Accurate Estimation of Gross Primary Production of Paddy Rice Cropland with UAV Imagery-Driven Leaf Biochemical Model

Abstract

Accurate estimation of gross primary production (GPP) of paddy rice fields is essential for understanding cropland carbon cycles, yet remains challenging due to spatial heterogeneity. In this study, we integrated high-resolution unmanned aerial vehicle (UAV) imagery into a leaf biochemical properties-based model for improving GPP estimation. The key parameter, maximum carboxylation rate at the top of the canopy ($V_{cmax,0}^{25}$), was quantified using various spatial information representation methods, including mean (μ_ref) and standard deviation (σ_ref) of reflectance, gray-level co-occurrence matrix (GLCM)-based features, local binary pattern histogram (LBPH), and convolutional neural networks (CNNs). Our models were evaluated using a two-year eddy covariance (EC) system and UAV measurements. The result shows that incorporating spatial information can vastly improve the accuracy of $V_{cmax,0}^{25}$ and GPP estimation. CNN methods achieved the best $V_{cmax,0}^{25}$ estimation, with an R of 0.94, an RMSE of 19.44 μmol m⁻² s⁻¹, and an MdAPE of 11%, and further produced highly accurate GPP estimates, with an R of 0.92, an RMSE of 6.5 μmol m⁻² s⁻¹, and an MdAPE of 23%. The μ_ref-GLCM texture features and μ_ref-LBPH joint-driven models also gave promising results. However, σ_ref contributed less to $V_{cmax,0}^{25}$ estimation. The Shapley value analysis revealed that the contribution of input features varied considerably across different models. The CNN model focused on nir and red-edge bands and paid much attention to the subregion with high spatial heterogeneity. The μ_ref-LBPH joint-driven model mainly prioritized reflectance information. The μ_ref-GLCM-based features joint-driven model emphasized the role of GLCM texture indices. As the first study to leverage the spatial information from high-resolution UAV imagery for GPP estimation, our work underscores the critical role of spatial information and provides new insight into monitoring the carbon cycle.

1. Introduction

Gross primary production (GPP) is the total amount of carbon compounds produced through plant photosynthesis and indicates the carbon sequestration capacity of ecosystems [1]. In cropland ecosystems, GPP is integral to yield formation [2]. Paddy rice, as an important staple food crop, occupies approximately 165.58 million hectares globally and sustains over half of the world population [3]. Therefore, accurate estimation of GPP in paddy rice cropland not only benefits the quantification of the carbon sequestration but also contributes to crop yield prediction [4].

Some classical models for GPP estimation have been developed over the past few decades, which can be categorized into the statistical model [5,6], the leaf biochemical properties-based model [7,8,9], the light use efficiency model [10,11], and the solar-induced chlorophyll fluorescence model [12,13]. These models have significantly advanced our understanding of the global carbon cycle. Among them, the leaf biochemical properties-based model, on the foundation of the Calvin cycle theory, elucidates the three steps of carbon sequestration in photosynthetic carbon reactions and has received much attention for its excellent physical interpretability.

In the leaf biochemical properties-based model, a major source of error in GPP estimation arises from the uncertainty in a key parameter, the leaf maximum carboxylation rate (normalized to 25 °C, $V_{cmax}^{25}$) [14]. $V_{cmax}^{25}$ indicates the enzyme activity of Rubisco. The substantial spatiotemporal variability of $V_{cmax}^{25}$ has long been observed [15,16]. Consequently, extensive efforts have been invested in quantifying the dynamics of $V_{cmax}^{25}$ to acquire an accurate GPP estimation. On the one hand, a number of biophysical and biochemical factors that mechanistically relate to $V_{cmax}^{25}$, such as leaf nitrogen, leaf phosphorus, specific leaf area, leaf chlorophyll content, and leaf area index, have been identified [17,18,19,20]. On the other hand, the spectral information, including vegetation index, multispectral reflectance, hyperspectral reflectance, and fluorescence spectral information, has also shown promising potential for retrieving $V_{cmax}^{25}$ [21,22,23,24,25]. The abovementioned factors have been incorporated into the leaf biochemical properties-based model and significantly enhance the accuracy of GPP estimation [14].

At the field-scale, $V_{cmax}^{25}$ is a macroscopic feature within the specific spatial range, theoretically upscaled from leaf-level $V_{cmax}^{25}$ and thus determined by the spatial distribution of $V_{cmax}^{25}$-related properties on underlying surfaces. However, the spatial information of $V_{cmax}^{25}$-related properties on underlying surfaces has rarely been considered in previous studies. Traditional methods for measuring biophysical and biochemical factors, whether destructive or non-destructive, make it challenging to obtain precise spatial distributions. While satellite remote sensing offers widely accessible spectral images at a medium and low spatial resolution (10–30 m resolution) from platforms like Sentinel-2, Landsat 8 OLI and TIRS, Landsat ETM+, and Moderate-resolution Imaging Spectroradiometer (MODIS), this resolution is often insufficient for capturing the detailed spatial distribution of spectral properties. Moreover, satellite images are frequently disrupted by clouds, causing spatiotemporal discontinuities. These gaps in spatial information likely impede the accurate estimation of filed-scale $V_{cmax}^{25}$. In contrast, the unmanned aerial vehicle (UAV) can capture high-resolution images even in cloudy conditions, offering better capability to describe the spatial distribution of spectral information. Therefore, UAV imagery holds great potential for extracting valuable spatial information to enhance field-scale $V_{cmax}^{25}$ estimation.

Due to the high-dimensional characteristics of UAV imagery, extracting meaningful information using traditional methods poses challenges. There are several approaches to extracting the spatial information from different perspectives. The standard deviation of the pixel values is the most straightforward indicator and is commonly used to quantify landscape heterogeneity [26]. However, it fails to provide other valuable information, such as the spatial distribution patterns. Texture analysis is a widely used approach in image processing. Texture is defined as the measure of coarseness, contrast, directionality, line-likeness, regularity, and roughness [27]. Various texture feature extraction methods have been proposed, including statistical approaches, structural approaches, transform-based approaches, model-based approaches, graph-based approaches, learning-based approaches, and entropy-based approaches [28]. Among these, the gray-level co-occurrence matrix (GLCM) is the most commonly used method [29] and has been widely applied in quantifying landscape patterns [30], as well as estimating various properties on underlying surfaces, such as leaf area index and biomass, from remote sensing images [31,32]. Other texture features, such as Fourier spectrum texture [33] and local binary pattern (LBP) texture [34,35], have also shown strong performance in quantifying specific landscape features, including leaf area index and orchard citrus fruits [36,37]. Despite the explicit and interpretable information extracted through texture analysis, this method may overlook large potential information in the raw images. The breakthrough in deep learning techniques, especially convolutional neural networks (CNNs), vastly enhances the capacity to extract image information. CNN utilizes the convolution and pooling operation to automatically detect the patterns across the entire image and thus can make full use of image information. CNN became a hot spot for its unparalleled performance regarding the image recognition issue [38]. Inspired by its success, the CNN technique is also employed to interpret remote sensing images. Notably, the combination of CNN and high-resolution UAV imagery shows a powerful capacity for retrieving valuable variables, such as crop yield [39] and leaf area index [40].

