Inversion Studies on the Heavy Metal Content of Farmland Soils Based on Spectroscopic Techniques: A Review

Abstract

In recent years, heavy metal pollution in farmland soil has become a crisis due to human activities or natural impacts, with particular emphasis on cases from China, where this issue is prominent, greatly affecting crop production and food safety. In the context of a low heavy metal (HM) content in farmland soil, which is difficult to monitor in real time, effective and rapid monitoring of soil plays a decisive role in subsequent targeted protection measures. To this end, this paper provides a narrative review of the application of spectral sensing technology on the basis of the quantitative inversion of heavy metal content in farmland soil using different platforms (ground, airborne, and spaceborne). The sensing process evaluates the mechanism by which soil produces different weak spectral features from the perspective of the heterogeneity of farmland soil. Different methods used for the quantitative inversion of heavy metals (by studying the correlation between soil heavy metals and organic matter, clay minerals, metal oxides, crop vegetation index, etc.) and their feasibility were clarified. At the same time, relevant research on key technologies used in various processes—such as follow-up pretreatment, spectral feature extraction, and the establishment of inversion models for spectral data of different farmland soil types—was summarized, with a primary focus on cases in China. Finally, the challenges, applications, and research directions related to heavy metal spectral inversion in farmland soil were discussed.

1. Introduction

The farmland soil system is a complex organic whole with a specific function, which is composed of a solid phase, liquid phase, and gas phase, and primarily includes organic matter and minerals [1]. As an important environment for crop growth, the soil system has a great impact on the growth and quality of crops and directly provides the necessary nutrients for crop growth. However, its function depends on the physicochemical properties of its structure, as well as the balance of its components [2]. In recent years, due to the impact of human activities [3] (the expansion of the industrial production scale, pollution of the urban environment, and the increase in agricultural chemicals), the problem of excessive heavy metal content in farmland soil has become increasingly serious. According to the National Soil Pollution Survey report, at least 10 million hectares of farmland soil in China have been affected by heavy metal pollution [4]. Heavy metals mainly refer to metal elements or metal-like elements that are highly toxic, difficult to degrade, persistent, and easy to accumulate in living organisms [5]. They pollute the soil system, affect the quality and yield of crops, and even endanger human health (Figure 1) [6].

It has been reported that heavy metals enriched in farmland soils include elements such as arsenic (As), cadmium (Cd), chromium (Cr), copper (Cu), lead (Pb), mercury (Hg), nickel (Ni), and zinc (Zn) [7,8]. The management of heavy metals mainly includes monitoring, blocking, and remediation, among which the real-time monitoring of soil heavy metal content is particularly important, which can be used to explore rapid and reasonable countermeasures for subsequent blocking and remediation processes. This process is of great significance for controlling subsequent pollutants, protecting the soil environment, and ensuring crop food safety.

Traditional methods for detecting heavy metals in farmland soils primarily rely on wet chemical analysis (HCA). In recent years, many emerging HM content monitoring methods, such as atomic absorption spectrometry (AAS), inductively coupled plasma emission spectrometry (ICP-OES), and inductively coupled plasma mass spectrometry (ICP-MS), have emerged to analyze extremely small variations of trace elements of heavy metals in terms of ppb (parts per billion) or ppt (parts per trillion), but they still require the use of different chemicals to chemical extraction for analysis. Despite the high accuracy of these methods, they are inefficient, costly, and time-consuming and do not allow for the rapid monitoring of soil constituents on a wide range of temporal and spatial scales [9]. In addition, HM detection methods represented by X-ray fluorescence spectrometry (XRF) that do not require preprocessing by chemical means have also emerged, but their monitoring efficiency is still low. With the emergence of spectral sensing technology, it has been widely used in the field of agriculture. Specifically, hyperspectral remote sensing makes use of narrow and continuous spectral channels as well as high-resolution spectral imaging of features, which can simultaneously acquire spectral information and image information of features and enables the rapid, wide-range, quantitative, and presentable detection of crops. At the same time, it has a great advantage in reflecting certain subtle changes in the physical properties of crops, vegetation, or soil, and it can capture the reflectance curves of different features. In addition, spectral sensing technology has good spatial continuity in detecting the distribution of heavy metals in soil [10], which can be important for developing a wide range of rapid, non-destructive testing methods for soil.

In recent years, researchers and scholars have widely studied the effectiveness of spectral information in predicting soil nutrient content and heavy metal content, of which ultraviolet (UV), visible (VIS), near-infrared (NIR), and short-wave infrared (SWIR) spectroscopy have been applied in the detection of heavy metal content in soils [11,12,13]. Some scholars have summarized some inversions of soil heavy metals using spectral data [14,15]; however, relevant studies and reviews on soils of agricultural nature are still very limited. There is an urgent need to research farmland soil. It is worth mentioning that at this stage, the quantitative studies of heavy metals in farmland soils are mostly focused on crop soils such as rice soil, vegetables soil, tea plantations soil, orchards soil, etc., and most of the studies on this type of soil are focused on the bands of the visible-near-infrared and short-wave infrared spectra (VNIR-SWIR: 350–2500 nm) [16]. Therefore, the use of VNIR-SWIR spectroscopy has become an effective means of predicting soil heavy metals (HMs), which is of great significance for heavy metal monitoring in farmland soils. Spectral techniques provide a new way of thinking about the content analysis of HMs by monitoring a comprehensive view of the features. To this end, we summarize the application, challenges, and development trends of spectroscopic techniques for monitoring heavy metals in farmland soil.

The research framework of this review is as follows:

First, in order to ensure the scientific rigor of the retrieved articles, we used the web of science and CNKI as databases to retrieve more than 140 pieces of relevant literature in the past 16 years (2010–2025), with the key words of (“farmland soil” or “agricultural soil”) AND (“spectroscopy” or “hyperspectral” or “spectral”) AND (“heavy metals”) AND (“inversion” or “prediction”). Then, the most relevant studies consistent with the scope of this review were selected for detailed inspection, excluding the literature using qualitative analysis, the literature on non-heavy metal elements (such as nutrients, organic carbon content, etc.), the literature related to the mapping of heavy metal concentrations in a wide range of soils, and the literature on non-agricultural soils (such as mining areas, urban soils, etc.). After that, we carefully read the full text of the retrieved articles and included them as references in this article to ensure that each article was related to this article. It is worth mentioning that most of the cases from China were used in this paper due to the bias of studies from China in this field, especially in Section 3.3, Modeling Methods.

The purpose of this paper is to investigate the following scientific issues:

(1)

What is the spatial distribution of heavy metals in farmland soil, and what is the difference from other soils?

(2)

How can we determine the physical and chemical values of heavy metals in farmland soil (i.e., how can we obtain the model output value of the established inversion model)?

(3)

What are the ways to obtain spectral data (i.e., how can we obtain the model input value of the established inversion model)?

(4)

How can we reduce the influence of noise and interference in the obtained spectral data on the establishment of the inversion model?

(5)

How can we mitigate the problem of redundant information and dimension disasters in spectral data?

(6)

What are the current modeling methods used for heavy metal content in farmland soil based on spectral technology? How can we judge the quality of the model?

(7)

What are the problems in the current field? What are the research trends?

For this reason, the narrative method of this paper emphasizes “theoretical depth over systematic exhaustiveness” to elaborate on the relevant technologies used in the spectral inversion process of heavy metals in farmland soil and gives priority to the relevant content between the retrieved literature and the key links in the inversion process (i.e., data acquisition, preprocessing, modeling, etc.) for an in-depth analysis, rather than seeking to list or summarize all the methods used in the retrieved literature. Based on the above seven core research issues, these issues are covered in subsequent chapters, forming a cohesive analysis chain: the Introduction introduces the background and significance of the current research in the field, the second chapter introduces the methods of obtaining soil physicochemical values and spectral data, as well as the advantages and disadvantages of different data acquisition platforms, describes the impact of farmland soil heterogeneity on the characteristic absorption characteristics of heavy metals and the subsequent processing of spectral data, and explains the basic mechanism of spectral inversion. Then, the third chapter narrates the causal relationship from preprocessing feature extraction model establishment, that is, it introduces in detail how to use the physical and chemical composition and spectral data to build the model, as well as the strategy to improve the accuracy of the model, and describes the evaluation index of the accuracy of the inversion model. Finally, the discussion part emphasizes the potential defects and future challenges of quantitative retrieval of heavy metal concentrations in farmland soil by spectral technology, and transforms the discrete research results into the discussion on the correlation of future soil monitoring. Figure 2 shows the conceptual framework of this study, describing the relationship between each stage of the inversion process and key modeling technologies.

