How Optical Sensors and Deep Learning Enhance the Production Management in Smart Agriculture

The advent of smart agriculture marks a paradigm shift from experience-driven to data-driven decision-making, fundamentally reshaping centuries-old farming practices. At the heart of this transformation lies the synergistic integration of optical sensors and deep learning (DL) technologies [1]. Modern optical sensors range from cost-effective RGB cameras and near-infrared-modified consumer devices to sophisticated multispectral, hyperspectral, and fluorescence imaging systems [2]. These sensors are now routinely deployed on unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs), fixed-wing aircraft, tractors, and stationary phenotyping platforms. These sensors deliver unprecedented spatiotemporal resolution and spectral richness, capturing subtle physiological signals (chlorophyll fluorescence, water stress indices, and anthocyanin accumulation) that remain invisible to the human eye [3]. Rich optical datasets collected by these sensors can be paired with DL architectures, particularly convolutional neural networks (CNNs), vision transformers (ViTs), and hybrid models. When combined with these powerful DL approaches, the datasets are transformed into actionable intelligence, enabling precise yield forecasting at the individual plant or plot level, automated phenotyping of thousands of genotypes per day, and real-time variable-rate application of water, fertilizers, and pesticides. By the end of 2024, an estimated 400,000 DJI Agriculture drones were in operation worldwide, representing a 90% increase since 2020 [4]. Meanwhile, the market for AI-enabled precision agriculture solutions is expanding at an extraordinary pace [5], driven by technological maturity as well as urgent global needs stemming from climate variability, labor shortages, and the necessity to feed a projected population of 10.3 billion by the mid-2080s [6].

Nevertheless, significant scientific and engineering hurdles continue to impede seamless transition from laboratory prototypes to robust, farmer-ready systems. Spectral saturation in dense canopies, bidirectional reflectance distribution function (BRDF) effects, mixed pixels in low-altitude imagery, and rapid illumination changes (such as cloud passage and diurnal variation in solar angle) routinely degrade data quality and model reliability. Additionally, occlusion by leaves, overlapping fruits, and dense weed canopies further complicates the detection of small targets. Perhaps most critically, the computational footprint of state-of-the-art DL models often exceeds the thermal and power constraints of edge devices carried by lightweight drones or solar-powered field robots [7]. These challenges are further compounded by the notorious difficulty of collecting large, well-annotated, and geographically diverse agricultural datasets. Unlike urban or indoor scenes, crop appearance varies dramatically across cultivars, growth stages, management practices, and stress conditions. In response, the research community has pursued multiple convergent strategies: (i) lightweight yet expressive architectures (Ghost, ShuffleNet, MobileViT, EfficientFormer); (ii) knowledge distillation and neural architecture search tailored to agricultural tasks; (iii) advanced feature selection and band-reduction techniques for hyperspectral data; (iv) domain adaptation and self-supervised pretraining on massive unlabeled satellite/drone imagery; and (v) hybrid physics-informed neural networks that embed radiative transfer principles to improve generalization under varying atmospheric and canopy conditions [8]. These innovations have collectively pushed mean average precision (mAP) beyond 90% in many real-world tasks while reducing inference latency to tens of milliseconds on embedded GPUs, bringing fully autonomous optical-DL pipelines within practical reach.

This Special Issue, entitled “How Optical Sensors and Deep Learning Enhance the Production Management in Smart Agriculture,” assembles thirteen high-quality contributions that exemplify the current state-of-the-art and illuminate viable paths forward (

Table 1. Summary of publications featured in this Special Issue.

