Comparison of machine learning techniques for identifying weeds in maize crops¹

Comparação de técnicas de aprendizado de máquina para identificar plantas daninhas na cultura do milho

ABSTRACT

The present study was conducted at the Federal Rural University of Rio de Janeiro, in the years 2024 and 2025, to evaluate the performance of three supervised classifiers - maximum likelihood, random forest, and support vector machine - for spectral discrimination between maize (Zea mays L.) and the weed (Cyperus rotundus), considering different phenological stages and two cropping seasons (main and second crop), using multispectral images acquired by remotely piloted aircraft. Overall accuracy, as well as class metrics of precision, recall, and F1-score, were used to evaluate the performance of the algorithms in classifying four targets: maize, weed, soil, and shadow. The performance of the classifiers improved as the phenological stage progressed. In the first year (second crop), the V8 phenological stage showed the greatest differentiation between the plants, and the highest-performing classifier was maximum likelihood, with an overall accuracy of 89%. For the maize class, precision was 0.88, recall was 0.81, and F1-score was 0.85, and for the weed class, precision was 0.68, recall was 0.89, and F1-score was 0.77. In the second year (main crop), the best differentiation occurred at the V6 stage, and the maximum likelihood classifier again presented the best performance, with an overall accuracy of 88%, precision of 1.00, recall of 0.78, and F1-score of 0.88 for the maize class, and precision of 0.52, recall of 1.00, and F1-score of 0.68 for the weed class. Thus, across the years analyzed, the maximum likelihood classifier performed best.

Key words:
remote sensing; supervised classification; vegetative stages; spectral analysis

HIGHLIGHTS:

Multispectral images obtained by remotely piloted aircraft are effective for detecting weeds in maize crops.

Phenological development improves spectral separability, enhancing pixel-based classifications.

The maximum likelihood classifier achieves the best overall performance across all growth stages.

RESUMO

O presente estudo foi realizado na Universidade Federal Rural do Rio de Janeiro, nos anos de 2024 e 2025, com o objetivo de avaliar o desempenho de três classificadores supervisionados - maximum likelihood, random forest e support vector machine - na discriminação espectral entre milho (Zea mays L.) e daninha (Cyperus rotundus), considerando diferentes estádios fenológicos e duas épocas de cultivo (safra principal e secundária), por meio de imagens multiespectrais adquiridas por aeronaves remotamente pilotadas. A acurácia global, assim como as métricas de precisão, recall e F1-score por classe, foram utilizadas para avaliar o desempenho dos algoritmos na classificação de quatro alvos: milho, daninha, solo e sombra. Observou-se a evolução no desempenho dos classificadores com o avanço do estádio fenológico. No primeiro ano (safra secundária), o estádio fenológico V8, apresentou maior diferenciação entre as plantas, e o classificador de melhor desempenho foi o maximum likelihood, com acurácia global de 89%. Para a classe do milho a precisão foi de 0,88, recall de 0,81 e F1-score de 0,85, e para a classe daninha a precisão foi de 0,68, recall de 0,89 e F1-score de 0,77. No segundo ano (safra principal), a melhor diferenciação ocorreu no estádio V6, e o classificador maximum likelihood, novamente, apresentou melhor desempenho, com acurácia global de 88%, precisão de 1,00, recall de 0,78 e F1-score de 0,88 para a classe do milho, e precisão de 0,52, recall de 1,00 e F1-score de 0,68 para a classe daninha. Assim, nos diferentes anos analisados, o classificador maximum likelihood apresentou melhor desempenho.

Palavras-chave:
sensoriamento remoto; classificação supervisionada; estádios vegetativos; análise espectral

INTRODUCTION

Maize is an essential commodity for global food security (Shah et al., 2025). According to FAO (2025), global cereal production is expected to exceed 3 billion tons, a historic milestone driven mainly by maize, which continues to show favorable prospects for yield and production growth compared to 2024. In this context, Brazil’s total maize production in the 2024/25 crop year is estimated at 128.3 million tons (CONAB, 2025).

Good crop management is fundamental for high yield, particularly regarding pest, disease, and weed control. Weeds are among the main causes of economic losses, as they reduce yield and increase production costs due to control measures (Marques et al., 2024).

The use of remotely piloted aircraft (RPAs) has emerged as an efficient strategy for weed detection and mapping, especially when combined with machine learning techniques (Bretas et al., 2024). Studies such as Tetila et al. (2025) highlight the potential of RPA imagery and classifiers for weed recognition in maize fields. Yet, they do not consider the tolerance limits of coexistence between crops and weeds. Machine learning techniques support information extraction from large image datasets, and classifiers such as support vector machine (SVM), random forest (RF), and maximum likelihood (ML) have demonstrated strong performance in remote sensing applications.

The SVM has strong generalization performance, even with small training sets, making it effective at separating spectrally similar classes (Gu & Congalton, 2025). RF is widely used due to its high robustness to spectral variability and the complexity of remote sensing data. This characteristic stems from its structure based on multiple decision trees, which reduces overfitting and enables efficient handling of nonlinear interactions among spectral bands and derived variables, as demonstrated by Liu et al. (2025) in maize mapping over mountainous areas with high spatial heterogeneity.