2. Data Acquisition and Platforms

2.1. Heterogeneity of Farmland Soils and Its Impact on Spectral Inversion

When collecting soil samples, considering the heterogeneity of farmland soil and its impact on subsequent spectral inversion is a necessary prerequisite and a key challenge for accurate heavy metal spectral inversion. As a soil suitable for crop growth after artificial cultivation or maturation (including tea garden soil, rice soil, orchard soil, etc.), farmland soil not only has certain differences in the distribution of soil and heavy metals compared to other soil types [17], but also has many physical and chemical properties, such as particle size, type of soil, stratification, pH, water content, etc. may change the adsorption characteristics of heavy metals [18], thereby affecting their spectral absorption properties [19], which is more significant for farmland soils that require frequent fertilization and irrigation operations. In many previous quantitative studies on soil heavy metals, samples were often collected using a uniform sampling method based on geostatistics, which did not adequately consider the spatial scale heterogeneity brought about by the uneven distribution of soil heavy metal content [20,21], resulting in soil samples collected that were not spatially gradient, despite the high accuracy of the model created, they could not be representative of the heavy metal contamination level of soil samples in the study area. Therefore, it is necessary to assess the environmental factors and spatial migration characteristics of heavy metals migration in farmland soils, for example, for mining areas or more heavily polluted soils only relatively low concentrations of heavy metals will stay on the surface, and most of the high concentrations of heavy metals will be deposited into the deeper regions of the soil along with the sewage, exhaust and other media, which will destroy the activity of soil microorganisms and impede the growth of the root system of the crops. However, due to the influence of fertilization and irrigation, the properties of heavy metal migration and adsorption in farmland soil may be completely different. There have been many studies on the adsorption characteristics of heavy metals by different farmland soil types. Huang, in his study on the adsorption characteristics of Cu and Zn on different particle sizes of red rice soil aggregates, found that soil aggregates with clay particle sizes showed faster adsorption rates and capacities of heavy metals, and possessed metal forms that seemed to be more stable [22]; Li’s study similarly found that clay aggregation was mainly resulted due to heavy metal cations being the most sensitive ions inducing clay aggregation, and Li pointed out that polarization-induced covalent bonding plays a crucial role in clay aggregation [23]. In addition, the adsorption capacity of different heavy metal ions on different soils varies greatly, resulting in a huge difference in their spatial distribution in the soil layer. Although it is difficult to fully grasp the spatial distribution of heavy metals in the target area, the inversion of multiple heavy metal contents should be carried out as much as possible according to the adsorption characteristics of the soil heterogeneity for each heavy metal element when the samples are collected or measured. This includes, for example, measuring the various physical and chemical properties of farmland soil before sampling in order to provide a reference for the collection method of soil samples, which is very important for obtaining accurate predictions and the subsequent analysis of the spectral data.

2.1.1. Soil Particle Size

According to the standards of the United States Department of Agriculture (USDA), farmland soils can be subdivided into coarse sandy soils (250–2000 μm), fine sandy soils (53–250 μm), coarse mud (20–53 μm), fine chalky soils (2–20 μm), and clay soils (<2 μm) [24], and according to stratification, they can be divided into the plow, plowpan, subsoil layer, and parent material layer. Previous studies have confirmed that the physicochemical properties of soils and the adsorption of heavy metals are highly correlated with soil heterogeneity [18]. Due to the influence of long-term anthropogenic and mechanical operations, soil structure and aggregate stability are constantly altered. The particle size of soil is closely related to the permeability of soil aggregates. Smaller soil particles have a larger specific surface area, and at the same time, they contain a higher content of organic matter, iron, manganese, and aluminum oxides to provide more adsorption sites [25,26], leading to differences in the adsorption capacity and spatial distribution of heavy metals in farmland soils with different particle sizes, which in turn affects spectral inversions.

2.1.2. Soil Layering

It has been shown that for farmland soil, due to long-term cultivation, the soil surface has more organic matter, which means that the surface soil has more adsorption sites for most heavy metals [27], especially for Zn. Zhong et al. found that the content of Zn in the soil layer was different: plow layer > plow pan > subsoil layer > parent layer [28]; Zhang et al. had a similar finding in that there was a greater content of Cd, Pb, and Zn in the surface layer than in the deep layer. This also makes it difficult to determine the absorption characteristics of different heavy metals [29].

2.1.3. Soil Organic Matter

Organic matter is one of the most important substances within the soil and adsorbs heavy metals and influences the spectral response. The surface soil of farmland is usually rich in organic matter (due to tillage and fertilization), while the situation in deep soil may be completely different. Many studies have shown that soils with a higher organic matter content tend to adsorb and immobilize heavy metals more efficiently, reducing their bioavailability. One study found that the adsorption capacity of soil for heavy metals or metal oxides was greatly reduced by removing organic matter from the soil [30]. In addition, a similar study found that the spectral response characteristics of organic matter can substantially mask the absorption peaks of iron oxides and clay minerals, which in turn affect the spectral response characteristics of heavy metals when the organic matter content of the soil was greater than 2% [31].

2.1.4. pH

pH is also considered to be one of the determining factors for heavy metal sorption in farmland soils. The long-term fertilization of crops significantly alters the pH of farmland soils. Although it does not directly produce spectral absorption features, it directly affects the hydrolysis of heavy metal ions and the formation of ion pairs, and several studies have reported a high correlation between pH and some specific trace heavy metals [32,33,34]. In general, the adsorption capacity of most heavy metal ions in farmland soils is positively correlated with pH, which means that acidified soils may adsorb more heavy metal ions, which are then taken up by crop roots [35]. For example, Huang et al. found that the adsorption capacity of Cu and Zn within farmland soils significantly increased with increasing pH [22]. Uchimiya et al. drew similar conclusions, suggesting that the adsorption capacity of Cr, As, and other heavy metals that exist as oxygenated anions in soils decreases with increasing pH [36].

2.2. Acquisition of Physical and Chemical Values

2.2.1. Pretreatment

In spectral inversion based on different platforms, the method of obtaining the physicochemical values of farmland soils is often carried out through a series of physicochemical experiments. Due to the large number of impurities and complex compositional pairs in farmland soils, as well as the small percentage of heavy metals in farmland soils, obtaining the physicochemical values of heavy metals in soils usually requires a series of preprocessing operations, such as air-drying, milling, sieving, the removal of impurities, digestion, and the production of standard samples of different concentrations [37]. In general, the physical and chemical values correspond to the ground truth of the subsequent inversion model; therefore, the correct acquisition of the physical and chemical values is closely related to the reliability of the inversion model, and the methods should be selected according to the properties of different heavy metal elements. It has been demonstrated that the effect of heterogeneity in soil samples on the accuracy of spectral quantification models can be substantially reduced, and the coefficient of determination of the pretreated model can often be increased by more than 0.2 [11].

2.2.2. Digestion

The digestion process is particularly important in a series of soil pretreatment processes. The sample needs to be processed via digestion before entering the instrument, enabling detection of true levels of heavy metals. The choice of digestion method directly affects the precision and accuracy of the analysis results [38]. Commonly used soil digestion methods at this stage are mainly classified as wet digestion, microwave digestion, and pressure tank digestion. Wet digestion refers to the destruction of impurities in the soil using an acidic liquid to extract the heavy metal content. Traditional digestion equipment, such as a water bath, digestion furnace, and electric hot plate, is more common in the laboratory. However, there are still some disadvantages in their application scenarios, such as the water bath only being superior in low-temperature digestion, which is more limited when dealing with high-temperature samples; the digestion furnace and the electric hot plate have the advantages of temperature control and stability. However, they are still insufficient in terms of energy consumption and the consistency of experimental results. The newly developed hole-type digestion instrument is regarded as an effective method for wet digestion, as it can simultaneously digest multiple samples and has good operational stability, solving the problems of inefficiency and poor reproducibility of the previous method. Akhter et al. successfully analyzed the concentrations of Cd, Cr, Cu, Co, Fe, Mn, Pb, and Zn in garlic soil by using the wet digestion method with an atomic spectrophotometer for the first time [39]. The microwave digestion method is used to extract heavy metals by heating and decomposing the samples, utilizing the effect of molecular polarization and ionic conductivity produced by microwaves. Microwave digestion can involve the “internal heating” of various areas of the sample so that the sample is heated more quickly and evenly. Common microwave digestion instruments are mainly divided into closed digestion instruments and automated focusing digestion instruments. Closed digestion instruments can achieve more thorough digestion of samples and have less acid loss, but due to the tanks being independent of each other, the batch digestion cost is higher. In high temperatures and high pressures, there is a certain safety risk; on the contrary, automated focusing digestion can achieve batch digestion but may cause a certain degree of acid loss. In a non-closed situation, it may cause the samples to decompose. On the contrary, automated focused digestion can achieve batch digestion, but it may cause some acid loss and may cause sample contamination under non-closed conditions. There have been cases in which microwave digestion and ICP-MS have been successfully used to detect the contents of Ni, Cu, As, Cd, Sn, Zn, Pb, and Hg in a corn field, and ideal results have been achieved [40]. Pressure tank digestion uses chemical reactions (acid–base neutralization, redox, etc.) under high temperatures and high pressures to digest heavy metals in samples, which has the characteristics of a complete digestion effect and a wide range of applications. Some scholars found that pressure tank digestion has superior effects on Se digestion and subsequent prediction by comparing pressure tank digestion with microwave digestion. However, pressure tank digestion instruments are often inefficient and difficult to use for large-volume samples.

Table 1. This is a table comparing the three digestion methods.

Digestion Method	Advantages	Deficiencies	Applicability	Reference
Electrothermal plate digestion	Easy to operate, wide range of applications, good stability	May have an effect on the results of the experiment	Most elements	[41]
Microwave digestion	High efficiency of digestion in a single pass, high sample recovery, precise control	Higher cost, difficult to achieve simultaneous digestion of large volumes of samples	Most elements	[40]
Pressure tank digestion	Easy to operate, wide range of applications, low sample contamination	Low digestion efficiency and high operational requirements	Volatile elements	[42]

shows a comparison of the three digestion methods mentioned above.

