations under NPK and PK treatments. Organic fertilizer treatments led to moderate nutrient losses but greater variability. Organic fertilizer treatments resulted in moderate nutrient losses but showed greater interannual variability. Meteorological drivers differed among crop types. Nitrogen enrichment was mainly associated with high temperature and precipitation, whereas phosphorus loss was primarily triggered by short-term extreme weather events. Linear regression models performed well under simple fertilization scenarios but struggled with complex nutrient dynamics. Crop-specific traits such as flooding in rice fields, irrigation in corn, and canopy coverage in citrus significantly influenced nutrient migration. The findings of this study highlight that nutrient losses are jointly regulated by crop systems, fertilization practices, and meteorological variability, particularly under extreme weather conditions. These findings underscore the necessity of crop-specific and climate-adaptive nutrient management strategies to reduce agricultural nonpoint source pollution. By integrating long-term field observations with machine learning–based analysis, this study provides scientific evidence to support sustainable fertilizer management, protection of water resources, and environmentally responsible agricultural development in subtropical regions. The proposed approaches contribute to sustainable land and water resource utilization and climate-resilient agricultural systems, aligning with the goals of sustainable development in rapidly urbanizing river basins.

Accurate county-level rice yield estimation is important for regional production monitoring, yet dense field-level observations are often unavailable and prediction across unseen years remains challenging. This study involved the development of a prior-guided environmental fusion network (PGEF-Net) for annual rice yield estimation in Guangdong Province, China. PGEF-Net uses a Random Forest-derived feature-importance prior to guide environmental attention and integrates environmental and general feature representations through gated fusion. Under leave-one-year-out (LOYO) validation of 447 county–year observations from 2020 to 2024, PGEF-Net achieved the highest pooled R² (0.883) and lowest RMSE (0.263 t ha⁻¹) among five learning-based models, but its RMSE advantage over the training-fold county-mean historical reference (0.268 t ha⁻¹) was not statistically significant. Ridge yielded the lowest MAE (0.154 t ha⁻¹) and MAPE (2.58%) among the learning-based models. Among environmental and remote-sensing predictors, thermal variables showed the strongest cross-method signal, with maximum temperature and GDD10 among the leading variables. Ablation analysis showed the largest numerical decreases in R² and increases in RMSE after removing Environmental Attention or the Environmental Branch. An additional yield anomaly analysis showed limited and year-dependent prediction performance after removing training-fold county means, indicating that persistent county-level yield differences were an important source of predictability under the present LOYO design.