In contrast, the ML classifier shows consistent performance in agricultural areas with greater intra-class spectral homogeneity, where spectral distributions approximate normality, resulting in high accuracy in pixel-based classification, particularly when there is adequate spectral separability between crops and weeds (Phiri et al., 2020; Kim et al., 2023).

Thus, this research evaluated the performance of three supervised classifiers - maximum likelihood, random forest, and support vector machine - for spectral discrimination between maize (Zea mays L.) and the weed (Cyperus rotundus), considering different phenological stages and two cropping seasons (main and second crop), using multispectral images acquired by remotely piloted aircraft.

MATERIAL AND METHODS

Thestudywasconductedintwodistinct 1-haareas(Figure 1) at the Universidade Federal Rural do Rio de Janeiro (UFRRJ), located in the municipality of Seropédica, RJ, Brazil. Although spatially distinct, both areas share similar soil characteristics, which are classified as Alfisols by the Soil Survey Staff (2022). Before the experiment, the areas were maintained under fallow conditions, with no recent crop cultivation. Each area had its crop evaluated in two distinct periods.

Figure 1
Aerial image of the first area (A) and the second area (B)

In the first area, the analyses were carried out during the autumn period, corresponding to the second crop, and covered april and may 2024 (first year). In the second area, analyses were conducted during the summer period, corresponding to the main crop, and encompassed january and february 2025 (second year). Conducting the experiment in different crop years allowed the evaluation of classifier performance under contrasting environmental and phenological conditions, reflecting natural variability in climate, crop development, and weed dynamics.

The analyses conducted in 2024 and 2025 corresponded, respectively, to the areas with geographic coordinates 22° 46’ 37” S; 43° 41’ 36” W at 7 m in altitude and 22° 47’ 33” S; 43° 41’ 20” W at 9 m in altitude, Datum World Geodetic System (WGS) 84, Zone 23 S. The predominant climate of the region, according to the Köppen classification, is Aw (Alvares et al., 2014), with a total annual precipitation of 1,274 mm and a mean annual temperature of 23.7 °C (Silva & Dereczynski, 2014).

Soil preparation was carried out in both experimental areas and for both planting seasons, consisting of one pass of a heavy disc harrow followed by a pass of a light disc harrow. In the second crop, the hybrid cultivar BM 3063 PRO 2 was sown on April 5, 2024, whereas in the main crop, BM 790 PRO 3 was sown on December 23, 2024, with population densities of 60,000 and 70,000 plants per hectare, respectively, and a row spacing of 0.80 m. The adoption of distinct hybrids was intended to allow evaluation of the classifier’s performance under different cultivar conditions.

The images were acquired by an RPA, the DJI Phantom 4 Multispectral, composed of a Red, Green, and Blue (RGB) camera and a multispectral camera (blue 450 ± 16 nm, green 560 ± 16 nm, red 650 ± 16 nm, red edge 730 ± 16 nm, and NIR 840 ± 26 nm). Radiometric calibration was automatically performed by the Phantom 4 Multispectral’s onboard system under field conditions, before flight operations, in accordance with the equipment’s standard procedure.

The flight dates were defined according to the phenological stages of each maize cultivar, ranging from V2 to V8, which are reached at different days after sowing (DAS) depending on the cultivar and growth conditions. These stages correspond to the estimated coexistence period between maize and the weed (Cyperus rotundus), beyond which crop yield losses may occur (EMBRAPA, 2006; Oliveira et al., 2024). At all collection stages, field surveys were conducted to collect ground reference data for validating weed occurrence locations and the regions identified in the images, following standard ground truthing procedures for remote sensing accuracy assessment (Congalton & Green, 2019). Cyperus rotundus (commonly known as “tiririca”) was the only weed species found in both experimental areas and was therefore considered the sole target species representing the weed class in this study.

In 2024, at the V2 stage, the images indicated the absence of weeds in the study area; therefore, this phenological stage was not analyzed. In 2025, rainfall conditions prevented the aerial survey scheduled for the V4 stage. Rainfall accumulation of 10.8 mm was recorded based on observational data from the local meteorological station Seropédica - Ecologia Agrícola (RJ), operated by the Brazilian National Institute of Meteorology. Such precipitation levels may affect flight safety and the radiometric consistency of the image. Therefore, flights were conducted only under dry conditions and after canopy drying to minimize spectral distortions. At the V8 stage, the high level of vegetation cover promoted by the maize canopy significantly reduced light penetration into the lower strata, inhibiting the emergence and growth of weed species and leading to the absence of this class. Therefore, this stage was not evaluated.

