Delineation of Cocoa Agroforests Using Multi-Season Sentinel-1 SAR Images: Low Grey Level Range Reduces Uncertainties in GLCM Texture-Based Mapping

Abstract

Delineating the cropping area of cocoa agroforests is a major challenge for quantifying the contribution of the land use expansion to tropical deforestation. Discriminating cocoa agroforests from tropical transition forests using multi-spectral optical images is difficult due to a similarity in the spectral characteristics of their canopy; moreover, optical sensors are largely impeded by the frequent cloud cover in the tropics. This study explores multi-season Sentinel-1 C-band SAR image to discriminate cocoa agroforests from transition forests for a heterogeneous landscape in central Cameroon. We use an ensemble classifier, random forest, to average SAR image texture features of GLCM (Grey Level Co-occurrence Matrix) across seasons; next, we compare classification performance with results from RapidEye optical data. Moreover, we assess the performance of GLCM texture feature extraction at four different grey level quantization: 32bits, 8bits, 6bits, and 4bits. The classification overall accuracy (OA) of texture-based maps outperformed that from an optical image; the highest OA of 88.8% was recorded at 6bits grey level. This quantization level, in comparison to the initial 32bits in SAR images, reduced the class prediction error by 2.9%. Although this prediction gain may be large for the landscape area, the resultant thematic map reveals the decrease and fragmentation of forest cover by cocoa agroforests. According to our classification validation, the Shannon entropy (H) or uncertainty provides a reliable validation for class predictions and reveals detail inference for discriminating inherently heterogeneous vegetation categories. The texture-based classification achieved a reliable accuracy considering the heterogeneity of the landscape and vegetation classes.

Integration of Sentinel optical and radar data for mapping smallholder coffee production systems in Vietnam

“…While our model targets coffee, we drew from mapping literature on similar tree crops: shade cocoa, coconut, rubber, and oil palm. While separating tree crop from surrounding land cover such as forest and other agriculture land, we would emphasize, in continuation of previous work, the added value of texture features in distinguishing landscape and cropping systems (Burnett et al, 2019;Gao et al, 2015;Gomez et al, 2010;Liu and Chen, 2019;Numbisi et al, 2019). We would also make the distinction between tree crops with a regular clearing rotation (such as rubber) and fruit and nut trees (such as a coffee and cocoa), and between monoculture dominated landscape, agroforestry dominated landscapes, and mixed monoculture and agroforestry.…”

“…Precipitation or temperature features (Cordero -Sancho an d Sader, 2007; Kelley et al, 2018) and texture features (Gaertner, 2017;Gomez et al, 2010;Tsai and Chen, 2017) are rapidly becoming key in difficult-to-map coffee regions. Studies that include texture features often use high-resolution proprietary data such as WorldView (Gaertner, 2017) or Quickbird (Gomez et al, 2010), or focus on mapping or distinguishing other tree crops such as cocoa, rubber, oil palm (Burnett et al, 2019;Gao et al, 2015;Nomura and Mitchard, 2018;Numbisi et al, 2019;Torbick et al, 2016). Notably, Gray Level Co-occurrence Matrices (GLCM) texture measures were derived from Sentinel-1 for cocoa mapping in Numbisi et al (2019).…”

“…Studies that include texture features often use high-resolution proprietary data such as WorldView (Gaertner, 2017) or Quickbird (Gomez et al, 2010), or focus on mapping or distinguishing other tree crops such as cocoa, rubber, oil palm (Burnett et al, 2019;Gao et al, 2015;Nomura and Mitchard, 2018;Numbisi et al, 2019;Torbick et al, 2016). Notably, Gray Level Co-occurrence Matrices (GLCM) texture measures were derived from Sentinel-1 for cocoa mapping in Numbisi et al (2019). Texture measures were demonstrated to have more use in SAR images compared to optical and are of added value in heterogenous landscapes (Mishra et al, 2019).…”

“…After testing moving windows of size 3, 5, and 7 in Ea H'Leo, a 5x5 pixel moving window was chosen to best capture the plot-level texture, particularly between sun and intercropped coffee. S1 SAR textural information has been shown to be particularly beneficial in heterogenous landscape (Mishra et al, 2019), distinguishing between smallholder and large-scale plantations (Oon et al, 2019), but has also been applied for agroforestry mapping, as is done for a cocoa landscape in Numbisi et al, (2019).…”