Farmland shelterbelts can be represented as typical narrow linear features in remote sensing imagery. However, the significant challenges remain in accurate segmentation, due to their strong global contextual dependencies and weak local features. In this study, a semantic segmentation model, named MFF-Net, was proposed to accurately and rapidly extract cultivated land and shelterbelts. Multi-feature fusion block (MFFB) and spatially gated fusion mechanism (SGFM) were also adaptively integrated with three complementary approaches: the long-range dependency modeling using a mamba-like linear attention (MLLA) operator, the local detail perception in convolutional networks, and the frequency-domain edge enhancement via fast Fourier transform (FFT). The feature representation was effectively balanced when segmenting elongated objects. Furthermore, a task- oriented super-resolution preprocessing was introduced into the network. This front-end step was designed to reconstruct high-resolution images. Thereby, the textual details were enhanced to sharpen the boundaries of subtle features, like shelterbelts. Superior input was provided for the subsequent segmentation model. A series of experiments was performed on the self-constructed dataset of the farmland shelterbelt. The results demonstrate that the superior performance of MFF-Net was achieved in the precision rates of 96.42% for cropland and 82.83% for shelterbelts, with a mean intersection over union (mIoU) of 83.45%, thus outperforming a range of advanced models, including RS³Mamba, DC-Swin, DeepLabV3+, and SegMAN. Ablation studies validated the effectiveness of each component within the multi-feature fusion. Frequency-domain features via FFTFormer contributed to a 2.80% increase in the shelterbelt IoU, indicating the enhanced boundary discrimination. The spatially gated fusion mechanism (SGFM) further boosted the overall mIoU by 1.71%, indicating the adaptive balance on the contribution rates from different feature domains. Compared with a baseline model, the full MFFB was improved mIoU of 5.14%. Super-resolution preprocessing was integrated with the highly effective auxiliary strategy. Taking 4x upsampled images as the input, there was a remarkable 6.61% increase in shelterbelt IoU and a 2.47% gain in overall mIoU. The difficulties with segmenting narrow targets were effectively mitigated to augment their pixel-width and edge clarity. A complete technical pipeline was established from pixel-level segmentation to vectorization and application. Accurate area estimation was achieved with average relative errors of 7.50% for cropland and 6.76% for shelterbelts, compared with the national survey data. An application analysis of shelterbelt closure degree was also made on individual farmland plots. The 91.95% of the plots met the national standard requirement (≥0.75). In conclusion, the MFF-Net model can be expected to effectively segment the narrow linear features in complex landscapes. Synergistic fusion of global, local, and frequency-domain features was combined with task-oriented super-resolution enhancement. This research can provide a robust technical pathway for the precise monitoring and dynamic evaluation of farmland shelterbelt networks in ecological conservation.

Simultaneous Localization and Mapping (SLAM) is one of the most crucial aspects of autonomous navigation in mobile robots. The core components of the SLAM system can be depth perception and pose tracking. However, the existing unsupervised learning visual odometry framework cannot fully meet the actual requirements of the visual odometry information, particularly on the scale uncertainty in visual odometry estimation. It is still lacking in the geometric constraints during the autonomous operation of greenhouse mobile robots. In this study, an unsupervised optical flow-based visual odometry was presented. An optical flow estimation network was trained in an unsupervised manner using image warping. The optical flow between stereo images (disparity) was used to calculate the absolute depth of scenes. The optical flow between adjacent frames of left images was combined with the scene depth, in order to solve the frame-to-frame pose transformation matrix using Perspective-n-Point (PnP) algorithm. The reliable correspondences were selected in the solving process using forward and backward flow consistency checking to recover the absolute pose. A compact deep neural network was built with the convolutional modules to serve as the backbone of the flow model. This improved network was designed, according to the well-established principles: pyramidal processing, warping, and the use of a cost volume. At the same time, the cost volume normalization in the network was estimated with high values to alleviate the feature activations at higher levels than before. Furthermore, the local geometric consistency constraints were designed for the objective function of flow models. Meanwhile, a pyramid distilling loss was introduced to provide the supervision for the intermediate optical flows via distilling the finest final flow field as pseudo labels. A series of experiments were conducted using a wheeled mobile robot in a tomato greenhouse. The results showed that the better performance was achieved in the improved model. The local geometric consistency constraints improved the optical flow estimation accuracy. The endpoint error (EPE) of inter-frame and stereo optical flow was reduced by 8.89% and 8.96%, respectively. The pyramid distillation loss significantly reduced the optical flow estimation error of the flow model, in which the EPEs of the inter-frame and stereo optical flow decreased by 11.76% and 11.45%, respectively. The EPEs of the inter-frame and stereo optical flow were reduced by 12.50% and 7.25%, respectively, after cost volume normalization. Particularly, the price decreased by 1.28% for the calculation speed of the optical flow network. This improved model showed a 9.52% and 9.80% decrease in the root mean square error (RMSE) and mean absolute error (MAE) of relative translation error (RTE), respectively, compared with an existing unsupervised flow model. The decrease was 43.0% and 43.21%, respectively, compared with the Monodepth2. The pose tracking accuracy of this improved model was lower than that of ORB-SLAM3. The pure multi-view geometry shared the predicting dense depth maps of a scene. The relative error of depth estimation was 5.28% higher accuracy than the existing state-of-the-art self-supervised joint depth-pose learning. The accuracy of pose tracking depended mainly on the motion speed of robots. The performance of pose tracking at 0.2 m/s low speed and 0.8 m/s fast speed was significantly lower than that at 0.4-0.6 m/s. The resolution of the input image greatly impacted the pose tracking accuracy, with the errors decreasing gradually as the resolution increased. The MAE of RTE was not higher than 3.6 cm with the input image resolution of 832×512 pixels and the motion scope of 1 m, whereas, the MAE of relative rotation error (RRE) was not higher than 1.3º. These findings can provide technical support to design the vision system of greenhouse mobile robots.