Images were acquired at different maize phenological stages (e.g., V2, V4, V6, V8), allowing evaluation of spectral separability over time. In 2024, the flights were carried out at the V4 (19 DAS), V6 (24 DAS), and V8 (31 DAS) phenological stages, at a height of 50 m, with a GSD of 2.6 cm, 75% front overlap and 70% side overlap, and a speed of 3.0 m s⁻¹. In 2025, the flights were performed at the V2 (13 DAS) and V6 (29 DAS) stages, at a height of 35 m, with a GSD of 1.62 cm, 80% front and side overlap, and a speed of 3.1 m s⁻¹. Still in the field, for image georeferencing, 30 control points were collected according to ASPRS (2023) recommendations, using Hiper VR GNSS receivers with the Real Time Kinematic (RTK) method, providing an average uncertainty of 4 mm.

The initial processing of the images was performed in Agisoft Metashape Professional Edition, version 2.0.0, to generate orthomosaics. Subsequently, these orthomosaics were imported into ArcGIS, version 10.8, using the Geocentric Reference System for South America (SIRGAS) 2000, Zone 23 S reference system for classification. In ArcGIS, pixel-based classification was adopted, considered the most suitable method for images with higher GSD. Four classes were established: maize, weed (Cyperus rotundus), soil, and shadow. Based on the recommendations of Congalton & Green (2019), 50 samples per class were collected for training. Then, the Multivariate tool from the Spatial Analyst Tools package was used to train the algorithm with the samples defined through Create Signatures. For generating the classified image, the following classifiers were used: maximum likelihood (ML), random forest (RF), and support vector machine (SVM).

The random forest classifier used predefined parameters, including the maximum number of trees (50), the maximum tree depth (30), and the maximum number of samples per class (1000). The values of these parameters were maintained with the aim of balancing precision, learning capacity, greater model generalization, and time optimization. In addition, for both RF and SVM, color and mean attributes were defined to differentiate the classes. Subsequently, samples were generated in random regions of the image to validate the classification against the actual land-cover class.

Confusion matrices were calculated to indicate the percentage of objects correctly or incorrectly classified across all images. The confusion matrix was generated using balanced accuracy, so the weights for the classifications were equal across all classes, regardless of their frequencies in the dataset. In this way, the model’s performance reflects its real ability to distinguish between classes, rather than simply the predominance of the majority class.

Classifier performance was expressed as overall accuracy (OA), precision, recall, and F1-score, as in García-Navarrete et al. (2025). Overall accuracy was used to determine the total proportion of correctly classified pixels in relation to the total number of reference pixels, as shown in Eq. 1. Precision was used to evaluate the probability that a pixel classified into a given class actually belongs to that class in the field, as indicated in Eq. 2. Recall was used to determine the proportion of reference pixels of a category that were correctly classified, as represented in Eq. 3, determining how frequently the classifier detected the samples of each class and the sensitivity of the model. Finally, the F1-score, presented in Eq. 4, computed the harmonic mean of precision and recall, ranging from 0 (worst performance) to 1 (best performance) (Liang et al., 2020).

(1) $$\text{Overall Accuracy}=\frac{\text{Total number of correctly classified pixels}}{\text{Total number of reference pixels}}\times 100$$

(2) $$\text{Precision}=\frac{\text{Number of correctly classified pixels in each class}}{\text{Total number of pixels classified in that class}}\times 100$$

(3) $$\text{Recall}=\frac{\text{Number of correctly classified pixels for each class}}{\text{Total number of reference pixels in that class}}\times 100$$

(4) $$F1-\text{score}=2\times \frac{\text{Precision}\text{×}\text{Recall}}{\text{Precision}+\text{Recall}}$$

These metrics were used to evaluate the quality of image classification and the performance of each classifier studied, using the same training samples for ML, RF, and SVM at each phenological stage. The confusion matrix and overall accuracy were obtained using the Spatial Analyst Tools package through the Compute Confusion Matrix tool in ArcGIS. Precision, recall, and F1-score were computed from the confusion matrix in RStudio using the caret library.

RESULTS AND DISCUSSION

The classification results for the first year, obtained with the ML, RF, and SVM classifiers at the V4, V6, and V8 maize phenological stages, are summarized in Tables 1 and 2. Table 1 presents the confusion matrix, whereas Table 2 reports the performance metrics used for evaluation, including overall accuracy and class-specific precision, recall, and F1-score, providing the basis for a comparative analysis of classifier behavior across the phenological stages.

Table 1
Confusion matrix of the maximum likelihood (ML), random forest (RF), and support vector machine (SVM) classifiers for the V4, V6, and V8 phenological stages of the maize crop in the first year

Classifiers				ML					RF				SVM
Stage	Class	Maize	Soil	Weed		Shadow	Maize	Soil	Weed	Shadow	Maize	Soil	Weed		Shadow
	Maize	20	0		4	1	21	0	2	2	21	0		4	0
V4	Soil Weed	0	24		0	1	1	24	0	0	0	23		0	2
		15	2		8	0	15	4	6	0	13	5		6	1
	Shadow	2	0		0	23	2	0	0	23	0	0		0	25
	Maize	23	0		1	1	17	1	7	0	17	0		4	4
V6	Soil Weed	0 15	25 0		0 9	0 1	1 15	23	0	1	0	25		0	0
								0	6	4	10	1		11	3
	Shadow	0	0		1	24	2	0	1	22	0	0		0	25
	Maize	22	0		2	1	19	0	6	0	16	0		3	6
V8	Soil Weed	0	25		0	0	0	25	0	0	0	25		0	0
		5	0		17	3	8	1	14	2	6	1		15	3
	Shadow	0	0		0	25	4	0	2	19	2	0		0	23