A few pioneering works have explored the retrieval of photosynthesis traits from images. Sun et al. [41] utilized several texture features, including the GLCM texture feature, LBP texture feature, and wavelet transform texture features, to monitor the leaf-scale net photosynthetic rate of potatoes. Deng et al. [42] proposed the optimal deep learning architecture to estimate the leaf-scale $V_{cmax}^{25}$ from hyperspectral images. These works suggest the potential of images for accurate estimation of photosynthetic traits. However, to the best of our knowledge, few studies have examined the effect of spatial information on field-scale $V_{cmax}^{25}$ estimation. As the first study that leveraged spatial information from high-resolution UAV imagery for $V_{cmax}^{25}$ and GPP estimation, we extracted the spatial information of the underlying surface from high-resolution images using various representation methods, including texture features and deep learning, to quantify the macroscopic feature, field-scale $V_{cmax}^{25}$. The UAV-imagery-driven $V_{cmax}^{25}$ model was integrated into a classical leaf biochemical properties-based model for better GPP estimation. Our findings demonstrate the critical role of spatial information and provide new insight into monitoring GPP. The specific objective of this study is to (1) evaluate whether spatial information can improve the accuracy of field-scale $V_{cmax}^{25}$, (2) build a UAV-imagery-driven hybrid framework for GPP estimation and validate its performance on the paddy rice field, and (3) see how spatial information contributes to field-scale $V_{cmax}^{25}$ estimation.

2. Materials and Methods

The research workflow is depicted in Figure 1, delineated into three distinct stages: (1) data preprocessing, including generating high-resolution UAV RGB orthomosaics and reflectance maps of UAV images, and half-hourly high-quality carbon flux measurements; (2) texture feature extraction from UAV images using four methods, including standard deviation features, GLCM-based texture features, LBP-based texture features, and CNN-based features; (3) model assessment, including evaluating the performance of 8 UAV imagery-driven $V_{cmax,0}^{25}$ models and their influences on GPP estimation, and analyzing the feature contribution of each $V_{cmax,0}^{25}$ model using the Shaley value method.

2.1. Experiment Site and Data Collection

A two-year experiment from 2022 to 2023 was conducted in an 18.93 ha paddy field, located in Hengsha Country, Shanghai City, China (Figure 2). Rice was transplanted on June 1st, 2022, and May 25th, 2023, and harvested on December 3rd, 2022, and October 20th, 2023. The soil texture was silt loam, with a mean dry bulk density of 1.37 g cm⁻³ and a mean saturated volumetric soil water content of 44.23%. The paddy field maintained a water layer with a 5–20 cm depth, except during the late tillering and full ripening stages. The plant density was 2.38 × 10⁵ plants ha⁻¹.

2.1.1. Eddy Flux and Micrometeorological Measurements

The eddy covariance system (IRGASON, Campbell Scientific, Logan, UT, USA) with an installation height of 2.85 m collected the 10 Hz three-dimensional wind speed and direction, ambient CO₂ and H₂O density, sonic temperature, and air pressure. The corresponding meteorological and soil elements, including air temperature and humidity (HMP155A, Vaisala Ojy, Vantaa, Finland), incoming outgoing shortwave radiations (CNR4-L, Kipp & Zonen, Delft, The Netherlands), four-layer soil moisture, and temperature at 5 cm, 10 cm, 20 cm, and 40 cm (TDR310N, Acclima, Meridian, ID, USA) were measured.

The 30 min eddy fluxes, including net ecosystem exchange (NEE), latent heat flux (LE), sensible heat flux (H), and friction velocity (u_*), were generated from 10 Hz raw data of the EC system using the EddyPro software v7.0.9 (www.licor.com/). The data quality of eddy flux was evaluated using the scheme proposed by Foken et al. [43], which classifies the flux data into 9 levels. The data labeled as levels 7–9 were regarded as low-quality and removed. In addition, the flux data were dropped when u_* was less than a specific threshold that indicated the period of low turbulent mixing. The u^* threshold was calculated for each season using the moving point method [43]. The seasonal u_* threshold is shown in Table S1 (see Supplementary Material Section S2). Then, the missing flux data were filled using the marginal distribution sampling method [44]. Finally, the 30 min NEE observation was partitioned into GPP and ecosystem respiration (R_eco) using the night-time flux partitioning method [44]. The processes of u_* filter, gap filling, and flux partitioning were conducted using the REddyProc package [45].

To determine the spatial extent of the EC system measurement, the flux footprint was calculated using the Kljun model [46], which requires the standard deviation of lateral velocity fluctuations (σ_y), friction velocity (u_*), Obukhov length (L), wind speed (u), and boundary layer height (blh) (see Supplementary Material, Section S2). σ_y, u_*, L, and u were measured from the EC system. There was no direct measurement of blh, so we adopted the blh data from ERA5 hourly data on singe-level products (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels, accessed on 17 October 2024) in this study. Then, an 80% footprint was regarded as the measurement range of the EC system.

2.1.2. UAV Data Collection

RGB and multispectral images were acquired using DJI Phantom 4 Multispectral (P4M) (SZ DJI Technology Co., Ltd., Shenzhen, China), equipped with one RGB sensor and five monochrome sensors to measure spectral response in the blue (450 nm ± 16 nm), green (560 nm ± 16 nm), red (650 nm ± 16 nm), red-edge (730 nm ± 16 nm) and nir (840 nm ± 16 nm) bands. Flight operations were conducted between 9:00 AM and 12:00 AM at an altitude of 120 m. The Pix4dMapper software v4.4.9 was employed to process the raw photos captured by UAV to generate the RGB orthomosaics and reflectance maps with a spatial resolution of 6.5 cm/pixel. The reflectance maps underwent correction using the images of the calibrated reflectance panel (MicaSense, RP06), which were captured during the UAV flight. Georeferencing was performed using ArcGIS software v10.8.2 to ensure pixel alignment in the UAV images. In total, 33 UAV observations were collected (

Table 1. The date of the UAV flight.

