Deep learning in multi-sensor agriculture and crop management

Elsevier eBooks, Год журнала: 2025, Номер unknown, С. 335 - 379

Опубликована: Янв. 1, 2025

Язык: Английский

---

Dynamic mutual training semi-supervised semantic segmentation algorithm with adaptive capability (AD-DMT) for choy sum stem segmentation and 3D positioning of cutting points

Computers and Electronics in Agriculture, Год журнала: 2025, Номер 232, С. 110105 - 110105

Опубликована: Фев. 17, 2025

Язык: Английский

---

Integration of convolutional neural networks with parcel-based image analysis for crop type mapping from time-series images

Earth Science Informatics, Год журнала: 2025, Номер 18(3)

Опубликована: Фев. 27, 2025

Язык: Английский

---

Systematic review on machine learning and computer vision in precision agriculture: Applications, trends, and emerging techniques

Engineering Applications of Artificial Intelligence, Год журнала: 2025, Номер 148, С. 110401 - 110401

Опубликована: Март 13, 2025

Язык: Английский

---

Exploring Nutrient Deficiencies in Lettuce Crops: Utilizing Advanced Multidimensional Image Analysis for Precision Diagnosis

Sensors, Год журнала: 2025, Номер 25(7), С. 1957 - 1957

Опубликована: Март 21, 2025

In agricultural production, lettuce growth, yield, and quality are impacted by nutrient deficiencies caused both environmental human factors. Traditional detection methods face challenges such as long processing times, potential sample damage, low automation, limiting their effectiveness in diagnosing managing crop nutrition. To address these issues, this study developed a deficiency system using multi-dimensional image analysis Field-Programmable Gate Arrays (FPGA). The first applied dynamic window histogram median filtering algorithm to denoise captured images. An adaptive integrating global local contrast enhancement was then used improve detail contrast. Additionally, combining threshold segmentation, improved Canny edge detection, gradient-guided segmentation enabled precise of healthy nutrient-deficient tissues. quantitatively assessed analyzing the proportion tissue Experimental results showed that achieved an average precision 0.944, recall rate 0.943, F1 score 0.943 across different growth stages, demonstrating significant improvements accuracy, efficiency while minimizing interference. This provides reliable method for rapid diagnosis lettuce.

Язык: Английский

---

Mdlis: An Integrated Model- and Data-Driven Method for Low-Light Instance Segmentation

Опубликована: Янв. 1, 2025

Язык: Английский

---

Construction and Enhancement of a Rural Road Instance Segmentation Dataset Based on an Improved StyleGAN2-ADA

Sensors, Год журнала: 2025, Номер 25(8), С. 2477 - 2477

Опубликована: Апрель 15, 2025

With the advancement of agricultural automation, demand for road recognition and understanding in machinery autonomous driving systems has significantly increased. To address scarcity instance segmentation data rural roads unstructured scenes, particularly lack support high-resolution fine-grained classification, a 20-class dataset was constructed, comprising 10,062 independently annotated instances. An improved StyleGAN2-ADA augmentation method proposed to generate higher-quality image data. This incorporates decoupled mapping network (DMN) reduce coupling degree latent codes W-space integrates advantages convolutional networks transformers by designing transfer block (CCTB). The core cross-shaped window self-attention mechanism CCTB enhances network’s ability capture complex contextual information spatial layouts. Ablation experiments comparing original demonstrate significant improvements, with inception score (IS) increasing from 42.38 77.31 Fréchet distance (FID) decreasing 25.09 12.42, indicating notable enhancement generation quality authenticity. In order verify effect on model performance, algorithms Mask R-CNN, SOLOv2, YOLOv8n, OneFormer were tested compare performance difference between enhanced dataset, which further confirms effectiveness module.