2.2.3. Instrumental Analysis

Instrumental analyses of heavy metal content are often performed after the digestion process described above. Currently, heavy metal content detection methods (i.e., ground true in spectral inversion modeling) are mainly performed using X-ray fluorescence spectrometry (XRF), atomic absorption spectrometry (AAS), atomic fluorescence spectrometry (AFS), inductively coupled plasma atomic emission spectrometry (ICP-AES), and inductively coupled plasma mass spectrometry (ICP-MS). In the context of a low heavy metal limit content in farmland soil, it is often necessary to use methods with higher sensitivity and lower detection limits. Therefore, Hg and As are often determined in soil using AFS. Some scholars have already used the AFS method to determine the content of Hg and As within the soil in a tea plantation, which demonstrates the use of AFS in farmland soil testing [43,44]. Heavy metal elements such as Cu, Zn, Ni, Cr, Cd, and so on, are considered to be more suitable for AAS and the ICP-AES method [45]. For example, Mahmood et al. successfully used flame atomic absorption spectrometry (AAS) to detect the content of Na, Ca, K, Mg, and Fe in a vegetable field [46]; Nolos et al. successfully determined the content of eight heavy metal elements in vegetable and soil samples using the ICP-AES method [47]. Most of these methods are inefficient despite their high measurement accuracy. Nevertheless, the process of obtaining the physico-chemical values through instrumental analyses is essential for most quantitative spectral inversions in order to obtain the ground truth of the spectral inversion model, the accuracy of which is crucial for training and validation during subsequent modeling.

2.3. Inversion Platform Based on Spectroscopy

Spectral inversion of heavy metals in farmland soils is mainly divided into three types, namely, proximal sensing spectroscopy, airborne and drone spectroscopy, and satellite-based spectroscopy. Proximal sensing spectroscopy provides detailed spectral information on the ground, airborne and drone spectroscopy enables rapid regional monitoring, and satellite-based spectroscopy has the capability of macro-monitoring on a large scale. Figure 3 illustrates several different spectral inversion platforms.

2.3.1. Proximal Sensing Spectroscopy

Proximal sensing spectroscopy, also known as in situ field spectroscopy, is a non-destructive detection method based on the development of visible–near-infrared spectroscopy (VNIR-SWIR, 350–2500 nm). The technique can directly obtain the fine features of reflectance spectra on the surface of the ground (a spectral resolution of up to 3–10 nm) through contact or non-contact measurements (a measuring distance of <200 cm) [48]; features a low cost, high accuracy, high resolution, and flexibility; has been widely applied to the quantitative inversion of heavy metals in soil of farmland on a regional basis; and is considered to be a reliable detection technique for heavy metal. Field spectroscopy has been successful in the inversion of soil active components such as clay minerals and organic matter [49,50] since the early 20th century. Meanwhile, the determination of heavy metals has been proven to have a close relationship with active soil components. With a certain degree of predictive method transferability, and compared with airborne spectroscopy and satellite spectroscopy, proximal sensing spectroscopy is considered to have a higher accuracy of inversion, so it has been widely used in the quantitative analysis of heavy metal content in farmland soils [11,51,52]. For example, Xu et al. successfully obtained the best validation accuracies of Hg, Cu, and Cr for an agricultural land in Liaoning province by using an ASD FieldSpec 3 handheld ground spectrometer through the combination of FOD, the optimal band combination algorithm, and GRNN, which were Hg (=0.70, RPD = 1.86), Cu (=0.65, RPD = 1.73) Cr (=0.69, RPD = 1.73), and Hg (=0.70, RPD = 1.86) respectively. 0.69, RPD = 1.83) [53]; Zhang et al. also inverted the Cd concentration of a farmland soil in Baoding using a PSR-3500 field spectrometer, and the RPD vs. values of the laboratory spectra of the established GA-PLSR model were improved from 1.919 and 0.707 to 3.727 and 0.923, respectively, and those of field spectra increased from 1.057 and 0.036 to 1.747 and 0.646, respectively [54]. Due to the limitation of the number of soil samples collected and the area of the soil samples collected, however, it is only suitable for measuring local small-scale areas. It cannot satisfy macroscopic monitoring on a large scale, and the model has poor transferability. We believe that in the future, the development trend of proximal sensing spectroscopy will focus on building collaborative monitoring networks, that is, deploying spectral monitoring stations in typical pollution areas to achieve the continuous monitoring of heavy metal concentrations in minutes. At the same time, combined with airborne and satellite-borne spectrometers for intensive measurements at the field scale, it is expected to break through the spatial limitations of proximal spectroscopy and accelerate its transformation from a research tool to an application-oriented monitoring tool.

Table 2. Some common proximal sensing-based spectral sensors.

Sensor	Mechanism	Bandrange/ Resolution	Manufacturer	Cost	Advantages	Shortcomings	Reference
ASD FieldSpec	Full-range VIS-NIR-SWIR reflectance	350–2500 nm; 3 nm @700 nm, 10 nm @1400/2100 nm; 25° FOV	Malvern Panalytical, Malvern, UK	USD 50,000–70,000	Gold-standard field accuracy; portable	Limited to point measurements; requires dark current calibration	[55]
PSR-3500	Handheld VIS-NIR-SWIR spectrometer	350–2500 nm/3.5 nm @700 nm	Spectral Evolution, Haverhill, MA, USA	USD 22,000–28,000	Low cost; portable	Requires sunlight optimization	[56]
AgroSpec Mobile	Vehicle-mounted NIR spectrometer	900–1700 nm/10 nm	Tec5, Steinbach, Germany	USD 28,000–35,000	Multi-point measurement; portable	Incomplete bands	[57]

shows some common proximal sensing-based spectral sensors.

2.3.2. Airborne and Drone Spectral Remote Sensing

Airborne spectral remote sensing, as an important part of the air-based Earth observation system, acquires sub-meter-level hyperspectral image data by carrying manned aircraft (such as fixed-wing aircraft and helicopters) or unmanned aerial platforms (UAVs). The core of the technology lies in the fact that airborne imaging spectrometers (e.g., HySpex, AVIRIS, CASI) can realize the simultaneous acquisition of hundreds of consecutive narrow bands (bandwidth of 5–20 nm) in the spectral range of 380–2500 nm, forming a three-dimensional data cube with the characteristic of “one picture in one spectrum” (space × space × spectrum). Compared with other platforms, the spatial resolution of airborne spectroscopy can be dynamically adjusted by flight altitude (usually 300–5000 m) and instantaneous field of view (IFOV). Therefore, airborne spectroscopy can flexibly adjust the flight altitude to improve the resolution according to the actual situation of the investigated soil area, and it can realize ultra-fine observation of 0.1–2 m in the key area, which is suitable for mesoscale (10–1000 km²) farmland pollution monitoring tasks.

Farrand first used AVIRIS airborne spectral data to establish an early soil heavy metal spectral monitoring platform [58], and then airborne spectroscopy gradually began to be applied to the regional-scale distribution of soil heavy metal contamination [16]. Along with the development of airborne sensors, the semi-quantitative/full-volume inversion of soil heavy metal content has also been gradually applied to airborne spectroscopy. For the first time, Tan et al. used HySpex VNIR-1600 and HySpex SWIR-384 airborne sensors to establish an inversion model for the Cr, Cu, and Pb content of agricultural land in a mining area in Xuzhou and found that the RF model had the highest prediction accuracy. It was found that the R² of Cr, Cu, and Pb in the RF model were 0.75, 0.68, and 0.74, respectively, and the RMSEs were 5.62, 8.24, and 2.81 (mg/kg), respectively. This demonstrates the potential of airborne spectroscopy for quantitative analysis under complex surface conditions [59]. However, airborne spectroscopic data are often difficult to obtain, mainly due to the high cost of flight operations of 500–2000 USD per square kilometer originating from airspace approvals, aircrew and fuel consumption, and the data acquisition cycle being strictly limited by meteorological conditions (cloudiness < 10%, wind speed < 8 m/s). In addition, radiometric calibration requires simultaneous ground-based spectral measurements (ASD FieldSpec), increasing manpower investment. These bottlenecks have spawned the rapid development of UAV hyperspectral technology. To this end, many studies have begun to make the shift to UAV spectroscopy.

The Unmanned Aerial Hyperspectral System (UAV-HRS) forms a new near-Earth observation paradigm by integrating a miniature spectrometer (e.g., GaiaSky-mini and Headwall Nano-Hyperspec) with a multi-rotor/vertical take-off and landing (VTOL) platform. Its technical advantages include the following aspects: First, it has a spatial resolution of up to 1–10 cm (flight altitude of 50–200 m), which can identify heavy metal enrichment features caused by farmland micro-geomorphology. Then, operation timeliness is improved to an hourly response, which supports the dynamic monitoring of key nodes. Finally, the cost of a single operation is only 1/10–1/20 of that of traditional aviation. Based on GaiaSky-mini spectra, Yi et al. first verified that low-altitude UAVs have the ability to roughly estimate the content of certain heavy metals in farmland soils, which provides a basis for the quantitative inversion of heavy metals in farmland soils by UAV spectroscopy [55]. How to solve the shortcomings, such as incomplete spectral bands of UAV sensors, insufficient endurance, and poor flight stability at the present stage, however, still limits the development of UAV spectroscopy [60], and it is expected that this series of problems will be solved in the future with the innovations of spectral sensors, the upgrading of on-board energy systems (e.g., hydrogen fuel-cell UAVs), and the fusion of intelligent algorithms.

Table 3. Some common airborne and drone spectral sensors.