2022	2023
June 6th, 2022	June 1st, 2023
July 20th, 2022	June 28th, 2023
July 27th, 2022	July 11th, 2023
August 3rd, 2022	July 19th, 2023
August 7th, 2022	July 24th, 2023
August 9th, 2022	July 27th, 2023
August 15th, 2022	August 11th, 2023
August 26th, 2022	August 16th, 2023
September 1st, 2022	August 24th, 2023
September 7th, 2022	August 31st, 2023
September 16th, 2022	September 6th, 2023
September 24th, 2022	September 9th, 2023
October 11th, 2022	September 17th, 2023
October 20th, 2022	September 26th, 2023
November 1st, 2022	October 1st, 2023
	October 6th, 2023
	October 11th, 2023
	October 16th, 2023

). To build the daily $V_{cmax,0}^{25}$ model, the daily UAV images were generated using linear interpolation for each pixel. Then, daily UAV images were clipped using the external rectangle of the flux footprint climatology during the rice growth season and masked using daily flux footprint climatology (see Supplementary Material, Section S3).

2.1.3. Canopy Parameters and Canopy Height Measurement

The plant area index (PAI) and clumping index (Ω) were measured using an LI-COR 2200 canopy analyzer. The leaf area index (LAI) was extracted from the PAI using an empirical method proposed by [47] (see Supplementary Material Section S4). Canopy height (h_c) was measured using a tape measure. PAI, Ω and h_c were all measured at fixed positions in the paddy field. There were 39 sampling positions in 2022 and 30 in 2023, respectively.

2.2. Leaf Biochemical Properties-Based Model

In this study, we employed the Boreal Ecosystem Productivity Simulator (BEPS) [48] for half-hourly GPP estimations. The BEPS model is based on leaf biochemical properties and follows a two-leaf scheme, which separated the canopy into sunlit leaves and shaded leaves using the following methods:

$$LA{I}_{sunlit}=2\cos \theta \left(1-e^{-0.5\Omega LAI/\cos \theta }\right)$$

(1)

$$LA{I}_{shaded}=LAI-LA{I}_{sunlit}$$

(2)

where θ is the solar zenith angle; Ω is the clumping index; LAI, LAI_sunlit, and LAI_shaded are the total leaf area index of the canopy, leaf area index of sunlit leaves, and leaf area index of shaded leaves, respectively.

The BEPS model calculates the instantaneous GPP as the sum of photosynthesis from the sunlit leaves and shaded leaves.

$$GPP=A_{sunlit}LA{I}_{sunlit}+A_{shaded}LA{I}_{shaded}$$

(3)

where A_sunlit and A_shaded are the instantaneous photosynthesis rates of the representative sunlit leaf and representative shaded leaf (μmol m⁻² s⁻¹), respectively. The instantaneous photosynthesis rate is calculated using the Farquhar–von Caemmerer–Berry (FvCB) biochemical model [7].

$$A=\min \left(W_c,W_j\right)$$

(4)

where W_c is the RuBP carboxylation rate limited by Rubisco activity (μmol m⁻² s⁻¹); W_j is the RuBP carboxylation rate limited by electron transport (μmol m⁻² s⁻¹); R_d is the leaf dark respiration during the daytime (μmol m⁻² s⁻¹).

$$W_c={\displaystyle \frac{V_{c\max }\left(C_i-{\Gamma }_{*}\right)}{C_i+K_c\left(1+O/K_o\right)}}$$

(5)

$$W_c={\displaystyle \frac{V_{c\max }\left(C_i-{\Gamma }_{*}\right)}{C_i+K_c\left(1+O/K_o\right)}}$$

(6)

$$R_d=0.015V_{c\max }$$

(7)

where C_i is the CO₂ concentration in the intercellular air space (μmol mol⁻¹); O is the O₂ concentration in the intercellular air space and is set as 210 (mmol mol⁻¹); K_c is the Michaelis–Menten constant of Rubisco for CO₂ (μmol mol⁻¹); K_o is the Michaelis–Menten constant of Rubisco for O₂ (mmol mol⁻¹); Γ^* is the CO₂ compensation point in the absence of leaf dark respiration (μmol mol⁻¹); J is the electron transport rate (μmol e⁻ m⁻² s⁻¹), and is a function of the incoming photon flux density (PPFD, μmol photon m⁻² s⁻¹) and J_max:

$$J={\displaystyle \frac{J_{\max }PPFD}{PPFD+2.1J_{\max }}}$$

(8)

The temperature dependence of K_o, K_c, Γ^* is described using the Arrhenius function [49]:

$$f(T_l)=k_{25}{e}^{{\displaystyle \frac{\left(T_l-25\right)E_a}{298R\left(T_l+273\right)}}}$$

(9)

The temperature dependence of V_cmax and J_max is described using the modified Arrhenius function [49]:

$$f(T_l)=k_{25}{e}^{{\displaystyle \frac{\left(T_l-25\right)E_a}{298R\left(T_l+273\right)}}}{\displaystyle \frac{1+e^{\frac{298\Delta S-H_d}{298R}}}{1+e^{\frac{\left(T_l+273\right)\Delta S-H_d}{R\left(T_l+273\right)}}}}$$

(10)

where k₂₅ is the parameter value at 25 °C; T_l is the leaf temperature (°C); R is the universal gas constant (8.314 J K⁻¹ mol⁻¹); E_a is the activation energy (J mol⁻¹); H_d is the deactivation energy (J mol⁻¹); ΔS is the entropy term (J K⁻¹ mol⁻¹).