Articles	Models	Optical Sensors	Platforms	Authors	References
1	YOLO	Zenmuse P1	UAV	Sun et al.	[10]
2	YOLO	Canon EOS 5D Mark IV	Field	He et al.	[11]
3	YOLO	Digital camera	Field	Sun et al.	[12]
4	DNN	HeadWall MV.C VNIR	UAV	Chen et al.	[13]
5	DCFM	ImSpector V10E	Field	Qi et al.	[14]
6	RF	GaiaSky-mini	UAV	Wang et al.	[15]
7	VisLAI	Mavic 3M	UAV	Fu et al.	[16]
8	LAINet	Sony IMX707	Field	Yang et al.	[17]
9	RF	Sentinel-2 MSI	Satellite	Lu et al.	[18]
10	RF, XGBoost	Sentinel-2 MSI	Satellite	Simeón et al.	[19]
11	CFD	-	UAV	Xu et al.	[20]
12	QY-SE-MResNet34	iPhone 14 Pro	Field	Li et al.	[21]
13	DRL	-	Review	Zhao et al.	[22]

). The collected works span staple crops (rice, maize, wheat), high-value horticultural species, and complex real-world field scenarios. They comprehensively address the entire pipeline, from sensor selection and flight planning to model optimization and on-farm validation. Methodologically, the collected works showcase the full spectrum of contemporary deep learning and machine learning approaches. These include heavily optimized YOLO variants for real-time detection and ripeness grading [9], attention-augmented lightweight CNNs for disease severity assessment, random forest and extreme gradient boosting ensembles fused with vegetation indices for yield prediction, dual-branch spectral–spatial networks for hyperspectral anomaly detection, and transfer-learning frameworks that dramatically reduce the annotation burden. Crucially, nearly all studies include rigorous field experiments conducted under commercial or near-commercial conditions across multiple continents, with quantitative metrics (R² > 0.85 for regression tasks, mAP@0.5 > 0.90 for detection, overall accuracy > 95% for classification) that meet or exceed the thresholds required for operational deployment. By bridging the gap between algorithmic sophistication and agricultural practicality, this Special Issue not only documents tangible progress but also establishes a benchmark and inspiration for the next wave of optical-DL innovations in global food production.

The YOLO series, renowned for its real-time detection capabilities, has been extensively adapted for agricultural tasks involving optical imagery. Researchers have introduced targeted enhancements to address issues like small-target detection [23], background complexity, and model efficiency. Sun et al. [4] proposed YOLOv8n-SSDW, a lightweight model incorporating separable residual CoordConv, spatio-channel enhanced attention, and dynamic upsampling for barnyard grass detection in rice fields. This achieved an mAP@0.5 of 0.851, with a 10.6% reduction in parameters, proving effective in occluded paddy environments during drone-based field tests. Similarly, He et al. [5] developed a sparse parallel attention mechanism integrated with CNN for passion fruit disease detection using optical sensing, attaining a precision of 0.93, a recall of 0.88, an accuracy of 0.91, an mAP@0.5 of 0.90, and outperforming baselines like Faster R-CNN and YOLO in handling multi-scale lesions under complex backgrounds. Sun et al. [6] advanced a GCSS-YOLO variant for classifying four tomato ripening stages in greenhouses, embedding RepNCSPELan for feature extraction and Shape_IoU loss, resulting in an mAP@0.5 of 0.853 and successful mobile deployment via NCNN for real-time monitoring.

Hyperspectral and multispectral optical data, combined with DL, have shown promise in quantitative trait estimation. Chen et al. [7] employed E2D-COS feature selection with DNN and transfer learning to estimate maize leaf chlorophyll content from UAV hyperspectral imagery, identifying key bands (e.g., 418 nm, 688 nm) and improving R² by up to 0.11 across growth stages. Qi et al. [8] introduced a dual-channel feature fusion model (DCFM) for hierarchical rice blast detection, fusing spectral features via successive projections and spatial features from MobileNetV2-CBAM, yielding an overall accuracy of 96.72% and a Kappa of 95.97%. Wang et al. [9] fused days after transplanting and meteorological factors with UAV hyperspectral data for rice LAI inversion, using RF on preprocessed spectra to achieve R² = 0.8015 and RMSE = 0.5745, enhancing stability under small-sample conditions.

