Machine learning and remote sensing for precision agriculture: Progresses and challenges

Crop monitoring and growth prediction

Using remote sensing imagery for crop monitoring represents a foundational application within precision agriculture. Optical remote sensing, radar remote sensing, and thermal infrared remote sensing can respectively capture spectral, structural, and temperature information related to plant attributes (Ennouri and Kallel, 2019). By analyzing these data, scientists can monitor crop growth dynamics with high accuracy (Chang et al., 2020; Mohammed et al., 2019; Zarco-Tejada et al., 2016). While traditional indices like NDVI are widely used, they tend to saturate at high biomass levels. Recent advancements have shifted towards Solar-Induced Chlorophyll Fluorescence (SIF) and Red-edge indices. Unlike structural indices, SIF directly reveals the photosynthetic activity of crops by capturing the emission of excess energy during photosynthesis, offering earlier detection of physiological stress (Wang et al., 2023; Zhang et al., 2022a). Similarly, optimized spectral indices leveraging the Red-edge band have been proven to estimate canopy nitrogen content more accurately across different growth stages, overcoming the saturation limitations of NDVI (Yang et al., 2023). Integrating multi-source remote sensing data into supervised learning models allows for precise monitoring of crop health. Researchers have trained models based on decision trees (Panella et al., 2024), random forests (Brewer et al., 2022), or support vector machines (Kok et al., 2021) to efficiently identify crop growth anomalies and issue early warnings. By integrating large-scale remote sensing imagery with historical meteorological data, researchers can use deep learning models to predict crop yields (Khan et al., 2024) and iteratively refine these models as new data becomes available. These applications make essential contributions to the precise management of crops and the rational planning of crop planting.

Soil health assessment and management

Soil health plays a vital role in supporting crop growth, and the integration of machine learning and remote sensing technologies has demonstrated significant potential in soil assessment (Zarei et al., 2021). By leveraging thermal infrared and radar remote sensing technologies, researchers can remotely evaluate key soil attributes based on distinct physical mechanisms. Thermal infrared remote sensing assesses soil moisture through thermal inertia principles—wetter soil has a higher heat capacity and thus heats up or cools down more slowly than dry soil (Anderson et al., 2024). In parallel, Synthetic Aperture Radar (SAR) utilizes the significant difference in dielectric constants between water (~80) and dry soil particles (~3–5) to quantify moisture content through backscatter intensity, enabling monitoring even under cloud cover (Quast et al., 2023). By conducting in-depth comprehensive analysis of these multidimensional datasets using machine learning algorithms, precise estimations of soil fertility dynamics and nutrient distribution can be achieved, providing a scientific basis for fertilizer and irrigation decisions in agricultural production (Abioye et al., 2022). For example, clustering algorithms are extensively employed in soil type classification studies, enabling the segmentation of farmland into distinct soil zones for precision management (Zhao et al., 2023). Moreover, neural networks and deep learning models have demonstrated their advantages in predicting dynamic soil changes, while accurately modeling temporal variations in soil characteristics (Zhang et al., 2022b).

Pest and disease monitoring and control

Pests and diseases represent critical challenges in agriculture, and remote sensing technology has the capability to detect early signs of diseases by monitoring spectral variations in crops. When combined with machine learning algorithms, remote sensing data can further enable the recognition of pest and disease types as well as their spatial extent. At the macro-level, multispectral or hyperspectral remote sensing imagery, integrated with supervised learning algorithms such as decision tree classifiers (Chopda et al., 2018) or random forest models (Iordache et al., 2020), allows for the early detection of disease-affected areas and the classification of different pest and disease types in the early stages of crop growth. At the micro-level, certain studies have investigated the combination of ground-based remote sensing with deep learning models to dynamically identify pest species and their developmental stages under laboratory or simulated field conditions, providing technical support for early warning of pest and disease outbreaks (Lu and Ye, 2020; Ye et al., 2020). In the domain of pest and disease control, CNNs have been applied to identify pests and diseases from high-resolution remote sensing images (Tetila et al., 2020). These models can automatically learn pest characteristics from large volumes of annotated image data, thereby significantly enhancing detection accuracy, enabling farmers to take timely preventive measures, and ultimately reducing pesticide use to achieve more environmentally sustainable pest management.