The classical BEPS model assumes that a linear relationship exists between the maximum carboxylation rate and leaf nitrogen content, and leaf nitrogen decreases exponentially from top to bottom in a canopy [48]. In this study, we assumed that the maximum carboxylation rate decreases exponentially from top to bottom in the canopy and eliminated the two parameters, namely nitrogen content at the top of the canopy and the relative change in $V_{cmax}^{25}$ with leaf nitrogen content.

$$V_{c\max }^{25}\left(LAI\right)=V_{c\max ,0}^{25}{e}^{-k_{n}LAI}$$

(11)

Therefore, the $V_{cmax}^{25}$ of representative sunlit and shaded leaves are calculated as follows (see the Supplementary Material Section S5):

$$V_{c\max ,sunlit}^{25}=V_{c\max ,0}^{25}{\displaystyle \frac{k\left[1-e^{-\left(k_n+k\right)LAI}\right]}{\left(k_n+k\right)\left(1-e^{-kLAI}\right)}}$$

(12)

$$V_{c\max ,shaded}^{25}=V_{\max ,0}^{25}{\displaystyle \frac{\frac{1-e^{-k_{n}LAI}}{k_n}-\frac{1}{k+k_n}\left[1-e^{-\left(k_n+k\right)LAI}\right]}{LAI-\frac{1}{k}\left(1-e^{-kLAI}\right)}}$$

(13)

where $V_{cmax,0}^{25}$ is the $V_{cmax}^{25}$ at the top of the canopy (μmol m⁻² s⁻¹); k_n is the leaf nitrogen content decay rate with increasing depth into the canopy and is taken as 0.3; k = G(θ)Ω/cosθ, and the projection coefficient G(θ) is taken as 0.5 for the assumption of spherical leaf angle distribution. The maximum electron transport rate for the sunlit and shaded leaves (J_{max, sunlit}, J_max,shaded, μmol m⁻² s⁻¹) is obtained using the following equation [50]:

$$J_{\max ,sunlit}^{25}=2.39V_{c\max ,sunlit}^{25}-14.2$$

(14)

$$J_{\max ,shaded}^{25}=2.39V_{c\max ,shaded}^{25}-14.2$$

(15)

The leaf stomatal conductance for CO₂ (g_sc) is coupled with photosynthesis using the Ball–Woodrow–Berry equation [51]:

$$g_{sc}=m{\displaystyle \frac{A{h}_s}{C_s}}+b$$

(16)

where h_s is the relative humidity at the leaf surface (%); C_s is the CO₂ concentration at the leaf surface (μmol mol⁻¹); m and b are the coefficients in the Ball–Woodrow–Berry equation, respectively. m and b are set as 6.08 and 0.019, respectively. These two values were obtained by fitting steady-state-condition A-C_i curves from Li-6800. More detailed information on the BEPS model can be found in the Supplementary Material, Section S6. The BEPS model also considers the influence of soil water stress on g_s [48]. Given that rice cultivation typically involves a flooded or saturated state, with minimal exposure to soil water stress, the impact of the soil water stress factor on g_s was not considered in this study.

2.3. Derivation of Field-Scale $V_{cmax,0}^{25}$ Measurement

This study employed a classical optimization algorithm, the differential evolution (DE) algorithm [52], to retrieve the daily $V_{cmax,0}^{25}$. The differential evolution algorithm contains three steps, including mutation, crossover, and selection. Detailed information about DE algorithm were displayed in the Supplementary Material Section S7. The differential evolution algorithm was implemented using the lmfit python package [53], with the mean square error (MSE) chosen as the objective function. Integrating the DE algorithm with BEPS, called BEPS-DE framework, was employed to derive the field-scale $V_{cmax,0}^{25}$ under the assumption of constant $V_{cmax,0}^{25}$ for 7 days. Only daytime GPP observation (when incoming shortwave radiation is larger than 0) were used to fit $V_{cmax,0}^{25}$ because the BEPS model sets nigh-time GPP as 0 while night-time GPP observation from the EC system are not 0.

2.4. Texture Extraction and $V_{cmax,0}^{25}$ Models

In this study, several types of texture features were evaluated for field-scale $V_{cmax,0}^{25}$ estimation.

2.4.1. Standard Deviation Texture

The first texture feature is the standard deviation of reflectance within the flux footprint climatology.

$${\sigma }_{ref}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N{\left(x_i-\overline{x}\right)}^2}}$$

(17)

where x_i is the reflectance value at each pixel within the flux footprint climatology; $\overline{x}$ is the mean reflectance value within the footprint climatology.

2.4.2. GLCM-Based Texture Features

GLCM is a square matrix that can examine the spatial distribution of the gray levels in the image. It describes how frequently a combination of pixels is present in an image in a given direction and distance. For a given image I with a size of K×K and maximum gray value of G, the element of a GLCM for a displacement vector d(dx, dy) is defined as follows:

$$P\left(m,n\right)={\displaystyle \sum _{i=1}^K{\displaystyle \sum _{j=1}^K\left\{\begin{array}{l}1, \mathrm{if} I\left(i,j\right)=m \mathrm{and} I\left(i+d_x,j+d_y\right)=n\\ 0, \mathrm{otherwise}\end{array}\right\}}}$$

(18)

$$P\left(m,n\right)={\displaystyle \frac{P\left(n,m\right)}{{\displaystyle \sum _{i=0}^G{\displaystyle \sum _{j=0}^{G}P\left(i,j\right)}}}}$$

(19)

where dx = dcosθ; dy = dsinθ; d and θ are the predefined distance and direction, respectively. The GLCM is typically interpreted using texture indicators. Six GLCM-based texture features (

Table 2. The list of GLCM-based texture features in this study.

Texture	Formula
Contrast (con)	$con={\displaystyle \sum _{i,j=0}^{M-1}{P}_{i,j}{\left(i-j\right)}^2}$
Correlation (cor)	$cor={\displaystyle \sum _{i,j=0}^{G-1}\left[\frac{\left(i-{\mu }_i\right)\left(j-{\mu }_j\right)}{{\sigma }_i^2{\sigma }_j^2}\right]}$
Dissimilarity (dis)	$dis={\displaystyle \sum _{i,j=0}^G{P}_{ij}\left|i-j\right|}$
Homogeneity (hom)	$hom={\displaystyle \sum _{i,j=0}^G\frac{P_{i,j}}{1+{\left(i-j\right)}^2}}$
Angular second moment (asm)	$asm={\displaystyle \sum _{i,j=0}^G{P}_{i,j}^2}$
Entropy (ent)	$ent={\displaystyle \sum _{i,j=0}^G-\ln \left(P_{ij}\right)P_{ij}}$

) were extracted from UAV images.

The computation of GLCM is influenced by the pixel pair distance offsets and pixel pair angles. We pre-specified a series of distance offsets (1,2,4,8,16) and angles (0, π/4, π/2, 3π/4). Then, 20 GLCMs, each representing a unique combination of distance offsets and angles, were employed to derive texture samples. Finally, the mean of 20 texture samples was set as the texture feature of the image.