Sensor	Mechanism	Bandrange/Resolution/Spatial Resolution	Manufacturer	Cost	Advantages	Shortcomings	Reference
HySpex (Airbone)	Push-broom imaging spectrometer	400–2500 nm/1800 px/5.3 nm/0.5–19 m	Norsk Elektro, Oslo, Norway	USD 320,000+	Full range of bands; high resolution	Requires ground calibration; high operational cost	[59]
CASI-1500 (Airbone)	Programmable imaging spectrometer	380–1050 nm/288 bands (2.2 nm FWHM)/1–5 m	ITRES, Calgary, AB, Canada	USD 400,000+	Band customization for specific HM features; lightweight (15 kg) suits UAV integration	Limited to VIS-NIR	[61]
AVIRIS-NG (Airbone)	Whiskbroom imaging spectrometer	380–2510 nm/425 bands (5 nm)/3–20 m	NASA/JPL, La Cañada Flintridge, CA, USA	USD 10 M+ (system)	Full-range coverage identifies metal oxide features	Requires NASA aircraft; high cost	[58]
GaiaSky-mini (UAV)	Grating-based VNIR-SWIR	400–1000 nm/3.5 nm/8 cm@300 m AGL	Dualix, Wuxi, China	USD 48,000	High stability; high resolution	Needs to be paired with drones; short battery life	[55]
Headwall Nano-Hyperspec (UAV)	Grating Spectral CMOS	400–1000 nm/6 nm/10 cm@100 m AGL	Headwall, Bolton, MA, USA	USD 80,000–120,000	High stability; high resolution	Wind vulnerability; high cost	[60]
DJI P4 Multispectral (UAV)	Filter multispectral camera	6 bands (450–860 nm)/40 nm/5 cm@50 m AGL	DJI, Shenzhen, China	USD 6500	Low cost; lightweight	Incomplete bands; low resolution; short battery life	[62]

shows some common airborne and drone spectral sensors.

2.3.3. Spaceborne Hyperspectral Remote Sensing

As a new generation of Earth observation technology, Spaceborne Hyperspectral Remote Sensing (SHRS) can obtain continuous spectral information on surface materials in hundreds or even thousands of bands through hyperspectral imagers installed on orbiting satellite platforms. Relying on typical multi/hyperspectral satellite data such as Gaofen series (GF), Zhuhai No. 1 (OHS), MODIS, Hyperion, Landsat, etc., and with its unique spectral diagnostic characteristics, this technology has become an important means for large-scale surface heavy metal pollution monitoring.

In terms of application scenarios, the outstanding advantage of satellite-borne hyperspectral is reflected in its ability to provide a sub-nanometer spectral resolution and kilometer-level spatial coverage. A single transit can obtain continuous spectral data for an area of hundreds of thousands of square kilometers, and with periodic revisiting characteristics (such as the 16-day revisiting cycle of Landsat), spatial and temporal dynamic monitoring of heavy metal pollution can be achieved. This macro-monitoring capability shows its unique value in the fields of land resources exploration, ecological environment assessment, and agricultural surface pollution monitoring. In particular, in the monitoring of heavy metals in farmland soil, the spatial inversion of heavy metal concentrations at the regional scale can be achieved by establishing a database of heavy metal feature spectra in combination with machine learning algorithms. Typical examples include the PP-LightGBM fusion model proposed by Lin et al. The multi-element pollution distribution map of As and Cd in farmland soil was successfully constructed by interpreting the GF-5 hyperspectral data, which verified its capability for large-scale mapping [63].

The on-orbit observation of the satellite platform, however, faces the influence of complex atmosphere–surface interactions (including factors such as the atmosphere, vegetation, water mist, and solar radiation angle) [64], which leads to a certain loss of accuracy in the on-board spectra [65]. Therefore, a series of preprocessing operations (geometric correction, atmospheric correction, remote sensing image stitching and cropping, etc.) are required for the quantitative inversion of heavy metals in farmland soils in order to eliminate signal aberrations caused by aerosol scattering, solar altitude angle variations (>±5°), vegetation cover (NDVI > 0.6), and thin-cloud interferences (cloudiness < 10%). Despite the loss of accuracy, recent studies have shown that reliable results can still be obtained by optimizing the preprocessing process and modeling methods. Guo et al., for example, constructed a Cr inversion model based on Zhuhai-1 data and combined it with SNV spectral normalization, UVE feature selection, and SVR regression. The validation set R² reached 0.97, and the RMSE was controlled to be less than 4.3 mg/kg [66]. Sun et al. presented a coupled high-precision inversion model for agricultural soil nickel content based on Zhuhai-1 satellite-borne hyperspectral images [67].

At present, the constraints on the development level of current inversion techniques for star-borne spectra primarily stem from the balance of the spatial–spectral resolution. Among the existing satellite-borne hyperspectral sensors, Hyperion has 242 bands (10 nm resolution), but its 30 m spatial resolution limits fine mapping. WorldView-3 contains only eight multispectral bands, although it achieves a 1.24 m spatial resolution [64]. This contradiction is breaking through with the research and development of a new generation of spectral sensors. The upcoming launch of the Environmental Disaster Reduction 2H star will achieve the synergistic enhancement of a 5 nm spectral resolution and 10 m spatial resolution. Combined with deep learning algorithms (e.g., 3D-CNN and Transformer Architecture) for the extraction of high-dimensional spectral features, the future of star-based spectroscopy for farmland monitoring of heavy metals is expected to greatly improve precision and efficiency in the future.

Table 4. Some common spaceborne spectral sensors.

Sensor	Mechanism	Configuration	Manufacturer	Advantages	Shortcomings	Reference
Landsat-9	MultiSpectral Imaging	433–2290 nm; 9 bands; 30 GSD	NASA, Washington, DC, USA	Free data access	Wide bandwidth, poor weak signal capture capability	[64]
Zhuhai-1 OHS	Prism Spectral Imaging	400–1000 nm; 150 bands; 10 m GSD	Orbita, Zhuhai, China	5-day revisit; 2.5 nm spectral resolution	Low signal-to-noise ratio for dark soils	[67]
GF-5 AHSI	Dual grating spectrometer	400–2500 nm; 330 bands; 30 m GSD	CAST, Beijing, China	Wide spectral coverage; free data access	Low resolution	[68]

shows some common spaceborne spectral sensors.

3. Analysis of Spectral Data

3.1. Preprocessing of Spectral Data

Raw spectral data from diverse platforms often fail to accurately reflect heavy metal content in farmland soils due to interference from soil heterogeneity, environmental factors, and instrument noise, resulting in spectral redundancy or distortion [69]. A series of preprocessing operations, such as spectral curves, can often eliminate or reduce these phenomena and effectively enhance the spectral features [70], as shown in Figure 4. The specific strategy for preprocessing must be closely tailored to the platform from which the data were acquired (see Section 2.3), as the sources of noise and interference introduced by different platforms are fundamentally different. For example, ground-based data are the first to address soil physical heterogeneity (grain size and moisture), while airborne/UAV data focus on atmospheric effects and geometric corrections. In contrast, satellite-based data need to overcome the challenges of mixed pixels and strong atmospheric influences.

For satellite and aerospace hyperspectral data, a series of calibration operations are often required, especially in combination with the characteristics of remote sensing data and the specificity of the farmland scene for optimization. For example, the absorption and scattering of electromagnetic waves by the atmosphere can significantly affect the realism of the spectral signals, so the systematic errors need to be eliminated by radiometric correction (e.g., atmospheric correction based on the MODTRAN model) and geometric correction (e.g., the correction of pixel displacement due to topography) as a matter of priority. In addition, there are often crop stubble, traces of cultivation and other local interference on the surface of farmland, and texture noise can be reduced by spatial domain filtering. Spectral unmixing technology can be used to distinguish the mixed pixels of soil and vegetation so as to extract the pure soil spectra. In the extraction of weak signals of heavy metals, the fusion of multi-temporal phase data can effectively inhibit random noise. It is worth noting that, for the specific response of different heavy metal elements in different bands (e.g., the indirect oxidation characteristics of arsenic in 500–600 nm), it is necessary to design element-oriented band combination strategy rather than relying on the general preprocessing process, which has higher requirements for the precise application of hyperspectral data. In the future, as the resolution of on-board sensors (e.g., GF-5 30 m hyperspectral) and the revisit frequency increase, how to realize adaptive preprocessing (i.e., real-time atmospheric parameter inversion and correction) in a dynamic environment will become a key breakthrough direction to improve the efficiency of heavy metal monitoring in farmland.

Regarding the spectral data acquired in the laboratory, unlike the satellite-borne and aerospace hyperspectral, soil samples often only need to be preprocessed by a series of preprocessing means (e.g., air-drying, milling, sieving, etc.) to eliminate the differences in the physical properties of the soil samples in terms of particle size and moisture, and then a series of mathematical transformations can be used to preprocess the raw spectral data acquired by different platforms for the purposes of noise suppression, feature enhancement, and scattering correction. The raw spectral data acquired by different platforms can then be preprocessed by means of a series of mathematical transformations to achieve noise suppression, feature enhancement, and scattering correction.

3.1.1. Smoothing

The spectral curve is affected by external factors affecting the different energy of each band, resulting in a series of “burr” phenomenon that may occur on the spectral curve. Spectral reflectance smoothing can be reduced to a certain extent, thereby improving the signal-to-noise ratio. It has been shown that "burrs" commonly occur in the 400–760 nm and 2300–2400 nm bands [71].

In spectral smoothing, a variety of methods can be chosen flexibly according to the type of noise and data characteristics. Savitzky–Golay smoothing (SG) balances noise suppression and waveform retention by local polynomials fitting within a sliding window, which is especially suitable for denoising the characteristic peaks in the spectrum of heavy metals in soils with high fidelity, but the width of the window and polynomial order need to be carefully chosen to avoid excessive smoothing [72]. Wavelet Transform (WT) utilizes multi-scale decomposition to separate high-frequency noise and low-frequency signals is good at dealing with non-smooth spectra (e.g., field data with random scattering interference) [73], and accurately removes noise in specific frequency bands (e.g., water molecule absorption interference in near-infrared wavelengths) through thresholding, but the number of decomposition layers and the choice of the parent function affect the reconstruction accuracy.