Table 2
Performance metrics obtained by the maximum likelihood (ML), random forest (RF), and support vector machine (SVM) classifiers for the V4, V6, and V8 phenological stages of the maize crop in the first year

Classifiers		ML			RF					SVM
Stage	Metric	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow
	Precision	0.80	0.96	0.32	0.92	0.84	0.96	0.24	0.92	0.84	0.92	0.24	1.00
	Recall	0.54	0.92	0.67	0.92	0.54	0.86	0.75	0.92	0.62	0.82	0.60	0.89
V4
	F1-score OA	0.65	0.94	0.43 0.75	0.92	0.66	0.91	0.36 0.74	0.92	0.71	0.87	0.34 0.75	0.94
	Precision	0.92	1.00	0.36	0.96	0.68	0.92	0.24	0.88	0.68	1.00	0.44	1.00
	Recall	0.61	1.00	0.82	0.92	0.49	0.96	0.43	0.81	0.63	0.96	0.73	0.78
V6
	F1-score OA	0.73	1.00	0.50 0.81	0.94	0.57	0.94	0.31 0.68	0.85	0.65	0.98	0.55 0.78	0.88
	Precision	0.88	1.00	0.68	1.00	0.76	1.00	0.56	0.76	0.64	1.00	0.60	0.92
	Recall	0.81	1.00	0.89	0.86	0.61	0.96	0.64	0.90	0.67	0.96	0.83	0.72
V8
	F1-score OA	0.85	1.00	0.77 0.89	0.93	0.68	0.98	0.60 0.77	0.83	0.65	0.98	0.70 0.79	0.81

Overall accuracy improved with maize development. At the V4 stage, the three classifiers ML, RF, and SVM showed similar performance, indicating limited class separability at this early phenological stage. At V6, classification performance became more differentiated, with the ML classifier outperforming the other methods, while RF showed a reduction in overall accuracy. The highest overall accuracy values were observed at V8, when canopy development and spectral contrast were more evident, with the ML classifier performing best across stages.

For the maize class, performance varied across both classifiers and phenological stages. At V4, RF and SVM exhibited higher precision than ML. However, at V6 and V8, ML outperformed the other classifiers, showing superior precision and recall, particularly at V8, when crop structure and spectral stability were more developed. SVM showed higher recall at the earlier stages, indicating greater sensitivity in identifying maize pixels. These trends were reflected in the F1-score: SVM showed better balance at V4, while ML achieved the best overall balance at V6 and V8.

The soil class was well classified across all phenological stages and classifiers. High precision and recall values indicate that bare soil was spectrally distinct from the other classes, especially at V6 and V8, when near-perfect classification was achieved. Among the classifiers, ML demonstrated the most stable performance across all stages for this class, as reflected in the F1-score values.

In contrast, the weed class presented the greatest classification challenge. Precision values were below 0.5 at the V4 and V6 stages, indicating that more than half of the weed class was confused with other classes, mainly maize. Nevertheless, recall values were comparatively higher, suggesting that most weed pixels were detected, albeit with limited specificity. An improvement was observed at V8 for all classifiers, especially for ML and SVM, reflecting increased canopy differentiation and reduced spectral overlap. The F1-score confirmed this trend, showing gains at V8 and highlighting the superior balance achieved by ML.

For the shadow class, all classifiers exhibited strong performance throughout the phenological stages. SVM showed consistently high precision at V4 and V6, while ML achieved the best balance between precision and recall at V6 and V8. At V8, overall performance remained high, as indicated by the F1-score, demonstrating that shadow areas were effectively distinguished across phenological stages.

The results indicate that the increase in canopy cover with phenological advancement reduced soil spectral noise and enhanced contrast between the crop and weeds, thereby improving classifier performance (Almeida-Ñauñay et al., 2023). For the maize class, SVM showed better balance at the initial stage (V4), indicating a greater ability to handle spectrally similar classes. In contrast, ML performed best at the more advanced stages (V6 and V8), when classes were morphologically and physiologically more distinct, making it the highest-performing classifier for the maize crop. This behavior is consistent with Medina & Atehortúa (2019) and Dash et al. (2023), who highlight the superiority of SVM in situations of high spectral similarity and the efficiency of ML in scenarios with greater differentiation between features.

In the case of the soil class, the high precision and stability across phenological stages reflect its distinct spectral signature compared with vegetation. Zhao et al. (2022) reported high performance (99.67%) for soils using a hybrid RF model, a result consistent with the present study. Although Foody (2023) notes that high metrics may mask sampling imbalances, the use of balanced accuracy in the present study mitigated this effect, thereby ensuring robust evaluations. The high separability of soil, particularly at early crop stages, is attributed to its distinct spectral signature compared to vegetation, as soil exhibits reflectance patterns that differ markedly from canopy reflectance across multiple spectral bands, enabling its discrimination in remote sensing analyses (Yang et al., 2025).