2.4.3. LBP Texture

The LBP represents each image pixel with a local binary pattern operator that generates a binary code while comparing a neighboring pixel and its center patch gray unit. The LBP codes are computed as follows:

$$LB{P}_{P,R}\left(q_c\right)={\displaystyle \sum _{p=0}^{p-1}s\left(q_p-q_c\right)2^p}$$

(20)

$$s\left(x\right)=\left\{\begin{array}{l}1, x\ge 0\\ 0, othersise\end{array}\right.$$

(21)

where q_c is the gray value of the center pixel; q_p is the gray value of P neighboring pixels with a specified radius R. The LBP is invariant against any monotonic transformation of image brightness. The local binary pattern histogram (LBPH) is a commonly used LBP-type texture feature. It employs the distribution of the LBP value to represent the image. A whole image is typically divided into several local regions, and the histograms from each subregion are extracted independently. Then, these subregional histograms are concatenated to build a global description. In this study, we divided the UAV images into 9 subregions (3 × 3), calculated the histogram with 10 bins, and then concatenated the 9 histograms together to obtain the LBPH texture with 90 features. Figure S3 shows the process of extracting the LBPH texture.

2.4.4. Convolutional Neural Network

In this study, we adopted two typically convolutional neural network architectures, including AlexNet [38] and ResNet [54], to extract spatial information from UAV images. ResNet is more complex and has a deeper structure than AlexNet (Figure S4). It is important to note that the input images should maintain the same size when fed into the CNN. Since images with a resolution of 6.5 cm/pixel are too large for the CNN to handle efficiently, UAV images were resampled to a 1m resolution and clipped by the external rectangle (see Section 2.1.2) to maintain the same size. Pixels outside the flux footprint climatology were set as 0. The five-band multispectral images were concatenated into an input tensor (430 × 420 × 5). The input images were preprocessed using the standardization and then fed into the CNN to estimate field-scale $V_{cmax,0}^{25}$.

The different combinations of texture features for $V_{cmax,0}^{25}$ were evaluated (

Table 3. The list of $V_{cmax,0}^{25}$ models proposed in this study.

	${\mathit{V}}_{\mathit{c}\mathit{m}\mathit{a}\mathit{x},0}^{25}$ Model	Description
M1	$V_{c\max ,0}^{25}=\mathrm{DNN}\left({\mu }_{ref}\right)$	μref: mean value of reflectance within the flux footprint climatology. M1 is driven by the mean value of the 5-band reflectance.
M2	$V_{c\max ,0}^{25}=\mathrm{DNN}\left({\mu }_{ref},{\sigma }_{ref}\right)$	σref: standard deviation value of reflectance within the flux footprint climatology. M2 is jointly driven by the mean and standard deviation of the 5-band reflectance.
M3	$V_{c,\max 0}^{25}=\mathrm{DNN}\left(GLCM\right)$	GLCM: GLCM-based texture features. M3 is driven by the GLCM-based texture features from the RGB images.
M4	$V_{c,\max 0}^{25}=\mathrm{DNN}\left({\mu }_{ref},GLCM\right)$	M4 is jointly driven by the mean value of the 5-band reflectance and GLCM-based texture features from the RGB images.
M5	$V_{c\max ,0}^{25}=\mathrm{DNN}\left(LBPH\right)$	LBPH: local binary pattern histogram. M5 is driven by the local binary pattern histogram feature from the RGB images.
M6	$V_{c\max ,0}^{25}=\mathrm{DNN}\left({\mu }_{ref},LBPH\right)$	M6 is jointly driven by the mean value of the 5-band reflectance and the local binary pattern histogram feature from the RGB images.
M7	$V_{c\max ,0}^{25}=\mathrm{AlexNet}\left(Imag{e}_{ref}\right)$	Imageref: standardized UAV reflectance maps. M7 and M8 are driven by the 5-band reflectance maps.
M8	$V_{c\max ,0}^{25}=\mathrm{ResNet}\left(Imag{e}_{ref}\right)$

). For the mean and standard deviation of the 5-band reflectance, GLCM-based texture features, and LBPH-driven models, the deep neural network (DNN) was employed to estimate field-scale $V_{cmax,0}^{25}$.

$$\begin{array}{l}{\mathit{h}}_1=f_0\left({\mathit{w}}_0\mathit{x}+{\mathit{b}}_0\right)\\ {\mathit{h}}_l=f_l\left({\mathit{w}}_{l-1}{\mathit{h}}_{l-1}+{\mathit{b}}_{l-1}\right)\\ \mathit{y}={\mathit{w}}_L{h}_L+{\mathit{b}}_L\end{array}$$

(22)

where x is the input layer, namely the combination of the 5-band reflectance and corresponding texture features; h_l is the l-th hidden layer; w_l and b_l are the weights and bias of the l-th layer; f_l is the activation function of the l-th layer. The performance of the deep neural network highly relies on hyperparameters, including the number of hidden layers, the number of neurons, the selection of activation function, and the dropout value. We used the genetic programming algorithm to optimize the hyperparameters. The search space was set as follows: (1) the maximum number of hidden layer was set as 4; (2) the number of neurons for each hidden layer was selected from the set [16,32,64,128,256,512,1024]; (3) the activation function for each hidden layer was selected as either ReLU or sigmoid; (4) the dropout rate for each hidden layer was selected from an interval of 0 to 0.5. The daily $V_{cmax,0}^{25}$ and corresponding multispectral surface reflectance for each land cover type were randomly apportioned into training and validation sets, with an 80–20 split. The former was used to train the weights and biases, and the latter was employed to evaluate the generalization error.

2.5. Feature Contribution Analysis Using the Shaley Value

The Shaley value (SHAP) was adopted to interpret $V_{cmax,0}^{25}$ models. In this study. SHAP [55] is a game-theoretic approach to explain the performance of the machine learning model. It can quantify the contribution of the input feature to the prediction for an individual sample. SHAP specifies the explanation using the additive feature attribution method:

$$f(x)=g(x^{\prime })={\varphi }_0+{\displaystyle \sum _{i=1}^M{\varphi }_i}{x}^{\prime }$$

(23)

where f(x) is the original model, g(x) is the explanation model with simplified input x′; x′∈{0,1}^M; x and x′ are related via a mapping function, x=h_x(x′); M represents the number of input features; ϕ_i is the feature attribution for the feature i. The explanation model g has a unique solution:

$${\varphi }_i(f,x)={\displaystyle \sum _{z^{\prime }\subseteq x^{\prime }}\frac{\left|z^{\prime }\right|!\left(M-\left|z^{\prime }\right|-1\right)!}{M!}\left[f_x\left(z^{\prime }\right)-f_x\left(z^{\prime }\backslash i\right)\right]}$$

(24)

where |z′| represents the number of non-zero entries in z′, and z′ ⊆ x′; f(x′) = f(h_x(z′)) = E[f(z)|z_S], and S is the set of non-zero indexes in z′.