The moving average method replaces the center point with the mean value of the data in the sliding window, which is simple, efficient, and suitable for spectra with a smooth baseline (e.g., organic matter in the near-infrared region), but the window is too large and leads to the broadening of the characteristic peaks. Gaussian filtering is based on the weighted average of the Gaussian function, which retains the symmetry of the peaks and is often used in quantitative analysis, where the morphology of the absorption peaks needs to be maintained (e.g., the 800 nm shoulder peak of Cu²⁺), and the intensity of the smoothing is controlled by the standard deviation. Median filtering takes the median of the data within the window and is robust to impulse noise (e.g., instrument transient anomalies), but it may obscure detailed information. Low-pass filtering truncates the high-frequency component by the Fourier transform, which is suitable for scenarios where electronic noise is obvious, but the frequency band of the noise needs to be pre-determined to avoid signal loss.

Localized Weighted Regression (LOESS) fits localized data through adaptive weighting, which is flexible in dealing with nonlinear noise (e.g., spectral drift under humidity gradient change), but the computational amount increases significantly with the increase in data volume. Empirical modal decomposition (EMD) adaptively disassembles the signal into multi-frequency intrinsic modal functions (IMFs) to strip noise layer by layer (e.g., overlapping peaks in multi-metal composite contamination), but it is sensitive to endpoint effects. Machine learning denoising (e.g., convolutional neural networks) learns noise patterns through training and can handle complex disturbances, but it relies on a large amount of labeled data and arithmetic support. Various methods can be used individually or in combination (e.g., SG + wavelet) to achieve an optimal balance between signal-to-noise ratio improvement and feature fidelity.

Although the current smoothing method can effectively suppress noise, it is overly dependent on manual parameter adjustment, which easily erases the weak absorption characteristics of heavy metals while eliminating “burrs”. Excessive smoothing may lead to the loss of key signals, and the attenuation of spectral details directly restricts the accuracy of the subsequent quantification of heavy metals.

3.1.2. Feature Enhancement

In the enhancement of the spectral characteristics of heavy metals, it is necessary to take into account their low concentration, weak absorption, and complex interaction with soil components, starting with the physical and chemical mechanisms of spectral signals. Among these, continuum removal (CR) eliminates the interference of the global trend of the soil parent material background reflectivity by mathematically fitting the envelope of the spectral curve and normalizing the original reflectivity to the baseline of the envelope so that the local absorption characteristics of heavy metals combined with minerals (such as the weak absorption shoulder peak of arsenic and iron oxides at 550 nm) can be highlighted. A series of spectral derivatives can also be used to amplify the subtle spectral changes of heavy metal elements. For example, the first derivative (FD) can enhance the sudden change in slope at the boundary of co-adsorption of heavy metals and iron–manganese oxides (for example, the inflection point in the reflectivity of copper near 700 nm caused by the surface complexation of Cu²⁺ and FeOOH) [74], and it is usually combined with a Savitzky–Golay filter (window width of 15–25 nm; polynomial order of 2–3) to suppress high-frequency noise. The second derivative (SD) can separate the overlapping absorption peaks of lead near 2200 nm and Al-OH lattice vibration (a half-peak width of about 20 nm) and quantitatively determine the differences in the occurrence forms of lead by the negative peak position and amplitude (for example, the second derivative response of exchangeable lead at 2215 nm is stronger) [75]. The Fractional-Order Derivative (FOD) breaks through the limitations of traditional integer-order derivatives (first-order and second-order) by introducing non-integer-order differentiation operations (such as 0.5th order and 1.5th order) and can achieve a more refined balance between noise suppression and feature resolution [76]. It is especially suitable for extracting the weak absorption characteristics of heavy metals. Some scholars have found that the full width at half maximum (FWHM) of the hydroxyl absorption peak near 2200 nm is about 20 nm. Using the ν = 0.5 derivative can effectively separate the overlapping peaks of Cd and clay minerals (the signal-to-noise ratio is improved by 1.8 times), while the traditional second derivative will lose 10% of the peak area information due to excessive smoothing [77].

3.1.3. Correction for Sample Heterogeneity

In the context of heterogeneity of farmland soils, the adaptability of heavy metal inversion models across particle size, pH, and humidity has become a focus of research. The multidimensional differences in the physicochemical properties of soil matrices often lead to spectral scattering base shifts, absorption peak shape distortions, and model migration failures, which need to be aligned using spectral baseline correction and normalization algorithms to achieve the mathematical spatial mapping of heterogeneous samples. The standard normal variable (SNV) eliminates the scattering intensity bias caused by particle size differences by centering the mean and normalizing the standard deviation of the sample-by-sample spectra [78]. Multivariate scattering correction (MSC) constructs a linear regression model based on the reference spectra to correct for the light-range effect caused by the differences in the particle size distributions [79], and the spiking algorithm optimizes the migration path of the model by introducing the feature samples of the target domain, as proposed by Guerrero et al. [80]. This effectively mitigates the spectral response bias across regional soil textures.

To address the spectral baseline drift caused by dynamic changes in soil moisture, external parameter orthogonalization (EPO) was used to construct moisture interference subspaces and remove them via projection so that the spectral features of the samples with different moisture contents were focused on the heavy metal response components [81]. The direct normalization (DS) algorithm achieves lossless transfer of wet and dry soil model parameters by building a linear transformation matrix between source and target domain spectra [82]. Chen et al. verified that it can reduce the wet soil organic carbon (SOC) prediction bias to a zero-value interval while improving the RPD index from 0.64–2.04 to 7.01 orders of magnitude [83]. It is worth noting that Knadel confirmed the universal advantage of EPO and DS in eliminating moisture interference through a multi-algorithm comparison, and the fluctuation of the R² values of its corrected model could be controlled within ±0.05 at each moisture gradient [84].

In terms of particle size heterogeneity correction, Tamburini et al. compared the effect of SNV, MSC, and Normalized by Closure (NCL) on the treatment of unground soil samples and found that the difference in RMSE among the three in SOC prediction was less than 0.15 g/kg, which verified the robustness of the scattering correction algorithm to the effect of particle size [85]. Xu’s team further revealed that in terms of the PLSR of the unground soil samples pretreated with MSC model, the coefficient of determination R² improved by 0.12–0.18 compared with the milled sample, and the variance inflation factor (VIF) of the characteristic band decreased by about 40%, confirming that spectral correction can replace mechanical milling to achieve mathematical abatement of particle heterogeneity interference [86].

Regarding the fusion modeling of heterogeneous data from multiple sources, Liu et al. innovatively constructed a ratio model of dry and wet soil reflectance, screened the Cd-sensitive bands (e.g., the C-H bond vibration region near 1650 nm) using the Boruta algorithm, and combined stepwise regression with VIF to control the multicollinearity to finally achieve a 23% increase in the inversion accuracy of Cd content in HJ-1A satellite data [77]. This method breaks through the limitation of traditional laboratory drying treatment and provides a spectral-domain solution for the rapid monitoring of heavy metals in wet soil in situ.

These cases provide new ideas for the migration of soil heavy metal spectral inversion models considering wet different grain sizes and moisture contents. Such algorithms can provide ideas for spectral model migration and avoid the problem of duplicated modeling of heterogeneous samples of soils in quantitative spectral inversion, thus saving a lot of manpower and material and financial resources and obtaining more rapid and reliable prediction results. However, differences in parameters such as resolution, sampling frequency, and the number of bands between spectral sensors make the migration of inversion models across platforms potentially ineffective, and we encourage more research to enable this challenge to be addressed.

Table 5. Comparison of several common preprocessing methods.

Preprocessing Methods	Advantages	Deficiencies	Applicability	Reference
SG	Effective waveform preservation, good noise reduction, high flexibility in parameter adjustment, and strong applicability	Possible data loss and overfitting; complex parameter selection	Most situations	[87]
FD, SD, FOD	Effectively highlights signal peaks, details information, and reduces reflectance baseline drift	May cause an increase in high-frequency noise	Most situations	[88]
WT	Spectral data can be analyzed on different scales to better match the characteristics of the signal and effectively extract information on different levels of spectral detail	High computational complexity; generates redundant information	Better effect on Zn, Cr, As, Cd, and Ni	[89]
CR	Improve the signal-to-noise ratio of spectral data, reduce spectral noise, and facilitate subsequent feature extraction	Possible data loss	Better effect on Zn, Cd, Cr, Cu, As, and Pb	[71]
SNV	Samples can be standardized, and interference from scattering effects, instruments, and soil sample heterogeneity can be reduced	Sensitivity to anomalous data	Suitable for eliminating differences in soil physical properties	[90]
MSC	Improves signal-to-noise ratio, enhances characteristic absorption peaks, and reduces scattering effects and interference from heterogeneity of instrument and soil samples	May increase standard errors between samples	Suitable for eliminating differences in soil physical properties	[91]
MC (Mean Centering)	Eliminates absolute spectral absorptions, removes scale differences between samples for analysis, and provides data visualization	May increase standard errors between samples	Most situations	[13]

shows comparison of some common preprocessing methods.