The low precision obtained for the weed class at all stages confirms the difficulty in distinguishing between crops and small-sized weeds due to spectral overlap and shadowing (Tetila et al., 2025). The observed limitation is mainly due to the weed growing very close to the base of maize plants and to progressive canopy closure, which reduces weed exposure and increases shading, thereby hindering spectral separability. However, the performance increase at the V8 stage indicates that greater vegetative development strengthens the species’ spectral signature, favoring its detection (Villiers et al., 2023). Regarding the shadow class, the alternation between SVM and ML as the best-performing classifier demonstrates that the spectral pattern of shadows is strongly influenced by canopy structure. At the early stages, SVM achieved higher precision due to its selectivity at pixel boundaries, whereas at V8, ML performed better because of its probabilistic ability to model variations in light intensity.

Overall, the results showed that at early stages (V4), both ML and SVM achieved higher performance than RF (Table 2). As phenological development progressed, ML outperformed the other classifiers, achieving higher overall accuracy and F1-score, especially at V8, when the greater canopy development increased the spectral contrast among the crop, weed, soil, and shadow classes.

Thus, it is concluded that the maximum likelihood classifier at the V8 stage was the most effective combination for class differentiation, since at this stage maize exhibits greater spectral homogeneity and canopy stabilization, which increase class separability and favor the statistical performance of the classifier, in agreement with the observations of Garofalo et al. (2015), who highlight the greater spectral separability at more advanced stages of crop development.

The classification results for the second year, obtained with the ML, RF, and SVM classifiers at the V2 and V6 maize phenological stages, are summarized in Tables 3 and 4. Table 3 presents the confusion matrix, whereas Table 4 reports the performance metrics used for evaluation, including overall accuracy and class-specific precision, recall, and F1-score, providing the basis for a comparative analysis of classifier behavior across the phenological stages.

Table 3
Confusion matrix of the maximum likelihood (ML), random forest (RF), and support vector machine (SVM) classifiers for the V2 and V6 phenological stages of the maize crop in the second year

Classifiers		ML					RF		SVM
Stage	Class	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow
	Maize	23	2	0	0	23	0	2	0	20	2	3	0
V2	Soil	0	25	0	0	0	25	0	0	0	24	0	1
	Weed	4	9	10	2	8	7	10	0	4	9	12	0
	Shadow	0	0	0	25	0	0	0	25	0	0	0	25
	Maize	25	0	0	0	25	0	0	0	24	0	0	1
V6	Soil	0	25	0	0	0	25	0	0	1	24	0	0
	Weed	7	4	13	1	3	11	11	0	4	8	13	0
	Shadow	0	0	0	25	0	0	0	25	0	0	0	25

Table 4
Performance metrics obtained by the maximum likelihood (ML), random forest (RF), and support vector machine (SVM) classifiers for the V2 and V6 phenological stages of the maize crop in the second year

	Classifiers			ML			RF					SVM
	Stage	Metric	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow	Maize	Soil	Weed	Shadow
		Precision	0.92	1.00	0.40	1.00	0.92	1.00	0.40	1.00	0.80	0.96	0.48	1.00
		Recall	0.85	0.69	1.00	0.93	0.74	0.78	0.83	1.00	0.83	0.69	0.80	0.96
	V2	F1-score	0.88	0.82	0.57	0.96	0.82	0.88	0.54	1.00	0.82	0.80	0.60	0.98
		OA			0.83				0.83				0.81
		Precision	1.00	1.00	0.52	1.00	1.00	1.00	0.44	1.00	0.96	0.96	0.52	1.00
		Recall	0.78	0.86	1.00	0.96	0.89	0.69	1.00	1.00	0.83	0.75	1.00	0.96
	V6	F1-score	0.88	0.93	0.68	0.98	0.94	0.82	0.61	1.00	0.89	0.84	0.68	0.98
		OA			0.88				0.86				0.86

Overall accuracy was high across all classifiers in the second year and increased from V2 to V6. At V2, ML, RF, and SVM exhibited similar overall classification accuracy, indicating limited differentiation among classifiers under early canopy conditions. At V6, classification performance improved for all classifiers, with ML showing the highest overall accuracy, followed by SVM and RF. This increase highlights the positive effect of crop development on class separability and classifier robustness.

For the maize class, precision at V2 was higher for ML and RF, while SVM showed lower discriminative capacity at this early stage. At V6, precision improved for all classifiers, reaching optimal or near-optimal performance. Recall values showed a shift in classifier behavior across stages: ML achieved higher recall at the V2 stage, whereas RF achieved the highest recall at the V6 stage. The F1-score showed that ML performed best at the V2 stage, whereas RF performed best at the V6 stage, although all classifiers maintained high values.

The soil class showed high performance across all classifiers and phenological stages, reflected in consistently high precision, recall, and F1-score values. This result is mainly attributed to the distinct spectral signature of soil compared to vegetation, which enhances class separability. Precision remained close to or equal to 1 for all classifiers at both stages, while recall increased at V6, particularly for ML. The F1-score increased from the V2 to the V6 stage for the ML and SVM classifiers, whereas RF showed a reduction at the V6 stage.

