Persimmons fruits, originating from China, boast a cultivation history exceeding 3 000 years. In actual production practice, the phenotypic characteristics of persimmon fruits serve as crucial criteria for variety identification. Attributes such as fruit shape, longitudinal groove, fruit apex, cross furrow, and fruit indent, constitute the foundation for naming and classifying new varieties, underscoring the significance of phenotypic traits for germplasm resource identification. Nevertheless persimmon species are diverse and resourceful, with large similarities between varieties, making manual identification time-consuming and labor-intensive. Moreover, the phenotypic parameters of naturally grown ripe persimmons exhibit different degrees of heterogeneity and diversity—such as inconsistent outline shapes and sizes, varying depths of longitudinal grooves, cross-grooves, and fruit indent, as well as differences in the spatial geometry of fruit apex—making it challenging to use a single phenotypic parameter as a baseline for accurately observing and discriminating among different varieties. Currently, researches on phenotypic multi-label recognition primarily focus on the extraction of fruit phenotypic parameters, with relatively few studies dedicated to phenotypic recognition of persimmon fruits. To achieve rapid and accurate identification of ripe persimmon varieties and their phenotypic traits, an enhanced multi-label recognition model, YOLOv8m-LCA, is proposed based on YOLOv8m. Here, LCA represents three newly optimized modules in the baseline network architecture of YOLOv8m: L stands the initial of LSKA (large separable kernel attention) module, C for that of CBAM (convolutional block attention module) and A for that of subsampling Adown module. To address the problem of irregular size and difficulty in accurately extracting the contours of naturally ripened persimmons, a large separable kernel attention module is newly added into the C2f (cross-stage partial-connection with 2 convolutions) module in the YOLOv8m network, enhancing the fruit edge feature details and reducing the computation load of the model. To improve the recognition rate of the model for key phenotypic features, four new convolutional block attention modules were newly added into the backbone network to strengthen its weighted attention processing capability in both spatial and channel dimensions. Furthermore, a dual-path subsampling module was adopted to replace the original convolutional subsampling module reducing the feature maps' high-frequency information loss. The classification loss function of the head network was optimized to enable the output of multi-labelled instances encompassing both varieties and phenotypic features. During the image acquisition process, objective factors such as variations in light intensity were fully considered to obtain a persimmon image dataset containing 30 varieties. Images that failed to accurately reflect the target features were removed, and enhancement operations were applied to the divided training and validation images, yielding a total of 5 060 images in the enhanced dataset. The results show that on the self-constructed dataset of 30 persimmon varieties and 14 types of phenotypic characters, in comparison with YOLOv8m, the proposed YOLOv8m-LCA model achieves an overall recognition precision of 93.0%, a recall of 92.0%, and an mAP@50 of 94.9%, representing respective increases of 7.4, 5.8 and 6.8 percentage points. Additionally, the model size, the floating point of operations, and the number of parameters stand at 27.6 MB, 41.3 G and 13.63 M, which are 47.02%, 47.59% and 47.29% lower than those of YOLOv8m, respectively. The method proposed in this paper outperform other current and classical YOLO series algorithms, such as that of YOLOv5m, YOLOv9m, YOLOv10m, YOLOv11m and YOLOv12m, respectively. It provids a model reference for the identification of phenotypic characteristics of persimmon and other fruit germplasm resources, and can also be applied to multi-label feature extraction for crops with more complex spatial structure, edges and textures in natural growing environment.

Accurate canopy area temperature prediction played a crucial role in intelligent agricultural production and climate regulation. As a key environmental factor, canopy temperature significantly influenced crop growth, physiological metabolism, and the interaction between plants and pathogens, ultimately affecting crop health, yield, and quality. However, traditional temperature forecasting methods struggled to capture multi-scale dependencies in long and short time sequences, leading to suboptimal predictive performance. To address these challenges, a fusion model based on the Informer architecture and long short-term memory neural networks (LSTM) was proposed for canopy temperature prediction by integrating multi-source environmental data. In the encoder layer, a sparse attention mechanism was employed to extract multi-scale information from the input features. This mechanism was particularly suited for long-term time series data, as it effectively captured the coupling relationships between environmental factors at different time scales. Sparse attention allowed the model to focus on the most relevant information, which was crucial for long-term temperature prediction, as it reduced noise and redundancy commonly present in traditional models. In the decoder layer, LSTM was used to extract short-term temporal dependencies, further enhancing the continuity and consistency within the time series data. The ability of LSTM to capture sequential information over short intervals enabled the model maintain prediction consistency by leveraging temporal patterns. This contributed to more reliable predictions, especially in dynamic environments where short-term fluctuations in temperature played a key role. Additionally, an improved residual feedforward network (IRFFN) was incorporated to further optimize the model structure. This module improved information flow through the network by allowing residual connections, alleviating the vanishing gradient problem often encountered during training. By enhancing the model's ability to learn complex non-linear relationships, IRFFN enabled improve prediction performance without significantly increasing computational complexity. The experiment was conducted in a kiwifruit-grape intercropping orchard, where a suspended inspection robot was deployed to collect canopy microenvironmental data from January 2021 to 2024, with measurements recorded at 30-minute intervals. Data preprocessing was performed using the isolation forest algorithm to eliminate outliers, followed by normalization to standardize feature values. A Spearman correlation analysis was conducted to identify key environmental variables affecting canopy temperature, such as canopy temperature, humidity, soil temperature, wind speed, atmospheric pressure, and soil moisture, which were selected as input features for the model. Experimental results demonstrated that the proposed Informer-LSTM model significantly outperformed mainstream models, including Times-Net, Autoformer, Informer, and Reformer, in terms of prediction accuracy. The model achieved a mean absolute error (MAE) of 0.166 ℃, a root mean square error (RMSE) of 0.224 ℃, and a coefficient of determination (R²) of 0.978. Compared to the baseline Informer model, the MAE and RMSE were reduced by 0.265 and 0.448 ℃, respectively. The stability of the model was further verified by testing predictions at different time steps (1, 6, 12 h). In the task of predicting the future 1 h, the MAE and RMSE were 0.468 and 0.583 ℃, respectively, while in the task of predicting the future 12 h, the MAE and RMSE increased to 1.368 and 1.527 ℃, respectively, and the average MAE and RMSE values were 31.75% and 29.62% lower than the baseline model Informer. The final validation experiment was conducted using actual canopy microenvironmental data from February 2024, confirming the model’s generalization capability and robustness, with an RMSE of 0.259 ℃, MAE of 0.178 ℃, and R² of 0.969. This study demonstrated that Informer-LSTM provided a high-precision and stable approach for short- and medium-term canopy temperature forecasting.