Modelos preditivos para a estimativa da produtividade da cultura da soja por meio do sensoriamento remoto proximal e aprendizado de máquina

Abstract

Estimating agricultural crop productivity remains subjective, mainly leading to inaccuracies and limitations in interpreting results. Given this scenario, there is a clear need to adopt more accurate, sustainable, and technically sound methodologies. In this context, Remote Sensing (RS) combined with Artificial Intelligence (AI) techniques stands out as a promising approach for obtaining more accurate and reliable estimates, Digital Agriculture and Precision Agriculture. Based on this assumption, the present study aimed to estimate soybean productivity by integrating Proximal Remote Sensing (PRS) and Machine Learning (ML) algorithms across 20 cultivars grown in Paraná. The cultivars evaluated were: TMG 23303 XTD, TMG 23306 I2X, TMG 7362 IPRO, TMG 2461, TMG 24402 I2X, TMG 24403 I2X, TMG 24404 I2X, TMG 24405 I2X, TMG 24430 I2X, TMG 24406 I2X, TMG 2165 I2X, BMX Fúria IPRO, BMX Fibra IPRO, Neo 610 IPRO, BMX Lotus IPRO, DM 66i68 IPRO, P 96R10 IPRO, P 95Y95 IPRO, P 96Y90, and P 96R29 IPRO. The experiment was conducted in the experimental area of the Federal Technological University of Paraná (UTFPR), Santa Helena campus, located in the municipality of Santa Helena, State of Paraná. The experimental design adopted was a completely randomized design (CRD) in continuous strips, with 20 treatments and nine replicates. The variables evaluated included biophysical plant characteristics (plant height, canopy width, productivity, and coverage area), as well as the NDVI (Normalized Difference Vegetation Index) and NDRE (Normalized Difference Red Edge) vegetation indices, obtained from Proximal Remote Sensing (PRS). Additionally, crop productivity was determined, enabling the correlation between spectral parameters and the agronomic performance of different cultivars. The database was divided into two stages: training (80%) and validation (20%). The input variables included biophysical plant characteristics and vegetation indices, while the output variable represented soybean productivity. For data analysis, ML Random Forest Classification algorithms were used to select the most relevant variables, and Random Forest Regression was used to develop and validate predictive models. The performance of the validated models was assessed using two error metrics: Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE), along with an accuracy metric represented by the Coefficient of Determination (R²). The results showed that the best performance was achieved by models that included all inputs combined, with an accuracy of 87.41%. This was followed by scenarios that considered the Normalized Difference Red Edge (NDRE) and the plant's biophysical characteristics, which achieved an accuracy of 71.70%. The combination of vegetation indices with canopy width resulted in a lower accuracy of 50.27%. These findings emphasize the potential of integrating spectral and biophysical variables with machine learning techniques to enhance the accuracy of soybean yield estimates.