The weed class showed lower precision at the V2 stage than at the V6 stage across the three classifiers, suggesting confusion with other classes. Recall values were high at both stages and reached maximum levels at V6, suggesting effective detection of weed pixels despite persisting limitations in class specificity. The F1-score was lower at the V2 stage than at V6, and ML achieved the best overall balance between precision and recall at V6.

The shadow class showed stable and high performance at both phenological stages. Precision was high across all classifiers, reflecting strong spectral separability from the other classes. Regarding recall, RF showed greater sensitivity in identifying shadow pixels at both the V2 and V6 stages when compared with the other classifiers. However, the F1-score indicates high performance in shadow classification across all three classifiers, with values close to or equal to 1.

The high performance of the classifiers at the initial stage (V2) can be attributed to specific experimental conditions, such as the lower flight altitude, which provided higher spatial resolution and enabled the identification of more distinct spectral patterns of the crop. In addition, the higher population density favored a more uniform canopy cover, reducing the influence of exposed soil and consequently minimizing spectral confusion among classes. Velumani et al. (2021) report that reducing flight altitude, thereby lowering GSD, significantly improves classifier performance, especially at early maize stages, consistent with the results of this study.

The evaluation of the results by class reinforces the high performance of the classifiers already at the early growth stage. For maize, the consistency of the F1-score values at both stages indicates high spectral separability even in early phases, reflecting the canopy homogeneity and the spectral definition of the crop. In the case of soil, the high metrics at V2 and V6 reflect its distinct spectral signature, which is easily recognized by supervised algorithms.

The weed class showed difficulty with discrimination at the initial stages, as evidenced by low precision and high recall, indicating overestimation of the class and confusion with maize and shadow. Overestimation and confusion of the weed class frequently occur in low-height and low-density species, which are prone to spectral mixing in transition regions. Similar results were reported by Dai et al. (2024), who highlighted that differences in canopy structure and pigment content can increase classification errors in boundary areas and mixed pixels.

The shadow class showed high metric values, indicating strong spectral separability and low radiometric variability. The stability of the soil class results is associated with its characteristic low-reflectance spectral response, which reduces class ambiguities even under variations in illumination or canopy density. This behavior was described by Xu et al. (2025), who reported high precision of supervised models in shadow identification when spectral separability is well defined.

The results reinforced that the V6 stage was the most suitable for class differentiation, presenting better spectral separability and greater consistency in the metrics, especially for maize and weed. Among the classifiers, ML demonstrated the most robust performance, with the highest overall accuracy (0.88) and the best F1-score for the weed class (0.68). Thus, the combination of ML at the V6 stage yielded the best classification performance in this cropping season.

In the first year, it was observed that the differentiation between maize (Zea mays L.) and the weed (Cyperus rotundus) was greater at the V8 stage, which marks the limit beyond which weeds no longer pose a risk to productivity. The differentiation of these classes at a more advanced stage is associated with the GSD used in this first survey, considering that the higher the value assigned to this parameter, the more difficult it is to identify the classes. In addition, in the first year, because it corresponded to the second crop period, the stand was lower, resulting in greater spacing between maize plants and, consequently, greater availability of areas for weed emergence and growth, both between rows and within planting rows, making plant differentiation more difficult.

In the second year, the classifiers showed better performance already at the initial stages of the crop, which may be associated with the lower GSD. In addition, because this period corresponded to the main crop, the stand was higher, and maize plants were closer to each other, reducing the space available for the emergence and growth of weeds as well as their competitive capacity. This favored greater differentiation between these classes, as crop growth was enhanced. This effect is evidenced by the exclusion of the V8 stage in the second year from this study, as weed growth was inhibited by shading from the main crop.

The variable performance of the classifiers indicates that no single model is universally optimal across all conditions; instead, classifier effectiveness depends on a combination of factors, including phenological stage, canopy structure, spectral variability, and the spatial and structural characteristics of the weed community, which may either enhance or constrain classification performance. The ML classifier showed greater stability under the proposed scenario, whereas RF performed better on spectrally stable targets such as soil and shadow. SVMs proved competitive for vegetation classes, especially at more advanced stages.

CONCLUSIONS

• 1. The results obtained across the different maize cropping seasons demonstrated the feasibility of using multispectral images generated by remotely piloted aircraft for weed detection, especially when combined with machine learning algorithms.

• 2. In the first year of evaluation, the V8 vegetative stage provided the best separability conditions between maize and weed. In contrast, in the second year, the V6 stage proved more suitable given the spectral and operational flight conditions.

• 3. In both cropping seasons, the maximum likelihood classifier stood out with the best performance in distinguishing the evaluated classes, demonstrating strong potential for pixel-based analyses.

• 4. The phenological development of the crop contributed to increased spectral separability, favoring more precise classifications throughout the vegetative stages.

• 1
  Research developed at the Universidade Federal Rural do Rio de Janeiro, Seropédica, RJ, Brazil
• Ref. 302881
• Financing statement:
  Financial support provided by Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brazil (CAPES) - Finance Code 001.