3.2. Spectral Response Characteristics

3.2.1. Mechanism of Spectral Inversion

Spectral inversion of farmland soil is mainly performed by using its own spectral response characteristics, which can usually be divided into direct inversion and indirect inversion [92]. Direct inversion refers to the use of the inherent spectral response characteristics or absorption peaks of heavy metal elements within the soil itself to directly predict the heavy metal content. However, the heavy metal content is often low in farmland soil and can be affected by other impurities in the soil, unless the detected samples are very seriously contaminated by heavy metals and the detected sample content is relatively homogeneous. The response of heavy metal elements is very weak when detecting them using hyperspectral detection. Therefore, the direct inversion of soil heavy metals is often difficult to achieve in real applications, and there are few previous cases of direct inverse research on direct inversion cases. Although the characteristics of trace heavy metal elements are difficult to identify in the spectrum, they are easily bound to clay, iron oxide, or organic matter in the VNIR-SWIR spectral region (350–2500 nm) [93]. This makes heavy metal elements closely associated with the soil in the VNIR-SWIR spectral bands, and the vibration of these molecular bonds results in an energy shift. The metal cations are then adsorbed on the hydroxylated surface site, so the increase in metal cations leads to a decrease in ROH and an increase in RO (FeO, etc.) on clay and oxide surfaces [94]. In addition, the overtone between NH, CO, and CH and other groups leads to the complexation of heavy metals, which results in a large number of combinations between organic matter and heavy metals. Therefore, the high correlation between the concentration of heavy metals and organic matter makes the indirect inversion of heavy metals possible [95]. At the present stage, there are many constructive studies that provide a theoretical basis for the inversion of heavy metal content in farmland soil. Metal oxides, for example, have been found to have obvious absorption peaks in the visible light–short-wave infrared band (350–1100 nm), and the chemical bonds such as O-H, C-H, C-C, C-N, and C-O in organic matter can form a complex or adsorption state and produce cations by reacting with heavy metals. Cations, when receiving electromagnetic radiation, produce overtone vibrations, resulting in unique absorption characteristics in specific bands. There are 400–1000 nm, 1100–1450 nm, 1850–2050 nm, 2200–2400 nm, and other bands, which have been widely used in the study of soil composition [93,96,97], including the heavy metal content of agricultural soils. Regarding the organic matter of sensitive bands, the absorption bands of organic matter may be different due to the different soil sample types and water contents, but most of the absorption peaks of organic matter are concentrated at 350–700, 900–1000, 1300, 1650, 1900, and 2035–2300 nm [98]. Many scholars have carried out related research following the above findings, such as using spectroscopic techniques to predict the correlation between As and Cr, mainly depending on Fe and Fe oxides. The correlation between Fe and Fe oxides is thought to be closely related to the content of Cu [15]. Fe-Mn compounds in soil are thought to react with the cations of heavy metals to reduce the O-H bonds and produce more structurally stable heavy metal oxides [99], which is thought to be advantageous for the indirect inversion of soils [94]. Moros et al. took advantage of the relationship between heavy metals and organic matter in soil, and they concluded that partial least squares regression is useful in the indirect inversion of Cd, As, Co, Cd, and Cu in soils [100]. Dong et al. found a strong correlation between soil Cu and Zn elemental content and organic matter and pH value when targeting orchard soil inversion [101]. The heavy metal inversion model based on the 3000–8000 nm mid-infrared (MIR) range is better than the prediction effect of the visible–near-infrared range [96]. Regarding the characteristic absorption bands of heavy metal elements, most of the elements have an overlap between the sensitive bands, but each element still has its own unique absorption bands (Figure 5). The bands of 410, 581–630, 670, 690, and 1270 nm, for example, have been proven to have better accuracy in predicting Cd content [74,102]; Pb shows spectral characteristics close to the ultraviolet wavelength, and the characteristic bands are considered to be mainly located in the range of 400–700 nm [103]. Although some progress has been made in the study of inversion mechanisms, due to the complex physicochemical properties of farmland soils and the fact that the heavy metals are mostly trace elements, the spectral characteristics of soils are susceptible to the influence of the inversion mechanism. Although some progress has been made on the inversion mechanism, due to the complex physicochemical properties of farmland soils and the fact that heavy metals are mostly trace elements, the spectral characteristics of soils are easily affected by various factors, and the adsorption states of soils for different heavy metals are also different. In addition, due to the continuous and high-resolution characteristics of spectral remote sensing data, the full-band data of soil have problems such as redundancy of information, overlapping absorption, and multiple covariance, which make the prediction effect of the model established by the full-band spectral data poor. Therefore, a series of feature recognition algorithms established using the above spectral response mechanism has been gradually applied to spectral inversion in order to seek and analyze the laws of the spectral response features, which is convenient for guiding the screening of spectrally valid information.

3.2.2. Feature Extraction

Feature extraction, a fundamental technique in spectral data processing, reduces feature dimensionality while preserving critical information. This approach enables the identification of spectral features that best reflect the heavy metal–carrier interaction mechanism detailed in Section 3.2.1. Regarding the complex and redundant spectral information of agricultural soils, how to select the necessary information and reduce the difficulty of data analysis has become a troubling problem, especially when mapping high-dimensional data projection to a low-dimensional space, which will inevitably result in the loss of part of the original data information [104]. Currently, spectral feature bands are often selected by choosing raw or preprocessed univariate feature bands or full-band spectra for correlation analysis, or by combining algorithms (i.e., spectral indices), but the use of full-band modeling often leads to serious “dimensional catastrophe” problems [105]. Therefore, a suitable selection of feature bands can effectively simplify the model, shorten the prediction time, and improve the accuracy and generalization ability of the model in soil heavy metal inversion.

When performing feature screening, the band with the highest correlation with spectral reflectance needs to be selected as the feature variable [106]. The correlation coefficient method (Pearson and Spearman) is simple and fast for assessing the linear relationship between spectral features and heavy metal content and has been widely used in the identification of characteristic bands in agricultural soils [107]. However, its inability to capture nonlinear correlations has the potential to ignore inter-variable interactions. On this basis, Uninformative Variable Elimination (UVE) assesses the contribution of each variable to the predictive power of the model by calculating the correlation coefficients between the variable and the target variable and maximizes the predictive power of the model through repeated iterations [108], but it is sensitive to the intensity and distribution of the noise, and the parameters need to be set carefully. Principal Component Analysis (PCA) is a good solution to the problem of high data dimensionality and multicollinearity. It often extracts uncorrelated main variables to explain the maximum variance, effectively reduces the dimensionality of the data, and extracts the effective components of the data to facilitate the identification of effective spectral bands [109]. It has been proven to be suitable for quantitative spectral inversion studies on most soil types, but the new variables generated by the method lack clear physical significance and are difficult to associate with specific absorption peaks of soil heavy metals. However, the new variables generated by the algorithm lack clear physical meaning and are difficult to correlate with specific absorption peaks of soil heavy metals, leading to a decrease in the interpretability of the model. In contrast, the Successive Projection Algorithm (SPA) can significantly reduce the problem of covariance of redundant spectral information of different substances within the soil by maximizing the orthogonality of the projection vectors to screen the characteristic bands, and the model has high interpretability [110]. However, its iterative process is sensitive to the initial variables and may omit band combinations with synergistic effects, resulting in poor model prediction accuracy. In recent years, some emerging feature selection methods have also emerged. For example, the genetic algorithm (GA) is an algorithm based on genetic mechanism to simulate natural selection, and a large number of studies have used GA feature selection combined with PLSR modeling to achieve good prediction results in the inversion of Cd, As, Zn, and other heavy metals’ content in agricultural soils [54,111,112], but it is time-consuming and needs to be finely adjusted for crossover and variation probabilities, which leads to its limited practicality. Competitive adaptive reweighted sampling (CARS) is a feature selection method based on Monte Carlo sampling. CARS is a feature selection method based on Monte Carlo sampling, which is used for the fine selection of significantly correlated feature bands, and solves the “combination explosion problem” between variables to a certain extent [113]. The advantages of the inversion of As and Cd content in agricultural soils have been mentioned several times [114,115]. However, it is more suitable for high-dimensional small-sample data, and the iterative process may converge prematurely, which leads to the loss of key bands. The Boruta algorithm, first proposed by Kursa [116], is a random forest-based feature extraction algorithm that distinguishes the importance of variable features by comparing the original features and their randomized versions of “shadow features”, and it can comprehensively assess the importance of bands, but the computational complexity is high.

Each of the above methods has its own limitations in the screening process of spectral feature bands in farmland soils, and the difficulty of the current research lies in how to combine different types of farmland soil data with the characteristics and needs of the trade-offs in choosing different methods. In the future, with the sharing of global soil databases and the integration of data from multiple sources, it is expected that the process of feature extraction will reduce a large amount of redundant information in the spectral information, which will help the subsequent modeling process.

3.2.3. Spectral Index

Spectral indices, as an important tool for characterizing vegetation, soil composition, and physical parameters in remote sensing, are mainly used to amplify weak correlation information, suppress environmental noise interference, and enhance effective spectral features through linear or nonlinear mathematical operations of multi-band reflectance. These indices are widely used to extract surface parameters (e.g., vegetation cover, soil moisture, and heavy metal content) in satellite and laboratory hyperspectral inversions.

Common types of spectral indices include difference indices (DI), ratio indices (RI), normalized indices (NDI), etc. Their construction is often combined with algorithms such as multiple linear regression (MLR), partial least squares regression (PLSR), and random forests (RF), which have been validated for their effectiveness in the identification of the spectral response of heavy metals [117,118,119].

