Protected agriculture plays a critical role in ensuring global food security by enabling stable crop production under controlled environmental conditions. However, climate change–driven extreme weather events have increased flood-related risks, posing growing threats to the safety, productivity, and sustainability of solar greenhouse cultivation. In practical greenhouse production, timely and accurate monitoring of waterlogging is essential for reducing crop losses, improving environmental regulation, and supporting scientific management decisions. Nevertheless, accurate segmentation of greenhouse waterlogging remains challenging because of limited monitoring visibility, uneven illumination, strong specular reflections, and the pronounced scale diversity of inundated areas. These adverse conditions often result in blurred boundaries, ambiguous textures, and unstable visual characteristics between flooded and non-flooded regions. In addition, satellite- and drone-based remote sensing is often unsuitable for greenhouse-scale applications because of resolution limitations and operational constraints. To address these challenges, this study developed an end-to-end image-based segmentation framework, termed the Residual Multiscale Cascade Encoder–Decoder (RMCED) network, for delineating waterlogging areas inside solar greenhouses. The proposed architecture adopted the middle four stages of ResNet-50 as the encoder to balance network depth and computational cost while extracting multi-level semantic features. This encoder design enabled the model to effectively capture both low-level texture patterns and high-level contextual representations, thereby improving discrimination between waterlogged areas and surrounding background regions. The decoder was designed as a multiscale cascaded attention decoder, which progressively restored spatial resolution through the hierarchical fusion of low- and high-level features. This design enhanced the network’s ability to identify fine spatial boundaries and accurately localize small-scale flooded regions that are often overlooked by conventional models. To ensure effective information flow, long skip connections were introduced between the encoder and decoder, thereby reducing feature degradation and semantic loss during feature transmission and improving the preservation of structural details. A key contribution of this work was the introduction of a Lighting-Aware Convolutional Block Attention Module (LA-CBAM). Compared with the original Convolutional Block Attention Module, LA-CBAM incorporated multi-kernel convolution filters and Sobel-gradient operations to improve both illumination robustness and edge sensitivity. The multi-kernel design enabled the model to capture contextual cues across different receptive fields and to better adapt to the diverse spatial scales of greenhouse waterlogging patterns. Meanwhile, the Sobel operation extracted gradient-based texture transitions and boundary-related information, strengthening boundary awareness under varying light intensities and reflective interference. As a result, the proposed network maintained strong discrimination capability even in complex greenhouse environments with reflections and uneven brightness, and it showed better adaptability to irregular and fragmented flooded regions. The model was trained and validated on a self-constructed dataset of 571 pixel-wise annotated images collected in a solar greenhouse under natural operating conditions. All images were resized and normalized using ImageNet statistics to match the pretrained ResNet-50 encoder. Comparative experiments with mainstream segmentation models, including U-Net, DeepLabV3+, and PSPNet, showed that RMCED outperformed representative baselines on the proposed dataset, achieving a precision of 93.22%. Ablation studies further demonstrated that LA-CBAM enhanced the model’s perception of key features in waterlogged areas, while the multiscale cascaded decoder enhanced edge continuity and object completeness. Overall, this study presented an efficient and robust framework for greenhouse waterlogging monitoring, providing a promising technical basis for flood risk assessment, waterlogging management, and related image-based monitoring tasks in protected agricultural systems.

Fresh corn is a high value-added agricultural product with high nutritional value and is deeply loved by consumers. Accurate water management is crucial for the yield and quality of fresh corn. This study was based on the online acquisition of water requirement and meteorological data for two consecutive crops of fresh corn using a large weighing lysimeter and meteorological monitoring equipment. Three machine learning algorithms, categorical boosting (CatBoost), long short-term memory (LSTM,) and random forest (RF), were used to construct a prediction model for fresh corn water requirement. The shapley additive explanations (SHAP) method was used to analyze the contribution of each feature factor to the prediction results in depth. Firstly, the study combined the boruta algorithm (Boruta) and least absolute shrinkage and selection operator (LASSO) feature selection methods to identify the key factors affecting the water requirements of fresh corn. Subsequently, the common meteorological factors selected by both feature selection methods were used as input variables for the model. The study then established CatBoost, LSTM, and RF models for predicting the water requirements of fresh corn. Concurrently, the SHAP analysis method was employed to perform an interpretability analysis of the prediction results, clarifying the contribution of meteorological features to the prediction outcomes and enhancing the transparency and reliability of the model. The research results indicate that through the Boruta and LASSO methods, it was discovered that daily mean temperature (T_ave), daily maximum temperature (T_max), daily minimum temperature (T_min), daily average relative humidity (R_H), daily average wind speed (W_ave), daily average atmospheric pressure (P₀), and accumulated daily sunshine hours (SD) are key factors affecting the water requirements of fresh corn. The coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE) of the CatBoost prediction model are 0.974 9, 0.018 9, 0.000 6, and 0.025 1, respectively, indicating significantly higher prediction accuracy compared to LSTM and RF. Considering its predictive efficiency, the CatBoost model has the highest number of training epochs (30 000) and the shortest training time (23.93 s), making it the optimal prediction model. Through SHAP analysis of the CatBoost model, it was found that daily minimum temperature, daily average temperature, and daily cumulative sunshine hours are key factors affecting the water demand of fresh corn, and all have a positive impact on the prediction results. Further local interpretability analysis revealed that the same meteorological features exhibit variations in their impact on water requirement prediction outcomes across different samples. The interaction between feature changes also affects the prediction results. This study conducted an interaction analysis on the three features (T_min, T_ave, SD) that contribute the most to the prediction results and found that the accumulated sunshine hours increase with rising air temperature, and the water requirement for fresh corn also increases accordingly. The study can provide a reference basis for precise water management of fresh corn.