Disaster monitoring and extreme weather prediction

Remote sensing technology has also demonstrated extensive applications in disaster monitoring and extreme weather prediction. For example, remote sensing imagery can effectively monitor agricultural disasters such as droughts and floods (Pan et al., 2017). By integrating remote sensing imagery into machine learning models, researchers can precisely forecast the timing and locations of potential extreme weather events (Salcedo-Sanz et al., 2024), thereby enhancing the accuracy of early warnings. Furthermore, by integrating remote sensing data with meteorological data and employing time-series analysis and regression models, it is possible to elucidate the impact mechanisms of complex extreme weather events—such as high temperatures, droughts, and low temperatures—on crop growth and predict crop yield variations under future climate scenarios (Benipal et al., 2021; Feng et al., 2020; Vogel et al., 2019). This information can offer early warnings to farmers, enabling them to effectively mitigate disaster impacts and reduce losses, thereby enhancing agriculture's resilience to climate change.

Agricultural resource optimization management

Agricultural resource optimization, facilitated by the integration of machine learning and remote sensing, has evolved from simple input management to a complex coordination of 'Machine-Material-Human' elements. While early studies focused on generating precise recommendations for irrigation and fertilization through classification and regression models (Filgueiras et al., 2020), current research emphasizes a more holistic management of the agricultural lifecycle. For example, machine learning models that integrate remote sensing-derived soil moisture data allow for the dynamic optimization of irrigation schedules, significantly reducing water wastage compared to traditional methods (Ahmad et al., 2010). Furthermore, the application of deep learning has expanded beyond monitoring to automate field division schemes (Xu et al., 2024) and optimize the spatial allocation of labor and machinery. This comprehensive approach not only enhances resource utilization efficiency but also mitigates environmental impacts such as nutrient leaching and soil degradation, fostering the transition toward sustainable agricultural intensification (Getahun et al., 2024; Ye et al., 2024b; Jiang et al., 2026).

Comparative analysis of sensing modalities for decision support

To provide actionable guidance for selecting appropriate technologies, we synthesized the trade-offs between different sensing modalities, platforms, and operational scales (Table 1). A critical gap often overlooked is the connection between "failure modes" and "decision-making constraints". For instance, while optical satellites (e.g., Sentinel-2) offer macro-scale coverage for yield forecasting, their high susceptibility to cloud cover failure renders them unreliable for time-critical decisions like irrigation scheduling during the monsoon season. In contrast, Synthetic Aperture Radar (SAR) provides all-weather capability essential for disaster assessment but suffers from speckle noise and geometric distortions that complicate the interpretation of soil texture.

OPEN IN VIEWER

Table 1.

Comparative analysis of sensing modalities, platforms, and their constraints in agricultural decision-making.

Sensing Modality	Key Platforms	Operational Scale	Primary Application	Key Failure Mode	Decision-Making Constraint
Optical (VIS/NIR/SWIR)	Satellite (e.g., Sentinel-2); UAV	Regional to Field	Crop Classification, Phenology Monitoring	Cloud Cover & Atmospheric Noise: Data gaps during critical growth stages.	Strategic: Good for yield forecasting and insurance; unreliable for real-time daily operations.
Thermal Infrared (TIR)	Satellite (Landsat); UAV	Field to Plot	Water Stress Detection, Irrigation Scheduling	Environmental Sensitivity: Affected by wind speed, air temp, and low spatial resolution.	Tactical: Critical for irrigation timing, but low resolution limits use in precision drip systems.
Synthetic Aperture Radar (SAR)	Satellite (e.g., Sentinel-1)	Regional (All-weather)	Soil Moisture, Structural Mapping, Disaster Assessment	Speckle Noise & Geometry: Signal confusion in complex topography; difficult to interpret.	Emergency: Essential for flood monitoring when optical fails; requires complex preprocessing.
Hyperspectral / Lidar	UAV; Proximal (Ground)	Plant to Leaf	Disease Diagnosis, Phenotyping, Biomass Estimation	High Data Redundancy & Cost: "The Curse of Dimensionality" and high hardware costs.	Limited to small-scale, high-value crop plotting due to extensive data processing and operational limits (e.g., battery life).

Furthermore, the choice of platform dictates the scale of decision-making. Satellite-based systems support strategic planning (e.g., regional crop planting structure adjustment) due to their extensive coverage but often lack the spatial resolution required for tactical management (e.g., variable-rate spraying for individual plants). Conversely, UAV and proximal sensing offer sub-centimeter resolution ideal for precision breeding and disease diagnosis but are constrained by battery life and data processing costs, limiting their application to small-scale, high-value plotting. Table 1 systematically outlines these trade-offs to guide researcher and practitioner selection.