3. Results

3.1. Seasonal Variation in Carbon Fluxes, Environmental Elements, and Biophysical Indicators on the Paddy Field

Figure 3a illustrates the seasonal variation in daily NEE, GPP, and R_eco during the rice growth period. The NEE displayed a one-valley pattern. The minimum value of daily NEE was −12.08 and −14.60 gC m⁻² d⁻¹ in 2022 and 2023, respectively. The two minimum values occurred on August 25th, 2022, and July 29th, 2023. The cumulative NEE in the growth period was 502 and 552 gC m⁻². The GPP exhibited a one-peak mode, with peak daily values of 21.6 and 20.16 gC m⁻² d⁻¹ in 2022 and 2023, respectively. The peak value occurred on August 21st, 2022, and July 29th, 2023. Total GPP during the growth period was 1429 gC m⁻² in 2022 and 1337 gC m⁻² in 2023. A strong fluctuation can be found in the dynamics of NEE and GPP, while daily R_eco changed smoothly.

The summer of 2022 witnessed a severe meteorological drought along the Yangtze River. Accordingly, the incoming shortwave radiation (SW_IN) and air temperature (T_a) in 2022 were significantly higher than those in 2023 (Figure 3b,c). The daily mean SW_IN and T_a during the summer of 2022 were 21.35 MJ m⁻² d⁻¹ and 28.4 °C. The two values at the same time in 2023 were 18.00 MJ m⁻² d⁻¹ and 27.1 °C. The difference in T_a on the same days between 2022 and 2023 during the summer was as large as 6 °C. Moreover, 2022 experienced lower precipitation (P) and relative humidity (RH) compared to 2023, with mean RH values of 71% and 76, and accumulated P values of 759 and 922 mm during the growth period for 2022 and 2023, respectively (Figure 3d,g). Precipitation was more frequent in 2023, particularly during the summer. Differences in wind speed (u) and atmospheric CO₂ concentration (C_a) between the two years were minimal (Figure 3e,f), with wind speeds of 2.93 m s⁻¹ in 2022 and 2.62 m s⁻¹ in 2023, and CO₂ concentrations of 426 μmol mol⁻¹ in 2022 and 439 μmol mol⁻¹ in 2023. Due to the presence of a persistent water layer in the paddy field and a shallow groundwater table (about 0.5 m), only surface soil moisture (SM) went through several drying processes, while deep SM remained stable (Figure 3h). Therefore, rice rarely suffered soil water stress during the growth period.

When it comes to biophysical indicators, both PAI and h_c reached the plateau at the end of August (at the early heading and flowering stage) (Figure 3i,l). LAI also reached its peak at the end of August with a peak value of 4.52 and 4.03 in 2022 and 2023, respectively (Figure 3j). Then, LAI gradually declined due to leaf senescence. The PAI in 2022 was higher than that in 2023. The former reached 5.84 and the latter 4.44. The gap between 2022 and 2023 h_c values was small. The maximum h_c was about 0.95 m. The dynamics of Ω were stable during the growth period, with values varying from 0.93 to 0.96 (Figure 3k).

3.2. Seasonal Variation of $V_{cmax,0}^{25}$ Observation

Figure 4 presents the seasonal dynamics of field-scale $V_{cmax,0}^{25}$. The field-scale $V_{cmax,0}^{25}$ generally presented a single-peak pattern during the growth period, reaching its maximum in mid-July and declining after mid-August. The peak field-scale $V_{cmax,0}^{25}$ was approximately 170 μmol m⁻² s⁻¹. There were two significant drops in field-scale $V_{cmax,0}^{25}$ in early July and in the end of July in 2022, likely due to weed removal activities. In 2022, weed prevention before rice sowing was ineffective, leading to the thriving of weeds. Consequently, a substantial portion of the carbon flux in the paddy field came from the contribution of weeds.

3.3. Accuracy of $V_{cmax,0}^{25}$ and GPP Estimation

Figure 5 shows the performance of eight $V_{cmax,0}^{25}$ models. Three metrics, including correlation (R), root mean square error (RMSE), and medium absolute percentage error (MdAPE), were employed to assess the model performance. When employing traditional reflectance values as input (Figure 5a), the R, RMSE, and MdAPE of $V_{cmax,0}^{25}$ estimation on the test dataset were 0.87, 30.08 μmol m⁻² s⁻¹ and 23%, respectively. In contrast, the $V_{cmax,0}^{25}$ model driven solely by texture information displayed a noticeable performance degradation. The GLCM texture-driven model achieved $V_{cmax,0}^{25}$ estimation with an R of 0.80, RMSE of 36.50 μmol m⁻² s⁻¹ and MdAPE of 27% on the test dataset (Figure 5c). The LBPH-driven model exhibited the poorest accuracy, with R, RMSE, and MdAPE values of 0.81, 39.31, and 31% on the test dataset, respectively (Figure 5e). However, it can be seen that combining the reflectance value and texture information can significantly improve the accuracy of $V_{cmax,0}^{25}$ estimation, while the contribution of differing texture information varies significantly. Two joint-driven models, including μ_ref -GLCM texture feature and μ_ref-LBPH texture feature models, substantially improved the $V_{cmax,0}^{25}$ estimations, with R values reaching 0.93–0.94, and RMSE and MdAPE falling from 20.08 to 23.40 μmol m⁻² s⁻¹ and 12% to 16%, respectively. However, the combination of μ_ref and σ_ref provided only a marginal improvement, with an R of 0.89, RMSE of 28.70 μmol m⁻² s⁻¹ and MdAPE of 22% (Figure 5b). Two CNN methods demonstrated excellent performance. ResNet is typically superior to AlexNet in the ImageNet classification challenge, whereas AlexNet slightly outperformed ResNet in this study. The former acquired the $V_{cmax,0}^{25}$ estimation with an R of 0.94, RMSE of 23.70 μmol m⁻² s⁻¹, and MdAPE of 11%. The latter obtained the $V_{cmax,0}^{25}$ estimation with an R of 0.93, RMSE of 27.30 μmol m⁻² s⁻¹, and MdAPE of 14%. A noteworthy observation is that the CNN methods showed an overwhelming advantage over other models on the training dataset, with R reaching 1, and RMSE and MdAPE being lower than 10 μmol m⁻² s⁻¹ and 10%. Although this result indicates that the CNN methods encountered a degree of overfitting, it still suggests a strong ability to discern and utilize spatial information embedded in UAV images. The ResNet, with a more complicated structure, displays significantly higher potential than the AlexNet. The overfitting issue of CNN methods can be alleviated by training with more data. As the dataset expands, there is a promising outlook for CNN methods to transcend these challenges and realize significant success.