Continuously and stably tracking fruit objects is required for tomato production in modern agriculture in recent years. Some difficulties also remained on the statistical counting accuracy under greenhouse production, due to the mode constraints of crop planting. In this study, a counting method was proposed for tomatoes of varying maturity, utilizing ultra-depth masking and an improved YOLOv8 model. Multi-head self-attention (MHSA) was introduced to construct a spatially heterogeneous convolution kernel. The global features were also integrated using the lightweight convolution operator Involution. A new convolution operator was optimized and designed, termed Global Attention-based Involution (GAInvolution). This operator formed the backbone network of the tomato detector, YOLOT, an improved YOLOv8 model. The model also incorporated the WIoU (wise intersection over union) loss to improve the robustness of the object labeling process. In addition, the depth information was predicted by the mono-depth estimation model, called Depth Anything. The distant fruit objects were dynamically filtered to avoid object tracking loss or duplicate tracking. This processing was referred to as ultra-depth masking. A tomato counter was also optimized using the BoT-SORT algorithm. Combining the tomato detector, depth estimation model, ultra-depth masking, and object counter, the comprehensive framework was constructed to identify and count tomatoes of different maturity levels. The experimental results showed that the mean average precision at IoU thresholds of 0.5 (mAP₅₀) of the improved tomato detector increased by 3.2 percentage points, and the recall rate increased by 3.7 percentage points, indicating the effectiveness of the improved YOLOv8 model, compared with the original. The GAInvolution convolution operator significantly improved the tomato object detector. The detector with GAInvolution as the main operator achieved the 2.7 and 2.8 percentage point increase in the mAP₅₀ and the recall rate, respectively, thereby significantly improving tomato detection performance. WIoU loss function was introduced to further improve the detection accuracy, where the mAP₅₀ increased by 1.1 percentage points, compared with the original. Ultra-depth masking greatly contributed to the accuracy of tomato fruit counting. The depth threshold was dynamically calculated and predicted using the depth map and the Depth Anything model. The highest counting accuracy was achieved by setting the average depth value of the depth map minus 0.5 times the depth standard deviation as the threshold. The average counting precision (A_CP) of tomato fruits increased by 12.63 percentage points using ultra-depth masking. The better performance of counting tasks was achieved in the tomato fruits, with an A_CP of 93.80% and a calculation speed of 23 frames per second. The accuracy of fruit counting was closely related to the performance of the object detector. Moreover, counting accuracy varied with different inspection viewports, with the vertical viewport achieving 2.59 percentage points higher accuracy than the parallel one. Tomatoes of different maturity levels were counted with A_CP of at least 91%. This finding offers valuable technical insights for predicting fruit and vegetable yields using visual technology.

As the foundation for implementing closed-loop detection within the realm of visual SLAM (simultaneous localization and mapping), visual place recognition (VPR) has great potential in various applications of greenhouse robot navigation and other fields. However, the existing VPR cannot fully meet the actual requirements of greenhouse scenes due to the complexity and constant variations in the greenhouse environment. In particular, the local feature aggregation paradigm strongly depends on the induction bias of training samples in VPR models, which leads to the issue of information redundancy during feature aggregation. In this study, a greenhouse VPR was presented, according to the optimal transport of local feature aggregation. The process of aggregating local features into a global descriptor was framed as an optimal transport problem, where the cost matrix was predicted through an MLP (multi-layer perceptron). Thus, a cost matrix was dynamically generated using the local features that was extracted from the greenhouse scene images. Additionally, a 'dustbin' cluster was introduced into the cost matrix to allocate the redundant features. Taking the cost matrix as the input, the Sinkhorn algorithm was employed to determine an optimal solution to the assignment matrix. Furthermore, the soft assignment of local features to various clusters was achieved through the assignment matrix. Ultimately, the assignment was concatenated to form a global descriptor for the scene image, which was used for place recognition. A deep neural network (DNN) was optimized and designed to serve as the backbone for local feature extraction of greenhouse scene images, by combining the advantages of CNN (convolutional neural network) and Transformer. Furthermore, cosine similarity was used as the metric function to calculate the similarity measure between scene image global descriptors, so as to perform descriptor matching. A series of experiments were conducted in a tomato greenhouse. The experimental results showed that the improved model achieved better performance. The top-1 recall rate (R_@1) for place recognition was achieved at 88.96%, which was 29.67, 2.97, and 2.89 percentage points higher than the those of NetVLAD, MixVPR, and EigenPlaces models, respectively. When compared to the aggregators employed in MixVPR and NetVLAD, our aggregator achieved improvements in R_@1 by 1.09 and 21.65 percentage points, respectively, showcasing its effectiveness. Compared with the CNN, the improved network achieved an increase of 5.45 percentage points in R_@1. There was even more pronounced R_@1 improvement (reaching 10.48 percentage points), compared with a Transformer network. Simultaneously, our network resulted in a 1.6-fold increase in computation speed compared to the previous Transformer. In addition, the experiments further demonstrated that the improved model exhibited excellent performance of place recognition and strong robustness when dealing with factors, such as small sampling distance shifts, small viewpoint shifts, and different sunlight intensities. The greenhouse VPR achieved a place recognition rate of no less than 81.94% in actual greenhouses, indicating its practical application potential. The method based on optimal transport of local feature aggregation and global descriptor generation was effective for place recognition, and the image local feature extraction network can boost the performance of place recognition. These findings can provide technical support to the visual systems of intelligent agricultural machinery in the greenhouse.