Acknowledgments:

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brazil (CAPES) and Postgraduate Program in Agricultural and Environmental Engineering.

Data availability statement:

The authors declare that there are no data underlying the text.

Literature Cited

• Almeida-Ñauñay, A. F. et al. Optimization of soil background removal to improve the prediction of wheat traits with UAV imagery. Computers and Electronics in Agriculture, v.205, 107559, 2023. https://doi.org/10.1016/j.compag.2022.107559
  » https://doi.org/10.1016/j.compag.2022.107559
• Alvares, C. A. et al. Köppen’s climate classification map for Brazil. Meteorologische Zeitschrift, v.22, p.711-728, 2014. https://doi.org/10.1127/0941-2948/2013/0507
  » https://doi.org/10.1127/0941-2948/2013/0507
• ASPRS - American Society for Photogrammetry and Remote Sensing. The ASPRS positional accuracy standards: the geospatial mapping industry guide to best practices. 2. ed., 2023. Available on: <https://www.asprs.org/wp-content/uploads/2024/03/October2023_HLA-Positional_Accuracy_Standards.pdf?utm_source=>. Accessed on: Feb. 2024.
  » https://www.asprs.org/wp-content/uploads/2024/03/October2023_HLA-Positional_Accuracy_Standards.pdf?utm_source
• Bretas, I. L. et al. Detection and mapping of Amaranthus spinosus L. in bermudagrass pastures using drone imagery and deep learning for a site-specific weed management. Agronomy Journal, v.116, p.990-1002, 2024. https://doi.org/10.1002/agj2.21545
  » https://doi.org/10.1002/agj2.21545
• CONAB - Companhia Nacional de Abastecimento. Boletim da safra de grãos - 9º Levantamento: Safra 2024/25. Available on: <https://www.gov.br/conab/pt-br/atuacao/informacoes-agropecuarias/safras/safra-de-graos/boletim-da-safra-de-graos/9o-levantamento-safra-2024-25/e-book_boletim-de-safras-9o-levantamento-2025>. Accessed on: Jun. 2025.
  » https://www.gov.br/conab/pt-br/atuacao/informacoes-agropecuarias/safras/safra-de-graos/boletim-da-safra-de-graos/9o-levantamento-safra-2024-25/e-book_boletim-de-safras-9o-levantamento-2025
• Congalton, R.; Green, K. Assessing the accuracy of remotely sensed data: principles and practices. 3.ed. Boca Raton: CRC Press, 2019. 346p.
• Dai, J. et al. Canopy-level spectral variation and classification of diverse crop species with fine spatial resolution Imaging spectroscopy. Remote Sensing, v.16, 1447, 2024. https://doi.org/10.3390/rs16081447
  » https://doi.org/10.3390/rs16081447
• Dash, P. et al. Improving the accuracy of land use and land cover classification of Landsat data in an agricultural watershed. Remote Sensing, v.15, 4020, 2023. https://doi.org/10.3390/rs15164020
  » https://doi.org/10.3390/rs15164020
• EMBRAPA - Empresa Brasileira de Pesquisa Agropecuária. Manejo de plantas daninhas na cultura do milho. Passo Fundo: Embrapa Trigo, 2006. 67p.
• FAO - Food and Agriculture Organization - FAOSTAT (2025). Global cereal production to exceed 3-billion mark, stocks-to-use ratio highest in decades. Available on: < https://www.fao.org/worldfoodsituation/csdb/en>. Accessed on: Jan. 2026.
  » https://www.fao.org/worldfoodsituation/csdb/en
• Foody, G. M. Challenges in the real world use of classification accuracy metrics: From recall and precision to the Matthews correlation coefficient. Plos One, v.18, e0291908, 2023. https://doi.org/10.1371/journal.pone.0291908
  » https://doi.org/10.1371/journal.pone.0291908
• García-Navarrete, O. L. et al. Evaluation of YOLOv5 models for weed classification in crops maize (Zea Mays L.). In: Rossit, D.; Torres-Aguilar, C.E.; Toncovich, A.A. Advancements in optimization and nature-inspired computing for solutions in contemporary engineering challenges. Singapore: Springer Nature, 2025. p.183-209. https://doi.org/10.1007/978-981-96-0706-8_10
  » https://doi.org/10.1007/978-981-96-0706-8_10
• Garofalo, D. F. T. et al. Análise comparativa de classificadores digitais em imagens do Landsat-8 aplicados ao mapeamento temático. Pesquisa Agropecuária Brasileira, v.50, p.593-604, 2015. https://doi.org/10.1590/S0100-204X2015000700009
  » https://doi.org/10.1590/S0100-204X2015000700009
• Gu, J.; Congalton, R. G. Assessing the impact of mixed pixel proportion training data on SVM-based remote sensing classification: A simulated study. Remote Sensing, v.17, 1274, 2025. https://doi.org/10.3390/rs17071274
  » https://doi.org/10.3390/rs17071274