Considering the texture information also contributes to the GPP estimation, it can be found that the superior $V_{cmax,0}^{25}$ models result in more accurate GPP estimation (Figure 6). Two CNN models acquired the best GPP estimation, with an R of 0.92, RMSE of 6.5–6.6 μmol m⁻² s⁻¹, and MdAPE of 23–24%. The u_ref -GLCM texture feature and the u_ref -LBPH texture feature joint-driven models also achieved excellent GPP estimation. The former obtained the lowest RMSE, with a value of 6.43 μmol m⁻² s⁻¹. The latter also obtained promising results, with an R of 0.92, RMSE of 6.74 μmol m⁻² s⁻¹, and MdAPE of 26%. The μ_ref-σ_ref joint-driven model and LBPH solely driven model performed worst for GPP estimation. The RMSE of the two models was larger than 7 μmol m⁻², and the MdAPE was over 30%. The μ_ref-GLCM-based texture feature joint-driven model achieved better performance in $V_{cmax,0}^{25}$ estimation in the high-$V_{cmax,0}^{25}$ range (Figure 5d). In contrast, the two CNN models showed an obvious underestimation issue when $V_{cmax,0}^{25}$ was high (Figure 5g–h), and the underestimation of $V_{cmax,0}^{25}$ further resulted in a significant bias for high GPP estimation.

4. Discussion

4.1. How Spatial Information Contributes to Field-Scale $V_{cmax,0}^{25}$ Estimation

This study assessed different forms of texture information for field-scale $V_{cmax,0}^{25}$ estimation and revealed significant differences in the ability of different texture features to describe field-scale $V_{cmax,0}^{25}$. We further elucidated the contribution of differing texture information using the SHAP method. Figure 7 shows the mean of the absolute SHAP (mean(|SHAP|)) for the six deep neural network-based models. Due to the high dimensionality of input data, only the top five most important features were presented. In the μ_ref-driven $V_{cmax,0}^{25}$ model (Figure 7a), the nir and red bands contributed most. This result is consistent with the current understanding of the relationship between $V_{cmax,0}^{25}$ and the spectrum. These two bands indicate the vegetation growth information and are the components of some classical vegetation indices, such as NDVI and EVI. In the GLCM-based texture-feature-driven-$V_{cmax,0}^{25}$ model (Figure 7c), the top 5 most important features were entropy in the green and red bands, dissimilarity in the blue band, contrast in the RGB band, and homogeneity in the green band. The five features exhibited nearly equal importance. In the LBPH-driven $V_{cmax,0}^{25}$ model (Figure 7e), the top five most important features were the histogram bars from subregions 1, 5, 1, 2, 3. In the μ_ref-σ_ref joint-driven model (Figure 7b), the top 5 most important features included mean value in red, green, and red-edge bands, and the standard deviation value in the green and nir bands. Here, the mean value information played a dominant role, and the standard deviation information played a subordinate one. Adding the standard deviation information diminished the importance of the nir band, which contributed most in the μ_ref -driven model. In the μ_ref-GLCM-based texture feature joint-driven model, the GLCM-based texture features contributed more than the mean reflectance. There were four GLCM textures in the top five most important features, including entropy in the blue, green and red-edge bands, and homogeneity. Conversely, the mean reflectance played a dominant role in the μ_ref-LBPH joint-driven model. The mean values of reflectance in the red, nir, green and blue bands were the top four most important features in this model. The most important LBPH feature was the histogram bar in subregion 8. The SHAP-based feature importance analysis suggests that entropy is a valuable feature for $V_{cmax,0}^{25}$ estimation, with dissimilarity and homogeneity also showing potential. The entropy, dissimilarity, and homogeneity describe the uncertainty of, local variation in, and smoothness of the image texture, respectively [56]. There perhaps exists some inherent link between $V_{cmax,0}^{25}$ and these GLCM-based texture features. While standard deviation is a simple and intuitive indicator of spatial heterogeneity, it offered limited information for $V_{cmax,0}^{25}$ estimation. The LBPH also failed to indicate the dynamics of $V_{cmax,0}^{25}$. Although this method is regarded as a powerful tool for describing local information and has achieved success in classification issues such as face recognition and weed detection, the histogram-format texture descriptor may be unsuitable for regression issues. LBPH failed to provide stable features for quantifying $V_{cmax,0}^{25}$. Despite this, the combination creating the μ_ref-LBPH joint-driven model outperforms the other two μ_ref-texture joint-driven model. A possible reason is that LBPH is a high-dimensional feature and contains more potential information. In summary, textures are important but may be better used as Supplementary Information to estimate field-scale $V_{cmax,0}^{25}$.

The SHAP method was also employed to interpret the AlexNet-based $V_{cmax,0}^{25}$. Figure 8 shows the spatial distribution of SHAP on 4 typical days, clearly revealing the spatial heterogeneity in the SHAP maps. The contributions from roadway and water bodies differed obviously from those from the paddy field. When the underlying surface was bare soil, the spatial variability in SHAPs across bands was relatively low. When the paddy field was covered by rice, the SHAP variability intensified, particularly in the nir and red-edge bands. Specifically, the regions with a high contribution in these bands were closely linked to areas with a mix of vegetation and bare soil or blends of well-growing and poor-growing vegetation. As the rice matured, the spatial variability in the SHAP in the nir band weakened, while its variability in the red band intensified. To sum up, the CNN model attached much importance to the subregion with high spatial heterogeneity in the nir and red-edge bands for $V_{cmax,0}^{25}$ estimation. This emphasis was particularly evident in regions characterized by a mix of vegetation and bare soil, and a mix of well-growing and poor-growing vegetation. The SHAP method was not used to analyze the ResNet-based model due to the complexity of ResNet, which made SHAP computation too resource-intensive. Therefore, the feature contribution of the ResNet-based model was not included in this study.