In the detection of trace heavy metals in farmland soils, direct measurements are limited due to high cost and poor timeliness. It was found that heavy metal-stressed crops (e.g., rice, wheat) triggered differences in spectral response due to changes in chlorophyll content, e.g., the phenomenon of abrupt changes in the chlorophyll absorption of red light and reflection of near-infrared light (i.e., the red-edge effect) in the red-edge interval of 680–750 nm was significantly correlated with heavy metal concentrations. The quantification of such differences with vegetation indices (e.g., the NDVI, SAVI, and RVSI) can indirectly invert the soil pollution status. For example, Liu et al. constructed a heavy metal vegetation stress index (HMSVI) in a study of peach forests in Pinggu District, Beijing, and successfully predicted the distribution of Cd, As, and Pb in the soil [120]; Shi et al. found that the reflectance of wheat leaves in the bands of 768 nm, 939 nm, and 953 nm was highly correlated with the arsenic content of the soil and established a PLSR model based on this [10]; and Liang used a genetic algorithm (GA) to screen the characteristic bands of rice leaves and achieve high-precision inversion of As and Cd.

The red-edge position (REP), as a key parameter of the red-edge effect, has been shown to be closely associated with the bioeffectiveness of heavy metals such as Cu, Cr, Cd, and As [121]. For example, the REP shift of rice leaves can reflect the accumulation level of heavy metals in soil [122], while differences in the spectral response (e.g., reflectance changes in the near-infrared band) of wheat chlorophyll content can effectively indicate the level of arsenic contamination [10]. These studies indicate that by analyzing the response mechanism of vegetation spectral features and heavy metal stress, the limitations of traditional detection can be overcome, providing an efficient and non-destructive technical path for monitoring soil pollution in farmland. Future studies need to further optimize the band selection and mathematical combination form of spectral indices, as well as integrate machine learning models to enhance the ability to capture nonlinear relationships, which is expected to promote the intelligent development of accurate management of the farmland environment. Especially for the research of farmland soil, spectral indexes more suitable for soil properties need to be developed, such as adding parameters, including the soil particle size, humidity, and temperature, to ensure the accuracy and comprehensiveness of prediction.

3.3. Machine Learning Modeling Methods

3.3.1. Linear Regression Models

Linear regression models are a commonly used model for estimating the soil heavy metal content, including partial least squares regression (PLSR), multiple linear regression (MLR), and principal component regression (PCR), which are characterized by a simple structure, few parameters, and a relatively simple process. Among them, PLSR is one of the most common and effective spectral inversion models, which is mainly used to predict the future trend of the data by analyzing the historical data. PLSR was initially applied to predict the heavy metal stress level in the floodplain in the spectral data analysis of soil heavy metals [123], and it was widely used in the monitoring of heavy metals in farmland [54,75,111,112,124], which has the advantages of dealing with the problem of multiple covariation between the independent variable and the dependent variable effectively. It can screen out the variables with high correlation with the model prediction when facing a small sample size so as to improve the accuracy of model prediction. Since the inversion of heavy metal content is often multivariate, MLR has been widely applied to the multivariate analysis of spectra, and Kemper et al. [37] first demonstrated the feasibility of MLR for soil heavy metal stress levels, which can capture the joint effects of multiple independent variables on the dependent variable and has a good effect of explaining the interactions between variables, allowing the model to adapt to the complex data patterns and relationships better, Therefore, a large number of studies have used MLR to evaluate the pollution characteristics and source analysis of heavy metal pollution in farmland soils [125]; however, MLR is sensitive to outliers and noise and may lead to the problem of multivariate covariance when there are multiple independent variables that are correlated, resulting in unstable model prediction results. The predictor variables were first converted to principal components using PCA, and then the variables were input to MLR for prediction [97], which is mostly used for the regression modeling of VNIR-SWIR spectra and has been shown to be suitable for predicting soil heavy metal stress levels in grassland, industrial, and agricultural areas [126,127]. Although PCR and PLSR have the advantage over MLR in dealing with multicollinearity of data, they can only estimate the linear relationship between spectra and soil properties, and it is difficult to fully capture the complex relationship between spectral features and heavy metal content [128]. On the contrary, some nonlinear models have emerged in recent years to effectively address such problems.

3.3.2. Nonlinear Models

Machine learning has shown a wide range of potential applications in soil physicochemical property research, especially driven by hyperspectral technology, and various types of models have been breaking through performance boundaries in tasks such as soil classification, pollution monitoring, and digital mapping. Especially when dealing with nonlinear problems, some models have shown considerable prediction accuracy. Random forest (RF), as a representative of integrated learning, achieves high generalization and anti-interference ability through the voting mechanism of multiple decision trees [129], and its advantage lies in the natural immunity to multivariate covariance and the efficiency of handling high-dimensional data [130]. For example, Hong et al. used RF to predict Cu, Zn, Cr, and other elements in soil, with the coefficient of determination (R²) exceeding 0.59 [131], while Tan’s RF model preprocessed by continuum removal exceeded 0.90 in the inversion of Cr, Zn, and As in farmland, which verified its adaptability to complex soil environments [132]. Support vector machine (SVM), on the other hand, transforms the nonlinear problem into one that is linearly differentiable in high-dimensional space through kernel function mapping [133], and it performs well in small-sample scenarios. Comparative experiments by Lv et al. show that the performance of SVM is significantly better than that of Bayesian Regularized Neural Networks (BRNNs) and partial least squares regression (PLSR) in star-borne hyperspectral inversion [74].

Regarding the model overfitting problem, Ridge Regression (RR) demonstrates stability in balancing complexity and fit by introducing L2 regularization terms to constrain the parameter weights, which is especially suitable for the scenario of multiple covariances between heavy metal content and spectral features in farmland soil [134]. XGBoost, as an efficient variant of gradient boosted decision tree (GBDT) [135], gradually optimizes the prediction accuracy through residual fitting strategy, and has been reported to have an R² of 0.65 in Cu content inversion [136], which confirms its potential to deal with nonlinear relationships.

With algorithmic evolution, deep learning models further expand the boundaries of soil heavy metal analysis. This type of model exhibits unique advantages in processing spaceborne hyperspectral data with massive samples, as it can capture complex spatial spectral patterns. Early studies such as that by Kemper et al. used artificial neural networks (ANNs) to mine the nonlinear associations between spectra and heavy metals [37], whose multilayered structure endowed the models with powerful feature abstraction capabilities; the subsequently developed back-propagation neural network (BPNN) optimized the weights by gradient descent to achieve higher accuracy in pollutant monitoring [137]. Extreme Learning Machine (ELM), on the other hand, significantly improves the training efficiency through the single hidden layer structure with parameter adaptivity [138], e.g., the prediction accuracies of As and Cd by ELM are close to those of traditional models in cinnamon soil type farmland. The iteration of these techniques not only enriches the methodological toolbox but also provides technical support for the large-scale and high-precision intelligent monitoring of soil environment.

Table 6. Some cases of inverse modeling of heavy metal content in farmland soils based on spectroscopic techniques (mainly based on Chinese farmland soils).