• Kim, K. et al. Deep learning of high-resolution unmanned aerial vehicle imagery for classifying halophyte species: a comparative study for small patches and mixed vegetation. Remote Sensing, v.15, 2723, 2023. https://doi.org/10.3390/rs15112723
  » https://doi.org/10.3390/rs15112723
• Liang, P. et al. A case-based method of selecting covariates for digital soil mapping. Journal of Integrative Agriculture, v.19, p.2127-2136, 2020. https://doi.org/10.1016/S2095-3119(19)62857-1
  » https://doi.org/10.1016/S2095-3119(19)62857-1
• Liu, L. et al. High-resolution mapping of maize in mountainous terrain using machine learning and multi-source remote sensing data. Land, v.14, 299, 2025. https://doi.org/10.3390/land14020299
  » https://doi.org/10.3390/land14020299
• Marques, R. F. et al. Profundidade de semeadura e peso específico das sementes na emergência da planta daninha Urochloa ruziziensis em campo. Revista Brasileira de Ciências Agrárias, v.19, p.1-9, 2024. https://doi.org/10.5039/agraria.v19i1a3458
  » https://doi.org/10.5039/agraria.v19i1a3458
• Medina, J. A. V.; Atehortúa, B. E. A. Comparison of maximum likelihood, support vector machines, and random forest techniques in satellite images classification. Tecnura, v.32, p.13-26, 2019. https://doi.org/10.14483/22487638.14826
  » https://doi.org/10.14483/22487638.14826
• Oliveira, M. N. et al. Dinâmica populacional de plantas espontâneas no sistema de plantio orgânico de milho consorciado com lab-lab. Cadernos de Agroecologia, v.19, p.1-7, 2024. Available on: <https://cadernos.aba-agroecologia.org.br/cadernos/article/view/7773/5635>. Accessed on: Oct. 2025.
  » https://cadernos.aba-agroecologia.org.br/cadernos/article/view/7773/5635
• Phiri, D. et al. Sentinel-2 data for land cover/use mapping: A review. Remote Sensing, v.12, p.1-35, 2020. https://doi.org/10.3390/rs12142291
  » https://doi.org/10.3390/rs12142291
• Shah, A. et al. Maize in 2034: a decade of growth, innovation, and sustainability in crop production. Indus Journal of Bioscience Research, v.3, p.62-68, 2025. https://doi.org/10.70749/ijbr.v3i4.1070
  » https://doi.org/10.70749/ijbr.v3i4.1070
• Silva, W. L.; Dereczynski, C. P. Caracterização climatológica e tendências observadas em extremos climáticos no estado do Rio de Janeiro. Anuário do Instituto de Geociências, v.37, p.123-138, 2014. http://dx.doi.org/10.11137/2014_2_123_138
  » http://dx.doi.org/10.11137/2014_2_123_138
• Soil Survey Staff. Keys to Soil Taxonomy. 13.ed. USDA - Natural Resources Conservation Service, 2022. 401p. Available at: https://www.nrcs.usda.gov/resources/guides-and-instructions/keys-to-soiltaxonomy Accessed on: Mar. 2026.
  » https://www.nrcs.usda.gov/resources/guides-and-instructions/keys-to-soiltaxonomy
• Tetila, E. C. etal. Deeplearningmodelsfordetectionandrecognition of weed species in corn crop. Crop Protection, v.195, 107237, 2025. https://doi.org/10.1016/j.cropro.2025.107237
  » https://doi.org/10.1016/j.cropro.2025.107237
• Velumani, K. et al. Estimates of maize plant density from UAV RGB images using faster-RCNN detection model: Impact of the spatial resolution. Plant Phenomics, v.2021, p.1-16, 2021. https://doi.org/10.34133/2021/9824843
  » https://doi.org/10.34133/2021/9824843
• Villiers, C. de. et al. Weed detection in rainfed maize crops using UAV and PlanetScope imagery. Sustainability, v.15, 13416, 2023. https://doi.org/10.3390/su151813416
  » https://doi.org/10.3390/su151813416
• Xu, H. et al. A shadow detection method combining topography and spectra for remote sensing images in mountainous environments. Applied Sciences, v.15, 4899, 2025. https://doi.org/10.3390/app15094899
  » https://doi.org/10.3390/app15094899
• Yang, P. et al. Separation of the direct reflection of soil from canopy spectral reflectance. Remote Sensing of Environment, v.316, 114500, 2025. https://doi.org/10.1016/j.rse.2024.114500
  » https://doi.org/10.1016/j.rse.2024.114500
• Zhao, Z. et al. Rapid and accurate prediction of soil texture using an image-based deep learning autoencoder convolutional neural network random forest (DLAC-CNN-RF) algorithm. Agronomy, v.12, 3063, 2022. https://doi.org/10.3390/agronomy12123063
  » https://doi.org/10.3390/agronomy12123063

Edited by

• Editors:
  Ítalo Herbet Lucena Cavalcante & Carlos Alberto Vieira de Azevedo

Publication Dates

• Publication in this collection
  21 Aug 2026
• Date of issue
  2026

History

• Received
  17 Nov 2025
• Accepted
  12 Apr 2026
• Published
  31 July 2026