In crop monitoring and yield prediction, DL models optimized for optical inputs have addressed phenological variations and environmental stressors. Fu et al. [10] developed VisLAI, a computer vision-based model using UAV visible light imagery and HSV optimization for maize LAI estimation, reaching R² values up to 0.92 across stages and surpassing machine learning alternatives like Gradient Boosting (R² = 0.88). Yang et al. [11] proposed LAINet, an improved CNN with triplet attention and growth-stage integration for corn LAI, attaining R² = 0.81 on a regional dataset. Lu et al. [12] integrated Sentinel-2 multispectral data with soil salt content using RF for winter wheat yield under saline stress, improving R² to 0.78 and reducing RMSE by incorporating salinity indices. Simeón et al. [13] applied machine learning classifiers (e.g., RF, XGBoost) to Sentinel-2 temporal spectra for rice variety mapping, achieving 0.94 accuracy during reproduction using red-edge and NIR bands.

Further innovations extend to specialized applications. Xu et al. [14] simulated rotor airflow and droplet deposition in plant protection UAVs, optimizing nozzle distances for maximal deposition (0.766 μL/cm² at 360 mm) via optical sensing validation. Li et al. [15] enhanced ResNet34 with SE blocks and kernel adjustments for wheat powdery mildew severity classification, reaching 89% accuracy on leaf images. Zhao et al. [16] reviewed deep reinforcement learning (DRL) trends in agricultural machinery, highlighting hybrid architectures for navigation and UAV operations.

Despite the remarkable progress achieved, several persistent challenges continue to hinder the large-scale deployment of optical sensors and deep learning in real-world agricultural systems. First, data scarcity and poor generalization remain critical bottlenecks: field-collected optical datasets are often limited in size, imbalanced across growth stages or stress conditions, and heavily influenced by site-specific factors (soil type, cultivar, management practices), making models prone to overfitting or failure when transferred to new regions. Second, spectral overlap and environmental interference significantly degrade the signal-to-noise ratio in multispectral and hyperspectral imagery. Key sources of this interference include cloud cover, varying solar angles, leaf inclination angles, and soil background. Third, computational constraints on edge devices mounted on drones, robots, or tractors demand models that are not only accurate but also extremely lightweight, low-power, and capable of real-time inference under strict latency requirements.

To tackle these obstacles, the contributions in this Special Issue comprise a rich portfolio of innovative solutions. Transfer learning and domain adaptation techniques, often initialized from large-scale natural image-pretrained backbones or simulated spectra, have dramatically reduced the need for large, labeled field datasets while improving cross-region robustness. Attention mechanisms (channel, spatial, triplet, and dynamic sparse attention) have been strategically inserted into both CNN and transformer architectures to suppress background noise, enhance discriminative spectral-spatial features, and focus computational resources on disease lesions, small fruits, or subtle phenological changes. Model lightweighting strategies, including structured/unstructured pruning, knowledge distillation, and dynamic inference, have successfully reduced parameter counts, enabling smooth deployment on resource-constrained platforms such as NVIDIA Jetson series and mobile CPUs [24]. Furthermore, multimodal data fusion has emerged as a powerful paradigm, integrating optical/hyperspectral imagery with meteorological records, soil salinity indices, and growth-day information, which can significantly mitigate the impact of occlusion, illumination variation, and spectral saturation [25].

The thirteen articles assembled in this Special Issue collectively provide evidence of the maturity and practical viability of optical sensor–deep learning pipelines in smart agriculture. Through rigorous field validation across diverse crops, geographic regions, and sensing platforms (UAVs, ground robots, fixed-wing aircraft, handheld spectrometers), these studies demonstrate how state-of-the-art systems can operate reliably in unstructured, dynamic environments with accuracy, speed, and robustness values that meet or exceed the requirements of commercial farming operations [26]. More importantly, they establish a clear technical roadmap for the next generation of agricultural intelligence: edge–cloud collaborative inference for ultra-low latency decision-making; meta-learning and few-shot adaptation for rapid deployment to new cultivars or climates; continual learning frameworks that evolve with seasonal data streams; and deeper interdisciplinary fusion of optical sensing, robotics, agronomy, and economics [27]. These advancements are expected to drive not only incremental improvements but also achieve transformative leaps in resource-use efficiency, early-warning capability, and climate resilience, ultimately contributing to global food security and agricultural sustainability in an era of increasing environmental uncertainty.