It should be noted that UAV images were resampled from 6.5 cm/pixel to 1 m/pixel before being fed into the CNN due to limitations to computation capacity. Resampling large images to a smaller, manageable resolution is the most common approach, and this was employed in this study. This approach retains the global spatial information but loses the fine-scale features. The loss of spatial detail will possibly affect the extraction of textural features. Alternative strategies, such as patch-based processing [57], and patch gradient descent [58], offer promising solutions to these challenges. Developing comprehensive methods to interpret ultra-high-resolution UAV images using deep learning is a key area for future exploration.

4.2. Spatial Heterogeneity: Implications for Parameters Conversion Across Different Scales

Our study underscores the significant influence of spatial heterogeneity field-scale $V_{cmax,0}^{25}$. The spatial heterogeneity originates from the diverse elements in paddy fields, including the bare soil, water body, and weeds, as well as rice in different growth states, all of which have distinct photosynthetic properties. Field-scale $V_{cmax,0}^{25}$, representing a macroscopic feature within the flux footprint climatology, is intricately linked to this spatial heterogeneity. In contrast, leaf-scale $V_{cmax,0}^{25}$ solely reflects the local feature. This inherent spatial heterogeneity contributes to the observed mismatch between field-scale and leaf-scale $V_{cmax,0}^{25}$, as depicted in Figure S6 (see Supplementary Material Section S11). Given the global prevalence of spatial heterogeneity, as evidenced in various studies [59], relying solely on the leaf-scale $V_{cmax,0}^{25}$ model for field-scale simulation introduces uncertainties into GPP estimation. Our study advocates for the direct use of field-scale parameters to model GPP, acknowledging the complexities introduced by spatial heterogeneity.

4.3. Limitations

In this study, we proposed high-resolution UAV-image-based models for field-scale $V_{cmax,0}^{25}$ and GPP estimation and demonstrated enhancement through texture information. However, some limitations remain. Firstly, the field-scale $V_{cmax,0}^{25}$ observation was retrieved from the BEPS-DE framework. The accuracy of $V_{cmax,0}^{25}$ observation was completely determined by the EC-based GPP observation and the structure of the BEPS model. We note that there are several different methods for conducting $V_{cmax,0}^{25}$ observation, including the data assimilation approach, sunlit GPP-based inverting method [60], and optimality photosynthesis hypothesis-based approach [15]. In addition, the BEPS model is highly parameterized, with many parameters predefined. The uncertainties in these predefined parameters will ultimately influence the retrieval of $V_{cmax,0}^{25}$. For example, Jiang et al. [61] revealed that the variation in the nitrogen distribution coefficient within the canopy (k_n) resulted in a large gap between retrieved $V_{cmax,0}^{25}$ values. Therefore, it is necessary to comprehensively evaluate the different $V_{cmax,0}^{25}$ retrieval methods. This study evaluates four typical spatial information extraction methods, including σ_ref, GLCM texture, LBPH texture, and CNN. Some newly proposed texture features, such as learning-based approaches and the entropy-based approaches, have garnered significant attention [28]. It is necessary to explore the capabilities of these new features for GPP estimation. Lastly, this study faced some challenges in sample size. The overwhelming advantage of CNN models in the training dataset demonstrates its strong ability to extract valuable spatial information. The performance of CNN algorithms primarily depends on big training data. The CNN-based $V_{cmax,0}^{25}$ models were trained and validated only using 281 daily RBG and five-band reflectance maps. Compared with the complex structure of the CNN, the sample size in this study is relatively insufficient, resulting in a degree of overfitting. It is urgent to expand our dataset to fully harness the potential of deep learning.

5. Conclusions

Accurate estimation of GPP estimation poses a considerable challenge due to the influence of spatial heterogeneity. High-resolution UAV images can capture spatial heterogeneity and offer new insights for GPP monitoring. In this study, we proposed high-resolution UAV-imagery-driven models to quantify $V_{cmax,0}^{25}$, a key parameter of the leaf biochemical properties-based GPP model. The reflectance and texture information were extracted from UAV images and employed to build $V_{cmax,0}^{25}$ models. A significant improvement in $V_{cmax,0}^{25}$ and GPP estimation was obtained. The main conclusions are drawn as follows:

(1)

Combining reflectance and texture features can vastly improve the field-scale $V_{cmax,0}^{25}$ (R = 0.94, RMSE = 19.44 μmol m⁻² s⁻¹, and MdAPE = 11%) and further result in high-accuracy GPP estimation (R = 0.92, RMSE = 6.5 μmol m⁻² s⁻¹, and MdAPE = 23%). The performance of different texture features varies significantly. The CNN method shows the strongest ability to monitor $V_{cmax,0}^{25}$. The μ_ref-GLCM texture features and μ_ref-LBPH joint-driven models also acquire promising results. However, the standard deviation of reflectance contributes less to estimating $V_{cmax,0}^{25}$.

(2)

The contribution of input features changes significantly among different models. The CNN model focuses on nir and red-edge bands and pays much attention to the subregion with high spatial heterogeneity. The μ_ref-LBPH joint-driven model mainly attaches importance to the reflectance information. On the other hand, the μ_ref-GLCM-based features joint-driven model emphasizes the role of GLCM texture indices, including contrast, dissimilarity and homogeneity.

(3)

The strong spatial heterogeneity results in the large gap difference between field-scale and leaf-scale $V_{cmax,0}^{25}$ values and hinders the conversion of parameters across different scales

Our study confirms the necessity of considering spatial heterogeneity information for monitoring $V_{cmax,0}^{25}$ and GPP, which is ignored in the current understanding of the GPP model. The following perspectives should be the focus of future work. First, the fusion of satellite and UAV data offers great potential. Medium-resolution satellites, such as Landsat and Sentinel-2, can capture remote sensing images with spatial resolutions of 10–60 m, providing valuable spatial information about the underlying surface. Additionally, integrating multi-scale remote sensing images can achieve information complementarity, which is likely to enhance the accuracy of GPP monitoring. Second, further exploration of new interpretation methods is needed to better elucidate the mechanistic link between spatial heterogeneity and field-scale parameters.