Scholarlink

<p>Earth Observation (EO), along with machine learning (ML), has been widely recognized as an effective approach for estimating and mapping aboveground biomass (AGB). However, large-scale biomass estimation often requires extending predictions beyond the environmental settings of the training sites. Variation in agricultural practices, climatic conditions, or ecological characteristics between training areas and untested regions reduces the predictive model’s accuracy. This study aimed to assess the transferability and upscaling of ML models for sugarcane AGB and carbon stock estimation in Northeast Thailand. Field data were collected from a total of 315 plots, comprising 180 plots in Chaiyaphum–Khon Kaen and 135 plots in Kalasin province. Furthermore, time-series imagery from Sentinel-1, Sentinel-2, and Landsat 8/9 were integrated along with Random Forest Regression (RFR) and Support Vector Regression (SVR) algorithms. Map evaluation showed that RFR achieved a superior performance (R² = 0.87, RMSE = 5.05 t ha⁻¹) compared to SVR (R² = 0.83, RMSE = 6.22 t ha⁻¹). Moreover, RFR is inherently robust to noise and outliers introduced by sampling datasets, mixed pixels, or atmospheric variability, resulting in high efficiency and spatial map consistency. Transfer learning further validated the result from the optimized RFR model in Kalasin province with high accuracy (R² = 0.81, RMSE = 5.42 t ha⁻¹), confirming its generalization capability under smallholder farming systems. The optimized RFR model was then applied to upscale AGB and carbon stock estimation across the entire Northeast region. The resulting maps revealed substantial spatial variability, with AGB ranging from 29 to 77 t ha⁻¹ and carbon stock from 14 to 36 t ha⁻¹, capturing both regional trends and field-level heterogeneity. This study demonstrated the high potential of transfer learning to produce reliable and credible maps for smallholder farming systems at a large spatial scale. The findings highlight the robustness of integrating multi-source EO data with RFR modeling for transferable and scalable crop biomass estimation. This framework provides reliable spatial information to support carbon stock monitoring and sustainable agricultural management.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Scalable Mapping of Sugarcane Aboveground Biomass and Carbon Stock in Smallholder Fields of Northeast Thailand Using Multi-Sensor Data and Transfer Learning

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Savittri Ratanopad Suwanlee,
  • Ittipon Khuimphukhieo,
  • Akkaneerut Watwaingkham,
  • Todtawan Sajan,
  • Vorraveerukorn Veerachitt,
  • Mohammad D. Hossain,
  • Narissara Nuthammachot,
  • Sarawut Ninsawat,
  • Jaturong Som-ard

摘要

Earth Observation (EO), along with machine learning (ML), has been widely recognized as an effective approach for estimating and mapping aboveground biomass (AGB). However, large-scale biomass estimation often requires extending predictions beyond the environmental settings of the training sites. Variation in agricultural practices, climatic conditions, or ecological characteristics between training areas and untested regions reduces the predictive model’s accuracy. This study aimed to assess the transferability and upscaling of ML models for sugarcane AGB and carbon stock estimation in Northeast Thailand. Field data were collected from a total of 315 plots, comprising 180 plots in Chaiyaphum–Khon Kaen and 135 plots in Kalasin province. Furthermore, time-series imagery from Sentinel-1, Sentinel-2, and Landsat 8/9 were integrated along with Random Forest Regression (RFR) and Support Vector Regression (SVR) algorithms. Map evaluation showed that RFR achieved a superior performance (R² = 0.87, RMSE = 5.05 t ha⁻¹) compared to SVR (R² = 0.83, RMSE = 6.22 t ha⁻¹). Moreover, RFR is inherently robust to noise and outliers introduced by sampling datasets, mixed pixels, or atmospheric variability, resulting in high efficiency and spatial map consistency. Transfer learning further validated the result from the optimized RFR model in Kalasin province with high accuracy (R² = 0.81, RMSE = 5.42 t ha⁻¹), confirming its generalization capability under smallholder farming systems. The optimized RFR model was then applied to upscale AGB and carbon stock estimation across the entire Northeast region. The resulting maps revealed substantial spatial variability, with AGB ranging from 29 to 77 t ha⁻¹ and carbon stock from 14 to 36 t ha⁻¹, capturing both regional trends and field-level heterogeneity. This study demonstrated the high potential of transfer learning to produce reliable and credible maps for smallholder farming systems at a large spatial scale. The findings highlight the robustness of integrating multi-source EO data with RFR modeling for transferable and scalable crop biomass estimation. This framework provides reliable spatial information to support carbon stock monitoring and sustainable agricultural management.

Graphical Abstract