Mapping algorithmic architectures to agricultural scenarios

While previous sections outlined machine learning applications from an agronomic perspective, matching these scenarios with appropriate algorithmic architectures is equally critical for effective remote sensing analysis. Different network architectures possess distinct inductive biases suited for specific spatial, temporal, and spectral complexities. For instance, Convolutional Neural Networks (CNNs) are the dominant architecture for spatial pattern recognition tasks, such as fine-scale field boundary delineation and localized disease detection using high-resolution imagery (Too et al., 2019). Conversely, for temporal sequence modeling—such as yield prediction and phenology tracking—Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks are preferred due to their ability to capture temporal dependencies from time-series indices and meteorological data (Schwalbert et al., 2020).

Recent advancements have introduced more sophisticated architectures to overcome the limitations of early deep learning models. Transformers, leveraging spatiotemporal self-attention mechanisms, have demonstrated exceptional capability for handling long-range dependencies in complex tasks such as crop classification from time-series data, bypassing the sequential processing bottlenecks of LSTMs (Wang et al., 2024b). Furthermore, emerging State Space Models, such as Mamba, offer a paradigm shift by maintaining global receptive fields with linear computational complexity. In agricultural remote sensing, Mamba architectures are showing immense potential for processing continuous, ultra-long time-series satellite data and complex hyperspectral datacubes without prohibitive memory overhead (He et al., 2024). Finally, the advent of Vision Foundation Models (VFMs) is transitioning agricultural remote sensing toward a “pre-train, fine-tune” paradigm. Models leveraging massive unannotated Earth observation data exhibit remarkable zero-shot and few-shot generalization capabilities, mitigating the pervasive data scarcity and annotation bottlenecks in agriculture (Osco et al., 2023). To systematically guide algorithm selection, Table 2 maps these core architectures to their optimal agricultural remote sensing scenarios.

OPEN IN VIEWER

Table 2.

Mapping of deep learning architectures to agricultural remote sensing scenarios.

Algorithmic Architecture	Key Capability / Inductive Bias	Optimal Agricultural RS Scenarios	Preferred RS Data Modality	Advantages & Limitations
CNNs (e.g., ResNet, U-Net)	Local spatial feature extraction; spatial hierarchies.	Field boundary delineation; Disease lesion detection; Weed mapping.	High-res RGB (UAV), Multispectral imagery.	Adv: Excellent at spatial textures. Lim: Weak at capturing long-term temporal dynamics.
RNNs / LSTMs	Temporal sequence modeling; memory of past states.	Yield prediction; Phenology tracking; Drought monitoring.	Time-series indices (NDVI, EVI), Meteorological data.	Adv: Effective for sequential crop growth stages. Lim: Prone to vanishing gradients on very long sequences.
Transformers (e.g., ViT)	Global self-attention; long-range spatial-temporal dependencies.	Time-series crop mapping; Multi-modal data fusion.	Dense time-series optical stacks, Hyperspectral data.	Adv: Captures global context perfectly. Lim: Computationally expensive; requires massive training data.
State Space Models (e.g., Mamba)	Continuous sequence modeling with linear complexity.	Hyperspectral image classification; High-frequency monitoring.	Hyperspectral datacubes, dense time-series imagery.	Adv: High computational efficiency for long sequences. Lim: Emerging technology; fewer established agricultural baselines.
Large Foundation Models (e.g., SAM)	Universal representation learning; Zero/Few-shot generalization.	Cross-regional crop mapping; Automated data annotation.	Massive unannotated multi-source imagery.	Adv: Solves data scarcity (annotation) problems. Lim: High inference cost; "black-box" nature hinders interpretability.

Cognizing the complexity of arable land systems

The fundamental challenge in precision agriculture lies in cognizing the arable land system not merely as a static surface for production, but as a complex adaptive system driven by the dynamic coupling of natural resources and human activities. As illustrated in the cascade process of Figure 2, human behaviors—ranging from excessive fertilizer input and heavy machinery operation to mono-cropping patterns—interact continuously with the arable land entity (soil, water, and biota) to drive nonlinear changes in ecological functions. While current remote sensing technologies have achieved high accuracy in depicting "explicit" characteristics such as land use types and landscape fragmentation, they face substantial methodological hurdles in characterizing the “recessive” attributes generated through this cascade. Specifically, quantifying invisible parameters like nitrogen input intensity, soil organic carbon dynamics, and biodiversity loss remains difficult because these attributes do not have unique spectral signatures but are indirect consequences of complex "Human-Land" interactions (Du et al., 2024; Ren et al., 2025). Unlike visible surface features, these recessive attributes often require inverting complex radiative transfer models that are easily confounded by environmental noise. Consequently, accurately capturing the boundary of abrupt changes in ecological functions necessitates constructing high-dimensional, long-term datasets that can decouple these complex interactions, a task that traditional optical sensing alone struggles to achieve.

OPEN IN VIEWER

Figure 2.