Язык: Английский

---

BHI-YOLO: A Lightweight Instance Segmentation Model for Strawberry Diseases

Applied Sciences, Год журнала: 2024, Номер 14(21), С. 9819 - 9819

Опубликована: Окт. 27, 2024

In complex environments, strawberry disease segmentation models face challenges, such as difficulties, excessive parameters, and high computational loads, making it difficult for these to run effectively on devices with limited resources. To address the need efficient running low-power while ensuring effective in scenarios, this paper proposes BHI-YOLO, a lightweight instance model based YOLOv8n-seg. First, Universal Inverted Bottleneck (UIB) module is integrated into backbone network merged C2f create C2f_UIB module; approach reduces parameter count expanding receptive field. Second, HS-FPN introduced further reduce enhance model’s ability fuse features across different levels. Finally, by integrating Residual Mobile Block (iRMB) EMA design iRMA, capable of efficiently combining global information local information. The experimental results demonstrate that enhanced diseases achieved mean average precision (mAP@50) 93%. Compared YOLOv8, which saw 2.3% increase mask mAP, improved reduced parameters 47%, GFLOPs 20%, size 44.1%, achieving relatively excellent effect. This study combines architecture feature fusion, more suitable deployment mobile devices, provides reference guide applications agricultural environments.

Язык: Английский

---

Advanced segmentation models for automated capsicum peduncle detection in night-time greenhouse environments

Systems Science & Control Engineering, Год журнала: 2024, Номер 12(1)

Опубликована: Дек. 16, 2024

This research addresses challenges in capsicum peduncle detection night-time greenhouse environments, including low light, uneven illumination, and shadows, using advanced computer vision models. A dataset of 200 images was curated, capturing diverse distances, heights, occlusion levels, lighting conditions, rigorously pre-processed augmented. Two YOLOv9 instance segmentation variants, YOLOv9c-seg YOLOv9e-seg, were custom-trained fine-tuned Google Colaboratory. (56.3 MB) achieved superior mean Average Precision (mAP) scores 0.751 (box) 0.725 (mask), outperforming YOLOv9e-seg (121.9 with mAP 0.674 0.658 (mask). Grounded SAM, a zero-shot model, maximum confidences 59% 49% positional prompts. Comparative testing on 50 containing 70 capsicums showed achieving precision, recall, F1-scores 0.93, 0.86, 0.89, respectively, SAM (0.86, 0.70, 0.77). study highlights the efficacy single-shot versus models for automated controlled agricultural offering insights into model performance future directions optimization expansion.

Язык: Английский

---

An In-Field Dynamic Vision-Based Analysis for Vineyard Yield Estimation

IEEE Access, Год журнала: 2024, Номер 12, С. 102146 - 102166

Опубликована: Янв. 1, 2024

Accurately predicting grape yield in vineyards is essential for strategic decision-making the wine industry. Current methods are labour-intensive, costly, and lack spatial coverage, reducing accuracy cost-efficiency. Efforts to automate enhance estimation focus on scaling fruit weight assessments. Machine learning, particularly deep shows promise improving through automatic feature extraction hierarchical representation. However, most of these have been developed analyses at a particular time point solutions able consider temporal information captured across sequential frames currently poorly developed. This paper addresses this gap by introducing system estimation, utilising publicly available data repositories, such as Embrapa WGISD, alongside an in-house dataset collected Blackmagic camera pre-harvest stage. We introduce that utilises bunch regression estimate yield. Bunch estimates obtained summing samples randomly drawn from distribution empirical calibration. Grapevine bunches identified segmented using Mask R-CNN with Swin Transformer, SiamFC-based tracking mechanism employed number unique per panel or row. The berries each tracked determined density approach known multitask supervision. In our experiments, we demonstrate effectiveness proposed achieving harvested errors less than 5% two three vineyard panels. Larger harvest (around 15%) were observed due inaccuracies certain caused dense concentration one panel. should be contrasted current practice error up 30%, highlighting potential machine vision hands-off scale.

Язык: Английский

---