Region	Heavy Metals	Soil Types	Bands	Preprocessing Methods	Modeling Methods	Model Accuracy	Reference
Fuyang, China	Zn, Cu, Ni, Cr, As, Cd, Pb, Hg	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	LT, SG, Boruta	PLSR, CUBIST	RMSE— As: 4.04 Cd: 0.19 Cr: 8.04	[124]
Calcutta, India	As	Vegetable soil	VNIR-SWIR (350–2500 nm)	SGFD	Elastic net	$R^2$— As: 0.97	[139]
Xinjiang, China	Cr	Chilli soil	VNIR-SWIR (350–2500 nm)	SG, SD, LT	PLSR	$R^2$— Cr: 0.903	[75]
Xuzhou, China	Cr, Cu, Pb	Wheat soil	VNIR-SWIR (400–2500 nm)	MODTRAN, VCA	RF	$R^2$— Cr: 0.75 Cu: 0.68 Pb: 0.74	[59]
Fushun, China	Hg, Cu, Cr	Rice and vegetable soils	VNIR-SWIR (350–2500 nm)	FOD	GRNN, RF	$R^2$— Cu: 0.65 Cr: 0.69 Hg: 0.70	[53]
Yushu, China	As, Cu	Maize, rice, soybean soils	VNIR-SWIR (400–2500 nm)	RF	PP–LightGBM	$R^2$— As: 0.73 Cu: 0.75	[63]
Yixing, China	Cd, As	Rice soil	UV-VIS (301–1145 nm)	AFD	GA-PLSR	$R^2$— As: 0.89 Cd: 0.77	[111]
Geum River, Korea	As, Cu, Pb	Farmland soils in mining areas	VNIR-SWIR (350–2500 nm)	SG, PCA	CACNN	$R^2$— As: 0.82 Cu: 0.74 Pb: 0.82	[140]
Baoding, China	Cd	Orchard, rice soil	VNIR-SWIR (350–2500 nm)	SG	GA-PLSR	$R^2$— Cd: 0.923 (Lab), 0.646 (Field)	[54]
Baoding, China	Zn	Traditional farmland soils	VNIR-SWIR (400–2500 nm)	SG	GA-PLSR	$R^2$— Zn: 0.75	[112]
Minya Governorate, Upper Egypt	Cd, Co, Cu, Cr, Pb, Zn	Wheat, maize, soybean, cotton, potato, and sugarcane soils	VNIR-SWIR (350–2500 nm)	SG, SD, UVE	UVE-PLS	$R^2$— Cr: 0.74 Pb: 0.72 Cd: 0.62 Cu: 0.59 Co: 0.52 Zn: 0.46	[141]
Urumqi, China	Hg	Arid Zone Farmland Soil	VNIR-SWIR (350–2500 nm)	LTFD, ATFD	RF	$R^2$— Hg: 0.856	[142]
Shaoguan, China	Cr	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	SG, SNV, UVE, MSC, FD, SD, DWT	SVMR	$R^2$— Cr: 0.97	[66]
Honghu, China	As	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	FD, GF, NOR, CARS	CARS-PSO-SVM	$R^2$— As: 0.9823	[115]
Wuhan, China	Cr, As, Cd	Rice soil	VNIR-SWIR (350–2500 nm)	SG, PCA	PLSR	RPD: — Cr: 2.70 As: 1.81 Cd: 1.63	[13]
Khuzestan, Iran	Ni, Cu	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	SG	PLSR, PCR	$R^2$— Ni: 0.905 Cu: 0.825	[126]
Inner Mongolia, China	Ni, Cr	Grassland soil	VNIR-SWIR (350–2500 nm)	MSC, SNV, SD	PLSR, PCR, SVMR	$R^2$— Ni: 0.98 Cr: 0.98	[127]
Xuzhou, China	Cr, Zn, As, Pb	Farmland soils in mining areas	VNIR-SWIR (350–2500 nm)	FD, CR, SD, SNV	RF	$R^2$— As: 0.9912 Cr: 0.9110 Zn: 0.9061 Pb: 0.9756	[132]
Jiangxi, China	Cd	Farmland soils in mining areas	VNIR-SWIR (400–2500 nm)	FD	RF	$R^2$— Cd: 0.61	[74]
Xuzhou, China	Cd, Cr, Cu, Pb, Zn	Farmland soils in mining areas	VNIR-SWIR (350–2500 nm)	FD, CR	GRNN, MLR, SMO-SVM	$R^2$— Cd: 0.8628 Cr: 0.8532 Cu: 0.7988 Pb: 0.7901 Zn: 0.7653	[143]
Guangdong, China	Cu	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	SG, MSC, CWT	SVR, PLSR, BPNN, XGBoost, RF	$R^2$— Cu: 0.77	[136]
Xuzhou, China	Cr, Zn, Pb	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	FOD	ELM	$R^2$— Cr: 0.77 Zn: 0.86 Pb: 0.63	[119]
Guizhou, China	Cu, Cr, Ni, Pb	Karst seed soil	VNIR-SWIR (500–2500 nm)	SNV, MSC, NOR, FD, AT	ELM	$R^2$— Ni: 0.861 Cu: 0.883 Cr: 0.880 Pb: 0.797	[90]
Hubei, China	As	Farmland soils in mining areas	VNIR-SWIR (350–2500 nm)	IRIV-SCA	SVMR	$R^2$— As: 0.97	[144]
Xi’an, China	Ni, Fe, Cu, Cr, Pb	Orchard, forestry, farmland soils	VNIR-SWIR (350–2500 nm)	MDPSO	RF	$R^2$— Cr: 0.872 Pb: 0.876 Fe: 0.906 Cu: 0.912 Ni: 0.913	[145]
Shaanxi, China	Fe, Ni	Wheat, fruit tree soils	VNIR-SWIR (350–2500 nm)	FD, SD, CR, CWT, SG	SVM, ELM, PLSR	$R^2$— Fe: 0.71 Ni: 0.69	[146]
Yunnan, China	Zn, Ni	Traditional farmland soils	VNIR-SWIR (350–2500 nm)	FOD, SPA	PLSR, RF	$R^2$— Cr: 0.77 Zn: 0.86	[147]

shows some cases of inverse modeling of heavy metal content in farmland soils based on spectroscopic techniques.

3.3.3. Evaluation of Model Accuracy

The performance of the established heavy metal inversion model for farmland soils mainly depends on the accuracy, generalizability, and robustness of the model. The commonly used accuracy evaluation indexes include the coefficient of determination R², residual prediction deviation RPD, root mean square error RMSE, and the ratio of performance to quartiles RPIQ. Among them, it is the most widely used parameter, which ranges from 0 to 1. The closer to 1, the higher the prediction accuracy of the model and the better the fit between the model and the data, and the coefficient of determination of the laboratory spectral modeling in agricultural soils has been reported to often be in the range of 0.62–0.94. The RMSE is obtained by averaging the squares of the prediction errors and squaring the squares; therefore, the smaller the value of RMSE, the higher the prediction accuracy of the model, and the better its performance against quartiles, such as RPIQ. Therefore, the smaller the RMSE value, the higher the prediction accuracy of the model, and the more sensitive it is to the extreme errors in the predicted values. The validation method is mainly cross-validation, which is a good method to comprehensively evaluate the generalization ability of the model by dividing the dataset into a training set and a test set (or more subsets) and repeating the process of training the model on the training set and evaluating the model’s performance on the test set in order to validate the model’s performance in the face of different samples. Robustness is then mainly evaluated through outlier sensitivity analysis and model stability analysis. It should be noted that it is usually difficult to explain the quality of the models solely through the horizontal comparison of accuracy. The evaluation standard still needs to be evaluated by comparing the size of model parameters, data quality, environmental variables, and platform differences and selecting the appropriate accuracy evaluation index.

4. Discussion

Spectroscopy has become an important technology in monitoring heavy metals in soils, allowing for rapid and non-destructive monitoring of the biochemical properties of soils and field crops in relation to heavy metal content and providing guidance for crop management.

By integrating different bands of near-earth, unmanned aerial, aerospace, and satellite spectral sensors and adopting appropriate preprocessing, feature selection, and modeling methods, accurate and efficient estimation of heavy metal content in agricultural fields can be achieved. Based on what we have summarized in this paper, we believe that most research at this stage still faces the following challenges.

(1) There are many combinations of spectral preprocessing and modeling methods, and it is difficult to make a suitable choice. Due to the unavoidable influence of spatial scale, the physical and chemical properties of soil; environmental details; and other natural and human factors, equipment, or operation can easily become a source of error, and the regarding variety of farmland soil types, there is no model with enough applicable scenarios and high generalization ability to meet the needs of heavy metal detection in farmland soils. Thus, the solution still requires the popularization of the global library of spectral preprocessing and modeling methods for the calibration and high standardization of data acquisition and operation, as well as the development of a global database of spectral preprocessing and modeling methods. Further improvement in the efficiency, high-precision heavy metal retrieval, and convenience of soil heavy metal spectral detection is still a great challenge.

(2) Most of the studies on quantitative inversion of heavy metals in farmland soil focus on the field or laboratory VNIR-SWIR spectra, so the research results have limitations and regional bias. At the same time, the accuracy measurement standard is difficult to judge, and it is difficult to achieve real-time monitoring at the macro scale. Other inversion methods still need to be popularized.

(3) Due to the lack of high-quality studies in other regions, our results are mainly based on farmland soil environments in China, and the results have some geographical limitations. At this stage, the geographical limitations of this field are mainly reflected in the over-reliance on Chinese farmland data for model construction, which leads to the doubt of its global applicability. The research cases are focused on red soil and rice soil in the monsoon region of East Asia, which lacks the validation of other typical agricultural regions, and global research is still needed in the future.

(4) The degree of automation of soil spectral heavy metal detection is low. The degree of automation is not high even in the laboratory analysis stage, which still requires human beings to carry out a series of sampling, measuring, calibrating, and calculating processes, and the degree of automation of the spectral inversion method still needs to be improved.

(5) There are few practical applications, and the cost of the equipment required for spectral inversion is relatively high, while requiring a certain technical basis for the operators, and there are few studies on the evaluation of the economic benefits of the spectral inversion model for soil heavy metals.

(6) It is difficult to integrate data acquired using different platforms (i.e., ground-based, airborne, and satellite-based spectra) at this stage, and although algorithms have been applied to calibrate field-collected spectra and laboratory-measured spectra from specific sampling points, calibration methods are still scarce. In addition, there are still differences in acquisition time, spatial resolution, acquisition band, and accuracy between them; therefore, model migration across instrument platforms is also a problem, and more research is needed to provide more effective ways to realize the spatial regional distribution of heavy metal contamination in agricultural soil hyperspectral images at a wide range of scales.

Therefore, in terms of future research prospects for heavy metal detection in agricultural soils, we believe that we should increase the promotion of the establishment of a global library of soil spectral data preprocessing and modeling and calibration methods, and by opening up soil sample data collected in different regions, we can provide suitable models for heavy metal detection in different application scenarios. We encourage the exploration of other spectral bands besides the VNIR-SWIR spectral range in the establishment of the inversion model in order to obtain more information about heavy metals in agricultural soils. At the same time, it is important to introduce a highly standardized data acquisition and operation process to ensure the quality and reproducibility of spectral data. Furthermore, it is importnat to explore more machine learning methods that are suitable for dealing with multivariate and nonlinear problems in the context of complex physical and chemical properties of agricultural soils so as to further improve the generalization ability and accuracy of the inversion models. In addition, we should find ways to improve the automation and intelligence of the spectral inversion process. In addition, we should look for ways to improve the automation and intelligence of the spectral inversion process, such as the construction of an intelligent analysis platform for spectral data. Finally, we should further develop multi-platform data fusion algorithms to achieve seamless data integration and adopt optimization algorithms to improve the temporal and spatial resolution and accuracy of the data so as to push forward the development of the monitoring of heavy metals in agricultural soils to a higher degree of precision and a wider range.

5. Conclusions

In conclusion, it is evident that spectral technology provides a novel approach to the analysis of crop moisture and soil heavy metal stress status. This article narratively reviews the research progress, process, and existing achievements of spectral technology in heavy metal inversion in farmland soils. The article summarizes the achievements of previous researchers using the methods of physicochemical value acquisition, soil heterogeneity and its influence on spectral inversion, spectral data preprocessing technology, spectral data dimensionality reduction, feature band extraction, modeling methods and evaluation indexes of accuracy, etc. This article also provides its own evaluation, which shows that the theoretical foundation of soil heavy metal spectral detection has been relatively perfect and has great potential for practical application.