Cascade process between human behavior - arable land entity - ecological function change in arable land ecosystem.

Spatiotemporal mismatch and data complexity

Remote sensing systems are inherently complex, focusing on data acquisition and analysis that involve high-dimensional, multi-temporal, and multi-spectral features capturing variables across diverse spatiotemporal scales (Im and Jensen, 2008; Lillesand et al., 2015). However, a critical constraint specific to agriculture is the spatiotemporal mismatch between satellite revisit cycles and rapid crop phenological transitions. Agricultural monitoring requires high-frequency observations to capture subtle physiological changes during critical growth windows, such as flowering or grain filling. Unfortunately, remote sensing imagery is frequently compromised by interference factors like atmospheric noise and persistent cloud cover during these peak growing seasons (Ju and Roy, 2008; Xu et al., 2019), resulting in the loss of critical spectral data that cannot simply be resolved by interpolation techniques. Although machine learning models, particularly deep learning, have shown promise in processing these heterogeneous datasets by learning non-linear representations (Yuan et al., 2020), fully utilizing such complex data requires powerful computing resources and optimized algorithms to reconstruct continuous growth trajectories from fragmented observations (Han et al., 2023). Future research must focus on fusing multi-source data to mitigate these interference factors, thereby bridging the gap between discrete static observations and the continuous dynamic nature of crop physiology.

Model generalization under heterogeneous agro-environments

The scalability of machine learning models in agriculture is currently hindered by the “Genotype × Environment × Management” (G×E×M) interaction, a challenge specific to biological systems that transcends standard computer vision problems. A model trained effectively in one region often fails to generalize to another, not merely due to statistical domain shifts, but because identical crop varieties exhibit varying spectral signatures under different soil textures, moisture levels, and microclimates. This issue is exacerbated in smallholder farming systems characterized by landscape fragmentation, where the prevalence of mixed pixels challenges the extraction of pure phenotypic traits. Improving the generalization ability of models to adapt to these diverse agricultural conditions is key to achieving precision management (Zheng et al., 2021). While emerging techniques such as transfer learning (Li et al., 2021) and semi-supervised learning (Amorim et al., 2019) offer new pathways to reduce dependence on labeled data by transferring knowledge across domains, their robustness in complex, fragmented agricultural landscapes remains to be fully validated. Developing domain-adaptive algorithms that can decouple crop-specific features from environmental background noise, or "invariant feature learning," remains a priority for future research to ensure models are robust across varying agro-climatic zones.

High-cost of agronomic ground-truthing and sampling

High-quality sample data is essential for machine learning models to make reliable predictions, yet data scarcity in agriculture differs fundamentally from other domains. Obtaining high-quality “ground truth” often necessitates destructive sampling and rigorous laboratory analysis (e.g., measuring leaf chlorophyll content or soil nutrient concentrations), which is far more labor-intensive, time-consuming, and costly than the simple visual bounding-box annotation used in general object detection (Russell et al., 2008). This reliance on manual intervention severely limits the efficiency and scale of agricultural data processing. To address this bottleneck, scholars are actively exploring methods to utilize unlabeled data (Zhang et al., 2023) or employ Generative Adversarial Networks (GANs) to automatically generate synthetic training samples (Lu et al., 2022). However, the quality of unlabeled data is difficult to guarantee, and discrepancies between synthetic data and the complex reality of field conditions often introduce noise that degrades model performance. Therefore, future research should prioritize integrating active learning strategies—which intelligently select the most informative samples for manual annotation—with small-sample learning techniques to maximize prediction accuracy while minimizing the cost and effort of agronomic ground-truthing.

Socio-technical barriers and institutional support

Despite the significant potential of integrating machine learning and remote sensing to transform agricultural production (Benos et al., 2021), the adoption gap is increasingly recognized as a structural socio-technical barrier rather than a purely cognitive one. The implementation of real-time monitoring and precise management is fundamentally constrained by the “digital divide,” characterized by insufficient rural network infrastructure and the high capital intensity of advanced sensing equipment (Mhlanga and Ndhlovu, 2023). Many farmers harbor doubts regarding the reliability and cost-effectiveness of these tools, stemming not solely from a lack of knowledge or "fear of the unknown," but from the absence of clear economic safety nets and return-on-investment models (Khaspuria et al., 2024). Consequently, widespread adoption necessitates a paradigm shift from placing the burden on individual farmers toward government-led institutional support. This includes expanding rural digital infrastructure, providing financial subsidies to mitigate adoption risks, and establishing socialized agricultural service platforms. Such institutional frameworks are essential to simplify technology operation and provide the long-term technical training mechanisms needed to bridge the gap between advanced algorithms and practical field management.